remote sensing

Article
Evaluation and Selection of Multi-Spectral Indices to Classify
Vegetation Using Multivariate Functional Principal
Component Analysis
Simone Pesaresi 1, * , Adriano Mancini 2 , Giacomo Quattrini 1 and Simona Casavecchia 1

1 Department of Agricultural, Food, and Environmental Sciences, D3A, Università Politecnica delle Marche,
Via Brecce Bianche 12, 60131 Ancona, Italy; [email protected] (G.Q.); [email protected] (S.C.)
2 Department of Information Engineering, DII, Università Politecnica delle Marche, Via Brecce Bianche 12,
60131 Ancona, Italy; [email protected]
* Correspondence: [email protected]

Abstract: The identification, classification and mapping of different plant communities and habitats is
of fundamental importance for defining biodiversity monitoring and conservation strategies. Today,
the availability of high temporal, spatial and spectral data from remote sensing platforms provides
dense time series over different spectral bands. In the case of supervised mapping, time series based
on classical vegetation indices (e.g., NDVI, GNDVI, . . .) are usually input characteristics, but the
selection of the best index or set of indices (which guarantees the best performance) is still based
on human experience and is also influenced by the study area. In this work, several different time
series, based on Sentinel-2 images, were created exploring new combinations of bands that extend
the classic basic formulas as the normalized difference index. Multivariate Functional Principal
Component Analysis (MFPCA) was used to contemporarily decompose the multiple time series. The
principal multivariate seasonal spectral variations identified (MFPCA scores) were classified by using
a Random Forest (RF) model. The MFPCA and RF classifications were nested into a forward selection
strategy to identify the proper and minimum set of indices’ (dense) time series that produced the
most accurate supervised classification of plant communities and habitat. The results we obtained
Citation: Pesaresi, S.; Mancini, A.; can be summarized as follows: (i) the selection of the best set of time series is specific to the study
Quattrini, G.; Casavecchia, S.
area and the habitats involved; (ii) well-known and widely used indices such as the NDVI are not
Evaluation and Selection of
selected as the indices with the best performance; instead, time series based on original indices (in
Multi-Spectral Indices to Classify
terms of formula or combination of bands) or underused indices (such as those derivable with the
Vegetation Using Multivariate
visible bands) are selected; (iii) MFPCA efficiently reduces the dimensionality of the data (multiple
Functional Principal Component
Analysis. Remote Sens. 2024, 16, 1224.
dense time series) providing ecologically interpretable results representing an important tool for
https://doi.org/10.3390/rs16071224 habitat modelling outperforming conventional approaches that consider only discrete time series.

Academic Editors: No-Wook Park,

Keywords: sentinel-2; time-series; functional data analysis; multivariate functional principal component
Kiwon Lee and Kwangseob Kim
analysis; habitat mapping; supervised classification; remote sensing; land surface phenology
Received: 20 February 2024
Revised: 27 March 2024
Accepted: 28 March 2024
Published: 30 March 2024 1. Introduction
Classifying and mapping plant communities and habitats are crucial for biodiversity
monitoring and defining conservation strategies for Natura 2000 sites in Europe [1,2].
Copyright: © 2024 by the authors.
Currently, vegetation mapping benefits from the growing availability of high-quality data
Licensee MDPI, Basel, Switzerland. from remote sensing platforms such as Landsat, MODIS and Sentinel [3,4]. These platforms
This article is an open access article offer multi-temporal and multi-spectral time series data enabling the capture of seasonal
distributed under the terms and variations in spectral reflectance related to the different phenological stages of vegetation
conditions of the Creative Commons (i.e., vegetation seasonality). These kinds of data are essential for an accurate supervised
Attribution (CC BY) license (https:// classification and mapping of plant communities and habitats [5–13]. Many studies have
creativecommons.org/licenses/by/ demonstrated the potential of direct machine learning applications for raw satellite multi-
4.0/). temporal data [14–18]. These models, which we can define as ‘Pure Machine Learning’

Remote Sens. 2024, 16, 1224. https://doi.org/10.3390/rs16071224 https://www.mdpi.com/journal/remotesensing

Remote Sens. 2024, 16, 1224 2 of 26

according to Durell et al. [19], usually use time series of individual spectral bands or
classic vegetation indices, such as the popular NDVI [20], consisting of a limited number of
scenes within a single year. However, these models rely on human experience and prior
knowledge of the best data acquisition time points and the most suitable set of indices to
capture habitats during their optimal phenological stages. Therefore, these models face
challenges in terms of transferability [21]. It is clear that recommending universal optimal
time points and indices for all habitats across diverse study areas with varying vegetation
and ecological characteristics is not feasible, despite the availability of indices tailored
for specific applications [22,23]. In this context it is necessary to develop adaptable and
transferable models that can autonomously select suitable indices and determine the ideal
times for data acquisition based on the specific vegetation and ecological characteristics of a
study area. A carefully selected set of area-specific indices offers significant advantages for
land management organisations in compliance with national and international guidelines,
such as the Habitats Directive [1,24,25]. These models should handle dense time series of
remotely sensed data. Such data, which, in a specific time window, provide a richer wealth
of information than multi-temporal data, are optimal for analysing seasonal changes in
vegetation and improving classification accuracy [26,27].
Recently, promising methods known as ‘Hybrid statistical-functional Machine
Learning’ [19], which combine machine learning with Functional Data Analysis (FDA) [28],
have been employed to classify and map vegetation and habitats in two Natura 2000 sites [29,30].
Exploring such hybrid models is essential because they are capable of efficiently analysing
dense time series of remote sensing data. The results are not only accurate but also facilitate
interpretations and provide support to phytosociologists and ecologists in understanding
the temporal spectral behaviour of plant associations (plant communities) [31–33]. The
efficiency of analysing dense time series by FDA lies in its fundamental philosophy, which
considers observed data functions as single entities, rather than merely as a sequence of
individual observations [34]. In practice, if the entire time series of a pixel is expressed
as a time function and considered as a single statistical unit, then a stack of remotely
sensed images (a cube with x, y and t axes) is considered as a single temporal archive [35],
essentially composed of as many functions as there are pixels in the area under test. The
pixel-based functions (times series) of remotely sensed data can be thought of as points (or
pixels) within a functional space [34]. The functional space can be univariate or multivari-
ate, depending on the number of metrics (band or indices) used to describe and track the
spectral variations within it (Figure 1).
Functional Principal Component Analysis (FPCA) is one of the most popular tech-
niques in FDA for reducing the amount of functional data [36,37]. FPCA adapts traditional
Principal Component Analysis (PCA) concepts to functions, allowing it to identify the
main modes of variation among observations (functions) within a univariate functional
space. It is evident that multivariate functional spaces are more natural and effective than
univariate ones when describing spectral variations in vegetation (Figure 1). This is because
seasonal patterns manifest differently across various spectral bands and vegetation indices,
depending on the phenological stages of vegetation [26]. Multivariate Functional Principal
Component Analysis (MFPCA) is well-suited for analysing multivariate functional spaces.
MFPCA decomposes the multivariate functional space into a set of orthogonal multivariate
functional principal components or modes of variation of functions (multivariate eigenfunc-
tions), together with corresponding functional principal component scores (FPC scores).
These FPC scores summarize the similarities between observations (functions), providing a
compact representation of the data (one score value per multivariate principal component
and per observation). In addition, these scores are uncorrelated by construction [38]. They
can then serve as a building block for further statistical analyses such as unsupervised
clustering, supervised classification methods or functional principal component regression
with multiple covariates [39].
Remote Sens. 2024, 16, x FOR PEER REVIEW 3 of 27

analyses such as unsupervised clustering, supervised classification methods or functional

Remote Sens. 2024, 16, 1224 3 of 26
principal component regression with multiple covariates [39].

