MGY-005

TECHNIQUES IN REMOTE
Indira Gandhi National Open University
School of Sciences SENSING AND DIGITAL
IMAGE PROCESSING

Block

IMAGE PRE-PROCESSING TECHNIQUES

UNIT 7
Image Statistics 153
UNIT 8
Radiometric Image Corrections 175
UNIT 9
Geometric Image Corrections 189
UNIT 10
Image Enhancement 205
UNIT 11
Image Filtering and Band Ratioing 221
UNIT 12
Image Fusion and Principal Component Analysis 239

Glossary 257
 BLOCK 2: IMAGE PRE-PROCESSING TECHNIQUES
In Block 1, you have studied the various techniques of remote sensing such as aerial
photography, multispectral, thermal, hyperspectral, microwave, LiDAR and UAV-based. A
sound understanding of all these remote sensing techniques is required to find suitable
remotely sensed data for a specific application. The raw data obtained from various remote
sensing techniques often contain distortions, errors and noises, which are entered during the
scanning processes. These errors are removed from the remotely sensed data during its
pre-processing phase. This phase is known as digital image pre-processing and it uses to
improve visual quality of data, which makes essay to extract useful information from the
data.
This is the second block of MGY-005 dealing with image pre-processing techniques. It
comprises six units, which introduce you to various aspects of digital image pre-processing
techniques such as image statistics, correction, enhancement and transformation.
Unit 7: Image Statistics deals with statistical properties of images. It introduces you to
univariate and multivariate image statistics.
Unit 8: Radiometric Image Corrections discusses the types and causes of radiometric
distortions in remote sensing images and methods of their corrections to make images able
to represent the terrain more closely.
Unit 9: Geometric Image Corrections deals with necessity of removing geometric errors,
types of geometric errors and methods used for their removal from the image data.
Unit 10: Image Enhancement deals with common image enhancement operations together
with linear and non-linear contrast enhancement. It also presents application of contrast
enhancement.
Unit 11: Image Filtering and Band Ratioing deals with digital image filtering approaches,
basics of band ratioing and significance of commonly used spectral indices.
Unit 12: Image Fusion and Principal Component Analysis introduces two types of image
pre-processing techniques such as image fusion and principal component analysis, which
are applicable on multispectral and hyperspectral images.
Expected Learning Outcomes________________________________
After studying this block, you should be able to:
 describe image statistics and its types such as univariate and multivariate image
statistics;
 types and causes of radiometric and geometric distortions and approaches used to
remove them;
 discuss linear and non-linear contrast enhancement techniques;
 discuss common types of image filters;
 elaborate on the principle of band ratioing;
 discuss various types of image fusion methods; and
 describe principal components analysis.

Wishing you success in this endeavour!

152
 UNIT 7

IMAGE STATISTICS

Structure____________________________________________________________________
7.1 Introduction 7.4 Types of Image Statistics
Expected Learning Outcomes Univariate Descriptive Image Statistics

7.2 Image Statistics Multivariate Image Statistics

7.3 Basics of Statistics 7.5 Activity

Data Types 7.6 Summary
Levels of Measurement 7.7 Terminal Questions
Population and Sample 7.8 References
Types of Statistical Analysis 7.9 Further/Suggested Readings
Significance 7.10 Answers

7.1 INTRODUCTION
You are aware that remote sensing is a technique for gathering information about any object or
feature from a distance without being making any physical contact with it. This information is
recorded in the form of raster data or images by using sensors or cameras mounted on the various
remote sensing platforms. The remote sensors detect terrain features by using reflected solar
radiation, heat emitted by objects and back-scattered energy in the microwave region of the
electromagnetic spectrum. Remotely sensed images give an objective record of the
electromagnetic spectrum pertaining to the physical and chemical properties of the terrain or
ground. Extracting information from the remotely sensed images requires the knowledge of the
image statistics.
You have learnt about basic principles, data formats, data products and applications of various
types of remote sensing techniques such as multispectral, thermal, hyperspectral, microwave,
LiDAR and UAV-based in the previous block. Images are the final product of remote sensing and
they can be microwave, panchromatic, multispectral and hyperspectral.
Block 2 Image Pre-processing Techniques
…………………………………………...............……………………….……………………….................……
In addition, images can be acquired in single or multiple bands. Image statistics
involves the study of remotely sensed images and their statistical properties.
These properties are useful in the description, classification, quantification,
quality assessment and reduction of images. They also guide us for choosing
suitable secondary treatments like filtering, restoring, coding and shape
detection for giving to the image. In this unit, we will discuss image statistics
and its types such as univariate and multivariate image statistics.

Expected Learning Outcomes_______________________

After studying this unit, you should be able to:
 define image statistics;
 discuss various types of image statistics;
 describe univariate descriptive image statistics and its measures; and
 explain multivariate image statistics and its visualisation.

7.2 IMAGE STATISTICS

As mentioned earlier that image statistics involves the study of remotely sensed
images and their statistical properties. It calculates raster cell statistics from one
or multiple selected images. It is used to compute the statistical measurements
of the images consisting of image histogram, mean, median, standard
deviation, dispersion, distribution, covariance and correlation. Image statistical
analysis gives us an easily interpretable visual format of an image. It helps us to
find typical problems in an image and to visually improve it. It also allows us to
determine whether an image is suitably exposed or not. The poorly exposed
images result in the loss of the information and this type of error should be
recognised at the earlier stage of image processing because it is not possible to
recover such loss at the later stages of image processing (Burger and Burge,
2022).
It is important to note that the volume of data available from remote sensors is
astounding both in spatial and temporal domains. For instance, while Terra and
Aqua satellites onboard the Moderate Resolution Imaging Spectroradiometer
(MODIS) provide satellite images at high temporal resolution of 1-2 days, others
from satellites like WorldView – 3 provide data at high spatial resolution of
almost 31 cm. Thus, making meaningful sense of the huge amount of remotely
sensed data requires knowledge of not just data handling but also its basic
statistical properties. Prior to discussing image statistics, it is important to
understand the basics of statistics especially different types of data and levels
of measurement for selecting appropriate statistical techniques.

7.3 BASICS OF STATISTICS

Statistics is a branch of applied mathematics that deals with the collection,
organisation, presentation, analysis and interpretation of data. Let us discuss
main concepts of statistics applicable to the statistical analysis of remotely
sensed images.

154 Contributor: Dr Manu Mehta and Mr. Akhilesh Kumar

Unit 7 Image Statistics
…………………….................……………..………...............………………………….…………………………
7.3.1 Data Types
In order to decide which statistics is well suited for a dataset, it is important to
first understand the type of data we are dealing with. Broadly speaking, data
can be categorised into two types such as qualitative and quantitative.
i) Qualitative Data
It mainly provides insights related to subjective opinions, preferences or
qualitative aspects of the phenomenon. The statistics pertaining to qualitative
data primarily revolve around generating descriptive statistics or summaries,
aiming to unravel patterns and trends or explore associations. Some of the
commonly used statistics for qualitative data includes frequencies, proportions
and modes.
ii) Quantitative Data
It represents numerical data and the statistics addressing the quantitative data
can be descriptive, providing information regarding the measures of central
tendency (e.g., mean and median), measures of dispersion or variability (e.g.,
range and standard deviation) or measures of distribution (e.g., skewness and
kurtosis). Quantitative data can also be used to generate inferential statistics,
using advanced statistical techniques like hypothesis testing, regression
analysis and analysis of variance (ANOVA), among others to draw inferences or
predictions based on the sample data.
Both qualitative and quantitative data provide broad understanding of the nature
of the information being measured or conveyed, level of measurements, order
of hierarchy or categories and the significance of zero points.
7.3.2 Levels of Measurement
The levels of measurement play an important role in selecting type of statistical
analysis for a particular image (Stommel and Katherine, 2014). There are some
statistical analyses which require more precise levels of measurement as
compared to the others. There are four levels of measurement, which are
described below:
i) Nominal Data
It represents labels or categories that do not have a numerical value or an order
assigned to it. Examples of such data could include land use classes (urban,
agricultural or vegetation) or colours (red, green, blue). The statistical
techniques pertaining to nominal data generally include frequency analysis,
where the focus is on counting and summarising the occurrences of different
categories or labels. Frequency, proportions and modes are some of the
commonly used measures for describing nominal data.
ii) Ordinal Data
It also represents labels or categories. However, these categories have a
meaningful order or hierarchy attached to them. Though these categories are
not numerically quantifiable, their relative position or rank has a meaning.
Common examples could include education levels (high school, graduate, post-
graduate) or literacy (illiterate, semi-literate, literate). Central tendency and
dispersion of ordinal data is generally summarised using median and quartiles.
Contributor: Dr Manu Mehta and Mr. Akhilesh Kumar 155
Block 2 Image Pre-processing Techniques
…………………………………………...............……………………….……………………….................……
iii) Interval Data
It is the third level of measurements and extends the characteristics of ordinal
data by exhibiting equal intervals between values. Examples include
temperature measured in Celsius or Fahrenheit scales. Statistical techniques
for interval data include measures of central tendency using mean or median
and measures of dispersion using range or standard deviation.
iv) Ratio Data
It is the highest level of measurement. In addition to possessing all the
characteristics of interval data, ratio data also have a meaningful zero point and
thus, it is possible to make statements about ratios and proportions. Examples
include age, height, income or weight. The statistical techniques that can be
applied to ratio data range from as basic as descriptive statistics to hypothesis
testing, regression, and other advanced modelling techniques.
It may be noted that the numerical analysis is possible for interval and ratio
data. However, it is not possible for nominal or ordinal data. Similarly, while
assessment of hierarchy is possible for ordinal or ratio data, it is not possible for
nominal or interval data.
7.3.3 Population and Sample
In statistics, population refers to the entire group of events for which we have
the information and sample is used to refer to a small subset of the population
which can be used as a smaller subset or representation of the entire
population. The importance of the concept of sample lies in the fact that often at
times, it is neither feasible nor appropriate to survey or collect information for
the entire population, either due to scale, resource constraints or some other
ethical reason preventing the testing or application to the entire population. In
such situation, sample can be used effectively to test the hypothesis and then
generalise the findings to the entire population.
Using a subset of the population to make generalisation about the entire
population often leads to uncertainty, commonly referred to as sampling error
and is generally measured by a confidence interval. For example, if a survey
states with 90% confidence that the average height in the age group 18-30 is
160 cm, it means that if a random person is picked up in this age group, 9 out of
10 times, his/her height will be around 160 cm.
While there are many formulae that can be used to determine the sample size
at a particular confidence threshold for a given population, one of them is the
Slovin’s formula, which states that–

where is the sample size, is the population size and is the error tolerance.
Let us calculate sample size (n) by taking a population size of 10,000 and 90%
confidence level (i.e., 10% or 0.10 tolerance level) and equation (1) as given
below:

156 Contributor: Dr Manu Mehta and Mr. Akhilesh Kumar

Unit 7 Image Statistics
…………………….................……………..………...............………………………….…………………………
7.3.4 Types of Statistical Analysis
Generally, statistical analysis is categorised into three main types – univariate,
which deals with a single variable, bivariate which deals with two variables and
multivariate statistics that deals with a number of variables.
Univariate Analysis: It focuses on one variable at a time and can be used to
examine the central tendency, variability or distribution of the data using various
measures like mean, median, mode, variance and standard deviation. It is
commonly used to summarise data, identify the outliers and gain insights about
the behaviour of a specific variable.
Bivariate Analysis: It seeks to study the relationship between two variables
and examines how the change in one variable impacts the other variables. It
can be done by using techniques like correlation, scatter plots or simple linear
regression to explore the cause-and-effect relationships, measure the direction
or strength of association and to make predictions based on the association.
Multivariate analysis: It extends the capabilities of bivariate analysis to three
or more variables and is focussed on understanding the complex relationships
and associations existing among them. Multivariate analysis relies on statistical
techniques like principal component analysis (PCA), multiple regression
analysis, and cluster analysis.
7.3.5 Significance
Understanding type of data and appropriate analysis technique is essential in
order to draw meaningful inferences and for making informed decisions. For
instance, while frequency and proportion are more suited to nominal data,
ordinal data can be better explained using measures of central tendency and
dispersion. Interval and ratio data, though, require more advanced statistical
techniques like correlation and regression.
The choice and outcome of statistical analysis is also dependent on number of
variables. As discussed earlier, where univariate analysis provides information
about individual variables, bivariate variables explore associations between two
variables. Multivariate analysis, though, can help unravel even more complex
dependencies and associations by taking all the variables into account.

7.4 TYPES OF IMAGE STATISTICS

In digital images, statistics mainly deals with the probability distribution function
of the brightness levels of a particular image and density function of the
brightness levels of the same image. It may be noted that majority of image
processing systems perform two types of statistical analyses such as univariate
and multivariate analysis on single and multiple band remotely sensed data
(Jensen, 2018). Let us discuss these two types of image statistics.
7.4.1 Univariate Descriptive Image Statistics
It describes and summarises individual image variable separately. It does not
work on cause-and-effect relationship. Its main aim is to describe image and
infer patterns that exist within the image. The most common univariate
descriptive statistics in remotely sensed data are described below.

Contributor: Dr Manu Mehta and Mr. Akhilesh Kumar 157

Block 2 Image Pre-processing Techniques
…………………………………………...............……………………….……………………….................……
i) Measure of Central Tendency in Remotely Sensed Data
A measure of central tendency of a given dataset describes the characteristics
of a dataset using a single value, typically the one that occupies the central
position within the dataset.
a) Mean
It is the arithmetic average of all the values for a variable. In terms of image
statistics, mean can be defined as the average value of all the pixel brightness.
If is the number of pixels and is the value of the pixel at position,
then mean is given by –

If sample size is considered, instead of the entire population, then mean ( ) is

given by –

For example, the pixel values in an image of 3x3 dimension , then

the mean pixel value will be calculated as –

While the simple mean gives equal weightage to all the observations, there may
be situations where there are multiple datasets but number of observations in
each of them is not same. In such cases, weighted mean can be calculated as:

where is the corresponding weight and other terms have the same meaning
as in Equation 2a.
Let us consider the following example. Suppose, we have an image array A with
the array of weights as W.

A= ,W=

Now, the mean of A = and the weighted mean of

A=

b) Median
Median represents the middle value of the dataset arranged in an order – either
ascending or descending and unlike mean, is unaffected by the extreme values
or outliers. To calculate the median, first arrange the observations in a dataset
in an order and then pick the middle value.
In order to calculate the median of a given image array below (2c), we first need
to arrange it into an order (2d) –

158 Contributor: Dr Manu Mehta and Mr. Akhilesh Kumar

Unit 7 Image Statistics
…………………….................……………..………...............………………………….…………………………

Since the number at the middle of the arranged order is 3, the median pixel
value of the image is 3. In case the number of data points is even, then median
value is taken as the average of the two middle values.
Let us take another example of an image array given at 2e for understanding
importance of median value over mean.

Image array =

The mean value

But the median value

For a dataset where there is huge difference between the values or where there
is possibility of having outliers, median values are recommended measures for
central tendency instead of mean.
c) Mode
Mode is defined as the value occurring most often in a dataset. It is a useful
statistic for determining the values that dominate the dataset.
Take an example of an image array matrix given below:

The pixel value 2 appears maximum number of times (3 in total), hence its
mode is 2.
It is worth mentioning here that a dataset can have more than one mode value.
In that case, the data is called as bimodal (2 modes), trimodal (3 modes) or
multimodal (more than 3 nodes). For instance, in the following case –

The mode values are = 4, 2, 7 and the dataset can be considered as trimodal.
ii) Measures of dispersion
Measures of the dispersion give us extent of variability, which helps us to
understand the distribution of the image data. In other words, it gives us
information on how stretched or squeezed the values are around their central
values. Let us discuss different measures of dispersion.
a) Range
It is the simplest measure of dispersion. The range of a band of imagery is
calculated as the difference between the maximum and minimum values.
Let us take the following image array –

Contributor: Dr Manu Mehta and Mr. Akhilesh Kumar 159

Block 2 Image Pre-processing Techniques
…………………………………………...............……………………….……………………….................……

Here the maximum valve = 30 and minimum value = 10, hence, the range = 30
– 10 = 20
But just like the mean, range is impacted by the presence of outliers in the
dataset. Consider the following dataset –

Now, the maximum value = 150 and minimum value = 10, hence, the range =
150 – 10 = 140.
But from the above dataset, it is evident that most values lie around 20±10 and
150 is the outlier. So, range is not a very reliable measure in case outliers are
present in the dataset.
b) Variance
Variance is described as the average of the squared distance of data points
from their mean. Just like the range, variance too is affected by the presence of
outliers. Variance for a band of imagery can be calculated as –

For sample, variance is calculated as –

Where, is the variance and is the standard deviation. The other terms have
the same meaning as in Equations 2a and 2b.
Let us calculate the variance by considering the following dataset

A=

In order to calculate variance, we first need to calculate the mean of the

dataset.

Mean = , then variance =

To understand the importance of variance, consider the following as an

example. There are two Images – A and B of size 3x3. With all pixel values
equal to 255, A represents a mostly white image, while in B the first two rows
are black and the third row is white.

A= , and B =

Now, Mean of A = = 50, and

160 Contributor: Dr Manu Mehta and Mr. Akhilesh Kumar

Unit 7 Image Statistics
…………………….................……………..………...............………………………….…………………………

Mean of B = = 85.Similarly, variance of A =

= 0, and

Variance of B = = 3211.11

So, while the means of both images are quite close, the variance indicates that
the variability in image B is much more than that in image A. As such, mean
alone could not provide the conclusive information about an image.
c) Standard Deviation
Standard deviation is simply the square root of the variance. Since standard
deviation has the same units, unlike variance, as the variable, a direct
comparison between the two is easily possible.
The standard deviation of the brightness value of pixels in band imagery is
given by –

If sample is considered instead of population, then standard deviation –

Where, the terms have same meanings as in Equations 2a, 2b, 3a, and 3b.

Considering the dataset from the previous section, the standard deviation =
. Hence, the average distance of the data points from
the mean is 2.58.

The standard deviation is routinely used in digital image processing for

performing linear contrast enhancement, spatial filtering, parallelepiped image
classification and for calculating accuracy assessment of a classified image
(Jenson, 2018). However, both mean and standard deviations are to determine
the degree of similarity or dissimilarity among neighbouring pixels in a digital
remote sensing image. If the standard deviation is closer to 1, then the pixels
are homogenous. Standard deviation farther away from 1 shows heterogeneity
in the pixels.
In most cases, the image statistics is commonly represented with the help of its
mean and standard deviation. The variance, though useful, is less widely used.
iii) Measures of Distribution (Histogram)
Measures of distribution provide information on the shape, symmetry and
concentration of individual data points in a dataset/histogram. Two of the most
used measures of distribution are –
a) Skewness
Skewness is the measure of asymmetry of a dataset or histogram. It is used to
determine the extent and direction of the deviation from the asymmetry
distribution. Skewness can be calculated using the following equation –
Contributor: Dr Manu Mehta and Mr. Akhilesh Kumar 161
Block 2 Image Pre-processing Techniques
…………………………………………...............……………………….……………………….................……

Here, is the number of observations, is the mean.

A skewness with value of 0 indicates a perfect symmetry, a positive skewness
indicates a longer tail on the right side of the distribution with more values
concentrated on the left, and negative skewness indicates a longer left tail with
values concentrated more on the right side of the distribution.
Let us calculate the value of skewness by considering the following dataset:

Its mean = , the skewness can be calculated as –

The skewness value of 1.39 indicates that the data points are positively skewed
towards the right side of the distribution.
b) Kurtosis
Kurtosis measures the sharpness of the peak and strength of the tails. Kurtosis
can be calculated as –

A kurtosis with value of 0 indicates a normal distribution, a positive kurtosis

(also called as leptokurtic distribution) indicates heavier tails, thereby indicating
presence of more extreme values, and a negative kurtosis (also called as
platykurtic distribution) is an indicative of lighter tails and a fewer extreme value.
Let us calculate the value of kurtosis by considering the following dataset:

Its mean = 6. So, kurtosis

The negative kurtosis value indicates that while the distribution has a relatively
flatter peak, the tails are lighter and the dataset does not appear to have
extreme values.
iv) Representation and Visualisation of Univariate Statistics
Univariate statistics are important for an exploratory analysis of the image or
dataset and some of the most common graphical ways of representing them
include the following:

162 Contributor: Dr Manu Mehta and Mr. Akhilesh Kumar

Unit 7 Image Statistics
…………………….................……………..………...............………………………….…………………………
 Histogram, which displays the frequency distribution of the variable, with
each bin representing the number of occurrences within that interval.
 Boxplot graphically summarises the distribution of the dataset using the
minimum, first quartile, median (second quartile), third quartile and
maximum range of the dataset and is very effective at visualising the
outliers in the dataset.
 Probability distribution function (PDF) can also be used to visualise shape of
the distribution.
The univariate statistics discussed above can also be calculated using python
or other programming languages. Fig. 7.1 shows the code snippets to create a
sample dataset of 1000 random points from a normal distribution and calculate
the common statistics. The outputs have also been displayed. The sample
dataset was also visualised using different graphical methods shown in Fig. 7.2.

Fig. 7.1: Calculation of common univariate statistics using python. Each band in
remote sensing images is mostly a 2-D array and the commands used
here can also be used on images, using suitable geospatial libraries like
rasterio or GDAL.

Contributor: Dr Manu Mehta and Mr. Akhilesh Kumar 163

Block 2 Image Pre-processing Techniques
…………………………………………...............……………………….……………………….................……

Fig. 7.2: Graphical representation of the sample dataset shown in Fig. 7.1 using
different visualisation techniques.

SAQ I
a) Define mode.
b) What does a data with negative skewness suggest?

