TERM PROJECT

Analyzing Land Use and Land Cover (LULC) Changes

in IIT Kanpur Over 10 Years

Group Members:
Adarsh Kumar 210048 [email protected]
Kshitij Ranjan 210538 [email protected]
Love Sinha 210565 [email protected]

A project presented for the course

CE671

Department of Civil Engineering

Indian Institute of Technology Kanpur
November 12, 2024
 CE 671: Introduction to Remote Sensing

1 Introduction
• In this project, we analyzed the changes in land use and land cover (LULC) within the
IIT Kanpur campus over the past 20 years.
• We downloaded and processed satellite imagery to classify the land into four primary
categories: forest, open land, buildings, and roads.
• We then created a video or GIF to visualize these changes over time.
• Additionally, we generated accuracy matrices to assess the classification’s precision and
conducted analytics based on the observed changes.

2 Methodology
2.1 Downloading Satellite Data
In this exercise, we used the Sentinel-2 (2A) data from 2015 to 2024, focusing on the Winter
Season during November.

1. We used the Copernicus Data Space Ecosystem to extract the required satellite data.
There, we explored the Sentinel-2 Data and browsed the IIT Kanpur Campus. We created
a polygon(square) around the Campus area, making it our region of interest.

Figure 1: Polygon for ROI

2. Before searching for the possible data available for our ROI, we filtered the output based on
the time range. We focused on November from the year 2014. Also, to avoid unnecessary
haze, we filtered those data that had more than 3% cloud cover interference.

Figure 2: Time Range filter

1
 CE 671: Introduction to Remote Sensing

3. From the list of 200+ footprints available, we chose 10 with the 2A corrections. The
moral behind this step was that it comes with the additional feature of Radiometric
Corrected imagery.

Figure 3: Mean and other Stats

4. In the downloaded data, we used the B2, B3, B4 and B8 bands, which are the default RBG-
NIR bands in Sentinel-2. The data was downloaded in WGS84 with UTM projection.

2.2 Collecting Ground Truth Data

1. For the collection of ground truth data, we used Google Earth PRO satellite images. There,
we marked 300+ points which minor changes as per ground truth for each and every
year and classified them into 4 classes, which are namely:

• Forest: Areas covered with dense vegetation, which may include trees, orchards, high
bushes etc.
• Open Land: Grasslands, parks, or any non-built-up area.
• Buildings: All built-up structures, including academic buildings, hostels, etc.
• Roads: Paved roads, brick or concrete pathways, etc.

2. The CSV files (initially .kmz files) collecting the class information, as well as the longitudes
and latitudes, were saved separately for each class and year. These will be used in 80:20
proportions for classification and cross-validation, respectively.

2
 CE 671: Introduction to Remote Sensing

Figure 4: Google Earth - Ground Truth

2.3 Image Classification

1. Pre-processing:

• Radiometric Correction: There is no further need to do this as the downloaded

images are of Sentinal-2_2A data, which are already Radiometrically corrected and
hence we get a haze-free and clear imagery.
• Image Cropping:
(a) For Image Cropping, we opened the QGIS software. There we imported the 4
bands of an year through Add Raster ribbon.

(a) Add raster (b) Bands imported from 2024

3
 CE 671: Introduction to Remote Sensing

• In the imported Band Image, we locate the IIT Kanpur Campus

Figure 6: Locating IIT Kanpur

• Now we went to the New Shapefiles Layers and chose the polygon type for the crop-
ping. We also ensured georeferencing by choosing the CRS as EPSG:4326 - WGS
84.

Figure 7: New Shapefile Toolbox

4
 CE 671: Introduction to Remote Sensing

• Futher, we created a Polygon focusing on the IIT Kanpur Campus, and created a
Shape file for cropping

(a) Polygon (b) Shape File

• Now we move to the SCP Plugin for clipping using this polygon. We select the band
layers for clipping and in pre-processing drop down,set them all in Band Set 1. The
vector for clipping will be our IIT Kanpur Shape File. Finally we run the setup to
clip all the four layers.

(a) Band Set 1 (b) Shape File chosen for clipping

• We get a cropped band image, which focuses on the IIT Kanpur Campus only. This
will next be used for classification.