Figure 1. Spectral variations in remotely sensed images over time. (a) Finite discrete time series:
this panel shows a typical representation of remotely sensed data captured at discrete points in time
Figuredata).
(raw Each point
1. Spectral on the in
variations graph represents
remotely dataimages
sensed from a specific moment.
over time. (b,c) discrete
(a) Finite Spectral variations
time series: this
in pixels as functions of time (smoothed representation of variations).
panel shows a typical representation of remotely sensed data captured at discrete These two panels show howin time
points
individual
(raw pixelpoint
data). Each spectral
on characteristics evolve over
the graph represents datatime,
fromsimplifying trend observation.
a specific moment. In detail
(b,c) Spectral variations
(b) defines a univariate functional space that describe the spectral variations in pixels
in pixels as functions of time (smoothed representation of variations). These two panels show how characterized
by a singlepixel
individual bandspectral
or index,characteristics
such as NDVI.evolve
This helps
overustime,
to understand
simplifyinghowtrend
one specific aspect In
observation. of detail
vegetation changes over time while (c) shows spectral variations in pixels characterized by
(b) defines a univariate functional space that describe the spectral variations in pixels characterized multiple
bybands or indices,
a single band orsuch as NDVI,
index, such GNDVI
as NDVI. and NDWI,
This helpsdefining multivariate how
us to understand functional space (this
one specific aspect of
allows us to
vegetation study how
changes overdifferent aspects
time while of vegetation
(c) shows spectralchange together
variations in over time.
pixels characterized by multiple
bands or indices, such as NDVI, GNDVI and NDWI, defining multivariate functional space (this
In this study, we develop new hybrid models that combine machine learning with
allows us to study how different aspects of vegetation change together over time.
MFPCA. MFPCA, the best of our knowledge, has not been previously used for supervised
classification of habitats and vegetation. We believe that these models are valuable for
In this study, we develop new hybrid models that combine machine learning with
analysing multivariate satellite dense time series, simultaneously considering seasonal
MFPCA.
spectral MFPCA,
variationsthe best
from of our knowledge,
different has not been
bands or vegetation previously
indices, used for supervised
and for evaluating new
classification of habitats and vegetation. We believe that these models
vegetation indices through combinatorial calculations using different formulas are valuable
to identifyfor an-
alysing multivariate
distinctive features forsatellite dense To
classification. time series,
further simultaneously
improve considering
classification performanceseasonal
and cre- spec-
tral
atevariations
interpretablefrom different
models, bandsaor
we include vegetation
selection indices,
strategy andonly
to retain for relevant
evaluating new
index vegeta-
time
series and exclude unnecessary ones. Our study was conducted in two Natura
tion indices through combinatorial calculations using different formulas to identify dis- 2000 sites in
central features
tinctive Italy, characterized by different
for classification. Toenvironmental
further improveconditions and vegetation
classification types. We
performance and cre-
configured three distinct hybrid models by varying input data types and feature
ate interpretable models, we include a selection strategy to retain only relevant index time selection
strategies and compared the results.
series and exclude unnecessary ones. Our study was conducted in two Natura 2000 sites
The objectives of this study aim to address the following questions:
in central Italy, characterized by different environmental conditions and vegetation types.
1. Do supervised hybrid classification approaches based on FDA produce a higher accu-
We configured three distinct hybrid models by varying input data types and feature se-
racy compared to machine learning methods directly applied to raw multi-temporal
lection strategies and compared the results.
data in both test sites?
2. The objectives
Among of this study
the examined hybridaim to address
approaches, the following
is there questions:achieves the
one that consistently
1. Do supervised
highest accuracyhybrid
in bothclassification
test sites? approaches based on FDA produce a higher ac-
3. curacy
Amongcompared
the explored
to machine learningone
formulas, is there that consistently
methods produces
directly applied tothe
rawhighest
multi-tem-
accuracy
poral dataininboth
bothtest sites?
test sites?
2.4. Among
Can an the
appropriate
examinedset hybrid
of indices be identifiedisfor
approaches, eachone
there study site?
that consistently achieves the
This work
highest is structured
accuracy as follows:
in both in Section 2 we introduce the materials and methods,
test sites?
focusing on the study area and the
3. Among the explored formulas, is there‘hybrid statistical–functional–machine learning’ models
one that consistently produces the highest
to analyse and classify dense remotely sensed time series. In Section 3 we present the results
accuracy in both test sites?
of our methodology applied to two different case studies. In Section 4 we discuss the results
4. Can an appropriate set of indices be identified for each study site?
 This work is structured as follows: in Section 2 we introduce the materials and meth-
ods, focusing on the study area and the ‘hybrid statistical–functional–machine learning’
Remote Sens. 2024, 16, 1224 models to analyse and classify dense remotely sensed time series. In Section 3 we present 4 of 26
the results of our methodology applied to two different case studies. In Section 4 we dis-
cuss the results and the impact of the developed approach, and in Section 5 we provide
and the impact
conclusions andofoutline
the developed approach, and in Section 5 we provide conclusions and
future work.
outline future work.
2. Materials and Methods
2. Materials and Methods
In this section we present two distinct approaches for classifying remotely sensed
data In
(seethis section
Figure we begin
2). We present
bytwo distinct
collecting approaches
Sentinel-2 for classifying
satellite time seriesremotely sensed
data, which can
data (see Figure 2). We begin by collecting Sentinel-2 satellite time series
be directly classified using Random Forest (first approach: ‘Pure Machine Learning’). data, which
Al-
can be directly classified using Random Forest (first approach: ‘Pure Machine
ternatively, spectral bands and indices created through combinatorial methods were Learning’).
Alternatively,
transformed into spectral bands functions
continuous and indices created
using through Additive
Generalized combinatorial methods
Models (GAM)were and
transformed into continuous functions using Generalized Additive
analysed with FDA (including FPCA and MFPCA). Random Forest can then be used Models (GAM) andto
analysed with FDA (including FPCA and MFPCA). Random Forest can then be used to
classify the FPCA-MFPCA scores (second approach: ‘Hybrid statistical-functional-Ma-
classify the FPCA-MFPCA scores (second approach: ‘Hybrid statistical-functional-Machine
chine Learning’). Further details are provided in the following sub-section. The developed
Learning’). Further details are provided in the following sub-section. The developed R code
R code is available in [40].
is available in [40].