7.4.2 Multivariate Image Statistics

In remote sensing, it is also important to measure of how much radiant energy
is emitted or reflected back from a feature/object in more than one band
(Jensen, 2018). The measures of multivariate image statistics among the bands
allow us to know how the measurements covary. In this section, we look at
different multivariate statistical techniques like covariance, correlation,
regression and principal component analysis.
a) Covariance
In many digital image processing techniques, understanding the relationship
between any two bands is an important aspect of image classification and
information extraction. This measure is known as covariance and it simply
refers to a measure of the joint variance between two variables. Covariance can
be positive or negative depending on the variability of one variable with respect
to other. If the change in both variables is in the same direction, covariance is
positive, else it is negative.
Just like variance, covariance is the average product of the differences in the
values of both variables from their respective means. The unit, too, is squared.
The covariance is generally expressed as a covariance matrix of dimension
with n being the number of variables, or bands. Covariance can be calculated
as –

164 Contributor: Dr Manu Mehta and Mr. Akhilesh Kumar

Unit 7 Image Statistics
…………………….................……………..………...............………………………….…………………………
While are mean values, and are pixel values at location
and is the number of observations.
Let us calculate the value of covariance of two image array A and B of size 2x2
as given below –

A= , and B =

The covariance matrix would look like –

Since the diagonal elements in the covariance matrix represent the covariance
between the same variables (x or y), it can also be said that the diagonal
elements of a covariance matrix are simply the variances of x or y and the
covariance matrix can also be written as –

Since and = 2.75, the covariance can be calculated as shown in Table 7.1.
Table 7.1: Calculation of terms used for determining covariance matrix

Pixel

(1,1) 5 2 2.25 -0.75 -1.69 5.06 0.56

(1,2) 4 3 1.25 0.25 0.31 1.56 0.06

(2,1) 1 5 -1.75 2.25 -3.94 3.06 5.06

(2,2) 1 1 -1.75 -1.75 3.06 3.06 3.06

Then, or = = 4.24,

or = = 2.91, and

= = = - 0.75

The covariance matrix can then be filled as

Covariance matrix suggests that A and B share a negative relationship.

b) Correlation coefficient
You are aware that covariance depends on the variance, but correlation
coefficient removes this dependency. It is a standardised measure of the
relationship of one variable with the other. Unlike covariance, correlation
coefficient provides insights on both the strength and direction of the
relationship between the variables and can be positive, indicating a change in
the same direction, or negative, if the change is in opposite direction.
Furthermore, correlation coefficient, being a scaled measure of covariance, is a
dimensionless quantity with values ranging between -1 and 1. The strength of

Contributor: Dr Manu Mehta and Mr. Akhilesh Kumar 165

Block 2 Image Pre-processing Techniques
…………………………………………...............……………………….……………………….................……
the dependency is indicated by how close the value is to either 1 or -1. There
are two common methods of calculating correlation coefficient as listed below:
 Pearson correlation coefficient: It works with continuous variables and
only evaluates the linear relationship between variables. It can be
calculated as –

 Spearman correlation coefficient: It can be used for both continuous and

ordinal variables. It is used to evaluate any monotonic relationships. Since
Spearman correlation is based on rank order, it can be calculated using the
following steps –
 Rank the values for both X and Y in some order. Generally, this is based
on the supposed hierarchy of one value over the other.
 Calculate the difference between the assigned rank of X and Y, and
square it.
 Use the equation below to calculate the Spearman correlation coefficient

While the sign of correlation coefficient indicates the nature of association, the
value indicates its strength.
Let us consider the following image arrays for calculating then Pearson
correlation –

A= and B = , then Pearson correlation coefficient can

be calculated by calculating various terms as shown in Table 7.2.
Table 7.2: Calculation of terms used for determining Pearson correlation
coefficient.

Pixel
number
(1,1) 7 12 49 144 84
(1,2) 6 8 36 64 48
(1,3) 8 12 64 144 96
(2,1) 5 10 25 100 50
(2,2) 6 11 36 121 66
(2,3) 9 13 81 169 117

So,

166 Contributor: Dr Manu Mehta and Mr. Akhilesh Kumar

Unit 7 Image Statistics
…………………….................……………..………...............………………………….…………………………

Hence, both images show a strong positive correlation of 0.76.

Let us consider two thematic images (or arrays), in which one represents
education levels (an ordinal data)and another gives income levels (a ratio data)
for calculating Spearman correlation coefficient.

If the educational level (X) = , and

income (Y) = , then Spearman correlation coefficient (Table 7.3)

can be calculated as –
Table 7.3: Calculation of Spearman correlation coefficient between two
images representing education and income levels.

Original Assigned Original Assigned

Pixels Rank Rank Rank Rank

(1,1) Preparatory 5 5.0 25 2 2.0 9.00

(1,2) Primary 4 4.0 10 4 4.5 0.25
(1,3) University 1 1.0 8 6 6.0 25.00
(2,1) Secondary 2 2.5 10 5 4.5 4.00
(2,2) Secondary 3 2.5 15 3 3.0 0.25
(2,3) Illiterate 6 6.0 50 1 1.0 25.00
So,

Hence, there appears to be a strong negative correlation between the level of

education and the income levels.
In order to calculate covariance and correlation between different bands of a
remotely sensed data, we can use image processing and GIS software like
ENVI, ArcMap or we can use programming languages like R and Python. While
the example shown below (Fig. 7.3) generates 1000 random pixels for 4
hypothetical bands (Fig. 7.4), they can easily be replaced by actual bands once
they have been converted to arrays using suitable geospatial libraries like
rasterio.

Contributor: Dr Manu Mehta and Mr. Akhilesh Kumar 167

Block 2 Image Pre-processing Techniques
…………………………………………...............……………………….……………………….................……

Fig. 7.3: Calculating covariance and correlation for multiple bands of a remotely
sensed image in python. To use actual bands, convert them to arrays
and substitute them in line 10-15.

The correlation and covariance matrices can be visualised using histograms

and heatmaps (Fig. 7.4).The histogram shows that while band 1 has similar
frequency distribution as band 3, the distribution of band 2 is similar to band
168 Contributor: Dr Manu Mehta and Mr. Akhilesh Kumar
Unit 7 Image Statistics
…………………….................……………..………...............………………………….…………………………
4(Fig. 7.4a-d). In the covariance heatmap, the diagonal elements represent
variances of each band and the off-diagonal elements represent the covariance
between pairs of bands. From Fig. 7.4e, it appears that while some bands
share a positive or direct relationship (e.g., band 1 and 3, 3 and 4, etc.), others
share a negative relationship (e.g., band 1 and 2, 2 and 4,
et

Fig. 7.4: Plot showing histogram of each sample band (a-d) and the heatmaps of
(e) covariance and (f) correlation matrix.

The correlation heatmap (Fig. 7.4f), on the other hand, indicates a mostly weak
or no correlation between bands, at least in a linear sense. This is expected
since the pixel values in the example has been generated randomly
c) Regression
Regression is a statistical measure that attempts to determine the relationship
between a dependent variable (generally referred to as Y) and one or more
independent variables (generally referred to as X). Regression can be used to
predict the dependent variables if the corresponding value of the independent
variable is known.
Contributor: Dr Manu Mehta and Mr. Akhilesh Kumar 169
Block 2 Image Pre-processing Techniques
…………………………………………...............……………………….……………………….................……
By using the least squares method, regression finds the best fitting line and the
corresponding regression equation. Slope and intercept are the two most
important measures that are obtained through regression. While the slope
indicates the rate of change of dependent variable with respect to the
independent variable, intercept provides information regarding the initial offset
value for the dependent variable. A regression equation is given as–

Here, is the y-intercept, , and is the slope given by the

equation(Chatterjee and Hadi, 2006) –

Let us consider an array X representing the independent variables

= , and an array Y representing the dependent variables

= ,

As shown in Table 7.4 for finding the regression equation, first, we need to
estimate the slope and y-intercept.
Table 7.4: Calculation of squared and product terms of each independent
and dependent variables used in the regression.

Pixels X Y
(1, 1) 7 12 49 84
(1, 2) 6 8 36 48
(1, 3) 8 12 64 96
(2, 1) 5 10 25 50
(2, 2) 6 11 36 66
(2, 3) 9 13 81 117
(3, 1) 4 9 16 36
(3, 2) 3 5 9 15
(3, 3) 10 15 100 150
(4, 1) 6 12 36 72
(4, 2) 6.4 10.7 40.96 68.48
(4, 3) 6.4 10.7 40.96 68.48
From the above data, 6.4, = 10.7, so –

170 Contributor: Dr Manu Mehta and Mr. Akhilesh Kumar

Unit 7 Image Statistics
…………………….................……………..………...............………………………….…………………………

And, taking x = 7 (i.e., first observation), , or

So,
Hence, the regression equation is y = 3.273 + 1.1604x. Using this equation, let
us try to predict the value y at x = 8.5.
i.e., .

The regression line along with the data points have been represented as a
scatterplot shown in fig. 7.5. The predicted value at x = 8.5 has also been
marked in Fig. 7.5. R is the Pearson correlation coefficient between X and Y.

Fig. 7.5: Scatterplot showing the distribution of data and the regression line. The
regression equation, correlation coefficient and the predicted value at x
= 8.5 has also been marked.

d) Visualisation of Multivariate Image Statistics

Some of the most common ways of representing the multivariate image
statistics have already been discussed. To summarise, here are some of the
possible choices of visualisations –
 Heatmap: It is a tiled plot which is extensively used to represent
covariances or correlations between two or more variables. Fig. 7.4 shows
two such heatmaps, one for covariance and another for correlation.
 Scatterplot: It is a useful tool to represent the pair-wise comparison of two
variables and is the preferred choice of visualisation for correlation or
regression. Fig. 7.5is an example of scatterplot showing the linear
regression between two variables.
Contributor: Dr Manu Mehta and Mr. Akhilesh Kumar 171
Block 2 Image Pre-processing Techniques
…………………………………………...............……………………….……………………….................……

SAQ II
a) If y=-3x; how are y and x correlated?
b) What is the covariance between two independent random variables?

7.5 ACTIVITY
Read the first four bands of Landsat 8 using python and calculate the following
statistics:
a) Univariate statistics – mean, variance and standard deviation
b) Find the correlation between green and red band as well as between red
and NIR band. Analyse the results.
Hint: Use `rasterio` library (ref.: https://rasterio.readthedocs.io/en/stable/) to
read and convert the data into `numpy` arrays and then use the code snippet
given in Fig. 7.1 and Fig. 7.3.

7.6 SUMMARY
 Statistics is a branch of applied mathematics that deals with the collection,
organisation, presentation, analysis and interpretation of data.
 In digital images, statistics mainly deals with the probability distribution
function of the brightness levels of a particular image and density function of
the brightness levels of the same image.
 Image statistics involves the study of remotely sensed images and their
statistical properties. It calculates raster cell statistics from one or multiple
selected images.
 Data can be categorised into two types such as qualitative and quantitative.
 There are four levels of measurement namely, nominal, ordinal, interval and
ratio.
 Univariate descriptive image statistics describes and summarises individual
image variable separately
 Though common univariate statistics like mean, variance and standard
deviation are affected by the presence of outliers; median and IQR are not.
 These methods provide a measure of the central tendency and dispersion
of the data. The distribution of a dataset is best captured through skewness
and kurtosis
 The measures of multivariate image statistics among the bands allow us to
know how the measurements covary.
 Multivariate statistics, on the other hand, use methods like covariance and
correlation to measure the strength and direction to explore associations,
patterns, and dependencies between more than one variable.
 The choice of an appropriate statistical technique is dependent not just on
data type or number of variables in the question, but also on the intended
application
172 Contributor: Dr Manu Mehta and Mr. Akhilesh Kumar
Unit 7 Image Statistics
…………………….................……………..………...............………………………….…………………………

7.8 TERMINAL QUESTIONS

1. Define image statistics. Discuss why it is important in digital image
processing.
2. Describe various types of statistical analysis.
3. Discuss various measures of central tendency in remotely sensed data.
4. Write notes on covariance and correlation.

7.9 REFERENCES
 Burger W and Burge MJ (2022) Histograms and Image Statistics. In: Digital
Image Processing. Texts in Computer Science. Springer, Cham.
https://doi.org/10.1007/978-3-031-05744-1_2
 Chatterjee S andHandi AS (2006) Regression analysis by example, John
Wiley & Sons.
 Jensen JR (2018) Digital Image Processing: A Remote Sensing
Perspective, 4th edition, Pearson India Education Services Pvt Ltd, Uttar
Pradesh, India.
 Stommel M andKatherine JD (2014) Properties of Variables: Levels of
Measurement. Springer Publishing Company,New York.

7.9 FURTHER/SUGGESTED READINGS

 Bhatta B (2008) Remote Sensing and GIS, Oxford University Press, USA.
 Datar R and Garg H (2019) Hands-On Exploratory Data Analysis with R:
Become an expert in exploratory data analysis using R packages, Packt
Publishing Ltd.
 Everitt B andHothorn T (2011) An introduction to applied multivariate
analysis with R, Springer Science & Business Media.
 Mukhiya SK and Ahmed U (2020) Hands-On Exploratory Data Analysis with
Python: Perform EDA techniques to understand, summarize, and
investigate your data, Packt Publishing Ltd.

7.10 ANSWERS
SAQ I
a) Mode is defined as the value occurring at the most in a dataset.
b) Negative skewness indicates a longer left tail with values concentrated
more on the right side of the distribution.

SAQ II
a) y and x are negatively correlated.

b) The covariance between 2 independent random variables is zero. At the

same time, a zero covariance between 2 random variables does not imply
that they are independent.

Contributor: Dr Manu Mehta and Mr. Akhilesh Kumar 173

Block 2 Image Pre-processing Techniques
…………………………………………...............……………………….……………………….................……
Terminal Questions
1. Please refer to section 7.2.
2. Please refer to subsection 7.3.4.
3. Please refer to subsection 7.4.1.
4. Please refer to subsection 7.4.2.

174 Contributor: Dr Manu Mehta and Mr. Akhilesh Kumar

 UNIT 8

RADIOMETRIC IMAGE CORRECTIONS

Structure_____________________________________________________
8.1 Introduction 8.5 Non-Systematic Radiometric Errors and
their Corrections
Expected Learning outcomes
Atmospheric Modelling
8.2 Concept of Image Distortion and Correction
Image Based Methods
Image Distortions
Direct Calibration Using Field-Derived
Image Corrections
Reflectance
8.3 Nature of Radiometric Errors
8.6 Activity
8.4 Systematic Radiometric Errors and their
8.7 Summary
Corrections
8.8 Terminal Questions
Random Bad Pixels (Shot Noise)
8.9 References
Line or Column Dropouts
8.10 Further/Suggested Readings
Line Start Problems
8.11 Answers
N-Line Striping

8.1 INTRODUCTION
You are aware that remote sensing images are commonly acquired by sensors mounted on remote
sensing platforms such as airplanes, unmanned aircraft or satellites. There is no perfect remote
sensing system that records absolutely correct, error free view of the ground surface. The
complexity of the interaction between earth and atmosphere with energy often introduces errors and
noise in the remote sensing image data during the scanning process. Thus, the obtained raw
images are susceptible to a variety of distortions, which degrade image quality and finally, reduce
accuracy of the image analysis. Therefore, it is required to preprocess the raw images in order to
make them suitable for application.
Block 2 Image Pre-processing Techniques
…………………………………………………………………….……………………….................……
Normally, distortions can affect the values recorded by the sensors differently in
different wavelengths. The geometry of the pixel as well as alignment of an
image with other images and also with reference maps is affected. It is
important to know the effects of these distortions on the images and apply
suitable corrections to minimise them. There are two types of errors namely,
radiometric and geometric that are present in the raw images and their removal
is known as image correction. In this unit, we will discuss the types and causes
of radiometric distortions in remote sensing images and methods of their
corrections to make images able to represent the terrain more closely. We will
discuss geometric corrections in next unit (i.e. Unit 9).

Expected Learning Outcomes______________________________

After reading this unit, you should be able to:
 describe the concept of distortions suffered by remotely sensed images;
 discuss types and causes of radiometric distortions; and
 explain the requirements and approaches of radiometric corrections.

8.2 CONCEPT OF IMAGE DISTORTION AND

CORRECTION
Radiometric and geometric distortions are two common types of distortions that
occur in remotely sensed image data. These distortions are developed due to
the variations in solar radiation incident on the curved earth’s surface, imperfect
transparency of atmosphere and sensor imperfections. First, let us first get
introduced to the concept of image distortion.
8.2.1 Image Distortions
Any kind of errors present in remote sensing images are known as image
distortions. Any remote sensing images acquired from either space-borne or
air-borne platforms are susceptible to a variety of distortions. These distortions
occur due to data recording procedure, shape and rotation of the Earth relative
to the remote sensor platform and environmental conditions prevailing at the
time of data acquisition.
Distortions occurring in remote sensing images can be categorised into two
types:
 radiometric distortions and
 geometric distortions.
Sensors record the intensity of electromagnetic radiation as digital numbers
(DNs). These digital numbers are specific to the sensor and conditions under
which they were measured. However, there are variations in the pixel intensities
(digital numbers) irrespective of the object or scene being scanned. The
recorded values get distorted due to one or more of the following factors:
 sensor ageing
 random malfunctioning of the sensor elements
 atmospheric interference at the time of image acquisition and
 topographic effects.
176 Contributors: Prof. B.K. Mohan and Dr. Omkar Verma
Unit 8 Radiometric Image Corrections
…………………….................……………..………………………………….…………………………
The above factors affect radiometry (variation in the pixel intensities) of the
images and resultant distortions are known as radiometric distortions.
As you know, an image is composed of finite number of pixels. The positions of
these pixels are initially referenced by their row and column numbers. However,
if you want to use images, you should be able to relate these pixels to their
positions on the earth surface. Further, the distance, area, direction and shape
properties vary across an image, thus these errors are known as geometric
errors/distortions. This distortion is inherent in images because we attempt to
represent three-dimensional earth surface as a two-dimensional image.
Geometric errors originate during the process of data collection and vary in type
and magnitude (Wilkie and Finn, 1996). There are several factors causing
geometric distortions such as:
 Earth’s rotation
 Earth’s curvature
 satellite platform instability and
 instrument error.
Let us now learn about the concept of image correction.
8.2.2 Image Corrections
As you now know that raw remote sensing images always contain significant
amount of distortions, therefore, they cannot be used directly for further image
analysis. The image correction involves image operations which normally
precede manipulation and analysis of image data to extract specific information.
The primary aim of image correction operations is to correct distorted image
data to create a more accurate representation of the original scene. Image
corrections are also known as a pre-processing of remotely sensed images. It is
a preparatory phase that improves quality of images and serves as a basis for
further image analysis. It may be noted that modeling of the radiometric and
geometric distortions and consequent corrections of distortions falls in the
category of preprocessing of remotely sensed imagery.
Depending upon the kinds of errors which are present in images, the image
correction functions are comprised of radiometric and geometric corrections.
Radiometric correction attempts to improve the accuracy of measurements
made by remote sensors pertaining to the spectral reflectance or emittance or
back-scatter from the objects on the earth surface.
Geometric correction is the process of correcting geometric distortions (pixel
position and shape errors) and assigning the properties of a map to an image.
Both of these corrections are made prior to actual use of remote sensing data in
resource management, environmental monitoring, disaster mitigation and
change detection by application scientists.
A complete chain of processing of remote sensing images is shown in Fig. 8.1.
It becomes clear from this figure that image correction forms an integral part of
processing of images.

Contributors: Prof. B.K. Mohan and Dr. Omkar Verma 177

Block 2 Image Pre-processing Techniques
…………………………………………………………………….……………………….................……

Fig. 8.1: Chain of broad steps in remote sensing data processing.

8.3 NATURE OF RADIOMETRIC ERRORS

You have read earlier that radiometric distortions relate to the distortions
suffered by the images in recorded values at different pixel locations. Let us
now discuss in detail the types of radiometric errors and their correction
procedures. The radiometric errors listed in subsection 8.2.1 can be broadly
categorised into the following two categories:
 internal errors and
 external errors.
Internal errors are introduced by the electronics or remote sensing systems
themselves. These kinds of errors are also known as systematic errors because
of their systematic or predictable nature. These errors can be modeled,
identified and corrected based on laboratory calibration or in-flight
measurements. For example, if a single detector has become uncalibrated, the
concerned row (in older satellites such as Landsat) or the column (in
pushbroom scanners like SPOT, IRS, IKONOS, WorldView-1-4) would appear
like a constant intensity stripe, that does not reflect the terrain changes on the
ground.
External errors are a result of phenomena that vary in nature through space
and time and hence are also known as non-systematic errors. External
variables such as atmospheric disturbances, steep terrain undulations can
cause remote sensor data to exhibit radiometric and geometric errors. It is
important to note that the identification of both internal and external errors is
required for correcting the image data.
Correction of radiometric errors requires knowledge about electromagnetic
radiation principles and the interactions that take place during data acquisition
process. The radiometric correction can benefit from the terrain information
such as slope and aspect and advanced information like bi-directional
reflectance distribution function (BRDF) characteristics of the scene. The BRDF
is a function which describes the magnitude of the upwelling radiance of the
target in terms of illumination angle and the angle of view of the sensor.
Radiometric correction procedures can be time consuming and at times
problematic.
We shall now discuss the two types of radiometric errors and their correction
procedures in the next two subsections.

178 Contributors: Prof. B.K. Mohan and Dr. Omkar Verma

Unit 8 Radiometric Image Corrections
…………………….................……………..………………………………….…………………………

SAQ I
a) List two types of image corrections.
b) What are internal errors?

8.4 SYSTEMATIC RADIOMETRIC ERRORS AND

THEIR CORRECTIONS
The systematic radiometric errors are also known as remote sensing system
induced radiometric errors. These errors are introduced due to earth rotation,
scan skew, velocity variation in platform and mirror scan, malfunctioning of
individual scanner/detector, aspect ratio and improper sensor calibration. Some
of the commonly observed systematic radiometric errors are listed below
(Jensen, 2018):
 Random bad pixels
 Line or column drop-outs
 Line start problems
 N-Line striping
We shall now discuss here these errors and their corrections.
8.4.1 Random Bad Pixels (Shot Noise)
Occasionally, an individual detector does not record received signal for a pixel.
This may result in random bad pixels. When there are numerous random bad
pixels found within the scene, it is called shot noise. Shot noise gives the
image an impression of having many dark poke marks. Generally, these bad
pixels contain values in the range of 0 or 255 (in 8-bit data) in one or more
bands. Shot noise is removed by identifying such pixels in a given band that are
either 0 (black) or 255 (white) in the midst of neighbouring pixel values that are
radically different. These noise pixels are then replaced by the average pixel
value of their immediate eight neighbouring pixels.
Take an example of the Landsat Thematic Mapper band 7 image, which is
given in Fig. 8.2. It has two of the pixels having zero gray levels along a bad
scan line, which is entirely different from their neighbouring pixels (Fig. 8.2a, b).
These are marked as shot noise pixels and are replaced by the average of their
eight neighbouring pixels (Fig. 8.2c).