Figure 10: Cropped Band Image

5
 CE 671: Introduction to Remote Sensing

2. Classification Methodology and Saving of TIFF files

• I chose the Google Earth Engine platform for classification and used the Gradient
Tree Boost classifier for supervised classification. This method enabled efficient
handling of large datasets and provided robust accuracy metrics. The code can be
accessed at the following link:
https://code.earthengine.google.com/25c80ddbbe59e05c4fb32ef169f520d8?n
oload=true.
• A summary of what was done goes as follows:
(a) Loaded Satellite Imagery: Imported Sentinel-2 bands (B2, B3, B4, and B8)
and created a combined image. A valid mask was applied to exclude invalid data
points.
(b) Labeled Ground Truth Data: Class labels were assigned to four feature col-
lections representing different land cover types i.e Forest, Buildings, Open Land
, Roads.
(c) Split Data: The data was split into training and testing datasets in an 80-20
ratio.
(d) Trained Classifier: A smileGradientTreeBoost classifier was trained on the
training data. The classifier’s performance was evaluated using a confusion ma-
trix, overall accuracy, and kappa coefficient.
(e) Classified Scene: The trained classifier was applied to the entire scene. The
results were visualized using class-specific colors, and pixel counts for each class
were calculated.
(f) Exported Results: The classified map was exported as a GeoTIFF file named
IITK_Classified_2024.tif, containing all classified zones with the specified
class labels. Link for tif files:
https://drive.google.com/drive/folders/1ngYs18OIgAqhOZASv8QrYAPOHU
IPlXw8?usp=sharing.
(g) A total of 296 points were used as training comprises of different land cover
feature (which is 80% of the total ground truth taken i.e. 370)

Table 1: Statistics of training input

Forests Buildings Open Lands Roads

Number of Points 56 96 72 72

2.4 Generating Accuracy Matrices

• The accuracy assessment was performed in Google Earth Engine. Each year’s accuracy
report, including the confusion matrix and the derived metrics, is documented in the PDFs
attached to the provided ZIP file.

6
 CE 671: Introduction to Remote Sensing

• A snippet of the code is shown as follows:

Figure 11: Confusion Matrix Code

• Discrepancies and their potential causes:

2.5 Aligning Data to a Common Reference

1. Map Alignment:
Since all classified images are in the same spatial reference system (WGS84 with UTM
projection), there is no need to align them again. This can be checked by right-clicking on
the layers in QGIS.

2. Resampling: All our images have the same resolutions of 10m, so there is no need to
resample them.
Still, one can go to layer properties -> pyramid and use bilinear or cubic resampling
methods to maintain data integrity.

Figure 12: Resampling (optional)

7
 CE 671: Introduction to Remote Sensing

2.6 Creating the GIF

1. Image Preparation: Each classified map was labelled to its corresponding year. The
following colour scheme was consistently used for class-identification:

• Yellow: Open Lands

• Green: Forests
• Gray: Buildings
• Black: Roads

2. The ezgif.com website was used to create a GIF file for the 10-year classification images.
The delays and cross-fade were selected meticulously.

Figure 13: GIF Editing

3. The titles and legends were also added frame-wise on the website.

8
 CE 671: Introduction to Remote Sensing

3 BONUS QUESTION
3.1 Change Detection:

Figure 14: Timeline of LULC in %

The pixel count after the classification details were used to create the % of Classes, which are
depicted above.

3.2 Trend Analysis:

The major trends observable over the years are as follows:

1. The general range of Land Covers is:

(a) Vegetation

4 Conclusions
The various conclusions are as follows:-

1. I analyzed spectral profiles to identify materials by comparing image data to known spectral
libraries.

2. I extracted mean spectra from ROIs, allowing me to compare average spectral features for
accurate material identification.

3. I designed color images using specific bands to enhance the discrimination of minerals
based on their unique spectral characteristics.

9
 CE 671: Introduction to Remote Sensing

References
[1] ENVI Tutorial: https://www.nv5geospatialsoftware.com/Support/Self-Help-Tools
/Tutorials

10