Figure 2. Starting from a set of Sentinel-2 images, we trigger a processing pipeline that extracts the
Figure 2. Starting from a set of Sentinel-2 images, we trigger a processing pipeline that extracts the
most relevant
most relevant vegetation
vegetation indices
indices that
that could
could be
be used
used to
to characterize
characterize the
the study
study area.
area.
2.1. Study Area
2.1. Study Area
This study focuses on two distinct areas of central Italy, specifically in the Marche
region,This studyare
which focuses
part ofonthetwo distinct
Natura 2000areas of central
network Italy,
(Figure specifically
3). The first area inofthe Marche
interest is
region, which are
Mount Conero, part ofinthe
situated theNatura
coastal2000
area network
of central(Figure
Marche3).(43The
◦ 33first area
′ 00′′ N, 13of
◦ 36interest
′ 00′′ E). is
It
Mount Conero, situated in the coastal area of central Marche (43°33′00″N,
is a Special Area of Conservation (SAC) known as ‘Monte Conero’ (code IT5320007) and 13°36′00″E). It
is a Special Area of Conservation (SAC) known as ‘Monte Conero’ (code
covers an area of 650 hectares. Mount Conero has an elevation of 572 m above sea level, withIT5320007) and
covers an area
an average of 650
annual hectares. Mount
precipitation Conero
of 710 mm and has an elevation
a mean of 572 m above
annual temperature of 14.9 sea level,
◦ C. The
with an average annual precipitation of 710 mm and a mean annual temperature
second study area is the ‘Gola di Frasassi’ (code IT5320003), also referred to as the Frasassi of 14.9
°C. The located
Gorge, second study
in thearea is the ‘Golaregion
mountainous di Frasassi’ (codeMarche’s
of central IT5320003),Apennines (43◦ 23
also referred to′ as
23′′the
N,
12 ◦ ′
Frasassi ′′
57 36 Gorge,
E). This located
SAC spans in anthearea
mountainous
of 728 hectaresregion of central
and reaches Marche’s
an altitude Apennines
of 935 m above
sea level. The average annual precipitation in this area is 1115 mm, while the mean annual
temperature is 12.7 ◦ C. According to the bioclimatic classification of Rivas-Martinez [41],
both study areas belong to the temperate sub-Mediterranean macrobioclimate. The first
area is characterised by a strong sub-Mediterranean level with pronounced summer aridity,
 (43°23′23″N, 12°57′36″E). This SAC spans an area of 728 hectares and reaches an altitude
of 935 m above sea level. The average annual precipitation in this area is 1115 mm, while
the mean annual temperature is 12.7 °C. According to the bioclimatic classification of Ri-
vas-Martinez [41], both study areas belong to the temperate sub-Mediterranean macrobi-
Remote Sens. 2024, 16, 1224 5 of 26
oclimate. The first area is characterised by a strong sub-Mediterranean level with pro-
nounced summer aridity, while the second area is characterised by a weak sub-Mediter-
ranean level
while the indicating
second lower summer
area is characterised by aaridity [42].
weak sub-Mediterranean level indicating lower
summer aridity [42].

Figure3.3.The
Figure Thetwo
twostudy
studyareas:
areas: (a)
(a)national
national and
and (b)
(b) regional
regionaloverview
overviewofofthethetwotwostudy areas;
study areas; S1 is
S1 is the Frasassi Gorge, and S2 is Mount Conero. (c) Panoramic image of the Frasassi
the Frasassi Gorge, and S2 is Mount Conero. (c) Panoramic image of the Frasassi Gorge area. (d) Gorge area.
(d) Panoramic
Panoramic imageimage of the
of the Mount
Mount Coneroarea.
Conero area.(e)
(e)Reference
Reference data
dataon
onthe
theDigital Elevation
Digital Elevation Model
Model with
with the boundary of the Frasassi Gorge Special Area of Conservation (SAC IT5320003).
the boundary of the Frasassi Gorge Special Area of Conservation (SAC IT5320003). (f) Reference (f) Reference
dataon
data onthe
theDigital
Digital Elevation
ElevationModel
Model with thethe
with boundary
boundaryof theofMount ConeroConero
the Mount area of interest.
area of interest.

2.2. Target Classes and Reference Data

Different vegetation types (recognised using the Braun-Blanquet approach) and the
corresponding 92/43/EEC habitats are present in the two study areas. In the Mount Conero
area, there are four different forest plant communities while the Frasassi Gorge area en-
compasses eight different vegetation typologies (four forests, two shrubs, one grassland
Remote Sens. 2024, 16, 1224 6 of 26

2.2. Target Classes and Reference Data

Different vegetation types (recognised using the Braun-Blanquet approach) and the
corresponding 92/43/EEC habitats are present in the two study areas. In the Mount
Conero area, there are four different forest plant communities while the Frasassi Gorge area
encompasses eight different vegetation typologies (four forests, two shrubs, one grassland
and a mosaic of garrigue and chasmophitic vegetation). Detailed descriptions are provided
in Table 1 and [29,30].

Table 1. Reference data for the study areas. Target classes for the supervised classification are listed.
For plant associations, we report the syntaxa name and the corresponding habitat code (Annex 1 of
the European Union Habitats Directive). The * denotes a priority habitat.

Class Plant Association (Syntaxa) Habitat Code Plots

Mount Conero area 172
Woods
Quercus ilex evergreen forest with a high occurrence of Mediterranean species
c1 9340 34
Cyclamino hederifolii-Quercetum ilicis [43].
Quercus ilex with deciduous trees mixed forest Cephalanthero longifoliae-Quercetum ilicis subass.
c2 9340 71
ruscetosum hypoglossy [43].
c3 Ostrya carpinifolia coastal deciduous forest Asparago acutifolii–Ostryetum carpinifoliae [44,45]. - 13
c4 Evergreen conifer forest plantations mostly dominated by Pinus halepensis and P. pinea [46]. - 54
Frasassi Gorge area 241
Woods
Quercus ilex (with deciduous trees) appenninic forest Cephalanthero longifoliae-Quercetum ilicis
v1 9340 34
subass. lathyretosum veneti [43].
v2 Quercus pubescens deciduous forest—Cytiso sessilifolii-Quercetum pubescentis [47,48]. 91AA * 28
v3 Ostrya carpinifolia deciduous appenninic forest—Scutellario columnae-Ostryetum carpinifoliae [49]. - 56
v4 Evergreen conifer forest plantations mostly dominated Pinus nigra ssp. nigra and P. halepensis Mill. [50]. - 31
Shrublands
Spartium junceum Shrub—Spartio juncei-Cytisetum sessilifolii Spartium junceum variant
v5 - 16
(Edoardo Biondi & Casavecchia, 2002).
v6 Junyperus oxycedrus shrub—Spartio juncei-Cytisetum sessilifolii Juniperus oxycedrus variant [51]. - 15
Grasslands
v7 Bromus erectus grassland—Asperulo purpureae-Brometum erecti [52]. 6210 * 16
Mosaic of garrigues and vegetation of rock and scree
Satureja montana Garrigues Cephalario leucanthae-Saturejetum montanae
(could include 6110 and 6220 habitats);
v8 Potentilla caulescens and Moehringia papulosa chasmophytic vegetation of shady and 6110, 6220, 8210 46
wet rocky gorge’s wall—Moehringio papulosae-Potentilletum caulescentis
(habitat 8210 “Calcareous rocky slopes with chasmophytic vegetation”) [52,53].

The collected reference data, distributed over the two study areas are presented in Figure 3.

2.3. Remote Sensing Data Collection and Generation of Vegetation Indices

Sentinel-2 L2A images were acquired using the Sen2r package version 1.6.0 [54]. A
total of 93 scenes (spanning from April 2017 to April 2020, as shown in Table A1) were
collected for the two study areas, ensuring a cloud cover below 25% within the training plots.
The images were pre-processed by masking the clouds and performing co-registration. A
spatial resolution of 10 m was used, with the bands at 20 m being resampled using the
nearest neighbours approach. Starting from the review of existing indices as in [55], we
tried to summarize basic formulas, but we also considered other mapping functions. We
considered up to 4 operands with basic rules to have a spectral order. The rules have been
introduced to ensure a link with well-known indices such as the NDVI (type #3 in Table 2).
The list of formulas is not related to a specific sensor/payload, and it could be applied to
data acquired using aerial and satellite platforms. We considered Sentinel-2 bands, but the
proposed approach can be applied to different types of platforms (e.g., Landsat-8).
Remote Sens. 2024, 16, 1224 7 of 26

Table 2. List of formulas for different types of indices. We analyse formulas with 2–4 operands and
constraints on band order. We considered the following Sentinel-2 bands: B2, B3, B4, B5, B6, B7, B8*,
B11, B12; * corresponds to B8–NIR (832.8 nm). More info of Sentinel-2 bands could be found here [56].