Contributors: Prof. B.K. Mohan and Dr. Omkar Verma 179

Block 2 Image Pre-processing Techniques
…………………………………………………………………….……………………….................……

(a)

(b) (c)
Fig. 8.2: Illustration of shot noise and its removal. a) a Landsat TM band 7 data
with shot noise (two black dots in the region within the white box); b)
zoomed image of the box portion showing the bad pixels along with DNs
of the neighbouring eight pixels for each bad pixel; and c) the same
image portion after the shot noise removal showing the pixel values
which has replaced the bad pixels. (Source: Lecture slides of Prof. J. R.
Jensen, University of South Carolina; used with permission)
If pixel at location (m,n) is a shot noise pixel, then f(m,n) is corrected as given
below:
f(m,n) = [f(m–1,n–1) + f(m–1,n) + f(m–1,n+1) + f(m,n-1) + f(m,n+1) + f(m+1,n–1)
+ f(m+1,n) + f(m+1,n+1)] / 8
where,
m and n are pixel locations in x and y coordinates, i.e. columns and
rows, respectively.
By replacing with average of neighbouring pixels like (15+17+16+13+16+16+
15+14) / 8 = 15 and (14+16+19+17+14+17+16+15) / 8 = 16), shot noise pixels
disappear after the correction, as seen in Fig. 8.2c.
180 Contributors: Prof. B.K. Mohan and Dr. Omkar Verma
Unit 8 Radiometric Image Corrections
…………………….................……………..………………………………….…………………………
8.4.2 Line or Column Dropouts
You may have noticed that a blank row containing no details of features on the
ground may be seen if an individual detector in an electro-mechanical scanning
system (e.g., Landsat MSS or Landsat 7 ETM+) fails to function properly. If a
detector in a pushbroom linear array (e.g., IRS-1C, QuickBird) fails to function,
this can result in an entire column of data with no spectral information. The bad
line or column is commonly called a line or column drop-out and contains
brightness values equal to zero or some constant value, independent of the
terrain changes. Generally, this is an irretrievable loss of information because
there is no way to restore data that were never acquired. However, it is possible
to improve visual interpretability of data by introducing estimated brightness
values for each bad scan row (or column) by replacing them with the average of
rows (or columns) above and below (or to the left and right). This concept works
because adjacent pixels often have similar pixel values.
8.4.3 Line Start Problems
There is another kind of problem encountered in earlier satellites in which the
scanner fails to start recording as soon as a new row starts. It may also happen
that the sensors place pixel data at inappropriate locations (with shift) along the
scan line. For example, all of the pixels in a scan line might be systematically
shifted just one pixel to the right. This is called a line-start problem. If line start
problem is always associated with a horizontal bias of a fixed number (say 50)
of columns, it can be corrected using a simple horizontal adjustment. It is found
difficult to restore the image data, if the quantity of the line start displacement is
random (Jensen, 2018).
8.4.4 N-Line Striping
Occasionally, a detector does not fail entirely but its calibration parameters
(gain and offset/bias) are disturbed. For example, a detector might record
signals over a dark, deep body of water that are almost uniformly 20 or 30 gray
levels higher than the other detectors for the same band. The result would be
an image with systematic, noticeable lines that are brighter than adjacent lines.
This is referred to as n-line striping. The affected line contains valuable
information but should be corrected to have approximately the same
radiometric quality as data collected by properly calibrated detectors associated
with the same band (Fig. 8.3).

(a) (b)
Contributors: Prof. B.K. Mohan and Dr. Omkar Verma 181
Block 2 Image Pre-processing Techniques
…………………………………………………………………….……………………….................……

(c) (d)

Fig. 8.3: Illustration of image destriping on a sample remote sensing data. a)

Original image data containing bad lines (stripes); b) zoomed portion of
the box region of the image (a) showing the stripes; c) sample data after
destriping; and d) zoomed portion of the box region of image c. (Source:
Lecture slides of Prof. J.R. Jensen, University of South Carolina; used with
permission)

To repair systematic n-line striping, it is necessary to identify mis-calibrated

scan lines in the scene. This is usually accomplished by computing a histogram
of values for each of the n detectors that collected data over the entire scene
(ideally, this would take place over a homogeneous area like a water body). If
one detector’s mean or median DN value is significantly different from the
others, there is a probability that this detector is out of adjustment.
Consequently, every line and pixel in the scene recorded by the maladjusted
detector may require a correction. This type of n-line striping correction:
 adjusts all the bad scan lines so that they have approximately the same
radiometric scale as the correctly collected data and
 improves visual interpretability of the data.
Let us now discuss non-systematic errors and their correction in the next
section.

8.5 NON-SYSTEMATIC RADIOMETRIC ERRORS

AND THEIR CORRECTIONS
It is noted that the radiometric errors also enter into the remotely sensing data
even when remote sensing systems work properly. These errors constitute non-
systematic errors and are developed due to i) atmospheric scattering and
absorption and ii) topographic attenuation. It is essential to carry out corrections
for non-systematic errors in the following circumstances:
 if you are trying to compare remote sensing images which have been
acquired at different times
 if you are modelling interactions between electromagnetic radiation and a
surface feature, or
 using band ratios for image analysis.
Correction of non-systematic errors include following three steps (Fig. 8.4):

182 Contributors: Prof. B.K. Mohan and Dr. Omkar Verma

Unit 8 Radiometric Image Corrections
…………………….................……………..………………………………….…………………………
Step 1: Involves the conversion of DNs to sensor spectral radiance. This step
requires information on the ‘gain’ and ‘bias’ of the sensor in each image band.
The ‘gain’ and ‘bias’ are the sensor calibration information. Bias is the spectral
radiance of the sensor for a DN of zero and gain represents the gradient of the
calibration. The sensor calibration is carried out before the launch of the sensor.
Sensor calibration includes procedures that convert digital numbers to physical
values of radiance. The relationship between DN recorded at a given location
and reflectance of the material making up the surface of pixel area changes
with time hence the coefficient values that are used to calibrate image data also
vary with time.

Fig. 8.4: Steps in non-systematic radiometric error correction process.

Step 2: Is the conversion of spectral radiance to apparent reflectance. It

converts DN values from radiance units to apparent reflectance. Apparent
reflectance defines as the reflectance at the top of the atmosphere.
Step 3: Involves the removal of atmospheric effects due to the absorption and
scattering of light (atmospheric correction). There are several methods for
atmospheric correction of the remotely sensed data. Some methods are
relatively straight forward while others are based on the physical principles of
interaction of radiation with atmosphere, and require a significant amount of
information pertaining to the atmospheric conditions to be effective. For your
convenience we can categorise atmospheric correction procedures into the
following three:
 atmospheric modelling
 image based methods and
 direct calibration using field-derived reflectance (empirical line method)..
8.5.1 Atmospheric Modelling
In this approach, atmospheric radiative transfer codes (models) are used that
can provide realistic estimates of the effects of atmospheric scattering and
absorption on satellite imagery. Once these effects have been identified for a
specific date of imagery, each band and/or pixel in the scene can be adjusted to
remove the effects of scattering and/or absorption. The image is then
Contributors: Prof. B.K. Mohan and Dr. Omkar Verma 183
Block 2 Image Pre-processing Techniques
…………………………………………………………………….……………………….................……
considered to be atmospherically corrected. Unfortunately, application of these
codes to a specific scene and date also requires knowledge of both the sensor
spectral profile and atmospheric properties at the same time. And, for most of
the historic satellite data, they are not available. Ideally, this approach is used
when scene specific atmosphere data (such as aerosol content, atmospheric
visibility) are available.
Most current radiative transfer based atmospheric correction algorithms can
compute much of the required information, if
 the user provides fundamental atmospheric characteristic information to the
programme or
 certain atmospheric absorption bands are present in the remote sensing
dataset.
For example, most radiative transfer based atmospheric correction algorithms
require that the user provides
 latitude and longitude of the image scene
 date and exact time of the image
 image acquisition altitude (e.g., 600 km above the ground level)
 mean elevation of the scene (e.g., 450 m above the mean sea level)
 an atmospheric model (e.g., polar summer, mid-latitude winter, tropical)
 radiometrically calibrated image radiance data
 data about each specific band (mean and full width at half-maximum
(FWHM)). The FWHM is the wavelength range defined by the two points at
which the intensity level is 50% of its peak value.
 local atmospheric visibility at the time of remote sensing data collection
These parameters are then input to the atmospheric model selected and used
to compute the absorption and scattering characteristics of the atmosphere at
the time of remote sensing data collection. These atmospheric characteristics
are then used to invert the remote sensing radiance to scaled surface
reflectance. Many of these atmospheric correction programmes derive the
scattering and absorption information they require from robust atmosphere
radiative transfer code such as
 MODTRAN 4
 ACORN
 ATCOR
 ATREM and
 FLAASH
8.5.2 Image Based Methods
One of the commonly used methods is known as dark pixel subtraction method
(Chavez, 1988). This method is generally used prior to band-ratioing a single
image and is not generally employed for image to image comparisons. This
method of radiometric correction is based on simple heuristics, used to reduce
the effect of haze in the image. The underlying assumption is that some image
184 Contributors: Prof. B.K. Mohan and Dr. Omkar Verma
Unit 8 Radiometric Image Corrections
…………………….................……………..………………………………….…………………………
pixels have a reflectance of zero and DN values of these “zero” pixels recorded
by the sensor result from atmospheric scattering (path radiance; see Fig. 8.5).
To remove path radiance, minimum pixel value for each band is subtracted
from all other pixels in different bands. DNs in pixels representing deep clear
water in near-infrared (NIR) band and dark shadows in visible bands are
assumed to result from atmospheric path radiance. Since histogram offset can
be used as a measure of path radiance, it is also known as histogram minimum
method. Bi-plots of NIR band against the other bands are generated for pixels
of dark regions and then regression techniques are used to calculate the y-
intercept which represents path radiance in other bands. The y-intercept value
is then subtracted from all pixels in the image.

Fig. 8.5: Histogram with an offset (DN = 31) in brightness values.

When multi-temporal imagery is being used for studying changes taking place
in a study area, common radiometric quality is required for quantitative analysis
of multiple satellite images of a scene acquired on different dates with same or
different sensors. The radiometric quality could also be rectified using one
image as a reference image. Transformed images appear to have been
acquired with the reference image sensor, under atmospheric and illumination
conditions nearly identical to those in the reference scene.
In order to achieve this, a few sets of scene landscape elements with a mean
reflectance which is (almost) time invariant are identified. These elements are
also known as pseudo-invariant features. The average gray level values of
these reference sets are used to calculate a mathematical mapping relating the
gray levels between the reference and the remaining images.

8.5.3 Direct Calibration Using Field-Derived Reflectance

This method is based on the assumption that reflectance measured in one
region for a particular feature is directly applicable to the same feature
occurring in other regions. The method requires some field work to measure
true ground reflectance of at least two regions/targets of area covered by the
image. The ground measurements are made using a spectral radiometer.
Sometimes, large areas on the ground are painted in white and black as seen
in Fig. 8.6 and recorded values over these areas are examined in different
Contributors: Prof. B.K. Mohan and Dr. Omkar Verma 185
Block 2 Image Pre-processing Techniques
…………………………………………………………………….……………………….................……
bands and calibration values are computed based on the relation between
expected and recorded values in different bands.

Fig. 8.6: Preparation of test sites for calibration purposes. (Source: Lecture slides
of Prof. J.R. Jensen, University of South Carolina; used with permission)

Besides atmosphere, topography of the Earth surface also induces errors,

which requires correction. The effects of topographic slope are:
 local variation in view and illumination angles and
 identical surface objects might be represented by totally different intensity
values.
The goal of topographic correction is to remove all topographically caused
variance, so that areas with same reflectance have same radiance or
reflectance. Ideal slope-aspect correction removes all topographically induced
illumination variation so that two objects having the same reflectance properties
show the same gray levels despite their different orientation to the Sun’s
position. This requires digital elevation data of the area covered by the entire
image as well as satellite heading along with Sun elevation and azimuth details.

SAQ II
a) List operations used for the elimination of the systematic radiometric
errors occurring in a satellite image.
b) What is FWHM?

8.6 ACTIVITY
 Prepare an illustration showing the various steps involved in non-systematic
radiometric error correction process.

8.7 SUMMARY
 There is no perfect remote sensing system that records absolutely correct,
noise free remotely sensed data.

186 Contributors: Prof. B.K. Mohan and Dr. Omkar Verma

Unit 8 Radiometric Image Corrections
…………………….................……………..………………………………….…………………………
 The raw images obtained from remote sensing sensors are susceptible to a
variety of distortions, which degrade image quality and finally, reduce
accuracy of the image analysis.
 Radiometric and geometric are two common types of distortions that occur
in remotely sensed image data.
 Any kind of errors present in remote sensing images are known as image
distortions.
 The image correction involves image operations which normally precede
manipulation and analysis of image data to extract specific information.
 The primary aim of image correction operations is to correct distorted image
data to create a more accurate representation of the original scene.
 Internal errors are introduced by the electronics or remote sensing systems
themselves and are known as systematic errors.
 External errors are a result of phenomena that vary in nature through space
and time and hence are also known as non-systematic errors.
 The removal of the systematic radiometric errors involves random bad
pixels, line or column drop-outs, line start problems and n-line striping
operations.
 The removal of the non-systematic radiometric errors involves atmospheric
modeling, image-based methods and direct calibration using field-derived
reflectance on the remotely sensed data.
 Radiometric corrections are the most important operations that help to
utilise information contained in the remotely sensed images in a standard
reference framework along with other geospatial datasets.
 Radiometric correction is extremely important for data processing and
analysis steps.

8.8 TERMINAL QUESTIONS

1. Differentiate between systematic and non-systematic radiometric errors
occurring in remotely sensed data.
2. Describe the operations used for eliminating systematic radiometric errors in
remotely sensed data.
3. Describe the process used for elimination of non-systematic radiometric
errors in remotely sensed data.

8.9 REFERENCES
 Chavez, P. S. Jr, (1988), An improved dark-object subtraction technique for
atmospheric scattering correction of multispectral data. Remote Sensing of
Environment, Vol 24, pp 459-479.
 Jensen, J. R., (2005), Introductory Digital Image Processing: A Remote
Sensing Perspective, Prentice Hall, New Jersey (for lecture slides of Prof. J.
R. Jensen, University of South Carolina).

Contributors: Prof. B.K. Mohan and Dr. Omkar Verma 187

Block 2 Image Pre-processing Techniques
…………………………………………………………………….……………………….................……
 Jensen, J. R., (2018), Introductory Digital Image Processing: A Remote
Sensing Perspective, 4th Edition, Pearson India Education Services Pvt.
Ltd.
 Wilkie, D. S. and Finn, J. T. (1996), Remote Sensing Imagery for Natural
Resources Monitoring: A Guide for First-Time Users, Columbia University
Press, USA.

8.10 FURTHER/SUGGESTED READINGS

 Jensen, J. R., (2005), Digital Image Processing: A Remote Sensing
Perspective, Prentice-Hall, New Jersey.
 Mather, P. M., (2004), Computer Processing of Remotely Sensed Images-
An Introduction, John Wiley & Sons Ltd, New Delhi.
 Schowengerdt, R., (2007), Remote Sensing: Models and Methods for Image
Processing, Academic Press, Burlington.

8.11 ANSWERS
SAQ I
a) The two types of image corrections are radiometric and geometric.
b) Internal errors are introduced by the electronics or remote sensing systems
themselves. These kinds of errors are also known as systematic errors
because of their systematic or predictable nature.

SAQ II
a) The removal of the systematic radiometric errors involves random bad
pixels, line or column drop-outs, line start problems and n-line striping
operations.
b) FWHM is the wavelength range defined by the two points at which the
intensity level is 50% of its peak value.

Terminal Questions
1. Please refer to section 8.3.
2. Please refer to section 8.4.
3. Please refer to section 8.5.

188 Contributors: Prof. B.K. Mohan and Dr. Omkar Verma

 UNIT 9

Geometric imaGe corrections

Structure_____________________________________________________
9.1 Introduction Hybrid Approach to Image
Rectification/Registration
Expected Learning Outcomes
9.5 Activity
9.2 Geometric Distortions
9.6 Summary
9.3 Types of Geometric Distortions
9.7 Terminal Questions
Internal Geometric Errors
9.8 References
External Geometric Errors
9.9 Further/Suggested Readings
9.4 Geometric Corrections
9.10 Answers
Image-to-Map Rectification

Image-to-Image Registration

9.1 INTRODUCTION
You are aware that raw remotely sensed image data obtained from various sensors often contains
several types of distortions (errors) and deficiencies, which decrease its quality. Distortions usually
occur in image data during the image acquisition process. Thus, it becomes essential to pre-
process the raw image data in order to remove distortions before using data for any application.
Distortions are of two types namely, radiometric and geometric. Their removal from the image data
involves the pre-processing operation on raw image data. The pre-processing of image data is
conducted to rectify and restore the image by removing both radiometric and geometric errors.
You have read the various methods used for removing radiometric errors in Unit 8. In this unit, we
will discuss necessity of removing geometric errors, types of geometric errors and methods used for
their removal from the image data.
Block 2 Image Pre-processing Techniques
…………………………………………………………………….……………………….................……
Expected Learning Outcomes______________________________
After studying this unit, you should be able to:
 define geometric distortions;
 discuss types and causes of geometric distortions; and
 outline the steps and methods for geometric correction as applied to the
images.

9.2 GEOMETRIC DISTORTIONS

They consist of pixel positional and shape related errors on the image data.
Geometric error represents a mismatch between the actual image acquisition
(sensor) coordinates and ideal image coordinates. It does not allow image
coordinate system to match with the geographic coordinate system. It is difficult
to visualise geometric errors in an image as compared with radiometric errors,
which can be easily seen and also relatively easy to remove from the image
data. The removal of geometric errors is a complex operation, but their removal
from the image data is essential. Failing to eliminate geometric errors has the
following disadvantages (Wilkie and Finn, 1996):
 minimise our ability to relate thematic features indentified in an image to the
same features on the ground;
 do not allow to integrate image data with map data;
 do not allow to make real area estimate or calculation; and
 do not allow to compare two different temporal images of the same area
pixel by pixel.
Success to remove geometric errors from image data has the following
advantages (Jensen, 2018):
 allow to perfectly relate remotely sensed information to other thematic maps
in the Geographic Information System (GIS), and to the field data or with
any other spatial data;
 allow to accurately extract area and direction information from the image
data;
 allow an image to match a specific projection surface or map;
 allow to produce reliable remote sensing-based product;
 allow to compare two images taken at different times and carry out change
analysis

9.3 TYPES OF GEOMETRIC DISTORTIONS

Any type of errors that occur in remotely sensed images are known as image
distortions. Remotely sensed imagery contains two types of geometric errors
namely internal and external. The knowledge of type of geometric error present
in image data is very important for its removal. Let us discuss types of
geometric errors and their major sources.

190
Contributors: Prof. B.K. Mohan and Dr. Omkar Verma
Unit 9 Geometric Image Corrections
…………………….................……………..………………………………….…………………………
9.3.1 Internal Geometric Errors
These errors are introduced by the sensor system itself and/or by the effects of
Earth’s rotation and curvature. These errors are predictable or computable and
often referred to as systematic that can be identified and corrected using pre-
launch or platform ephemeris and information of internal sensor distortion.
Ephemeris refers to the geometry of the sensor system at the time of imaging
as well as that of the Earth surface.

Reasons of geometric distortions causing internal geometric errors in remote

sensing images include the following:

 scan skew caused by the Earth’s rotation effects

 scanning system induced variation in ground resolution cell size and

dimensional relief displacement and

 scanning system tangential scale distortion.

Earth Rotation Effect: You know that Earth rotates on its axis from west to
east. Earth observing sun-synchronous satellites are normally launched in fixed
orbits that collect a path (or swath) of imagery as the satellite makes its way
from the north to the south in descending mode. As a result of the relative
motion between the fixed orbital path of satellites and the Earth’s rotation on its
axis, the start of each scan line is slightly to the west of its predecessor which
causes overall effect of skewed geometry in the image.

Variation in Ground Resolution Cell Size and Dimensional Relief

Displacement: An orbital multispectral scanning system scans through just a
few degrees off-nadir as it collects data hundreds of kilometers above the
Earth’s surface (between 600 and 700 km above the ground level). This
configuration minimises the amount of distortion introduced by the scanning
system. In case of low altitude multispectral scanning systems, numerous types
of geometric distortion may be introduced in the image that can be difficult to
correct.

Tangential Scale Distortion: It occurs due to the rotation of the scanning

system itself. When a scanner scans across each scan line, the distance from
scanner to ground increases further away from the centre of the ground swath.
Although scanning mirror rotates at a constant speed, the instantaneous field of
view of the scanner moves faster and scans a larger area as it moves closer to
the edges. It causes in the compression of images at points away from the
nadir as shown in Fig. 9.1. This distortion is known as tangential scale
distortion. This happens in case of the Whiskbroom sensors.

191
Contributors: Prof. B.K. Mohan and Dr. Omkar Verma
Block 2 Image Pre-processing Techniques
…………………………………………………………………….……………………….................……

Fig. 9.1: Tangential scale distortion. (Source: modified after

http://sar.kangwon.ac.kr/etc/fundam/chapter2/chapter2_10_e.html)

9.3.2 External Geometric Errors

External geometric errors are mainly caused by phenomena that vary in nature
through time and space (Jensen, 2018). The most common external factors that
can cause geometric error in remotely sensed data are random movements by
the spacecraft or platform at the time of scanning or data collection. External
errors are referred to as non-systematic (random) distortions and are caused
due to the following:
 altitude changes and/or
 attitude changes (yaw, roll and pitch).
Altitude Changes: Generally, for obtaining remote sensing data at a uniform
scale, the remote sensing system flies at a constant height above ground level
along the defined flight line. Any change in the height of the remote sensing
system gradually changes scale of imagery. The increase in altitude will
produce small scale imagery and decrease in altitude will generate large scale
imagery. In addition, the topographic nature of the ground also increases or
decreases the scale of imagery generated by the remote sensing system. The
remote sensing systems are not capable themselves to adjust to such changes
in elevation. Therefore, geometric rectification algorithms are used to minimise
altitudinal changes in the remotely sensed images.
Attitude Changes: The satellite-based remote sensing platforms are largely
stable as compared to airplane because they are not influenced by atmospheric
turbulence, but their rotation and orientation may introduce distortions in terms
of roll, pitch and yaw when a platform rotates randomly in three separate axes
(Fig. 9.2; Jensen, 2018).
192
Contributors: Prof. B.K. Mohan and Dr. Omkar Verma
Unit 9 Geometric Image Corrections
…………………….................……………..………………………………….…………………………
In a 3-dimensional object, roll of a remote sensing platform such as aircraft
refers to its orientation within yz plane or rotating around x-axis. It develops
when aircraft fuselage keeps stability in direction but its wings move up or down
(Fig. 9.2b). Pitch refers to its orientation within xz plane or rotating around y-
axis. It develops when wings of aircraft remain stable but tail of fuselage moves
up or down (Fig. 9.2c). Yaw refers to the direction of its orientation within xy
plane or rotating around z-axis. It occurs when wings of aircraft remain parallel
and its fuselage moves to some angle from the intended flight line to left and
right (Fig. 9.2d). It may be noted that pitch and yaw effects are considered as
constant errors and are removed during the pre-processing. However, the effect
of roll may differ significantly on a line-to-line basis and can be removed using
measurements from a gyroscope.