Formula #id Formula # of Operands Constraint #1 Constraint #2 # of Combinations

0 A 1 - - 9
1 A−B 2 A > B - 36
2 A/B 2 A > B - 36
3 ( A − B)/( A + B) 2 A > B - 36
4 ( A − B)/C 3 A > B C>B 84
5 ( A − B)/(C + B) 3 A > B C>B 84
6 ( A − B)/(C − B) 3 A > B C>B 84
7 ( A − B)/( A + B) 3 A > B A>C 84
( A + C )/( A − C )
8 (( A − B)/( A + B))(( D − C ) /( D + C )) 4 A>B D>C 126
9 A/B(C − D )/(C + D ) 4 A>B C>D 126
10 A/B( A − C )/( A + C ) 3 A>C - 84
11 A/B( B − C )/( B + C ) 3 B>C - 84
12 A/B·C/D 4 - - 126
13 ( A − B)/( A + B + C + 1e4) 3 A>B - 84
14 (( A − C ) − ( B − D ))/(( A − C ) + ( B − D )) 4 A>C B>D 126
15 ( A − B)/( A + B + C ) 3 A>B B>C 84
16 ( A − B)/(( A + B − C ) + 1e4) 3 A>B B>C 84
17 (2A − B − C )/(2A + B + C ) 3 A>B B>C 84
18 ( A − ( B + C ))/( A + ( B + C )) 3 A>B A>C 84
19 log( A/B) 2 - - 36
20 ( A − B)·C 3 A>B - 84

2.4. Time Series as Functional Data

We arranged the 93 Sentinel-2 images chronologically by Day of the Year (DoY), Refs. [57–59]
addressing outliers using the clean.ts() function from the R package forecast version
8.12 [60,61]. DoY values were aggregated into weekly averages (1–52 weeks) (e.g., Figure 4a).
We interpolated and smoothed the weekly values using a GAM model with cyclic penalized
cubic regression spline smooth (with default settings) [62]. GAMs have the advantage that
they do not require measurements (like those of spectral bands) to be uniformly distributed,
which is useful since clouds and other data issues cause random gaps in the data [63].
This process generated a weekly functional cubic cyclic spline representation of spectral
variations in the plots (e.g., Figure 4b), and we applied it to all index formulas listed in
Table 2. As mentioned in [36], the original discrete data were then set aside and the
Remote Sens. 2024, 16, x FOR PEER REVIEW 8 ofesti-
27
mated curves (Figure 4b) were used for the rest of the analysis. The example R code for
time series smoothing is available in [40] (repository ‘habitatmapmfpca’).

Example
Figure4.4.Example
Figure of derived
of derived timetime series
series considering
considering meanmean
weekly weekly
annualannual Sentinel-2
Sentinel-2 GNDVI GNDVI
vari-
variations
ations (2017–2020)
(2017–2020) of theof theplots
172 172 of
plots
theof the Mount
Mount ConeroConero
study study
area. On area.
theOnleftthe
(a) left
the (a) the discrete
discrete mean
mean weekly
weekly time series,
time series, while
while on theon the (b)
right rightthe(b) the weekly
weekly functional
functional cycliccyclic
cubiccubic
splinespline representation
representation of
of the
the spectral
spectral plotplot variations.
variations. TheThe letters
letters at the
at the toptop correspond
correspond to to
thethe initials
initials ofof
thethemonths
monthsofofthe
theyear.
year.

2.5. Analysis of Functional Data Using FPCA and MFPCA

FPCA is a widely used FDA technique to reduce the amount of functional data
[36,37]. It adapts traditional PCA concepts to functions, while preserving the functional
structure (i.e., chronological order) of the observations (curves) [64]. FPCA extracts prin-
Remote Sens. 2024, 16, 1224 8 of 26

2.5. Analysis of Functional Data Using FPCA and MFPCA

FPCA is a widely used FDA technique to reduce the amount of functional data [36,37].
It adapts traditional PCA concepts to functions, while preserving the functional structure
(i.e., chronological order) of the observations (curves) [64]. FPCA extracts principal compo-
nents (eigenfunctions representing the main modes of data variation) from the estimated
curves, providing eigenvalues to quantify the captured variation and FPC scores to quantify
curve similarities [32]. It is suitable for exploring and decomposing univariate functional
spaces defined by a single variable. MFPCA extends FPCA to multivariate functional data,
such as multiple bands or vegetation indices (Figure 1). It captures joint variations between
functions, decomposing the data into orthogonal multivariate functional principal compo-
nents (multivariate eigenfunctions) with eigenvalues and component scores. This provides
a parsimonious data representation, with one score value per multivariate principal com-
ponent per observation. The MFPCA scores, uncorrelated by construction, could be used
for further statistical analyses (e.g., unsupervised functional clustering, supervised func-
tional classification) [38] and graphical representation of the results for interpretation [32].
Univariate FPCA used the fdaPace R package version 0.5.5 [65] while MFPCA used the
approach from [38] implemented in the associated R package version 1.3.6 [66].

2.6. Random Forest Classifier

Random Forest (RF) is a powerful ensemble learning classifier commonly used in
habitat mapping studies based on remote sensing data [67]. We optimized RF performance
by adjusting two key parameters: ntree (set to 1500) and mtry (evaluated from 1 to the
square root of input variables) [68]. Imbalanced training and validation data can bias RF
models in vegetation-related studies, over-predicting majority classes and under-predicting
minority classes. To address this, we employed down-sampling in RF to balance class
frequencies [29,69]. Additionally, we applied Recursive Feature Elimination to select
important predictors and reduce input data dimensionality, enhancing model efficiency.
These settings were maintained for all different supervised classification approaches (see
following section).

2.7. Supervised Classification Approaches

We conducted supervised vegetation classification using Sentinel-2 temporal spectral
variations through two approaches: ‘Pure Machine Learning’ and ‘Hybrid Statistical-
Functional Machine Learning’ [19]. In the ‘Pure Machine Learning’ approach, we directly
applied the RF classifier to raw Sentinel-2 multi-temporal imagery. The ‘hybrid’ approach
integrated RF with FDA of dense time series, utilizing FPCA and MFPCA analyses for
supervised classification. Specifically, we designed three hybrid models, each generating
distinct input datasets for the classifier, consisting of separate FPCA and MFPCA scores.
Details of these models are provided in the following subsections.