(a) (b) (c) (d)

Fig. 9.2: Geometric distortions owing to change in aircraft rotation: a) Nominal

and actual coverage in an idealised condition; b) Nominal and actual
coverage in roll; c) Nominal and actual coverage in pitch; and (d)
Nominal and actual coverage in yaw. Red boundaries show nominal
coverage; black boundaries indicate actual coverage. (Source modified
after Schott, 2007)

It is important to recognise the source of internal and external error and whether
it is systematic (predictable) or non-systematic (random). As like systematic
radiometric errors, systematic geometric error is generally easier to identify and
correct than non-systematic or random geometric error.

SAQ I
a) What are internal geometric errors?
b) Define roll, pitch and yaw.
c) Name the sources of geometric distortions in a satellite image.

9.4 GEOMETRIC CORRECTIONS

You are already aware that remotely sensed images are not maps. These
images often have a mix of systematic and non-systematic geometric
distortions. These distortions make it impossible to integrate information
193
Contributors: Prof. B.K. Mohan and Dr. Omkar Verma
Block 2 Image Pre-processing Techniques
…………………………………………………………………….……………………….................……
extracted from remotely sensed images with the map data in GIS. Geometric
corrections involve the transformation of the remotely sensed image into a map
with scale and projection parameters. These corrections are undertaken in a
remotely sensed image when the image is to be used for any of the following
tasks:
 transforming an image to match with a map projection
 plotting points of interest on map and image
 sequential overlaying of the temporal images of the same area obtained by
varied sensors
 integrating remotely sensed images with GIS
You have been briefly introduced to the causes of geometric distortions in
section 9.3 of this unit. It is usually necessary to pre-process remotely sensed
data and remove geometric distortion so that individual pixels are in their proper
map locations. This allows remote sensing derived information to be related to
other thematic information in GIS or Spatial Decision Support Systems (SDSS).
Geometrically corrected imagery can be used to extract accurate distance,
polygon area and direction (bearing) information.
There are various terms which are used to describe geometric correction of
remote sensing images, and it is worthwhile to understand their definitions
before proceeding.
Geometric correction is the process of correction of raw remotely sensed data
for errors of skew, rotation and perspective.
Rectification is the process of alignment of an image to a map (map projection
system). In many cases, the image must also be oriented so that the north
direction corresponds to the top of the image. It is also known as
georeferencing.
Registration is the process of alignment of one image to another image of the
same area not necessarily involving a map coordinate system.
Geocoding is a special case of rectification that includes geographical
registration or coding of pixels in an image. Geocoded data are images that
have been rectified to a particular map projection and pixel size. The use of
standard pixel sizes and coordinates permits convenient overlaying of images
from different sensors and maps in a GIS.
Orthorectification is the process of pixel-by-pixel correction of an image for
topographic distortion. Every pixel in an orthorectified image appears to view
the Earth from directly above, i.e., the image is in an orthographic projection.
Geometric correction is usually necessary but it is not required if the purpose of
the study is not concerned with the precise positional information and rather
with the relative estimates of areas. A flow chart of geometric correction is
shown in Fig. 9.3.

194
Contributors: Prof. B.K. Mohan and Dr. Omkar Verma
Unit 9 Geometric Image Corrections
…………………….................……………..………………………………….…………………………

Fig. 9.3: Flow chart of geometric correction. (Source: http://sar.kangwon.ac.kr/etc/

rs_note/rsnote/cp9/9-4-1.gif)

You now realise that a geometrically uncorrected image is not of much use.
There are following two common geometric correction procedures which are
often used:
 image-to-map rectification
 image-to-image registration
 Hybrid approach to image rectification/registration
9.4.1 Image-to-Map Rectification
Image-to-map rectification refers to a process by which the geometry of an
image is made proper planimetric (i.e. map coordinates which are referred to as
x and y) map locations. In this process, a set of ground control points (GCPs)
are selected from corrected topographical maps representing the same
landscape of the area that is shown in the image to be corrected for making that
image planimetric. A ground control point represents a specific position which
consists of two pairs of known coordinates (reference and source coordinates).
The road intersections, bridge edges, dam edges, building corners and other
clearly identified permanent objects can be served as best GCPs. This
rectification process is used when we want to make accurate area, direction
and distance measurements. It may not, however, remove all distortions
caused by highly undulating terrain heights, leading to what are known as relief
displacement in images. This process normally involves selecting some image
pixel coordinates (both row and column) with their map coordinate counterparts
(e.g., meters northing and easting in a standard map projection such as
Universal Transverse Mercator projection). It is a most frequently used
approach, in which correction can be made within pixel levels of precision. The
following alternatives are used to obtain correct GCPs coordinate information
for image-to-map rectification (Jensen, 2018):
195
Contributors: Prof. B.K. Mohan and Dr. Omkar Verma
Block 2 Image Pre-processing Techniques
…………………………………………………………………….……………………….................……
 Hard copy planimetric maps
 Digital planimetric maps
 Geometrically rectified digital ortho-photo-quads
 Global positioning system points
To rectify a remotely sensed image to a map coordinate system during the
image-to-map rectification procedure, the following two operations must be
exercised:
 Spatial interpolation
 Intensity interpolation

Spatial Interpolation: It deals with the pixel position correction and involves
the identification of geometric relationship between the pixel coordinates
(column and row which are represented as x and y) of original input image and
the coordinates of its corresponding point (x and y) on the map (Jensen, 2018).
In this operation, numerous GCPs are employed to infer the nature of geometric
coordinate transformation that essentially be used to rectify each pixel in the
output image (x and y) with a value from a pixel in the unrectified input image
(x and y)

Intensity Interpolation: It deals with assigning pixel brightness value from

input image to its corresponding output rectified pixel position. In this operation,
pixel brightness value from an input image is extracted to place it in its
corresponding location within the output rectified image.

9.4.2 Image-to-Image Registration

It is the translation and rotation alignment process by which one image is
aligned to be coincident with respect to another image of similar geometry,
thereby allowing the user to select a pixel (i.e. GCP) in one image and its
positionally exact counterpart from the other image. The same general image
processing principles are used in both image rectification and image
registration. In case, an image that is already rectified to a map reference
system is used as base image then the second image also retains all geometric
errors present in the base image. However, this approach is more appropriate
when images of multiple dates are used for observing changes on the ground.
This is because if two images are separately rectified to the map reference
system each may have the same overall error but may be of a different nature,
resulting in twice the individual errors when two rectified images are used
together.

The general rule of thumb is to rectify remotely sensed data to a standard map
projection whereby it may be used in conjunction with other spatial data in a
GIS to solve problems. Therefore, most of our discussion here will focus on
image-to-map rectification. However, irrespective of the procedure used, the
process of geometric correction involves five steps as shown in Fig. 9.4.

196
Contributors: Prof. B.K. Mohan and Dr. Omkar Verma
Unit 9 Geometric Image Corrections
…………………….................……………..………………………………….…………………………

Fig. 9.4: Steps in geometric correction process.

Let us now discuss in detail the steps in geometric correction process.

Step 1: Collection of Ground Control Points
You have read that geometric distortions introduced by sensor system attitude
(roll, pitch and yaw) and/or altitude changes can be corrected using GCPs and
appropriate mathematical models. The best GCPs are the intersections of
roads, airport runways, large buildings, edges of dams and other stable
features. You would take a simple approach to achieve this in order to align the
image with a national reference map. You should learn that an important
concept here is GCP which is a location of features on the surface of the Earth
(e.g., a road intersection) that can be identified on the imagery and also located
accurately on a map. The image analyst must be able to obtain following two
distinct sets of coordinates associated with each GCP:
 image coordinates specified in i rows and j columns, and
 map coordinates (e.g., x, y measured in degrees of latitude and longitude,
or meters in a polyconic or Lambert conformal conic projection or a more
widely adopted Universal Transverse Mercator (UTM) projection.
Step 2: Solving a Polynomial Equation Using GCPs
After identifying GCPs, the next step is to transform row and column
coordinates of each pixel into coordinates of the reference map projection. This
is achieved using a transformation matrix which is derived from GCPs. The
transformation matrix contains coefficients calculated from a polynomial
equation.
The paired coordinates from many GCPs (e.g., 20) can be modeled to derive
geometric transformation coefficients relating to geometry of the image
coordinate system and the reference map coordinate system. These
coefficients may be used to geometrically rectify remotely sensed image to a
standard datum and map projection.
Image-to-map rectification involves computing coefficients of transformation
between the image and map coordinate system using GCPs. Each GCP
197
Contributors: Prof. B.K. Mohan and Dr. Omkar Verma
Block 2 Image Pre-processing Techniques
…………………………………………………………………….……………………….................……
contains row-column coordinates of a clearly identified point in the image and
its corresponding location in latitude-longitude or metres (East) and metres
(North) with reference to some origin on map. The transformation is normally
represented by a polynomial equation whose coefficients are determined by
solving the system equations formed using GCPs.
Step 3: Transformation of the Image to the Geometry of the Reference
Map/Image (Spatial Interpolation)
An important consequence of correcting images for geometric distortions is that
the raster grid for master dataset (map or image) and that of slave (image)
would not match. In such a case, gray level for each cell needs to be
recomputed for slave image after that it is geometrically rectified. The
determination of locations of slave pixels in the reference grid is a spatial
transformation from one coordinate system to another.
Mathematically, the relation between the ground and image coordinate systems
can be expressed in terms of the following two equations:
x’ = a1x + b1y + c1
y’ = a2x + b2y + c2
Here, x’ and y’ refer to the image coordinates and x and y to the ground
coordinates. This type of transformation is known as an affine transformation
that maps a plane to plane. Parallel lines stay parallel after affine
transformation.
This type of transformation can model following kinds of distortions in the
remote sensor data:
 translation in x and y
 scale changes in x and y
 skew and
 rotation.
Using six coordinate transform coefficients that model distortions in the original
scene, it is possible to use the output-to-input (inverse) mapping logic to
transfer (relocate) pixel values from the original distorted image (x, y) to the
grid of the rectified output image, (x, y).
Where,
x and y are positions in the output-rectified image or map and
x and y represent corresponding positions in the original input image.
Step 4: Assessment of Error
Though geometric correction process corrects much of the geometric error
present in the original image, not all of the errors are removed. Hence, prior to
creating output rectified image, it is essential to test quality of fit between the
two coordinate systems based on coefficients. In other words, it is required to
assess the accuracy of the polynomial transformation. Each GCP has influence
in value of coefficient in the transformation matrix and some GCPs would
introduce large amounts of error. The method used most often for calculating
198
Contributors: Prof. B.K. Mohan and Dr. Omkar Verma
Unit 9 Geometric Image Corrections
…………………….................……………..………………………………….…………………………
total error involves the computation of root-mean-square error (RMSE or
RMSerror) over all GCPs. It is important to remember that the units of RMS (root-
mean-square) error are the units of the source coordinate system. The error
being referred to here is the distance between GCP original source (measured
by the collector of GCPs) and re-computed coordinates according to the affine
transformation above. It is calculated using Pythagoras’s theorem as

 x  x    y   y 
2 2
GCPerror  orig orig

Where, xorig and yorig are the original row and column coordinates of GCP in the
image and x΄ and y΄ are the computed or estimated re-transformed coordinates
in the original image when we utilise six coefficients. The square root of the
squared deviations represents a measure of accuracy of each GCP. By
computing RMSerror for all GCPs, it is possible to see which GCPs contribute the
greatest error and sum all RMSerror.
The RMSerror computed over all GCPs is given by

1 n
RMSerror  
n i 1
{( xi'  xi' (computed )}2  {( yi'  yi' (computed )}2

Where the terms inside the square root are the image coordinates as measured
and as re-computed using the affine transformation. The GCPerror is computed
for each GCP and those that yield a large error or eliminated and the affine
transformation is recomputed. This is facilitated by selecting a well distributed
set of GCPs throughout the image. It may be noted that each GCP yields two
equations (one equation for the x-coordinate and one for the y-coordinate) and
there three equations are minimum to obtain a single solution. However, many
more than three GCPs are required to ensure a good geographic spread of the
GCPs in the image, and to have the option of eliminating incorrectly selected
GCPs.
You may note that not all of the original GCPs selected are used to compute
final six-parameter coefficients to rectify input image. This involves several
cycles, wherein all GCPs are initially used to find coefficients. RMSerror
associated with each of these initial GCPs is computed and summed. Then, the
individual GCPs that contributed the greatest amount of error are determined
and deleted. After the first iteration, this would leave you with a reduced number
of GCPs. A new set of coefficients is then computed using these GCPs. The
process continues until RMSerror reaches an acceptably low value (e.g., <1 pixel
error in the x-direction and <1 pixel error in the y-direction). When the
acceptable threshold is reached, the final coefficients and constants are used to
rectify input image to an output image in a standard map projection as
previously discussed.
Step 5: Resampling (Intensity Interpolation)
The next step is to create output image by assigning values to pixels on new
grid of the transformed image by interpolation from the unrectified image. This
process is called resampling or intensity interpolation. You have read earlier
that when pixel positions are recomputed through transformations controlled by
GCPs, a pixel in the reference image would overlap with several pixels in the
199
Contributors: Prof. B.K. Mohan and Dr. Omkar Verma
Block 2 Image Pre-processing Techniques
…………………………………………………………………….……………………….................……
unrectified input image. Pixels are resampled so that new pixel values for the
transformed image can be calculated. Therefore, it is necessary to find pixel
values in the transformed image through intensity interpolation.
Intensity interpolation (resampling) involves the computation of pixel value at a
pixel position x, y in the original (distorted) input image and its relocation to the
appropriate x, y coordinate location in the rectified output image. This process
is used to produce output image row-by-row and pixel-by-pixel. Most of the
time, x and y coordinates to be sampled in the input image are real numbers
(floating point), which do not coincide with pixel positions. In such a case, you
can employ one of the methods of intensity interpolation (resampling)
mentioned below:
 nearest neighbour (zero order) method
 bilinear interpolation (first order) method and
 cubic convolution (second order) method
Nearest neighbour method uses the closest pixel’s intensity value to assign to
output pixel value. In the bilinear interpolation method, values of nearby four
pixels in a 2 x 2 window are used to calculate an output pixel value that falls in-
between these four pixels in the uncorrected image with a bilinear function.
And, in the cubic convolution method, values of sixteen pixels in a 4 x 4 window
are used to calculate an output pixel value with a cubic function. Of the three
methods, the nearest neighbour method is the least computationally intensive
whereas cubic convolution method is the most computationally intensive
method. Nearest neighbour method is used for a majority of applications
involving digital interpretation because it transfers original pixel values without
averaging them whereas in the other two methods, the pixel values are altered.
In all the three methods, number of rows and columns of pixels in the output
image is calculated from the dimensions of output image, which is determined
by the geometric transformation and cell size.
9.4.3 Hybrid Approach to Image Rectification/Registration
In hybrid approach, both image-to-map rectification and image-to-image
registration procedures of geometric correction are used together. The
approach is mostly used when we want to undertake a rigorous change
detection analysis remotely sensed data involving two or more dates. This
approach makes easy to select GCPs for rectifying an image.
When a series of images is to be used to monitor changes happening in an
area over a period of time, we require that the images have a reference to the
ground locations and also an image-to-image reference. This kind of hybrid
approach is an advantage over a seemingly simpler approach of
georeferencing every image to the ground directly. If each image is handled
independently of the others, then there is a likelihood that their georeferencing
errors (even if they are small, of the order of fraction of a pixel size) may pull
each image in a different direction (Fig. 9.5).

200
Contributors: Prof. B.K. Mohan and Dr. Omkar Verma
Unit 9 Geometric Image Corrections
…………………….................……………..………………………………….…………………………

Fig. 9.5: Effect of individual geo-referencing: Black: first georeferenced image;

Red: second image after georeferencing; Green: third image after
georeferencing.

Given that the coordinate systems are very different between ground and the
image, the kind of misalignment between individually georeferenced images is
very likely. In case of image to image registration, since both have row-column
coordinate systems, this kind of problem is less likely. Moreover, if the spatial
resolution of all the images is same, then simple translation/rotation for
geometry and nearest interpolation for intensity calculations will suffice.
At present, we exclude the possibility of the images acquired at different times
coming from different satellites and different resolutions (spatial / spectral /
radiometric) and terrain within the study area being undulating instead of
planar. In other words, it is assumed that the study area has a flat terrain, and
all the images in the time series have identical spatial/spectral/radiometric
resolutions. If this assumption is not satisfied, then the georeferencing can
become much more complex, which is outside the scope of these notes.

SAQ II
a) Define rectification and registration.
b) What are ground control points?
c) List common types of procedure used for making images geometrically
correct.

9.5 ACTIVITY
Try to download an image from the Bhuvan portal (https://bhuvan.nrsc.gov.in)
and get its corresponding map from Survey of India or any other source. Use
QGIS or any image processing software you have and try to find whether image
is geometrically rectified/corrected.

9.6 SUMMARY
 Distortions usually occur in remotely sensed image data during the image
acquisition process. Thus, it becomes essential to pre-process the raw
image data in order to remove distortions before using for any application.
 Distortions are of two types namely, radiometric and geometric.
 Geometric distortions are errors, which represent a mismatch between the
actual image acquisition (sensor) coordinates and ideal image coordinates.

201
Contributors: Prof. B.K. Mohan and Dr. Omkar Verma
Block 2 Image Pre-processing Techniques
…………………………………………………………………….……………………….................……
 Geometric errors are of two types such as internal and external.
 Internal errors are introduced by the sensor system itself and/or by the
effects of Earth’s rotation and curvature.
 External geometric errors are mainly caused by phenomena that vary in
nature through time and space.
 Geometric corrections involve the transformation of the remotely sensed
image into a map with scale and projection parameters.
 Image-to-map rectification and image-to-image registration are two common
procedures used for making geometric correction in the image data.
 Image-to-map rectification refers to a process by which the geometry of an
image is made proper planimetric.
 Image-to-image registration is the translation and rotation alignment
process by which one image is aligned to be coincident with respect to
another image of similar geometry.
 The hybrid approach of geometric correction used both image-to-map
rectification and image-to-image registration procedures of geometric
correction together.
 Geometric corrections are the most important operations that help to utilise
information contained in the remotely sensed images in a standard
reference framework along with other geospatial datasets.
 Geometric correction places result on its correct location on a geographic
reference system whereby it can be used, for instance, to detect changes
between the content of a map and image, and also to update earlier date
maps with new information as observed in the latest image.

9.7 TERMINAL QUESTIONS

1. Describe geometric distortion and their types.
2. Describe the causes of external geometric errors.
3. Discuss image-to-map rectification procedure of geometric correction.
4. Discuss image-to-image registration procedure of geometric correction.

9.8 REFERENCES
 Jensen, J. R. (2005), Introductory Digital Image Processing: A Remote
Sensing Perspective, Prentice Hall, New Jersey (for lecture slides of Prof. J.
R. Jensen, University of South Carolina).
 Jensen, J. R. (2018), Introductory Digital Image Processing: A Remote
Sensing Perspective, 4th Edition, Pearson India Education Services Pvt.
Ltd.
 Schott, J.R. 2007. Remote Sensing: The Image Chain Approach. 2nd Ed.,
Oxford University Press, New York.
 Wilkie, D. S. and Finn, J. T. (1996), Remote Sensing Imagery for Natural
Resources Monitoring: A Guide for First-Time Users, Columbia University
Press, USA.

202
Contributors: Prof. B.K. Mohan and Dr. Omkar Verma
Unit 9 Geometric Image Corrections
…………………….................……………..………………………………….…………………………
 http://sar.kangwon.ac.kr/etc/fundam/chapter2/chapter2_10_e.html
 http://sar.kangwon.ac.kr/etc/rs_note/rsnote/cp9/9-4-1.gif

9.9 FURTHER/SUGGESTED READINGS

 Jensen, J. R. (2005), Digital Image Processing: A Remote Sensing
Perspective, Prentice-Hall, New Jersey.
 Mather, P. M. (2004), Computer Processing of Remotely Sensed Images-
An Introduction, John Wiley & Sons Ltd, New Delhi.
 Schowengerdt, R. (2007), Remote Sensing: Models and Methods for Image
Processing, Academic Press, Burlington.

9.10 ANSWERS
SAQ I
a) Internal geometric errors are introduced by the sensor system itself and/or
by the effects of Earth’s rotation and curvature.
b) In a 3-dimensional object, roll of a remote sensing platform such as aircraft
refers to its orientation within yz plane or rotating around x-axis. Pitch refers
to its orientation within xz plane or rotating around y-axis. Yaw refers to the
direction of its orientation within xy plane or rotating around z-axis.
c) Earth rotation and its curvature, satellite altitude and perturbations in
satellite nominal altitude are the main sources of geometric distortions in a
satellite image.

SAQ II
a. Rectification is the process of alignment of an image to a map (map
projection system). Registration is the process of alignment of one image to
another image of the same area not necessarily involving a map coordinate
system.
b. A ground control point represents a specific position which consists of two
pairs of known coordinates (reference and source coordinates).
c. Image-to-map rectification and image-to-image registration are two common
types of procedure used for making images geometrically correct.

Terminal Questions
1. Please refer to sections 9.2 and 9.3.
2. Please refer to subsection 9.3.2.
3. Please refer to subsection 9.4.1.
4. Please refer to subsection 9.4.2.

203
Contributors: Prof. B.K. Mohan and Dr. Omkar Verma
Block 2 Image Pre-processing Techniques
…………………………………………………………………….……………………….................……

204
Contributors: Prof. B.K. Mohan and Dr. Omkar Verma
 UNIT 10

IMAGE ENHANCEMENT
Structure_____________________________________________________
10.1 Introduction 10.4 Contrast Enhancement
Expected Learning Outcomes Linear Contrast Enhancement

10.2 Basics of Image Enhancement Non-Linear Contrast Enhancement

10.3 Common Image Enhancement Operations 10.5 Application of Contrast Enhancement

Image Reduction 10.6 Activity
Image Magnification 10.7 Summary
Transects (Spatial Profiles) 10.8 Terminal Questions
Spectral Profiles 10.9 References
Level Slicing 10.10 Further/Suggested Readings
10.11 Answers

10.1 INTRODUCTION
In the previous unit, you have learnt about the different errors affecting the quality of remotely
sensed data and image correction techniques. You can extract usable information from an image
that has been corrected for radiometric and geometric errors
Visual interpretation of remote sensing images is one of the most powerful ways to retrieve
information. Depending on the quality of image and human expertise in image interpretation, many
of the attributes are readily identified by human eyes. However, at times it is required to highlight
(enhance) certain features of interest in images to make the image better interpretable. This is
achieved by techniques known as image enhancement that is required to increase the quality of
remote sensing data.
In this unit, we will discuss about common image enhancement operations together with linear and
non-linear contrast enhancement and spatial image enhancement techniques.
Block 2 Image Pre-processing Techniques
…………………………………………………………………….……………………….................……
Expected Learning Outcomes___________________________
After studying this unit, you should be able to
 define image enhancement;
 describe the concept of commonly used image enhancement strategies;
 discuss linear and non-linear contrast enhancement techniques; and
 spatial image enhancement techniques.

10. 2 BASICS OF IMAGE ENHANCEMENT

The term enhancement is used to define the changes carried out in an image to
improve its appearance for better visual interpretation. There are number of
image enhancement techniques available but no single technique can be said
to be best because needs of users vary and also a technique which works for a
certain image for a particular area may not work for other images for the same
or different areas.
Image enhancement refers to the techniques used to increase the human
interpretability and ability to derive meaningful information from remotely
sensed images. It improves the quality and clarity of images. The enhancement
operations include removing blurring and noise, increasing contrast and
revealing details. For example, image enhancement of a blurred and low
contrast satellite image of the Himalayan Mountains could be made clear by
increasing contrast range and reducing blurring in the original image. The
original image might have areas of very high and very low intensity, which mask
details. The enhanced image will have improved contrast providing details of
the masked image regions.
It may be noted that removal of blur is a mathematically modelled process,
where the blur is sometimes represented as the point spread function of a linear
space-invariant system and requires inversion techniques of the digital image
restoration. Image enhancement is more often subjective, according to the user
preferences it is used to make the visual appearance of the image better.
Image enhancement techniques are also known as radiometric enhancement
techniques (Liu and Mason, 2009).
The primary goal of image enhancement is to improve the visual interpretability
of the images. This is done by increasing the apparent distinction between
features in the image. Image enhancement involves use of a number of
statistical and image manipulation functions available in an image processing
software. Image features are enhanced by the following two operations:
 Point Operations: In this operation, value of pixel is enhanced independent
of characteristics of neighbourhood pixels within each band.
 Local (Neighbourhood) Operations: In this operation, value of pixel is
enhanced based on neighbouring brightness values.
Image enhancement techniques based on point operations are also called as
contrast enhancement techniques and the enhancement techniques based on
neighbourhood operations are also known as spatial enhancement techniques.

206
Contributors: Dr. Omkar Verma and Dr. Manu Mehta
Unit 10 Image Enhancement
…………………….................…….………..………………………………….…………………………
The main aim of image enhancement is to enhance the contrast between the
landscape background and the object of interest to the observer because
contrast allows human to visually detect and identify objects in a landscape
(Wilkie and Finn, 1996; Schowengerdt, 2006). Furthermore, it can be used to
remove noise, brighten or sharpen an image and enhance its perceptual
aspects, which make an interpreter to easily identify objects. The image
enhancement techniques change the original values of the digital number
permanently. These techniques can be applied on the single band image and
on the individual bands of a multi-band image data.
The digital remotely sensed images are usually composed of picture elements
(pixels). Each pixel has a variable usually in the form of a binary integer ranging
from 0 to 255 levels. A single pixel may have several digital number variables
corresponding to different bands recorded in an image. The number associated
with each pixel is known as Digital Number or Brightness Value or Gray Level.
Digital number represents the average radiance of a small area within a scene.
A smaller value/number indicates low average radiance and a higher number
indicates high radiance from the area.

10.3 COMMON IMAGE ENHANCEMENT

OPERATIONS
Let us discuss the common image enhancement operations very briefly.
10.3.1 Image Reduction
Majority of available remotely sensed data are comprised of more than 3000
rows x 3000 columns and several bands. Many digital image processing
systems cannot display this data at a single time (Jenson, 2018). Thus, it
becomes necessary to reduce size of original image dataset and bring down its
size to such a level, which can be seen on the computer screen at one time
mainly for orientation uses. For reducing an image to 1/m2 of the original size,
every mth row and mth column of the image are properly chosen and shown. In
this way, for m = 2, the reduced image dataset bears only one-quarter (i.e.,
25%) of the pixels of the original dataset. It can be represented as a simple 2x
integer reduction (Fig. 10.1).
It may be noted that you may observe loss in information or pixels while
comparing original image with the reduced image and it usually happens
because many pixels are removed during the process of image reduction. This
is the reason that image reduction is mainly used for orienting within a scene
and finding coordinates of the row and column of a particular area of interest
(Jenson, 2018). Image reduction is also known as site-specific pixel
characteristic and allows us to get a regional view of the remotely sensed data.

Contributors: Dr. Omkar Verma and Dr. Manu Mehta

207
Block 2 Image Pre-processing Techniques
…………………………………………………………………….……………………….................……

Fig. 10.1: Image matrix showing the logic pertaining to 2x image integer
reduction. The output raster image is new and represents only 25% of
the original input image. (Source: redrawn after Jenson, 2018).

10.3.2 Image Magnification

It is also known as zooming in and is often undertaken during the following
stages:
i) For enlarging size of an image/map for getting detailed information of a
relatively small area of interest;
ii) For matching the size of another image/map.
Image magnification involves replication of row and column of a digital image.
In order to magnify an image, an integer factor mx is used, where each pixel in
the original image is replaced by an m x m block of pixels, where all pixels have
the same value as the original input pixel (Jensen, 2018). It can be represented
as a simple 2x magnification (Fig. 10.2).

Fig.10.2: Image matrix showing the logic pertaining to 2x image integer

magnification. (Source: redrawn after Jenson, 2018).

10.3.3 Transects (Spatial Profiles)

The preparation of spatial profile along the user-specified transect between two
or more points in a single-band or multi-band colour composite image by
extracting brightness values or percent reflectance values is useful for making
image interpretation (Jenson, 2018). Spatial profile is usually prepared in x-y
plot style, where y-axis values of the plot refer to the magnitude of the
brightness value at each pixel along the selected transect (Fig. 10.3).

208
Contributors: Dr. Omkar Verma and Dr. Manu Mehta
Unit 10 Image Enhancement
…………………….................…….………..………………………………….…………………………

(a) (b)
Fig. 10.3: a) Predawn imagery showing transect along the point A to B; and b) x-y
plot showing spatial profile along the points A to B. (Source: modified
after https://slideplayer.com/slide/3723856/)

10.3.4 Spectral Profiles

Spectral profile involves the extraction of full brightness values of the
electromagnetic spectrum in n bands for an individual pixel (Jensen, 2018). In
this profile, x-axis shows the number of the individual bands in the dataset and
y-axis represents the brightness value of the pixel for each band considered for
investigation. If the reflectance of the image data is converted into percent
reflectance, then y-axis is used to represent percent reflectance. Spectral
profile is used to enhance the differences between various objects/targets and
its importance depends on the quality of information present in the spectral data
(Jensen, 2018). Fig 10.4 illustrates the spectral reflectance profiles for
atmospherically corrected data of Airborne Visible Infrared Imaging
Spectrometer - Next Generation (AVIRIS–NG) mission (credit: NASA/ISRO).
Reflectance values lie in the range of 0-1 and for ease of visualisation, they are
multiplied by 10000.

Fig. 10.4: Spectral profile of vegetation from hyperspectral data; x-axis

represents the wavelength in micrometers and y-axis represents the
spectral reflectance values scaled by 10000. (Source: Data taken from
AVIRIS-NG)

10.3.5 Level Slicing

Level or density slicing is an enhancement technique whereby digital number
(DNs) distributed along the x-axis of an image histogram are divided into a

Contributors: Dr. Omkar Verma and Dr. Manu Mehta

209
Block 2 Image Pre-processing Techniques
…………………………………………………………………….……………………….................……
series of analyst specified intervals or slices and thereby creating a new image
showing select number of features (classes) in it.

SAQ I
a) Define image enhancement.
b) What is point operation?

10.4 CONTRAST ENHANCEMENT

Before discussing about the contrast enhancement, let us first recognise the
meaning of ‘contrast’. Contrast is the difference in pixel values of two different
objects at the same wavelength. It can be defined as the ratio of maximum
intensity to the minimum intensity over an image.
C = Lmax / Lmin
Where,
C = contrast,
Lmax = maximum intensity value present in an image, and
Lmin = minimum intensity value present in the same image.
It is important to note that the contrast ratio has profound implication on the
usefulness of an image as it aids in visual interpretation of the image. It is easy
to interpret the images having larger contrast ratio as compared to the images
with lower contrast ratio. High contrast in the images is generated when one
feature of the landscape reflects a high amount of radiation in a definite
wavelength and second feature reflects little/low radiation in the same
wavelength. Due several reasons, many satellite images lack required contrast.
Some of the reasons, which are responsible for low contrast in images are
listed below:
i. Landscapes consisting of individual features and background that do not
contain terrain covering full dynamic range of brightness recorded by the
remote sensing sensors will show low contrast.
ii. There are several different landscape features that reflect uniform amounts
of visible and infrared radiation and produce low contrast images.
iii. Some sensors lack sufficient sensitivity to detect and record the contrast of
the landscape.
iv. Scattering of radiation by atmosphere can also reduce the contrast of the
landscape.
It may be noted that the images usually acquired by a digital sensor might have
a limited range of digital numbers (DN) to accommodate all the values of
reflectance resulting in a poor quality display of the image. Due to this reason,
the image gray scale distribution is skewed and it is very difficult to distinguish
between different features in the image. In order to exploit the capability of the
image display system, the dynamic range of the image pixel values is increased
(Fig. 10.5). This technique is known as contrast enhancement. It involves

210
Contributors: Dr. Omkar Verma and Dr. Manu Mehta
Unit 10 Image Enhancement
…………………….................…….………..………………………………….…………………………
changing the original pixel values so that more of the available range is used
which increases contrast between features and their backgrounds.

(a) (b)
Fig. 10.5: Contrast enhancement operation on a satellite image: a) Original image; and
b) Enhanced image. (Source: Anand, 2012)

Commonly applied contrast enhancement techniques are the followings:

 Linear contrast enhancement
 Non-linear contrast enhancement

10.4.1 Linear Contrast Enhancement

Image display and recording devices often operate over a range of 256 gray
levels (maximum range of numbers represented in 8-bit image data). Image
data in a single image rarely extend over this entire range. The contrast
enhancement (also known as contrast stretching) technique expands the
narrow range of pixel values typically present in an input image to a wider range
of pixel values.
A linear contrast stretching occurs, when pixel values of an image are uniformly
stretched across a larger display range. By expanding the range of original
input values of the image, the total dynamic range of display device can be
utilised. One major disadvantage of a linear stretch is that it assigns an equal
number of display levels for infrequently occurring values as it does for
frequently occurring values. This can result in certain features being enhanced,
while other features, which may be of greater importance, will not be sufficiently
enhanced. There are three types of linear contrast enhancement technique as
given below:
 Minimum-maximum contrast stretch
 Percentage linear and standard deviation contrast stretch
 Piece-wise linear contrast stretch

Contributors: Dr. Omkar Verma and Dr. Manu Mehta

211
Block 2 Image Pre-processing Techniques
…………………………………………………………………….……………………….................……
i) Minimum-Maximum Contrast Stretch
In this technique, the original minimum and maximum values of the image are
assigned to a newly specified set of values based on the full range of available
brightness values of that image. For example, if maximum pixel value in the
original image is 158 and minimum pixel value is 60 as illustrated in Fig. 10.6a,
the ‘stretched’ values would be set at 255 and 0, respectively (Fig. 10.6b). The
intermediate gray level values of original image would be similarly and
proportionately set to new values. These types of enhancements are best
applied to remotely sensed images with normal or near-normal distribution,
meaning, all the brightness values fall within a narrow range of the image
histogram and only one mode is apparent. This approach of linear contrast
stretching is known as minimum-maximum contrast stretch.
Now, there are several digital image processing systems that contain built-in
functions, which automatically expand the minimum and maximum values to
optimise the whole range of available brightness values. This technique of
image enhancement is best suited to remotely sensed images with Gaussian or
near-Gaussian histograms.

(a)

(b)

(c)
Fig. 10.6: Illustration of principle of contrast stretch. Note: with a typical
application of a linear stretch, areas that were dark-toned will appear
darker and areas that were bright toned will appear even brighter. Linear
contrast enhancement also makes subtle variations within data more
obvious. (Source: redrawn after Jensen, 1996)

Ii) Percentage Linear and Standard Deviation Contrast Stretch

The percentage linear contrast stretch technique uses precise minimum and
maximum values that be positioned in a certain percentage of pixels from the
mean of the histogram. On the other hand, the standard deviation contrast
stretch uses a percentage of the selected minimum and maximum values of
212
Contributors: Dr. Omkar Verma and Dr. Manu Mehta
Unit 10 Image Enhancement
…………………….................…….………..………………………………….…………………………
pixels that coincides with a standard deviation percentage of pixels from the
mean of the histogram (Jensen, 2018).
It may be noted that the standard deviation from the mean is also used to move
the tails of the histogram away from the original minimum and maximum values.
It has been noted that for a normal distribution, around 68% of the observations
fall within ± 1 standard deviation of the mean, nearly 95.4 % of all observations
fall within ± 2 standard deviations and 99.73 % all observations fall within ± 3
standard deviations of the mean (Jensen, 2018). Let us take an example of the
Landsat Thematic Mapper band 4 images for the linear contrast stretching
applying a ± 1 standard deviation. It would give minimum value 12 and
maximum value 43. By apply above technique, all vales lie between 12 and 43
will be stretched to fall within the range of 0 to 255 (Jenson, 2018). As a
consequence, the values lie between 0 and 11 become 0 and those values
between 44 and 255 become 255. It increases the contrast of an image. The
main disadvantage of the technique is that the pixels information saturated
between at 0 and 255 is normally missed out. Fig. 10.7 shows the standard
deviation contrast enhancement operation applied on satellite data.

(a) (b)
Fig. 10.7: Standard deviation Contrast enhancement operation on a satellite
image: a) Original image; and b) Enhanced image. (Source: Quickbird)

One of the important uses of standard deviation stretch is when the input image
has a small number of very dark pixels (brightness values near 0 and a small
number of very bright pixels (brightness values near 255). In this case
maximum-minimum stretch obviously fails input minimum matches with the
display device minimum and input maximum matches with the display device
maximum and a simple linear contrast enhancement method will have no effect
on the input image contrast. However, the standard deviation stretch will
definitely work.
iii) Piece-wise Linear Contrast Stretch
The piece-wise linear contrast stretch is used when we have an image
histogram belonging to bimodal or trimodol distribution not to the Gaussian
distribution. In this stretch, numerous linear enhancement functions are used for
enhancing the brightness ranges in the modes of the histogram. It also involves

Contributors: Dr. Omkar Verma and Dr. Manu Mehta

213
Block 2 Image Pre-processing Techniques
…………………………………………………………………….……………………….................……
the setting of a series of small minimum-maximium stretches within a single
histogram. It is more useful to enhance any typical features in the image.
The following steps are undertaken while doing piece-wise linear contrast
stretch on an image (after Jensen, 2018):
i. View the raw image to be contrast stretched
ii. Display of the histogram of the raw image
iii. Superimpose histogram with an input-output line. This line will be visible in
the display as a line that runs diagonally from the lower left to upper right.
iv. Finally, inter-actively adjust the length and slope of n mutally exclusive
contrast stretches along the line.
As an illustration, Fig. 10.8 illustrates the piecewise linear contrast
enhancement on a satellite dataset.

(a) (b) (c)

Fig. 10.8: Piecewise linear contrast enhancement operation on a satellite image;

a) Original image; b) Enhanced image; and c) Piecewise linear function
used; original histogram in grey colour, modified histogram in yellow
colour. (Source: Quickbird; Image software: Erdas Imagine 2014)

10.4.2 Non-Linear Contrast Enhancement

Contrary to the linear stretch, a non-linear stretch expands one portion of gray
scale while compressing other portions. This sort of enhancement is typically
applied with purpose of highlighting one part of the brightness range. While
spatial information is preserved, quantitative radiometric information can be
lost. Three of the most common of non-linear contract stretching methods are
discussed below:
i) Histogram Equalisation
Histogram equalisation is one of the most useful forms of non-linear contrast
enhancement. When an image’s histogram is equalised, all pixel values of the
image are re-distributed so there are approximately an equal number of pixels
to each of the user-specified output gray scale classes (e.g., 32, 64 and 256).
The contrast is increased at the most populated range of brightness values of
histogram (peaks). In Fig. 10.5c, you can see how histogram equalisation at
different thresholds increases contrast of the image. It automatically reduces
contrast in very light or dark parts of the image associated with the tails of a
normally distributed histogram where the pixel population is sparse. Histogram

214
Contributors: Dr. Omkar Verma and Dr. Manu Mehta
Unit 10 Image Enhancement
…………………….................…….………..………………………………….…………………………
equalisation can also separate pixels into distinct groups, if there are very few
output values over a wide range. Histogram equalisation is effective only when
original image has poor contrast to start with, otherwise histogram equalisation
may degrade the image quality. Fig. 10.9 shows the histogram equalisation
applied on a satellite dataset.

(a) (b)
Fig. 10.9: Histogram equalised contrast enhancement operation on a satellite
image: a) Original image; and b) Enhanced image. (Source: Quickbird).

ii) Logarithmic Stretch

A logarithmic or square root stretch compresses the higher brightness values
within an image and disproportionately expands the darker values. It is applied
by taking the log (square root) of original brightness values. Logarithmic
approach stretches DN values in darker part of histogram whereas inverse
logarithmic approach stretches DN values in brighter part.
Fig. 10.10 shows the difference between a linear and a logarithmic stretch.
Thus, the original darker values are given a greater relative contrast than
original values at the higher end of the brightness scale. This would be done in
order to allow better visual discrimination among the darker features of the
image, while maintaining some limited recognition of brighter features. The
logarithmic stretch is a more emphatic version of the square root stretch. Each
of these can be inverted (log to inverse log; square root to squared), so as to
favour the bright end of scale instead.
The main utility of logarithmic contrast enhancement is to reveal details in the
low and very low reflectance portions of the image that are of interest to the
user who may not be interested in the higher reflectance areas.

Contributors: Dr. Omkar Verma and Dr. Manu Mehta

215
Block 2 Image Pre-processing Techniques
…………………………………………………………………….……………………….................……

Fig. 10.10: Logic of linear, logarithmic and inverse log contrast stretch. BV
indicates brightness value. (Source: redrawn after Jensen, 1996)

iii) Gaussian Contrast Stretch

It may be noted that the Gaussian contrast stretching can be performed in the
similar manner as we operate histogram equalisation. It is mostly used when an
image analyst wants to have the information of the dark and bright regions of
the image because this information is usually missed while applying other
contrast stretching techniques. This stretching is based on histogram of pixel
values and involves fitting of observed histogram of the input image to a normal
or Gaussian histogram. It also highlights the contrast in the tail parts of the
histogram. A histogram matching technique is involved in the Gaussian contrast
stretching to map the input image values onto the output Gaussian distribution.
As an illustration, Gaussian enhancement is shown on a satellite image taken
from Linear Imaging Self-Scanning Sensor (LISS) III data onboard Resourcesat
(Fig. 10.11).

(a) (b)
Fig. 10.11: Guassian Contrast enhancement operation on a satellite image: a)
Original image; and b) Enhanced image. (Source: LISS-III / Resourcesat
1)

The type of non-linear enhancement technique that is chosen to be performed

on an image depends on the nature of original data and what features the
analyst is interested in. For example, water features, which are typically darker
in tone, would benefit from log or square root stretches, while the land features,
which are brighter, would still be somewhat recognisable.

216
Contributors: Dr. Omkar Verma and Dr. Manu Mehta
Unit 10 Image Enhancement
…………………….................…….………..………………………………….…………………………

SAQ II
a) Define contrast.
b) List types of linear contrast stretch.

10.5 APPLICATION OF CONTRAST ENHANCEMENT

Satellite data usually lack adequate contrast and in general, contrast
enhancement is required to interpret the remotely sensed image. This low
contrast could arise due to several reasons. For instance, sometimes, the
scene itself has a poor contrast ratio for different features whose spectral
responses might closely resemble each other. Further, atmospheric scattering
may also decrease the image contrast especially in the shorter wavelength
regions. In addition, the sensitivity of the remote sensing instrument might be
low, leading to poor contrast of the image.
Contrast enhancement is a type of image enhancement, which is used to
highlight (enhance) certain features of interest in images to make the image
better interpretable. This technique is quite useful for improving the contrast of
the satellite and aerial images, which make them easier to interpret for an
interpreter. It is also important to note that an enhancement technique might not
be suitable for all images and their applications and therefore one has to
choose an appropriate enhancement technique based on the imaging sensor
type and the application. Image histograms help in the decision for proper
choice of contrast enhancement techniques. For instance, one might would like
to enhance only the lower part of the histogram, so accordingly, piecewise
linear contrast stretch might be used. On the other hand, in order to have a
uniform contrast enhancement, one may opt for histogram equalisation. Apart
from contrast enhancement, there are other several techniques for image
enhancement such as density slicing, edge enhancement and spatial filtering.
Overall, image enhancement including contrast enhancement is one of the
most interesting and visually attractive fields in aerial and satellite image
processing. In fact, contrast enhancement is also an emerging and sought-after
field in the medical image processing, where it is used to enhance the contrast
of the specific features for detecting several diseases and deciding their
treatments. In medical imaging, contrast enhancement is used to enhance the
contrast of the images obtained from X-ray based devices, Ultrasound (US),
Magnetic Resonance Imaging (MRI), Computed Tomography (CT) and Positron
Emission Tomography (PET). By using various techniques of contrast
enhancement, the image analysts improve the visual quality of satellite images
by increasing dominance of some features or by reducing ambiguity between
different parts of the image.