2.7.1. Pure Machine Learning Approach

Applying RF (or other machine learning methods) directly to raw satellite multi-
temporal imagery data from discrete time series is a common method for vegetation and
habitat mapping. These time series, typically based on a limited number of cloud-free
scenes (e.g., <15%) selected within one year, can be constructed using individual spectral
bands or predefined vegetation indices chosen by the authors [6,14,15,17,18,70–72]. In our
study we used Sentinel-2 spectral bands discrete time series as input data for RF, avoiding
an uncritical pre-selection among various available vegetation indices. We selected cloud-
free images from 2019 according to the criteria discussed above, providing the broadest
temporal coverage across different months for our study areas. For the Frasassi Gorge
study area we selected 9 images (excluding January, May, November, and December due to
cloud cover), and for the Mount Conero Area we selected 12 images (excluding January,
September and December due to cloud cover) (see Table A1). This approach, considered as
a baseline model, is referred to as B.
Remote Sens. 2024, 16, 1224 9 of 26

2.7.2. Hybrid Statistical–Functional–Machine Learning Approach

The first hybrid model used is the one proposed in [29], and referred to as mF. It
involves analysing Multivariate Functional Spaces using multiple univariate FPCAs, one
for each weekly vegetation index time series. The input data for RF consists of all uni-
variate FPCA component scores. While mF models can be effective in terms of Overall
Accuracy, it is important to note that the dimensionality of the input data can increase
rapidly since univariate FPCA can extract about 6–7 components from each weekly veg-
etation index time series. The R code was developed in [29] and is available in [40]
(repository ‘habitatmapfrasassi’).
For the second hybrid model, we applied MFPCA to simultaneously analyse and com-
press all weekly vegetation index time series generated by specific formulas (e.g., 36 indices
for formula id #3—Table 2). We decided to extract a maximum of 36 multivariate func-
tional principal components, balancing computational efficiency with effective vegetation
characterization and classification. This decision was guided by the fact that, as previously
mentioned, univariate FPCA typically only extracts about 6–7 components [29]. The result-
ing MFPCA components (multivariate eigenfunctions) and their scores offer a concise data
representation [38]. The MFPCA scores for these 36 components served as input for the RF
model, and this approach is denoted as M.
The third strategy aims to enhance vegetation classification accuracy by selecting a
reduced set of time series indices specific to the study areas. This approach combines FPCA,
MFPCA and RF through forward selection. For each iteration, an index time series was
added and classified by RF (initially decomposed with univariate FPCA and subsequently
with MFPCA). This process continued until no additional time series improved the model,
with improvement assessed using the Overall Accuracy metric. As in the case of the M
models, we limited MFPCA to extract a maximum of 36 components. The MFPCA scores
from the selected index time series served as RF input data. This strategy is labelled Ms.
The R code is available in [40] (repository ‘habitatmapmfpca’).

2.8. Accuracy Evaluation and Models Comparison

We assessed model accuracy using Overall Accuracy (OA), Producer Accuracy (PA),
User Accuracy (UA) and the κ coefficient [73,74]. More details are reported in Table S1.
To ensure robust estimates and minimize bias, we conducted 10-fold cross-validation
five times, resulting in a cross-validated confusion matrix. RF models and accuracies
were evaluated using the R caret package version 6.0.86 [75]. To compare all models
simultaneously in terms of accuracy and complexity, we recorded OA, PA, the number of
selected predictors (pr) and the final mtry of the RF model as columns in a data matrix.
Each model (B, Ms, M, mF applied to each formula) was represented as a row in the matrix.
Subsequently, we conducted a standardized Principal Component Analysis (PCA) on the
data matrix.

3. Results
3.1. Models Performance and Comparison
The OA of the models is presented for both study areas, categorized into Pure Machine
Learning and Hybrid Machine Learning approaches. Within the Hybrid Machine Learning
category, the results are further detailed based on the different modelling strategies and
indices formula ids. See Table 3 and Figure 5 for a summary of the results.
In the Mount Conero area, the baseline B model achieved an OA of 81.8%. Among
the hybrid models, mF models exhibited an average OA of 84.3%, with the highest OA of
86% achieved using formula id #11 and the lowest at 81.6% with formula id #1. The M
models had an average OA of 78.6%, with the highest OA of 85.6% obtained with formula
id #18 and the lowest at 66.3% with formula id #8. The Ms models achieved an average OA
of 84.4%, with the highest OA of 87.2% linked to formula id #15 and the lowest at 77.9%
with formula id #4.
Remote Sens. 2024, 16, 1224 10 of 26

For the Frasassi Gorge area, the B model achieved an OA of 76.9%. Among the hybrid
models, the mF models showed an average OA of 80.9%, with the highest OA of 82.9%
achieved using formula id #3 and the lowest at 77.3% with formulas ids #0 and #1. The
M models had an average OA of 74.2%, with the highest OA of 82.3% using formula id
#7 and the lowest at 63.4% with formula id #17. Additionally, the Ms models obtained an
average OA of 83.1%, with the highest OA of 86.5% linked to formula id #15 and the lowest
at 81.1% with formula id #19.
Table 3. Comparison of model and formula performances in the two study areas based on Overall
Accuracy. B—baseline model (Pure Machine Learning approach). mF, M, Ms—RF models based on
Functional Data Analysis (Hybrid statistical—functional–Machine Learning approach). Formula id
represents the different formulas used to generate indices detailed in Table 2. CO—Mount Conero
area. VM—Frasassi Gorge area. In grey if the accuracy exceeds that of B. In bold, the best performance
for each distinct hybrid approach.

Mount Conero Frasassi Gorge

Formula #id B mF M Ms B mF M Ms
0 0.818 0.826 0.812 0.812 0.769 0.773 0.785 0.812
1 0.816 0.838 0.835 0.773 0.778 0.845
2 0.844 0.768 0.839 0.816 0.675 0.824
3 0.849 0.825 0.849 0.829 0.817 0.829
4 0.857 0.790 0.779 0.811 0.733 0.832
5 0.857 0.793 0.859 0.819 0.731 0.842
6 0.841 0.675 0.842 0.808 0.646 0.815
7 0.854 0.802 0.860 0.818 0.823 0.836
8 0.831 0.663 0.838 0.816 0.644 0.840
9 0.856 0.797 0.848 0.816 0.754 0.840
10 0.835 0.778 0.840 0.792 0.667 0.811
11 0.860 0.790 0.860 0.825 0.708 0.828
12 0.842 0.732 0.851 0.825 0.668 0.814
13 0.828 0.826 0.844 0.784 0.764 0.832
14 0.844 0.819 0.838 0.802 0.778 0.840
15 0.847 0.838 0.872 0.813 0.810 0.865
16 0.832 0.814 0.843 0.783 0.787 0.828
17 0.845 0.671 0.847 0.798 0.634 0.856
18 0.845 0.856 0.857 0.806 0.798 0.835
19 0.850 0.829 0.850 0.805 0.813 0.811
20 0.852 0.794 0.851 0.820 0.764 0.822
mean 0.818 0.843 0.786 0.844 0.8 0.806 0.742 0.831

In both study areas, the Ms and mF models consistently outperformed the M and
B models, achieving a higher Overall Accuracy of 9.6 percentage points in the Frasassi
Gorge area, and 5.4 percentage points in the Mount Conero area (see Figure 5 and Table 3).
Furthermore, using indices (formula ids #1–#20 in Table 3) in the Ms and mF models
demonstrated superior performance compared to using individual bands (formula id #0 in
Table 3). In both study areas, the highest OA was achieved by the Ms models applied to
vegetation indices with formula id #15 (see Tables 2 and 3 for its definition).
 area, and 5.4 percentage points in the Mount Conero area (see Figure 5 and Table 3). Fur-
thermore, using indices (formula ids #1–#20 in Table 3) in the Ms and mF models demon-
strated superior performance compared to using individual bands (formula id #0 in Table
Remote Sens. 2024, 16, 1224 3). In both study areas, the highest OA was achieved by the Ms models applied to vegeta-
11 of 26
tion indices with formula id #15 (see Tables 2 and 3 for its definition).

5. Comparison
Figure 5.
Figure ComparisonofofOverall Accuracy
Overall Accuracy(OA) among
(OA) different
among modelmodel
different strategies for the two
strategies study
for the two study
areas. The
areas. The dashed
dashedline
linerepresents thethe
represents OAOAachieved by the
achieved bybaseline B model
the baseline using ausing
B model Pure Machine
a Pure Machine
Learning approach.
Learning approach. M,M,mFmF andand
Ms are
Msthree
are hybrid model strategies
three hybrid combiningcombining
model strategies Random Forest with Forest
Random
Functional Data Analysis (Hybrid statistical–functional–Machine Learning approach).
with Functional Data Analysis (Hybrid statistical–functional–Machine Learning approach). (a) (a) Mount
Conero Conero
Mount area. (b)area.
Frasassi
(b)Gorge area.
Frasassi Gorge area.

Tables A2 and A3 offer a comprehensive overview of all models for both the Mount
Tables A2 and A3 offer a comprehensive overview of all models for both the Mount
Conero and Frasassi Gorge areas providing accuracy (OA and PA), and complexity metrics
Conero
(numberand Frasassipredictors,
of selected Gorge areas pr, providing
and the final accuracy
mtry of(OA and PCA
the RF). PA), ofand complexity
these tables met-
rics (number
(Figure of selected
6) allows for a visualpredictors, pr, and
representation the
that final mtry
facilitates of the
model RF). PCAbased
comparison of theseon tables
(Figure 6) allows(inter-
their multivariate for aandvisual representation
intra-group) that
variability. facilitates
Similar modelsmodel comparison
are close together, and based on
their multivariate
dissimilar models are(inter-
furtherand intra-group)
apart. variability.
The properties Similarare
of the models models are by
indicated close
blacktogether,
arrows.
and The B model
dissimilar models is represented
are further by a redThe
apart. triangle, while the
properties mF, models
of the M and Ms aremodels
indicated by
applied
black to different
arrows. The B formulas
model are represented in
is represented byspider
a redplots with while
triangle, distinctthe
colours.
mF, M Theand first
Ms mod-
principal component (PC1) axis, accounting for 49.5% and 43.8% of the
els applied to different formulas are represented in spider plots with distinct colours.total variation in the The
Mount Conero and Frasassi Gorge areas, respectively, indicates an increasing gradient of
first principal component (PC1) axis, accounting for 49.5% and 43.8% of the total variation
accuracy among the models. It clearly shows that the Ms and mF models outperform the B
in the
and MMount
models Conero
in both OA and (asFrasassi
shown inGorge
Table 3areas, respectively,
and Figure 5) and PA.indicates
The second anprincipal
increasing gra-
dient of accuracy
component among
(PC2) axis, the accounts
which models. forIt clearly
22.5% andshows thatofthe
17.0% theMs
totaland mF models
variation in the outper-
form the B and M models in both OA (as shown in Table 3 and
Mount Conero and Frasassi Gorge areas, respectively, is directly related to the increasingFigure 5) and PA. The
second
numberprincipal component
of predictors used as input(PC2) axis,
data (pr)which
and theaccounts
mtry value.for 22.5% and 17.0% of the total
PCA in
variation analysis revealsConero
the Mount that theandMs models
Frasassiare the most
Gorge parsimonious,
areas, respectively, achieving
is directlythe related
highest OA and PA accuracy while using the fewest predictors and
to the increasing number of predictors used as input data (pr) and the mtry value. mtry (Figure 6).
Tables S2 and S3 provide details from the forward selection procedure used by Ms
models. These tables outline the selected bands and indices that constitute the minimal
set needed to optimize model performance in each formula and study area. The number
of time series (bands or indices) selected ranged from 1 to 9 (1 to 7 for the Mount Conero
area and 2 to 9 for the Frasassi Gorge area). The most frequently involved bands in the
selected indices (in descending order) for the Frasassi Gorge area were B7, B5, B11, B4, B3,
B12, while band B8 was the least utilized. For the Mount Conero area, the most utilized
bands were B7, B6, B11, while bands B8 and B5 were less utilized.
 Remote Sens. 2024, 16, x FOR PEER REVIEW 12 of 27

Remote Sens. 2024, 16, 1224 PCA analysis reveals that the Ms models are the most parsimonious, achieving 12
theof 26
highest OA and PA accuracy while using the fewest predictors and mtry (Figure 6).

Figure6.6.Principal
Figure PrincipalComponent
Component biplot
biplot relating
relatingproperties
propertiesofofaccuracy
accuracy and model
and modelcomplexity
complexity(black
(black
arrows)
arrows) toto the
the different
different supervised
supervised classificationmodels
classification models(B,(B,mF,
mF,M,M, Ms)
Ms) applied
applied to to
allall distinct
distinct for-
formulas.
mulas. (a) Mount Conero Area. PCA axis 1 accounts for 49.5% of the multivariate variation and axis
(a) Mount Conero Area. PCA axis 1 accounts for 49.5% of the multivariate variation and axis 2 for
2 for 22.5%. (b) Frasassi Gorge Area. PCA axis 1 accounts for 43.8% of the multivariate variation and
22.5%.
axis 2 (b) Frasassi
for 17.0%. GorgeOA–Overall
Labels: Area. PCAAccuracy;
axis 1 accounts for 43.8%
sd–standard of the pr–number
deviation; multivariate ofvariation and axis
input variables
2 selected;
for 17.0%. Labels: Random
mtry–final OA–Overall
ForestAccuracy; sd–standard
mtry parameter; deviation;
v1–v8 and pr–number
c1–c4 are of input of
Producer Accuracy variables
veg-
selected; mtry–final
etation types (listed Random
in Table 1)Forest mtry Gorge
for Frasassi parameter; v1–v8Conero
and Mount and c1–c4
areas,are Producer Accuracy of
respectively.
vegetation types (listed in Table 1) for Frasassi Gorge and Mount Conero areas, respectively.
Tables S2 and S3 provide details from the forward selection procedure used by Ms
3.2. Best Models
models. These tables outline the selected bands and indices that constitute the minimal
set needed
The Ms to optimize
models model
applied performance
to formula id #15in(see
each formula
Tables and
2 and 3) study
achievedarea.
theThe number
highest OA in
of time
both studyseries (bands
areas. Below,orwe
indices) selected
summarise theranged
accuracy from 1 toof
results 9 these
(1 to 7models
for theand
Mount Conero
compare them
area and 2 to 9 for the Frasassi Gorge area). The most frequently involved bands
to the B models by showing the error matrices (Tables 4 and 5). In the Supplementary Materials, in the
selectedgraphical
detailed indices (inrepresentations
descending order) for two
of the the Frasassi Gorge
Ms models arearea were B7,
provided B5, B11,S1B4,
(Figures B3,S2),
and
B12, while band B8 was the least utilized. For the Mount Conero area,
illustrating the selected time series and functional decomposition via MFPCA with the the most utilized
bands
most were B7, B6, B11,
discriminating while bands
components B8 andvariation)
(seasonal B5 were less
forutilized.
the different vegetation types.