10.6 ACTIVITY
 Download a multispectral satellite image (e.g., Landsat 8 Operational Land
Imager). Note the number of spectral bands. Plot the histogram for
individual bands. Apply linear contrast enhancement (minimum-maximum)
and histogram equalisation for each band. Tabulate the differences
observed.
Contributors: Dr. Omkar Verma and Dr. Manu Mehta
217
Block 2 Image Pre-processing Techniques
…………………………………………………………………….……………………….................……
10.7 SUMMARY
 Visual interpretation of remote sensing images is one of the most powerful ways to
retrieve information.

 Enhancement is used to define the changes carried out in an image to improve its
appearance for better visual interpretation.

 Image enhancement refers to the techniques used to increase the human

interpretability and ability to derive meaningful information from remotely sensed
images.

 Image features are enhanced by the following two operations: point and local.

 Image reduction, image magnification, transects (spatial profiles), spectral profiles

and level slicing are common image enhancement operations.

 Contrast is the difference in pixel values of two different objects at the same
wavelength.

 Contrast enhancement techniques are two types namely, linear and non-linear.

 A linear contrast stretching occurs, when pixel values of an image are uniformly
stretched across a larger display range.

 There are three types of linear contrast enhancement technique such as minimum-
maximum contrast stretch, percentage linear and standard deviation contrast
stretch and piece-wise linear contrast stretch.

 A non-linear stretch expands one portion of gray scale while compressing other
portions.

 Histogram equalisation, logarithmic stretch and Gaussian contrast stretch are three
common types of non-linear contrast enhancement.

10.8 TERMINAL QUESTIONS

1. Why are image reduction and magnification operations important in image
enhancement?
2. What is contrast enhancement? Discuss the factors responsible for
generating low contrast satellite images.
3. Describe various techniques of linear contrast enhancement.
4. Explain the techniques of non- linear contrast enhancement.

10.9 REFERENCES
 Anand A (2012). Unit 12 Image Enhancement and Transformation. In:
MGY-002: Remote Sensing and Image Interpretation. Indira Gandhi
National Open University, New Delhi, pp. 23-39.
 Jensen JR (1996) Introduction to Digital Image Processing: A Remote
Sensing Perspective, Prentice Hall, New Jersey.
 Jensen JR (2018) Digital Image Processing: A Remote Sensing Perspective, 4th
edition, Pearson India Education Services Pvt Ltd, Uttar Pradesh, India.

218
Contributors: Dr. Omkar Verma and Dr. Manu Mehta
Unit 10 Image Enhancement
…………………….................…….………..………………………………….…………………………
 Liu JG and Mason PJ (2009) Essential Image Processing and GIS for
Remote Sensing, Wiley-Blackwell, U.K
 Schowengerdt RA (2006) Remote Sensing: Models and Methods for Image
Processing, Academic Press, San Diego.
 Wilkie DS and Finn JT (1996), Remote Sensing Imagery for Natural Resources
Monitoring: A Guide for First-Time Users, Columbia University Press, USA.

 https://slideplayer.com/slide/3723856/

10.10 FURTHER/SUGGESTED READINGS

 Campbell JB (2006) Introduction to Remote Sensing, Taylor and Francis,
London.
 Gonzales RC and Woods RE (2018) Digital Image Processing, 4th Edition, Pearson
Education.

 Lillesand TM and Kiefer R (2007) Remote Sensing Image Interpretation,

John Wiley, New York.

10.11 ANSWERS
SAQ I
a) Image enhancement is defined as a process, which is used to define the
changes carried out in an image to improve its appearance for better visual
interpretation.
b) In point operation, the value of pixel is enhanced independent of
characteristics of neighbourhood pixels within each band

SAQ II
a) Contrast is the difference in pixel values of two different objects at the same
wavelength.
b) The types of linear contrast stretch are:
 Minimum-maximum contrast stretch
 Percentage linear and standard deviation contrast stretch
 Piece-wise linear contrast stretch

Terminal Questions
1. Please refer to subsections 10.3.1 and 10.3.2.
2. Please refer to section 10.4.
3. Please refer to subsection 10.4.1.
4. Please refer to subsection 10.4.2.

Contributors: Dr. Omkar Verma and Dr. Manu Mehta

219
Block 2 Image Pre-processing Techniques
…………………………………………………………………….……………………….................……

220
Contributors: Dr. Omkar Verma and Dr. Manu Mehta
 UNIT 11

IMAGE FILTERING AND

BAND RATIOING
Structure_____________________________________________________
11.1 Introduction Atmospherically Resistant Vegetation Index
Expected Learning Outcomes Normalised Difference Snow Index
11.2 Image Filtering 11.5 Significances of Spectral Indexes
Spatial Filtering 11.6 Activity
Frequency Filtering 11.7 Summary
11.3 Band Ratioing 11.8 Terminal Questions
11.4 Commonly Used Spectral Indices 11.9 References
Simple Ratio Index 11.10 Further/Suggested Readings
Normalised Difference Vegetation Index 11.11 Answers
Soil Adjusted Vegetation Index

11.1 INTRODUCTION
Image filtering and band ratioing are image enhancement processes. In the previous unit, you have
learnt that image enhancement is a technique used to improve the readability and capacity of
remotely sensed images for the extraction of meaningful data by humans.One of the image
enhancement techniques, known as contrast enhancement (also called radiometric enhancement)
is used to increase contrast and brightness of an image for extracting useful information (mostly) by
human interpreters. This technique is discussed in detail in unit 10.
Image filtering and band ratioing are other important techniques of image processing. A satellite
image or picture may be enhanced or modified through the use of filtering. Filtering includes the
image processing steps of edge enhancement, sharpening, and smoothing.
The digital image filtering makes images suitable for land use studies and for extracting many
geological structures.
Block 2 Image Pre-processing Techniques
……………………………….……….............…………………………….……………………….................……
The band ratioing technique is used to enhance the spectral difference
between electromagnetic bands by dividing the measure of reflectance for a
pixel in one image band by its geometrically corresponding pixel from another
image band. It is largely used to minimise illumination impacts in
landscape/terrain variation or to highlight objects of interest. In this unit, we will
discuss digital image filtering approaches, basics of band ratioing and
significance of commonly used spectral indices.

Expected Learning Outcomes_______________________

After studying this unit, you should be able to:
 define image filtering and band ratioing;
 discuss common types of image filters;
 apply spatial domain filtering on images;
 apply frequency domain filtering to images;
 elaborate on the principle of band ratioing.
 explore the importance and purpose of different spectral indices.

11.2 IMAGE FILTERING

In simple term, filtering is used to enhance the appearance of an image.Image
filtering is the technique of enhancing or suppressing the tonal changes in an
image, in specific ranges or frequencies of the pixel values. Or, to put it
another way, filtering is the action of enhancing or suppressing certain spatial
variations or pixel digital number (DN) values in a certain way inside a
picture/image. The following are two methods for digitally filtering photographs
that are frequently used:
 Spatial filtering
 Frequency filtering
11.2.1 Spatial Filtering
It is frequently employed for noise reduction and detail extraction. Contrary to
contrast stretching, which affects each individual pixel separately, filtering does
not do this. Each pixel in an image is impacted by its neighbours while filtering.
Information is really moved about the image as a result. The pixels of a picture
are directly utilised in the spatial filtering approach. A mask or kernel is typically
thought of as weights associated with a neighbourhood of spatially adjacent
pixels so that each neighbourhood has a unique central pixel. The mask is
moved across the image so that its centre traverses every pixel.
Moving a "window" of a few pixels in size (e.g., 3x3, 5x5, etc.) across each
pixel in the image, executing a mathematical computation using the pixel
values beneath that window, and replacing the centre pixel with the new value
are common spatial filtering techniques. The calculation is repeated as the
window is advanced by one pixel in both the row and column dimensions,
filtering the entire image and creating a "new" image in the process. Different
filters can be created to enhance or suppress various characteristics by

222 Contributor: Prof. B. Sathya Bama

Unit 11 Image Filtering and Band Rationing
…………………….................……………..………………………………….…………………………..............
changing the values of the window. By effect, spatial filters can be divided into
the following types:
 Smoothing filters,
 Edge enhancement filters and
 Edge filters
i) Smoothing (Low Pass) Filters
These are often referred to as smoothing spatial filters, and they work by
emphasising bigger, more uniform areas of a same tone while suppressing the
minor details in an image. Low-pass filters so typically help to improve the
appearance of an image. Further divisions include order-statistics non-linear
filters and averaging linear filters.
a. Average Linear Filter: It is a linear low pass filter that serves as the
foundation for many smoothing techniques. By averaging close pixels, it is
possible to reduce the inter-pixel variation in values within an image. When
using a low pass filter, the high-frequency information, i.e., rapid intensity
change over short distance in an image is often reduced but the low-
frequency information (minimal intensity variation over several pixels) is
kept in-tact. These filters are employed for noise reduction and rendering a
smooth appearance to the image. An array of ones split by the kernel's
number of components, such as the following 3 by 3 kernel (Fig. 11.1), is an
illustration of an averaging filter.

Fig. 11.1: Averaging filter kernel.

This filter's response is just the average of the pixels in the vicinity of the filter
mask. Consider the following picture patch, which is depicted in Fig. 11.2a. If
the aforementioned kernel is applied to the image's central 3x3 block, we obtain
the low pass filtered output, which is depicted in Fig. 11.2b. 5x5, 7x7 and other
sizes are also possible for the kernel.

10 20 30 40 50 10 20 30 40 50

20 20 60 80 70 20 31 47 52 70

30 60 30 90 20 30 47 61 55 20

40 70 100 40 10 40 62 64 47 10

50 100 80 10 50 50 100 80 10 50

(a) (b)
Fig. 11.2: The logic of applying an average linear filter in an image: a) Input image
before filter; and b) Output image after using an average linear filter.

Contributor: Prof. B. Sathya Bama 223

Block 2 Image Pre-processing Techniques
……………………………….……….............…………………………….……………………….................……
Fig. 11.3 demonstrates the effects of applying an averaging filter to a noisy
image. The averaging filter reduces the amount of noise in the image through
smoothing. The majority of noise in an image is an undesired high-frequency
component.

(a)

(b) (c)
Fig. 11.3: Images showing :a) Original image; b) Noise added input image; and c)
Averaging filtered output image.(Source:https://bhuvan-
app1.nrsc.gov.in/imagegallery/bhuvan.html)

b. Order-Statistics Non-linear Filters: These are non-linear spatial filters and

their response relies on rating the pixels in the neighbourhood in order to
identify their order, and then storing the result in an output file / array
corresponding to the position of the centre pixel in the input image. Median
filters are a typical illustration of these filers.

10 20 30 40 50

20 20 60 80 70

30 60 30 90 20

40 70 100 40 10

50 100 80 10 50

Fig. 11.4: A 5x5 image.

224 Contributor: Prof. B. Sathya Bama

Unit 11 Image Filtering and Band Rationing
…………………….................……………..………………………………….…………………………..............
A median filter is used to store the neighbourhood’s median pixel value, or the
centre element once the elements have been sorted, in the output file or array.
In particular, median filters are helpful in reducing impulsive noise, commonly
referred to as salt-and-pepper noise. Grey levels for salt are 255 and for pepper
they are 0. Consider the 5x5 image in Fig. 11.4 as an illustration.
If we apply a 3x3 median filter on the shaded salt noise-affected pixel (255), the
filtered images will be obtained as shown in Fig. 11.5. i.e., its neighbours are
sorted (20, 20, 30, 40, 60, 70, 100, 255), and its median value (40) is obtained
and replaced for the original pixel.

10 20 30 40 50 10 20 30 40 50

20 20 60 80 70 20 20 60 80 70

30 255 30 90 20 30 40 30 90 20

40 70 100 40 10 40 70 100 40 10

50 100 80 10 50 50 100 80 10 50
a) B
(a) (b)
Fig. 11.5: Filtered image obtained after applying a 3x3 median filter on image
shown in Fig. 11.4: a) Neighbourhood of shaded pixel and b) Median is
replaced as 40.

ii) Sharpening Filters

Sharpening filters highlight edges by removing blur. It enhances the grayscale
transition of an image, which is the opposite of image smoothing. The
arithmetic operators of smoothing and sharpening, also verifies this.
Smoothing is based on the weighted summation or integral operation on the
neighbourhood, the sharpening is based on the derivative or finite difference.
There are a couple of filters that can be used for sharpening (Fig. 11.6). One
of the most popular filters is Laplace operator. It is based on second order
differentiation.

(a) (b)

Contributor: Prof. B. Sathya Bama 225

Block 2 Image Pre-processing Techniques
……………………………….……….............…………………………….……………………….................……

(c)
Fig. 11.6: Filtering of an image: a) Laplacian filter kernel; b) Input image; and c)
sharpening filter output image (Source: https://bhuvan-app1.nrsc.gov.in/
image gallery/bhuvan.html for a)

11.2.2 Frequency Filtering

Frequency filtering approaches translate a picture/image into the frequency
domain using Fourier analysis, and then apply the filter to the changed image
(Fig. 11.7). It gives a plot of frequencies at every pixel. The main filter
application in the frequency domain is to produce a frequency-enhanced image
(Fig. 11.7).

Fig. 11.7: Schematic showing filtering in frequency domain. (Source: modified after
Gonzales and Woods, 2018)

Frequency filtering is usually of three types as listed below:

 Low pass filter
 High pass filter
 Band pass filter
i) Low Pass Filter
A low pass filter keeps the low-frequency components while filtering out the
high-frequency ones. It is used to amplify low-frequency components while
attenuating high-frequency components to smooth the image (Fig. 11.8).

226 Contributor: Prof. B. Sathya Bama

Unit 11 Image Filtering and Band Rationing
…………………….................……………..………………………………….…………………………..............
According to Fig. 11.8, low pass filtering in the frequency domain works as
follows:
G(u,v) equals H(u,v).F(u,v)
where H(u,v) is the filter transform function and F(u,v) is the Fourier transform
of the original image

Where, where D0 is a positive constant (acts as a cut-off frequency of the filter).

(a) (b) (c)

(d) (e)
Fig. 11.8: Schematics showing: a) Perspective view of frequency domain low-
pass filter; b) 2-D image; c) Low pass filter transfer function radial
cross-section; d) Noise added input image; and e) Low pass filtered
output image in frequency domain. (Source: Gonzales and Woods, 2018)

ii) High Pass Filters

A high pass filter keeps the high-frequency components while filtering out the
low-frequency ones. It is used to make the image sharper. In order to
preserve high-frequency components while attenuating low-frequency
components, it sharpens the image.
The following provides the frequency domain high pass filtering mechanism:
H(u,v) = 1 - H'(u,v)
where the high pass filtering Fourier transform is H(u, v) and the low pass
filtering Fourier transform is H'(u, v).
In Fig. 11.9, the low pass output is subtracted from the original image to get
the high pass filter output.
Contributor: Prof. B. Sathya Bama 227
Block 2 Image Pre-processing Techniques
……………………………….……….............…………………………….……………………….................……

(a) (b)
Fig. 11.9: Satellite images showing: a) Noise-added input image; and b) Output of
frequency domain high pass filter . (Source:https://bhuvan-
app1.nrsc.gov.in/imagegallery/bhuvan.html)

iii) Band Pass Filter

It is used to maintain the middle range band of frequencies while removing the
extremely low-frequency and extremely high-frequency components. Edges
are improved while noise is also diminished via band pass filtering.

SAQ I
a) Describe various types of image filters.
b) What will happen if you make the low pass kernel larger, from 3x3 to 7x7?

11.3 BAND RATIOING

As stated previously that the band ratioing technique is used to enhance the
spectral difference between electromagnetic bands by dividing the measure of
reflectance for a pixel in one image band by its geometrically corresponding
pixel from another image band (https://support.esri.com/en-us/gis-
dictionary/band-ratio). The pair of bands is so chosen that the reflectance in
one band is much higher than the reflectance in the other band for the same
pixel. It basically involves dividing the digital number of pixels in one picture
band by the digital number of pixels in the other image band, which is an image
modification technique used to improve the contrast between features. There
are certain factors which caused the spectral differences in the brightness
values of the similar surface features as listed below:
 Terrain illumination
 Topographic slope and aspect
 Ground albedo and shadow
 Seasonal variation in the angle and intensity of sunlight
Collectively, these conditions are termed as environmental conditions and may
create serious obstruction to an interpreter or classification algorithm to detect
surface features or land use land cove classes accurately from the satellite
images (Jensen, 2018). In some cases, ratio transformations of images are
used to minimise the effect of the environmental conditions.

228 Contributor: Prof. B. Sathya Bama

Unit 11 Image Filtering and Band Rationing
…………………….................……………..………………………………….…………………………..............

(a) (b)

(c) (d)
Fig. 11.10: True colour and band ratio images. a) Landsat TM true colour
composite (RGB); b) Thematic mapper band (TM) 3/TM4: This ratio
(d)
has defined barren lands and urban area uniquely but it could not
define water body, forests and croplands ;c) TM2/TM3: this ratio has
distinguished croplands, barren lands sharply but it has not
separated croplands, forests and water body. Both forests and water
body have appeared as lighter tone and barren land appeared as dark
tone and it did not enhance urban area. Since chlorophyll has strong
reflectance in the band 2 (0.52 -0.60 um) region and strong absorption
in the band 3 (0.63 -0.69 um) region, vegetation has appeared as
brighter tone; and d) TM3/TM5: This ratio enhanced barren lands,
highways, street patterns within the urban areas and urban built-up or
cemented areas. It could not enhance clear water but it enhanced
turbid water. This ratio is useful for observing differences in water
turbidity. Barren lands, highways, urban and built-up areas have
appeared as lighter tone and forests, water body and croplands
appeared as dark tone.

Ratio also gives unique information which is not present in any single band and
is useful for differentiating among various features such soils and vegetation.
Mathematically, the ratio function can be expressed as:
BVi,j,r = BVi,j,k/BVi,j,l
Where BVi,j,r is the value of output ratio for the pixel at row i and column j
Contributor: Prof. B. Sathya Bama 229
Block 2 Image Pre-processing Techniques
……………………………….……….............…………………………….……………………….................……
BVi,j,kand BVi,j,l represent the value of brightness at the same location in band k
and l.
If BVi,j,l is equal to zero, then it would become difficult to compute the ratio. In
this case there are some available alternatives, which used by assigning some
value to BVi,j,l between 0 and 1.
Band ratioing technique is generally applied on a multispectral image. Ratio is
an effective technique for selectively enhancing spectral features (Fig. 11.10).
Ratio images derived from different band pairs are often used to generate ratio
colour composites in a Red, Blue and Green (RGB) display. Many indices, such
as the normalised difference vegetation index (NDVI) have been developed
based on both differencing and ratio operations.
Band ratio is an effective technique for suppressing topographic shadows. For a
given incident angle of solar radiation, the radiation energy received by a land
surface depends on the angle between the land surface and incident radiation
(Fig. 11.11). Therefore, solar illumination on a land surface varies with terrain
slope and aspect, which results in topographic shadows. In a remotely sensed
image, the spectral information is often occluded by sharp variations of
topographic shadowing. DNs in different spectral bands of a multi-spectral
image are proportional to the solar radiation received by the land surface and
its spectral reflectance.

Fig. 11.11: Ratio of pixel values in NIR region to the corresponding pixel value in
the visible red region of the spectrum. The ratios for the illuminated and
shaded slopes are very similar, although pixel value differs by a factor
of more than two. Hence, an image made up of NIR:R ratio values at
pixel positions will exhibit a much reduced shadow or topographic
effect. (Source: redrawn after Liu and Mason, 2009)

11.4 COMMONLY USED SPECTRAL INDICES

The spectrum characteristics of features are highlighted by spectral indices,
which are mathematical combinations or transformations of bands. By
emphasising feature differences, the indices make them stand out from other
image features. There are many different types of indices, but vegetation
indices are the most used (Table 11.2). This comprises indices to denote water,
230 Contributor: Prof. B. Sathya Bama
Unit 11 Image Filtering and Band Rationing
…………………….................……………..………………………………….…………………………..............
minerals and burned areas, as well as indicators of man-made features. A
minimum of two distinct bands (wavelengths) are needed for the majority of
indices. Two or more of the image's spectral reflectance bands are used as
inputs in a mathematical procedure to calculate the spectral indices. The
derived spectral index represents a new image that emphasises specific land
surface characteristics or features, such as plant, soil, and water, which appear
more clearly than the original input bands.
Table 11.2: Some commonly used vegetation indices.

Vegetation Index Description

Ratio of green leaf scattering in NIR, chlorophyll
Simple ratio index
absorption in red
Normalised difference Normalised difference of green leaf scattering in
vegetation index NIR, chlorophyll absorption in red
Soil adjusted vegetation It is used in areas where vegetative cover is low
index (< 40%)
Atmospherically
An enhancement of NDVI to better account for
resistant vegetation
atmospheric scattering
index
Normalised difference It is used to identify snow cover and is often used
snow index for hydrological forecasting.
11.4.1 Simple Ratio Index
The Ratio Vegetation Index (RVI) or Simple Ratio (SR) is the most basic ratio-
based measure. This index is just the near-infrared (NIR) band reflectance
divided by the red band reflectance, as shown below:
Simple Ratio Index = Near-Infrared band / Red band
Larger SR values represent healthy vegetation, whereas smaller values
represent ice, water or soil. It typically gives a broad indicator of the vegetation,
but it also has a few related issues. First, there's the problem with dividing by
zero. It is not possible to generate an SR value for this pixel if the reflectance
for the red band is 0. The amount of red reflectance can also affect a wide
variety of possible values. It could be challenging to compare values as a result.
The Normalised Difference Vegetation Index (NDVI), developed in response to
these issues.
11.4.2 Normalised Difference Vegetation Index
One of the most widely used vegetation indexes is the Normalised Difference
Vegetation Index (NDVI). It is a numerical indicator that uses visible and NIR
bands of the electromagnetic spectrum. It is mainly adopted to analyse remote
sensing measurements and assess the ‘greenness’ of the target. NDVI is a very
important tool to study vegetation from remote sensing satellite as most of the
multispectral satellite sensors have visible and infrared channels, which can be
utilised to calculate NDVI. For example, NDVI can be calculated using the red
band and infrared band of LISS III data, by applying the formula given below:
Contributor: Prof. B. Sathya Bama 231
Block 2 Image Pre-processing Techniques
……………………………….……….............…………………………….……………………….................……
NDVI = (NIR – Red) / (NIR + Red)
NDVI works in the following ways. The pigment in plant leaves, chlorophyll
strongly absorbs visible light (from 0.4 to 0.7 µm) for use in photosynthesis. The
cell structure of leaves, on the other hand, strongly reflects NIR light (from 0.7
to 1.1 µm). The more leaves a plant has, more these wavelengths of light are
affected, respectively. We can measure intensity of light coming off the Earth in
visible and NIR wavelengths and quantify photosynthetic capacity of vegetation
in a given pixel of land surface. In general, if there is much more reflected
radiation in NIR wavelengths than in visible wavelengths then vegetation in that
pixel is likely to be dense and may contain some types of forest. If there is very
little difference in the intensity of visible and NIR wavelengths reflected, then
vegetation is probably sparse and may consist of grassland, tundra, or desert.
Since we know the behaviour of plants across EM spectrum, we can derive
NDVI information by focusing on the satellite bands that are most sensitive to
vegetation information (NIR and red). Therefore, bigger the difference between
NIR and red reflectance, more vegetation there has to be.
NDVI algorithm subtracts red reflectance values from NIR and divides it by the
sum of NIR and red bands as mentioned in the equation above.
This normalised index formulation allows us to cope with the fact that two
identical patches of vegetation could have different values if one were, for
example in bright sunshine, and another under a cloudy sky. The bright pixels
would all have larger values, and therefore, a larger absolute difference
between the bands. This is avoided by dividing by the sum of reflectance
values.
Theoretically, calculations of NDVI for a given pixel always result in a number
that ranges from minus one (–1) to plus one (+1); however, in practice, extreme
negative values represent water, values around zero represent bare soil and
values over 0.6 represent dense green vegetation.
NDVI has found a wide application in vegetative studies as it has been used to
estimate crop yields, pasture performance and rangeland carrying capacities
among others. It provides a crude estimate of vegetation health and a means of
monitoring changes in vegetation over time (Fig. 11.12). It can be used to
detect seasonal changes in green biomass but can also be used to detect
changes to human activities (logging) or natural disturbances such as wild fire.
NDVI has been found to be useful for continental or global scale vegetation
monitoring because it can compensate for changing illumination conditions,
surface slope and viewing aspect.
It is often directly related to other ground parameters such as percent of ground
cover, photosynthetic activity of the plant, surface water, leaf area index and the
amount of biomass.