3.2. Best
Table Models
4. Cross-validated confusion matrix (10-fold, repeated five times) for predicted target classes
in the Mount
The MsConero
modelsarea. The table
applied includesid
to formula Overall Accuracy,
#15 (see Tables Producer Accuracy, the
2 and 3) achieved Userhighest
Accuracy
OA in both
(expressed study areas.and
in percentage) Below,
the κwe summarise
statistic. the accuracy
The rows results
and columns of these
(c1–c4) models
represent and
the plant
compare them
associations to the B listed
and habitats models in by showing
Table the errormodel
1. B—baseline matrices (Tables
(Pure Machine4 and 5). In the
Learning Sup-
approach).
plementary
Ms-F15 Materials,
(Ms model with thedetailed
Formulagraphical representations
id #15) is the top-performingofmodel
the two Ms models
in terms areAccuracy
of Overall pro-
videdthe
among (Figures S1 and
RF models S2), on
based illustrating
Functional the selected
Data time(Hybrid
Analysis series and functional decomposi-
statistical–functional–Machine
tion viaapproach).
Learning MFPCA with Pred the most
stands for discriminating
prediction. components (seasonal variation) for the
different vegetation types.
B Ms-Formula id #15
Table 4. Cross-validated confusion matrix (10-fold, repeated five times)Reference
Reference for predicted target classes
in the Mount Conero area. The table includes Overall Accuracy, Producer Accuracy, User Accuracy
c1 c2 c3 c4 UA c1 c2 c3 c4 UA
(expressed in percentage) and the 𝜅 statistic. The rows and columns (c1–c4) represent the plant as-
c1 16.2 3.2 0.0 2.1 75.5 c1 39.2 3.7 3.1 3.4
sociations and habitats listed in Table 1. B—baseline model (Pure Machine Learning approach). Ms- 79.4
Pred model 4.0
F15 (Ms c2 36.2
with the 3.9
Formula 3.0
id #15) 76.9
is the Pred
c2
top-performing 1.3
model 16.9
in terms 0.0
of Overall 0.7 89.7
Accuracy
c3 0.0 0.3 3.5 0.0 91.2 c3 0.0 0.0 4.3 0.0 100.0
c4 0.9 0.8 0.0 25.8 93.8 c4 0.1 0.6 0.0 26.7 97.5
PA 76.8 89.3 47.7 83.7 PA 96.6 80.0 58.5 86.7
OA 81.79 (±9.50) OA 87.18 (±7.82)
K 0.72 (±0.14) K 0.80 (±0.11)
Remote Sens. 2024, 16, 1224 13 of 26

Table 5. Cross-validated confusion matrix (10-fold, repeated five times) for predicted target classes
in the Frasassi Gorge area. The table includes Overall Accuracy, Producer Accuracy, User Accuracy
(expressed in percentage) and the κ statistic. The rows and columns (v1–v8) represent the plant
associations and habitats listed in Table 1. B—baseline model (Pure Machine Learning approach).
Ms-F15 (Ms model with the Formula id #15) is the top-performing model in terms of Overall Accuracy
among the RF models based on Functional Data Analysis (Hybrid statistical–functional–Machine
Learning approach). Pred stands for prediction.

B
reference
v1 v2 v3 v4 v5 v6 v7 v8
v1 11.7 0 1.32 0.74 0 0 0 0 84.9
v2 0 5.87 1.49 0 1.07 0 0.17 0 68.3
v3 0.58 4.96 18.6 0.41 1.16 0 0 0 72.3
v4 1.4 0 0.33 11.7 0 0.83 0 0 82.0
pred
v5 0.17 0.74 0.17 0 2.07 0 0.25 0.41 54.3
v6 0 0 0 0 0.66 4.38 0 0.83 74.6
v7 0 0 0.41 0 0.33 0 5.37 0.25 84.4
v8 0.25 0 0.83 0 1.32 0.99 0.83 17.5 80.6
PA 82.9 50.7 80.4 91.0 31.3 70.7 81.3 92.2
OA 76.99 (±7.07)
K 0.72 (±0.08)
Ms-Formula id #15
reference
v1 v2 v3 v4 v5 v6 v7 v8 UA
v1 13.4 0.0 0.6 0.1 0.0 0.0 0.0 0.0 95.3
v2 0.0 6.8 0.9 0.3 0.6 0.0 0.0 0.0 78.8
v3 0.4 4.3 21.3 0.4 0.2 0.4 0.0 0.0 78.9
v4 0.2 0.0 0.3 12.0 0.0 0.0 0.0 0.0 95.4
pred
v5 0.0 0.2 0.0 0.0 3.8 0.0 0.4 0.0 85.2
v6 0.0 0.0 0.0 0.0 0.2 4.8 0.0 0.0 95.1
v7 0.0 0.0 0.0 0.0 0.1 0.4 5.9 0.4 86.6
v8 0.0 0.2 0.0 0.0 1.7 0.6 0.3 18.6 86.5
PA 95.3 58.6 92.1 93.5 57.5 77.3 88.8 97.8
OA 86.51 (±6.99)
K 0.83 (±0.08)

3.2.1. Mount Conero Area

The Ms model (applied to time series indices obtained with formula id #15) selected
six time series for the Mount Conero area (A, B, C operators of the formula id #15 index),
which were: (B12, B11, B03); (B07, B06, B04); (B11, B08, B07); (B08, B05, B04); (B07, B06, B03);
(B12, B08, B06) (Table S2). Their seasonal variations and functional decomposition are
depicted in Figure S1. With an OA of 87.18%, this model outperformed model B, which
achieved 81.7%, and demonstrated a higher PA for the target classes c1, c3 and c4, as well
as better UAs in all classes (Table 4).

3.2.2. Frasassi Gorge Area

The Ms model (applied to time series indices obtained with Formula id #15) selected
nine time series for the Frasassi Gorge area (A, B, C operators of the Formula id #15 index),
which were: (B10, B07, B04); (B08, B03, B02); (B07, B04, B02); (B07, B03, B02); (B10, B05, B04);
(B11, B10, B04); (B06, B05, B04); (B08, B07, B04); (B07, B04, B03) (Table S3). Their seasonal
variations and functional decomposition are depicted in Figure S2. With an OA of 86.5%,
this model outperformed the 76.9% achieved by the B model. Furthermore, all PAs and
UAs were higher for the Ms model compared to the B model (Table 5).
Remote Sens. 2024, 16, 1224 14 of 26

4. Discussion
4.1. Main Results
This study highlights the effectiveness of the ‘Hybrid statistical–functional–Machine
Learning’ approach, which combines RF with an FDA of dense multispectral time series.
The approach outperforms conventional methods that directly use RF on raw satellite
multi-temporal images. Dense time series, when properly analysed and compressed, offer
crucial information for characterizing seasonal spectral changes in vegetation, improving
classification accuracy [26,27]. Ms models, which were the most accurate in both study
areas, could be suitable tools with important practical implications for accurate classifi-
cation, mapping and monitoring of vegetation and habitats included in Annex I of the
92/43/EEC Directive. Indeed, these models not only effectively process dense time series
(increasingly accessible through web platforms like Google Earth Engine [76,77]) with
FDA, but also independently identify sets of indices specific to the study area (through
the forward selection strategy). The selection of location-specific indices plays a key role
in optimizing the land management [24,25]. Thus, these models are adept at capturing
vegetation and habitats during their optimal phenological stages without requiring prior
knowledge of the best times for data acquisition or the most appropriate index sets, thus
making them more transferable than conventional models [21]. In addition, the results
of these models are graphically interpretable, contributing to a better understanding of
critical seasonal multispectral variations among different plant communities and habitats
(Figures S1 and S2).
Furthermore, the Ms models allowed us to employ new vegetation indices derived
from a combinatorial approach and evaluate their effect on classification accuracy. The
results revealed two aspects of particular interest. In both study areas, the most accurate
models were the Ms models based on the formula id #15, an original index. In addition,
rarely used indices based only on visible spectral bands played a significant role, confirming
that classifications based only on known indices such as NDVI may not always be the most
effective choice for classification purposes [20,78] or for characterizing plant communities.
These results agree that specific plant communities and vegetation types have their own,
specific multispectral profiles [24,26,79].