232 Contributor: Prof. B. Sathya Bama

Unit 11 Image Filtering and Band Rationing
…………………….................……………..………………………………….…………………………..............

Fig. 11.12: Healthy vegetation absorbs most of the visible light that hits it and
reflects a large portion of NIR light. Unhealthy or sparse vegetation
reflects more visible light and less NIR light

The most popular and used satellite instrument for collecting NDVI is US
National Oceanic and Atmospheric Administration’s Advanced Very High
Resolution Radiometer (AVHRR) satellite. It is sensitive to wavelengths from
0.55 - 0.7 µm and 0.73 - 1.0 µm, both of which are idealised in NDVI
calculation. AVHRR’s detectors measure intensity of light being reflected from
the different bands. Landsat TM is also used to calculate NDVI but because its
band wavelengths differ (uses bands 3 and 4), it is most often used to create
images with greater detail covering less area. Moderate Resolution Imaging
Spectroradiometer (MODIS) sensor of NASA also has an NDVI standard
product.
11.4.3 Soil Adjusted Vegetation Index
Similar to the NDVI, the Soil Adjusted Vegetation Index (SAVI) is employed in
regions with low (40%) vegetative cover. In order to calculate SAVI, the
following is used:
SAVI = (NIR – Red) / (NIR + Red) x (1 + L)
Where, L is a correction factor, which varies from 0 for very high vegetation
cover to 1 for very low vegetation cover. For intermediate vegetation cover, a L
value of 0.5 is commonly used; when L is equal to zero, the SAVI becomes the
same equation as the NDVI.
11.4.4 Atmospherically Resistant Vegetation Index
An enhanced index called the Atmospherically Resistant Vegetation Index
(ARVI) is used to account for the influence of the atmosphere. It works best in
places with a lot of atmospheric aerosol, like tropical areas with soot pollution.
The benefit of this index is that it uses the blue band in addition to the red and
NIR bands. A self-correction method for the atmospheric influence on the red
band gives the ARVI greater tolerance to atmospheric effects than other
indices. The difference in radiance between the blue and red bands is
performed in order to adjust the radiance in the red channel (Kaufman and
Contributor: Prof. B. Sathya Bama 233
Block 2 Image Pre-processing Techniques
……………………………….……….............…………………………….……………………….................……
Tanre, 1992).Green vegetation typically has ARVI values between 0.20 and
0.80, with a range of -1 to 1. The formula below is used to compute ARVI:
ARVI = (NIR – RB) / (NIR + RB)
RB = Red – γ(Blue band – Red band)
Where RB = red band and γ = 1
11.4.5 Normalised Difference Snow Index
Normalised Difference Snow Index (NDSI) is used to identify snow cover and is
often used for hydrological forecasting. It is calculated by using the following:
NDSI = (Visible – SWIR)/(Visible + SWIR)
Where SWIR = Shortwave infrared
Snow cover reflects strongly and appears as brilliant as clouds in the visible
spectrum. Thus, it is visually challenging to discern from cloud cover. Snow
does not reflect much light in the shortwave infrared region of the spectrum,
allowing clouds and snow cover to be distinguished from one another. While the
threshold may change based on the imagery and kind of snow or ice, higher
NDSI values (> 0.4) often indicate the presence of snow.

11.5 SIGNIFICANCE OF SPECTRAL INDEXES

You are now aware that spectral indexes are fundamentally derived from the
spectral characteristics of the interest features. For instance, spectral indices
for vegetation state are created based on the idea that healthy vegetation
absorbs and reflects significantly in the visible red and NIR regions,
respectively. These characteristics are utilised to create more intricate spectral
indices for tracking phenology factors and vegetation conditions.
Under real-world conditions, the NDVI ranges from values of around -0.2 to 0.6
or 0.7. To discriminate principal land cover classes such as water, non-
vegetationand vegetation the following thresholds in the continuous range are
used (11.13 and 11.14):
i) NDVI with value less 0 indicates water
ii) NDVI with value ranges from0 to 0.2 shows non-vegetation (soil, sealed
surfaces or bare rock)
iii) NDVI with value ranges from 0.2 to 0.6 represents vegetation.
It should be noted that these class limitations are only a very rough
approximation because the mixed pixels effect, canopy reflectance, the
presence of water plants, and the illumination effect of atmospheric or
topographic circumstances can all affect NDVI values. The NDVI can be used
to conceal plants from other types of land cover and to infer the dynamism and
health of the vegetation. Additionally, it is useful for observing the connection
between vegetative development and environmental factors.

234 Contributor: Prof. B. Sathya Bama

Unit 11 Image Filtering and Band Rationing
…………………….................……………..………………………………….…………………………..............

Fig. 11.13: Sentinel 2A Satellite image of Marudhoor area, Thiruvaadhavoor,

Madurai District, Tamil Nadu showing vegetation, soil and other
features (Source:https://bhuvan-app1.nrsc.gov.in/imagegallery/bhuvan.html)

Fig. 11.14: NDVI of the satellite image given in Fig. 11.13 highlighting vegetation
rich areas in green.

It may be noted that when a considerable portion of the soil surface is exposed,
the soil reflectance can manipulate the NDVI values (Fig.11.15). Light reflected
from the soil can have a significant effect on NDVI values and can the values by
up to 20%. Compared to other indices, the ARVI resists the atmospheric effects
and helpful to identify vegetation at different levels (Fig. 11.16).

Fig.11.15: SAVI of a satellite image highlighting the regions of high soil content in
brown.(Source:https://bhuvan-app1.nrsc.gov.in/imagegallery/bhuvan.html)

Contributor: Prof. B. Sathya Bama 235

Block 2 Image Pre-processing Techniques
……………………………….……….............…………………………….……………………….................……

Fig. 11.16: ARVI of the satellite image shown in Fig. 11.19 highlighting vegetation
rich areas in green.

SAQ II
a) List out the any four spectral Indices.
b) What is the use of NDVI?

11.6 ACTIVITY
Write the formulae of the following:
i) BVi,j,r =
ii) ARVI =
iii) NDSI =

11.7 SUMMARY
 Image filtering and band ratioing are image enhancement processes.
 Filtering could be a procedure for modifying or enhancing a satellite image
or picture.
 The digital image filtering makes images suitable for land use studies and
for extracting many geological structures.
 Spatial and frequency filtering are two widely used approaches to digitally
filter images.
 The band ratioing technique is used to enhance the spectral difference
between electromagnetic bands by dividing the measure of reflectance for a
pixel in one image band by its geometrically corresponding pixel from
another image band.
 It is largely used to minimise illumination impacts in landscape/terrain
variation or to maximise phenomenon in objects of interest.
 It is an effective technique for suppressing topographic shadows.

 Spectral Indices are mathematical combinations or transformations of bands

that accentuate the spectral properties of features.

236 Contributor: Prof. B. Sathya Bama

Unit 11 Image Filtering and Band Rationing
…………………….................……………..………………………………….…………………………..............
 The calculated spectral index represents a new image that highlights
particular land surface features or properties such as vegetation, soil and
water, which appear better than the original input bands.
 Simple ratio index, Normalised difference vegetation index, Soil adjusted
vegetation index, Atmospherically resistant vegetation index and
Normalised difference snow index are commonly used indices.

11.8 TERMINAL QUESTIONS

1. What is image filtering? Discuss in detail types of filters used in image
filtering.
2. Describe in detail band ratioing. Add a note on its significance.
3. Write short notes on the following: i) Simple Ratio Indexii) Normalised
Difference Vegetation Index

11.9 REFERENCES
 Jensen J R (2018) Digital Image Processing: A Remote Sensing
Perspective, 4th edition, Pearson India Education Services Pvt Ltd, Uttar
Pradesh, India.
 Kaufman YJ and Tanre D (1992) Atmospherically Resistant Vegetation
Index (ARVI) for EOS-MODIS. IEEE Transactions on Geoscience and
Remote Sensing, 30, 261-270.
 Liu JG and Mason P J (2009) Essential Image Processing and GIS for
Remote Sensing, Wiley-Blackwell, U.K.
 https://support.esri.com/en-us/gis-dictionary/band-ratio

11.9 FURTHER/SUGGESTED READINGS

 Lillesand T M and Kiefer RW (2015) Remote Sensing and Image
Interpretation, 7thEdition, John Wiley.
 Schowengerdt R A (2006) Remote Sensing – Models and Methods for
Image Processing, Academic Press.
 Gonzales R C and Woods R E (2018) Digital Image Processing, 4th Edition,
Pearson Education.
 NPTEL course on Remote Sensing, IISc Bangalore at
https://nptel.ac.in/courses/105108077

11.10 ANSWERS
SAQ I
a) Spatial and frequency domain filters.
b) If we increase the size of the low pass kernel from 3x3 to 7x7, the image get
further smoothed and you will get a blurred appearance.

SAQII
a) SR, NDVI, SAVI, ARVI
Contributor: Prof. B. Sathya Bama 237
Block 2 Image Pre-processing Techniques
……………………………….……….............…………………………….……………………….................……
b) Normalized Difference Vegetation Index (NDVI) enhances the vegetation
and more specifically the healthy vegetation.

Terminal Questions
1. Please refer to section 11.2.
2. Please refer to section 11.3.
3. \Please refer to subsections 11.4.1 and 11.4.2.

238 Contributor: Prof. B. Sathya Bama

 UNIT 12

IMAGE FUSION AND PRINCIPAL

COMPONENT ANALYSIS
Structure_____________________________________________________
12.1 Introduction 12.7 Quality Assessment of Fused Image
Expected Learning Outcomes Qualitative Analysis
12.2 Image Fusion Quantitative Analysis
12.3 Satellite Images Used for Fusion 12.8 Principal Component Analysis
Panchromatic Images Principle of PCA
Multispectral Images Steps for PCA Transformation
12.4 Levels of Image Fusion 12.9 Activity
12.6 Methods of Image Fusion 12.10 Summary
Intensity, Hue and Saturation (IHS) 12.11 Terminal Questions
Transformation Technique 12.12 References
Brovey Method 12.13 Further/Suggested Readings
Wavelet Fusion Technique 12.14 Answers

12.1 INTRODUCTION
You are now well aware about pre-processing and processing operations of image. These
operations include image corrections, enhancement and transformation. These operations are
discussed in details in the previous units. Basically, image transformation (also termed as spectral
transformation) involves manipulation of multiple bands of data, from a single multispectral image or
from two or more images of the same area acquired at different times. In this process, a raw image
or several sets of images undergo some mathematical treatment to generatea new imagery. As a
consequence, new resultant transformed image generated from two or more sources highlights
particular features or properties of interest better than original input images. Hence, the transformed
image may have properties that make it more suited to a particular purpose than original input
images.
Block 2 Image Pre-processing Techniques
…………………………………………………………………….……………………….................……............
Image transformation techniques are useful for compressing bands having
similar kind of information into fewer bands and also to extract new bands of
data that are more interpretable to human eye. Apart from band ratioing, image
fusion and principal component analysis are other types of image
transformationoperations. These operations are much useful in the analysis of
multispectral and hyperspectral images. In this unit, we will discuss two types of
image transformation techniques such as image fusion and principal
components analysis.

Expected Learning Outcomes_______________________

After studying this unit, you should be able to:
 define image fusion and principal components analysis;
 discuss various types of image fusion methods;
 describe principal components analysis; and
 explain the need and significance of eigen values.

12.2 IMAGE FUSION

Image fusion is the process of combining relevant information of two or more
remotely sensed images of a scene into a highly informative single image.
Image fusion has gained prominence in remote sensing application is based on
the fact that the remote sensing instruments have design or observational
constraints, therefore, a single image is not sufficient for visual or machine level
information analysis. Thus, image fusion is an emerging field for generating an
informative image with the integration of images obtained by different sensors
for decision making.
In satellite imaging, two types of images are obtained. The panchromatic
images acquired by satelliteshave higher spatial resolution and multispectral
data have coarser spatial resolution. These two image types are merged
(fused) in order to getinformation of both the images into one single image. The
image fusion technique, thus, allows integration of different information sources
and fused image can have complementary spatial and spectral resolution
characteristics. In other words, fused image will have spatial information of the
panchromatic image and spectral information of multispectral image.
Image fusion is also known as panchromatic sharpening in which a
panchromatic image is fused with a low-resolution multispectral image for
getting a higher spatial resolution image yielding good quality spectral
information. More precisely, image fusion is used to generate high resolution
multispectral image by merging of high resolution panchromatic and low
resolution multispectral images of identical geographical coverage.The fused
images can be used in various applications. While fusing the images the
following factors are taken into consideration (Schowengerdt, 2007; Fonseca et
al., 2008):
 Spectral range of panchromatic image should cover the spectral range of all
multispectral bands involved in the fusion process for preserving the image
colour. It will avoid the colour distortion in the fused image.

240
Contributor: Prof. B. Sathya Bama
Unit 12 Image Fusion and Principal Component Analysis
…………………….................……………..………………………………….…………………………...............
 The high resolution image should be as similar as possible to that of the
replaced low resolution image during the fusion process.
 High resolution image should be globally contrast matched to the replaced
image in order to reduce residual radiometric artefacts.
 Panchromatic and multispectral images must be registered with a precision
of less than 0.5 pixel so that artefacts do not appear in the fused image.
 Ratio between the spatial resolutions of both images should not be greater
than 4, otherwise merging operation becomes difficult due to the registration
and re-sampling.
Image fusion involves transforming a set of low or coarse spatial resolution
multispectral images to high spatial resolution colour images by fusing a co-
registered fine spatial resolution panchromatic image. Usually, three low-
resolution images in the visible spectrum are used as main inputs to produce a
high-resolution natural (true) colour image as shown in Fig. 12.1, where the
image 12.1b is a natural colour image with a spatial resolution of 29 m and
image 12.1a is a panchromatic image with a spatial resolution of 4 m. By
combining these inputs, a high-resolution colour image is produced (Fig.12.1c).
The fused output retains spectral signatures of input colour image and spatial
features of input panchromatic image, and usually the best attributes of both
inputs. The final output with its high spectral and spatial resolution is often as
good as high-resolution colour images. In is important to note that if both inputs
are aerial imagery, resultant is high resolution aerial image, else if both inputs
are satellite imagery, resultant is a high resolution colour satellite image.

(a) (b)

Contributor: Prof. B. Sathya Bama 241

Block 2 Image Pre-processing Techniques
…………………………………………………………………….……………………….................……............

(c)
Fig. 12.1: Illustration of image fusion (merging) of IKONOS and Landsat
multispectral image. a) High spatial resolution IKONOS image, (b)
Landsat multispectral image of the same area and (c) Fused image
having spatial information of IKONOS image and spectral (colour)
information of the Landsat image. Note the improved clarity of features
in the fused image

12.3 SATELLITE IMAGES USED FOR FUSION

There are three types of satellite images namely, panchromatic, multispectral
and hyperspectral are available. Out of which, image fusion is mainly used for
merging panchromatic and multispectral images.
12.3.1 Panchromatic Images
Panchromatic images cover the entire visible region of the electromagnetic
spectrum and are captured as a single band image, giving a higher spatial
resolution to the image (https://eos.com/make-an-analysis/panchromatic/). The
resulting image does not yield any information belonging to a specific
wavelength. It is usually displayed as a gray scale image i.e. the displayed
brightness of a particular pixel is proportional to the pixel digital number which
is related to the intensity of radiation reflected by the targets in the pixel and
detected by the detector. Thus, when it comes to detecting brightness
differences, a panchromatic image is superior to its multispectral counterpart.
12.3.2 Multispectral Images
These are the main type of images acquired by remote sensing sensors. These
sensors are sensitive to different wavelength bands of electromagnetic
spectrum. Hence, multispectral images are also considered opposite of the
panchromatic images that record only the total intensity of radiation falling on
each pixel. Normally, satellites have three or more radiometers. Each one
acquires one digital image in a small band of visible spectra, ranging from 0.7 to
0.4 µm, called red, green, blue (RGB) region, and going to infrared wavelengths
of 0.7 to 10 µm or more, classified as near infrared, middle infrared and far
infrared (https://en.wikipedia.org/wiki/Multispectral_imaging).
242
Contributor: Prof. B. Sathya Bama
Unit 12 Image Fusion and Principal Component Analysis
…………………….................……………..………………………………….…………………………...............

12.4 LEVELS OF IMAGE FUSION

It may be noted that depending on the processing levels, the image fusion is
performed at three levels, which are listed below (Fig. 12.2):
 Pixel level
 Feature level
 Symbol or decision level

Fig. 12.1: Schematic showing levels of image fusion. (Source: modified after Pohl et
al., 1998; Zoran, 2009)

Pixel level image fusion is the simplest technique of image fusion where
fusion of is done at the lowest level. It processes pixels by combining the
intensities and values of two input images based on their average and generate
a new single resultant or output image. It uses raster data that is geo-coded or
at least co-registered. This technique gives more accurate results when
compared with other methods of image fusion.
Feature level image fusion is an intermediate level fusion and used to fuse
extracted salient features such as pixel intensities, edges or textures. Fusion at
feature level requires merging of input images from multiple data sources so
that additional composite features can be created.
Symbol/decision/interpretation level image fusion is a technique of higher
level image fusion. It includes a method that uses value-added data where the
input images are processed separately for information extraction (Pohl and Van
Genderen, 1998). The obtained information is then combined applying decision
rules to reinforce common interpretation and resolve differences and furnish a
better understanding of the observed objects. It may be noted that decision
level fusion is commonly used in applications where different sources or
sensors provide complementary information. In the following discussion, the
pixel level image fusion is presented.

12.6 METHODS OF IMAGE FUSION

You are now aware that image fusion allows us to combine the images of
different spectral and spatial resolution and simultaneously also permits to keep
the radiometric resolution as such. Several efforts have been made to develop

Contributor: Prof. B. Sathya Bama 243

Block 2 Image Pre-processing Techniques
…………………………………………………………………….……………………….................……............
suitable image fusion methods in order to preserve the spectral detail and
increase spatial information of the fused images. There are many methods
which are used for image fusion. These methods can be broadly divided into
two categories:
i) Spatial domain fusion methods: These methods directly deal with pixels
and manipulate pixel values to get desired result/image. The following are
the commonly used spatial domain fusion techniques:
 Intensity,Hue and Saturation (IHS) transformation
 Broveytransform or intensity modulation
 Principal Component Analysis (PCA)
 High pass filtering
ii) Transform domain fusion methods: These methods use the Fourier and
other transforms to first transfer an image into transform domain.
Subsequently, image fusion operations are performed on the transformed
image and then an inverse transform is used to generate an output image.
The following are the commonly used transform domain fusion techniques:
 Wavelet transform
 Laplace pyramid
 Curvelet transform
The commonly used methods are described below:
12.6.1 Intensity, Hue and Saturation (IHS) Transformation
Technique
As you are aware that the colour system RGB is usually used by computer
monitors to display a colour image. However, there is another colour system,
which is widely used to describe a colour is the system of Intensity, Hue and
Saturation (IHS).Basically, IHS transformation technique is based on RGB to
IHS colour space transformation. As you know, an image is displayed in a
colour monitors through its three colour guns which correspond to the three
additive primary colours i.e. red, green, and blue (RGB). When we display three
bands of a multiband image data set, the viewed image is said to be in RGB
space. However, it is also possible to define an alternate colour space that uses
three parameters namely, intensity (I), hue (H) and saturation (S) instead of
RGB. IHS colour space is advantageous in that it represents colours more
similar to as perceived by the human eye.
Intensity is the overall brightness of the scene and varies from black to white
and saturation represents the purity of colour and also varies linearly from 0 to
1, while hue represents colour or dominant wavelength of incident radiation.
Intensity (I) component is similar to the panchromatic image. This peculiarity is
used to produce fusion between panchromatic data having high spatial
resolution and multispectral data characterised by a high spectral resolution
and less spatial detail. The procedure of IHS based image fusion can be
summarised in the following steps (Fig. 12.3):

244
Contributor: Prof. B. Sathya Bama
Unit 12 Image Fusion and Principal Component Analysis
…………………….................……………..………………………………….…………………………...............

Fig. 12.3: RGB-IHS encoding and decoding for image fusion. (Source: redrawn
from Schowengerdt, 1997)

 resampling of the multispectral image (having RGB components) to the

same pixel size of the panchromatic image, using nearest neighbour
resampling criteria. This method, unlike others (Bi-linear and Cubic), does
not introduce radiometric distortion in the images
 transformation of resampled multispectral image from RGB to IHS colour
space. This step transforms RGB values to IHS values
 replace intensity (I) component of IHS image by high resolution
panchromatic image and
 perform reverse IHS-RGB transformation. This step transforms IHS values
to RGB values.
The resultant fused image is a mixture of spectral information from the low-
resolution colour composite and high spatial resolution information from the
panchromatic image, which better shows the image features. Despite being
useful, the major limitation of these image fusion techniques is that they can
distort spectral information of multispectral data while merging. Another major
disadvantage of fusion approaches is that they produce spatial distortion in the
fused image. Spectral distortion becomes a negative factor while we go for
further processing, such as classification.
12.6.2 BroveyTransform or Intensity Modulation
The Brovey transform technique was introduced by an American scientist, Bob
Brovey. It is a shortcut to image fusion and based on direct intensity modulation
(Liu and Mason, 2009). It involves combining of data from various sensors. It
uses a combination of arithmetic operations that multiplies each resampled,
multispectral pixel by the ratio of the corresponding panchromatic pixel intensity
to the sum of all the multispectral intensities (https://desktop.arcgis.com/en/
arcmap/10.7/manage-data/raster-and-images/fundamentals-of-panchromatic-
sharpening.htm). It preserves the corresponding spectral feature of each pixel
and transforms all the luminance details into a high resolution panchromatic
(PAN) image.
In this transformation, the equation uses RGB and panchromatic bands as
inputs to obtain output or fused new RGB bands. The fused R, G, and B images
are defined by the following equations:

Contributor: Prof. B. Sathya Bama 245

Block 2 Image Pre-processing Techniques
…………………………………………………………………….……………………….................……............

The Brovey transform usually yields similar to that of IHS image fusion
operation, but Brovey transform gives a colour distortion to the fused image.
12.6.3 Wavelet Fusion Technique
This technique represents a mathematical tool used in the field of signal
processing for detecting local features (Yang et al., 2009). It also can be
employed to decompose two-dimensional signals consisting of two-dimensional
gray scale image signals into dissimilar resolution levels based on elementary
functions such as wavelets for multi-resolution analysis. By using wavelets, a
signal can be decomposed into several levels, in which each level corresponds
to a coarser resolution (i.e. lower frequency bands) and finer level (i.e. higher
frequency bands) with wavelet coefficients. The wavelet coefficients comprise
spatial differences between two successive resolution levels for each level. The
wavelet fused image gives high spectral quality and less spatial information.
However, both high spectral and spatial resolutions are required for using
image data for various applications. There are two types of wavelet fusion
technique such as discrete and continuous.
i) Discrete Wavelet Fusion Technique
It is the simplest and easiest type of image fusion. It involves multiple-level
decomposition of the input image. In this technique, firstly, the input image is
divided into lower and higher frequency sub-bands with wavelet coefficients for
each sub-band. Secondly, the lower resolution images are replaced with
multispectral band at the same spatial resolution level. Lastly, a reverse wavelet
transform is performed based on the combined transform coefficientsfor
converting the decomposed and replaced divided sets of the image into a fused
image. Fig. 12.4 shows main steps involved in discrete wavelet fusion
technique.

Fig. 12.4: Schematic showing main steps involved in discrete wavelet fusion
technique. (Source: redrawn after Maranur, 2018).

246
Contributor: Prof. B. Sathya Bama
Unit 12 Image Fusion and Principal Component Analysis
…………………….................……………..………………………………….…………………………...............
ii) Continuous Wavelet Fusion Technique
In this technique, a continuous wavelet function is used to determine the
detailed coefficients of a continuous signal. It uses wavelets of different scales
for analysing signals. The continuous wavelet transform represents the sum
over all time of the signal multiplied by scaled and shifted versions of the
wavelet function. It produces many wavelet coefficients that are a function of
scale and position. The wavelets of the original signal can be obtained by
multiplying each coefficient by the appropriately scaled and shifted wavelet
(Misiti et al., 1996).

12.7 QUALITY ASSESSMENT OF FUSED IMAGE

It is important for us to know the quality or accuracy of a fused image so that it
can be used for specific application. The quality of a merged image can be
evaluated by comparing the merged image with a reference image. The
comparison can be done both visually and quantitatively. It is important to note
that we can perform a quantitative and blind evaluation of the fused images
when reference image is not available (Fonseca et al., 2011).
There are two ways for assessing the quality of the fused image namely,
qualitative (or subjective) and quantitative (or objective) analysis.
12.7.1 Qualitative Analysis
It involves visual comparison between a reference image and fused image. In
this analysis, an interpreter based on tone, texture, contrast, saturation, and
sharpness properties of the fused images evaluates the fused image by
comparing it with the reference image. In case of absence of reference mage,
the interpreters use their personal expertise and skill to evaluate the fused
image (Fig. 12.5). While evaluating fused image, the interpreters usually assign
grades to the fused image (Table 12.1)
Table 12.1: Qualitatively grading of a fused image. (Source: after Fonseca
et al., 2011)

Grade Absolute measure Relative measure

1 Excellent Best in group
2 Good Better than the average level in group
3 Fair Average level in group
4 Poor Lower than the average level
5 Very poor Lowest in the group

Contributor: Prof. B. Sathya Bama 247

Block 2 Image Pre-processing Techniques
…………………………………………………………………….……………………….................……............

(a) (b)

(c)
Fig. 12.5: Satellite images of the Worldview 2; (a) High resolution panchromatic
image at 0.46 m. (b) A 1.84 m low resolution multispectral image
interpolated at 0.46 m. (c) Fused image. Note that the fused image can
qualitatively be evaluated by using elements of image interpretation

12.7.2 Quantitative Analysis

It uses quality indicators that measures spectral and spatial similarity between
multispectral and fused images.
i) Correlation Coefficient (CC): The similarity between the fused and
reference images can be calculated using correlation coefficient. The unity
of correlation coefficient shows that both images are the same. It is defined
as -

248
Contributor: Prof. B. Sathya Bama
Unit 12 Image Fusion and Principal Component Analysis
…………………….................……………..………………………………….…………………………...............
Where, and are the fused image and pan-sharpened image, respectively.
The mean values of and are and , respectively. and are the
pixel values corresponding to the pixel of the images and
respectively.
ii) Root Mean Square Error (RMSE): It is frequently used to compare the
difference between the original and pan-sharpened images by calculating
the changes in pixel values. It is defined as:

Where denotes reference image and is the fused image, i and j are the row
index and column index, X and Y denote the horizontal and vertical dimensions.
The smaller value of RMSE indicates better performance.
iii) Peak Signal to Noise Ratio (PSNR): It is widely used metric and is
computed by the number of gray levels in the image divided by the
corresponding pixels in the reference and fused images. When the value is
high, the fused and reference images are similar. A higher value indicates
superior fusion. PSNR is defined as -

Where, represents the reference image and is the fused image, i and j are
the row index and column index.

SAQ I
a) What is a pan sharpening?
b) What are the various processing levels of image fusion?
c) List out any four evaluation metrics used for image fusion.

12.8 PRINCIPAL COMPONENT ANALYSIS

It may be noted that some of the bands of a multispectral or hyperspectral
image are most often highly correlated with several other bands containing
nearly the same information. Therefore, it is often required to reduce such
redundant information. Principal Component Analysis (PCA) is one of the
multivariate techniques that is widely used for data dimensionality reduction to
decompose the highly correlated bands into uncorrelated bands that can be
sorted on decreasing variance. In case of image processing, PCA is an image
transformation technique and is much useful in the analysis of multispectral and
hyperspectral images (Liu and Mason, 2009; Jensen, 2018). It is defined as a
technique that transforms the original remotely sensed image data into a set of
uncorrelated variables. The uncorrelated variables represent which ones
contain most of the information of original images and are easier to interpret
and which ones can be dropped due to less information.
Contributor: Prof. B. Sathya Bama 249
Block 2 Image Pre-processing Techniques
…………………………………………………………………….……………………….................……............
12.8.1 Principle of PCA
The goal of performing PCA involves the rotation or translation of the
orthogonal coordinate system of the input or original image bands in order to
match the axes of ellipsoidal cluster or to produce a new set of uncorrelated
variables of the image data. It re-distributes the brightness (pixel) values of the
original image and gives new values to the axes upon rotation for getting the
best information. Consider a multispectral thematic mapper (TM) image, in
which bands 1, 2, and 3 are highly correlatable and bands 5 and 7 are least
correlatable (Table 12.2). The correlation indicates redundancy in information
between the bands (Liu and Mason, 2009). Further, it probably shows that the
original axes arrangement of these bands may not be the best arrangement to
analyse the image data. In this case, we use PCA for getting best information
from the satellite image data. The PCA transformation operation on to the
correlated multispectral satellite image data gives an uncorrelated dataset
having ordered variance characteristics.
Table 12.2: Correlation matrix of TM image consisting of band 1 to band 5
and band 7. (Source: Liu and Mason, 2009).

Bands of TM image
Correlation
1 2 3 4 5 7
1 1.00 0.962 0.936 0.881 0.839 0.850
Bands of TM image

2 0.962 1.000 0.991 0.965 0.933 0.941

3 0.936 0.991 1.000 0.979 0.955 0.964
4 0.881 0.965 0.979 1.000 0.980 0.979
5 0.839 0.933 0.955 0.980 1.000 0.993
7 0.850 0.941 0.964 0.979 0.993 1.000

Fig. 12.6: Diagrammatic representation of the PCA (Source: redrawn after

Fonseca et al., 2011). PAN stands for panchromatic image, MS for
multispectral image, FUS for fused image, PCA for principal component
analysis and PC for principal component.
250
Contributor: Prof. B. Sathya Bama
Unit 12 Image Fusion and Principal Component Analysis
…………………….................……………..………………………………….…………………………...............
While performing PCA, several uncorrelatable variables are created and these
variables are described as principal component (PC), ranging from PC1, PC2,
… to PCn (Fig. 12.6). These variables are sometimes used in addition or in
place of the original variables in PCA. This process can be visualised by taking
two-dimensional distribution of brightness values of the two bands (say band1
and band 2) of a TM image.
Let us consider Fig. 12.7 shows an elliptic cluster of the image data points of
band 1 and band 2. The main goal of PCA is to translate or rotate the
coordinate system of band 1 and band 2 to match PC1 and PC2 representing
the two axes of the ellipsoid (Liu and Mason, 2009). The digital number values
at each point of the image represent the coordinate system. PC1 is shown by
the longest axis of the ellipsoid, PC2 by the second longest axis and so on.
PC1 represents the maximum amount of variation (or variance) in the image
data and gives maximum information, PC2 shows second most variation and
finally, PCn covers the least variation in the image data and shows little or no
information. It is interesting to note that the most of the information is present in
the first few components, like PC1 and PC2.

Fig. 12.7: Two-dimensional representation of image data points of band 1 and

band 2 for PCA transformation. (Source: redrawn after Liu and Mason,
2009)

12.8.2 Steps for PCA Transformation

i) Data Preparation
It includes collection and preprocessing of the remote sensing data. The
preprocessing comprises radiometric calibration, geometric correction and
atmospheric adjustment. The data should be prepared in a format that can be
used for PCA such as a matrix where each row represents a pixel and each
column represents a spectral band.
ii) Standardisation of Data
Since PCA is highly sensitive to the scale of the data, it is often important to
standardise the data to have mean = 0 and variance = 1. If there are only two
variables, univariate statistics like mean and variance is used. In case of
multiple variables, multivariate statistics like covariance and correlation is used.
While the former is used for data at different scales, the latter is preferred when
scales are the same.

Contributor: Prof. B. Sathya Bama 251

Block 2 Image Pre-processing Techniques
…………………………………………………………………….……………………….................……............
iii) Covariance Matrix Calculation
It involves calculation of the covariance matrix of the data. The covariance
matrix shows the relationships between different spectral bands and helps
identify the dominant patterns in the data.
iv) Eigenvalue Decomposition
It includes the performing of eigenvalue decomposition or singular value
decomposition on the covariance matrix. This step yields the eigenvalues and
eigenvectors of the matrix. The eigenvalue describes the percentage of the
variance that is contained within each PC. The higher the eigenvalue as a
fraction of the sum of the eigenvalues, the more relative information is
contained in the corresponding PC. Eigenvectors shows the direction of
maximum variance.
v) Principal Component Selection
The eigenvectors created in the previous steps are arranged in descending
order and the first few are selected depending on the number of components
required. Each multispectral or hyperspectral pixel vector is projected onto the
selected eigenvectors (dot products taken between pixel vector and
eigenvectors) and the resultant dot product outputs are the pixel values in the
corresponding principal components. For example, the dot product of the pixel
vector with the eigenvector corresponding to the highest eigenvalue yields the
pixel value in the first principal component. Likewise, the dot product of the pixel
vector with the eigenvector corresponding to the second highest eigenvalue
yields the pixel value in the second principal component, and so on.
vi) Dimensionality Reduction

It represents selection of a subset of the principal components that retain most

of the variance in the data. The first few principal components are selected, as
they contain the most important information. By reducing the dimensionality of
the data, it becomes easier to analyse and visualise the dataset.
The reduced data obtained through PCA can be further analysed or used for
various remote sensing applications, such as land cover classification, change
detection, or anomaly detection. Additionally, PCA can support in the
visualisation of the data by representing it in a lower-dimensional space while
preserving as much information as possible.
vii) Reconstruction
It is possible to reconstruct an approximation of the original multispectral /
hyperspectral image from the reduced dimensional principal component
dataset. This step involves multiplying a transformation matrix with the reduced
dimensional principal component pixel vector. Columns of the transformation
matrix will be the eigenvectors used to generate the principal components. After
the reconstruction, rescaling the elements to the dynamic range of the original
multispectral bands is required since the eigenvectors are real numbers, with
positive and negative elements in the second principal component onwards.
252
Contributor: Prof. B. Sathya Bama
Unit 12 Image Fusion and Principal Component Analysis
…………………….................……………..………………………………….…………………………...............

SAQ II
a) What are the steps to be followed in PCA?
b) What is the need of PCA in hyperspectral image analysis?

12.9 ACTIVITY
 You take two images (one panchromatic and one multispectral) of the same
with the same scale from any online source. Try to merge or fuse them and
evaluate the land use land coves features from the fused image.

12.10 SUMMARY
 Image transformation involves manipulation of multiple bands of data, from a single
multispectral image or from two or more images of the same area acquired at
different times.

 Image fusion is the process of combining relevant information of two or more

remotely sensed images of a scene into a highly informative single image.

 Panchromatic and multispectral images are commonly used in image fusion.

 Depending on the image processing levels, the image fusion is performed at

three levels such as pixel, feature and decision.
 Broadly, there are two types of methods namely; spatial domain fusion and
transformdomain fusion are used for image fusion.
 There are two ways for assessing the quality of the fused image namely, qualitative
(or subjective) and quantitative (or objective) analysis.

 Principal component analysis technique that transforms the original remotely

sensed image data into smaller set of uncorrelated variables.The
uncorrelated variables represent most of the information of original images
and are easier to interpret.
 In PCA, the most of the information is present in the first few principal
components, like PC1 and PC2.

12.11 TERMINAL QUESTIONS

1. Define image fusion and discuss spatial domain image fusion methods.
2. What are panchromatic and multispectral images?
3. Describe principle of principal component analysis.

12.12 REFERENCES
 Fonseca L, Namikawa L, Castejon E, Carvalho L, Pinho C, Pagamisse A
(2011) Image fusion for remote sensing applications. Image Fusion and Its
Applications. InTech. https://www.intechopen.com/chapters/15846.
 Jensen J R (2018) Digital Image Processing: A Remote Sensing
Perspective, 4th edition, Pearson India Education Services Pvt Ltd, Uttar
Pradesh, India.

Contributor: Prof. B. Sathya Bama 253

Block 2 Image Pre-processing Techniques
…………………………………………………………………….……………………….................……............
 Liu J G and Mason P J (2009) Essential Image Processing and GIS for
Remote Sensing, Wiley-Blackwell, U.K.
 Maranur J R (2018) Image fusion using the discrete wavelet transform
method. International Journal of Scientific Development and Research, 3(9),
220-226.
 Misiti, M, Misiti Y, Oppenheim G and Poggi J-M (1996) Wavlet Toolbox. The
MathWorks, Inc. USA.
 Pohl C and Van GenderenJ L (1998) Multisensor image fusion in remote
sensing: Concepts, methods and applications, International Journal of
Remote Sensing, 19:5, 823-854.
 Yang Y, Park DS, Huang S and Rao N (2010) Medical image fusion via an
effectivewavelet-based approach. EURASIP Journal on Advances in Signal
Processing. 579341, 1-13
 Zoran L F (2009) Quality evaluation of multiresolution remote sensing
images fusion. UPB Science Bulletin Series C, 71(3), 37-52.

12.13 FURTHER/SUGGESTED READINGS

 Lillesand T M and Kiefer RW (2015) Remote Sensing and Image
Interpretation, 7th Edition, John Wiley.
 Schowengerdt R A (2006) Remote Sensing – Models and Methods for
Image Processing, Academic Press.
 Gonzales R C and Woods R E (2018) Digital Image Processing, 4th Edition,
Pearson Education.
 http://www.ce.yildiz.edu.tr/personal/songul/file/1097/principal_components.p
df

12.14 ANSWERS
SAQ I
a) Pansharpening is a process of merging high-resolution panchromatic and
lower resolution multispectral imagery to create a single high-resolution
multispectral image.
b) There are three levels of image fusion such as pixel, feature and decision.
c) The four evaluation metrics used for image fusion include correlation
coefficient, root mean square error, peak signal to noise ratio and degree of
distortion.

SAQ II
a) The PCA steps include the following:

 Get data.

 Compute the mean vector (µ).

 Calculate the covariance matrix.

 Calculate the eigen vectors and eigen values of the covariance matrix.

 Choosing components and forming a feature vector.

254
Contributor: Prof. B. Sathya Bama
Unit 12 Image Fusion and Principal Component Analysis
…………………….................……………..………………………………….…………………………...............
b) PCA is used as a dimensionality reduction technique in hyperspectral image
analysis.

Terminal Questions
1. Please refer to section 12.1 and subsection 12.6.1.
2. Please refer to section 12.3.
3. Please refer to subsection 12.8.1.

Contributor: Prof. B. Sathya Bama 255

Block 2 Image Pre-processing Techniques
…………………………………………………………………….……………………….................……............

256
Contributor: Prof. B. Sathya Bama
 GLOSSARY
Algorithm : It specifies how to determine higher level data from
lower level source data.
Azimuth : It refers to the direction parallel to the path of flight of an
orbit of a satellite.
Band : It is a segment of the electromagnetic spectrum.
Brightness : It is the gray level value of a pixel within an image
corresponding to energy intensity. The high valve of the
gray level indicates the higher brightness.
Calibration : It is a process of comparing accuracy of an instrument to
the known standards.
Contrast : It is the amount of gray level variation within an image.
Error matrix : It refers to a table used to know accuracy of the thematic
map by comparing the map classification with the
reference data.
Filtering : It is the process by which the tonal variations in an
image in selected ranges or frequencies of the pixel
values are enhanced.
Image Classification : It is process of using characteristics of image data to
assign classes to a particular region.
Image Distortion : It is an error present in remote sensing images.
Image Enhancement : It refers to the techniques used to increase the human
interpretability and ability to derive meaningful
information from remotely sensed images.
Earth Observations : It is process of gathering of information about physical,
chemical and biological components of earth through
satellites.
Geometric Correction : It refers to the process of correcting geometric
distortions and assigning properties of a map to an
image.
Grayscale : It is a range of gray level or shades corresponding to the
pixel values of an image.
LiDAR : Light Detection And Ranging.
Masking : A process of removing some parts of an image.
Mean : It is the average of a set of data values.
Multispectral Image : It is the main type of image acquired by remote
sensing radiometers. Dividing the electromagnetic
spectrum into many bands, multispectral is the opposite
of panchromatic, which records only the total intensity of
radiation falling on each pixel.

257
Near Infrared : It refers to the band of the electromagnetic spectrum
from about 700 to 2500 nm.
NDVI : Normalised Difference Vegetation Index.
Panchromatic Image : It is an image uses a single band that combines red,
green and blue bands, allowing for a greater spatial
resolution. The resulting image does not contain any
wavelength specific information. It is usually displayed
as a gray scale image.
Pan Sharpening : It is a radiometric transformation where a high resolution
panchromatic image is fused with a low resolution
multispectral image.
Polarisation : It refers to the orientation of an electromagnetic wave.
Pixel : It is the smallest element of an image that has been
electronically coded in an array.
Pixel Value : Digital number showing information stored in one pixel
of a raster.
Principal Component : It is a technique that transforms the original vector
Analysis image data into smaller set of uncorrelated variables.
Principal components are derived such that the first
principal component accounts for much of the variation
of the original data.
Radar : It is an active sensor that uses its own radio waves to
measure the distance to an object.
Radiance : It is a measure of the light or energy emitted from a
target object.
Ratio Bands : These are a derivative band generated by dividing one
original band by another.
Reflectance : It is a measure of the light or energy reflected off a
target object.
Resolution : it refers to the smallest spatial feature and gray level
value that an image system can resolve.
Scattering : It refers to the dispersion of electromagnetic radiation as
a consequence of its interaction with molecules in the
atmosphere.
Threshold : It refers to the value used to segment the gray level
values of an image into two different regions.

258
Dear Learner,
While studying the theory course MGY-005 (Techniques in Remote Sensing and Digital Image
Processing), you may have found certain portions of the text difficult to comprehend. We wish to
know your difficulties and suggestions, in order to improve the course. Therefore, we request you to
please fill out and send us the following questionnaire, which pertains to this course (i.e., MGY-005).
If you find the space provided is insufficient, kindly use a separate sheet.

Please mail to:

Course Coordinator MGY-005
Discipline of Geology
School of Sciences
IGNOU, Maidan Garhi
New Delhi – 110068 (India)
Questionnaire

1. Name: _______________________________________________________________________

2. Enrolment No.

3. Regional centre code and name: ___________________________________________________

4. Study centre code and name: _____________________________________________________

5. How many hours did you spend for studying the each unit?

Course MGY-005 (Techniques in Remote Sensing and Digital Image Processing)

Unit no. 1 2 3 4 5 6 7 8 9 10 11 12

No. of
hours

6. Please give your reactions (by ‘√’ mark) to the following items based on your reading of the theory
course:
Items Excellent Very Good Good Poor Give specific examples,
if poor
Presentation Quality
Language and Style
Illustrations used
(Diagrams, Tables etc.)
Conceptual Clarity
Self Assessment
Questions (SAQ)
Feedback to SAQ

7. Your specific comments on the course MGY-005

……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………