4.2. Models Comparison

4.2.1. Pure Machine Learning Approach: B Models
The B models demonstrated a lower accuracy, with a difference of up to 9.6% compared to
the Hybrid statistical–functional–Machine Learning approach (see Table 3 and Figures 5 and 6).
This lower performance was expected for several reasons. Model B typically employs input
data based on time series of images selected for their cloud-free and low-cloud-cover
conditions in a single reference year, reducing the data processing complexity, e.g., [14].
However, this approach often results in a limited number of images being available, with
missing data for specific months. In our case, nine images were available for the Frasassi
Gorge area and twelve for the Mount Conero area, covering different months depending
on local weather conditions (e.g., excluding January, May, November and December for
the Frasassi Gorge area and January, September and December for Mount Conero area due
to cloud cover). This data gap may negatively impact the description of plant phenology
and thus the accuracy of vegetation classification [80]. These models can be defined as
“image-dependent” [81] since the timing and quality of image acquisition significantly
impact classification accuracy [24]. Another crucial aspect to consider is that B models often
skip important pre-processing steps aimed at noise detection, removal and reduction in
time series, despite recommendations from [78,82], with a negative impact on accuracy.

4.2.2. Hybrid Statistical–Functional–Machine Learning Approach

Hybrid models that combine RF with FDA, overcoming the limitations of the B model,
demonstrate a higher accuracy. The FDA approach treats temporal spectral variations
as curves (smoothed functions) (e.g., Figures 1 and 4), allowing dense time series to be
Remote Sens. 2024, 16, 1224 15 of 26

analysed and offering richer information within a specific time window [32] than the
B models for the classification stage. Unlike B models, hybrid models can be called
‘image-independent’ [81]. In these models, it is the quality of the functional data, which
must adequately represent seasonal spectral variations in vegetation (e.g., Figure 4), that
significantly influences the accuracy of the classification, rather than the timing and quality
of the individual images used to create it. During the transformation of the raw data into a
functional data using the GAM approach, it is essential to perform pre-processing steps to
identify and remove outliers and reduce noise [83]. Another advantage over B models is
that, to create pixel-based functions, it is better to exploit as much information as possible
for each pixel. Thus, even images with only small areas without clouds or even one pixel
without clouds can be used. In other words, if a part of an image is covered by clouds, this
does not prevent the use of the part without clouds, whereas this is usually not the case for
B models. We can assert that, if using dense time series data is an ideal choice for analysing
seasonal variations in vegetation and achieving more accurate classifications [26,27,58],
then FDA serves as an ideal tool for compressing and analysing dense time series data.
Ms, mF and M models have different characteristics and levels of accuracy. The Ms
models are consistently better than the others in terms of Overall Accuracy for both study
areas (Figure 5, Table 3). The superior performance is particularly evident, especially
when applied to indices generated with formula id #15, in a more complex study area,
such as the Frasassi Gorge, which has a higher number of target classes (Table 3). These
models also performed better compared to previous studies. In the Mount Conero area,
they achieved an 87.2% accuracy, exceeding the 83.2% accuracy in [30], which used only
NDVI seasonal variation data. In the Frasassi Gorge area, these models achieved an 86.5%
accuracy, exceeding the 82.1% accuracy in [29], obtained with mF models based on six time
series of preselected indices (see Table 3). It is important to note that the Ms models are
parsimonious. They achieved such a high accuracy with the smallest number of predictors
and mtry (Figure 6, Tables A2 and A3), and this means that they can select a tailored and
mutually complementary set of indices that best align with area-specific characteristics
by capturing crucial seasonal multispectral variations. The key to this capability lies
in the incorporation of two wrapper methods within Ms models, operating at distinct
levels. Forward selection works on the entire index time series, while Recursive Feature
Elimination focuses on individual MFPCA components extracted from the progressively
selected time series. In summary, Ms models improve the characterization and distinction of
various plant communities and habitats, enabling more accurate and detailed classifications.
Their parsimonious nature makes them interpretable, contributing to a better understanding
of critical seasonal multispectral variation among different plant communities and habitats
(Figures S1 and S2). These hybrid models can complement species-based approaches
in plant community ecology [30,32,33,38,84]. Besides their strengths, Ms models have
some limitations. Indeed, forward selection does not guarantee the identification of the
best model since the final set of selected indices is highly dependent on the first index
chosen [85]. Moreover, they may require long computation times for evaluation, especially
when dealing with many time series, such as those generated by formula id #15 (126 time
series of indices). However, to improve the efficiency of these models and reduce the
number of models to be evaluated, a preliminary filtering method could be implemented in
future analyses. This method aims to identify and remove strongly correlated time series,
allowing Ms models to process a smaller and more focused set of candidate time series.
The mF models, in line with prior research [29], demonstrated their effectiveness by
achieving high accuracies. However, they also exhibited complexity and a lack of parsi-
mony due to the utilization of many predictors (see Figure 6, Tables A2 and A3). This
complexity arises from the limitation of multiple separate FPCAs in adequately addressing
joint variations among different time series, resulting in the extraction of numerous cor-
related and redundant components. This redundancy makes the interpretation of results
complicated [38]. Each vegetation plot has multiple scores associated with different univari-
ate FPCA analyses which cannot be synthesized into a single functional reduced-ordination
Remote Sens. 2024, 16, 1224 16 of 26

space [29]. Consequently, while effective, these models are not very efficient and do not
facilitate the understanding of crucial seasonal multispectral variation among different
plant communities and habitats.
Finally, among the hybrid models, the M models proved to be less accurate. Their
accuracies were modest and highly variable, consistently lower than the mF and Ms models,
and often inferior to the B models as well (Table 3, Figures 5 and 6). The M models compress
all the time series of vegetation indices associated with a specific formula using a single
MFPCA, and the corresponding scores serve as input data for RF. It is likely that the
established number of components extracted (k = 36) proved inadequate and too low,
probably discarding useful seasonal variations for RF. To increase the accuracy of the model,
one solution would be to increase the number of MFPCA components. However, this
approach, as in mF models, hinders the identification of the minimum set of time series and
indices specific to the vegetation of the study area. This limitation prevents us from fully
capturing the crucial seasonal multispectral variations among different plant communities
and habitats. In contrast, this method is suitable when the time series and indices specific
to the study area are few and known.

4.3. Formula Comparison

Ms models performed best in both study areas using formula id #15. Surprisingly, this
formula performed better than the well-known and widely used normalized difference
(NDVI, Formula id #3) and simple difference (DVI, Formula id #1) formulas (Table 3). To
our knowledge, formula id #15 is an original index that has not been found in the literature
or common databases. It can be considered an extension of the normalized difference index,
as it uses the difference between two bands in the numerator and the sum of the same
two bands plus a third one in the denominator.
The results presented in Tables S2 and S3 show the final indices selected from the Ms
models in the two study areas. In particular, the frequent use of Red Edge spectral bands
(B5 and B7), SWIR (B11, B12) and, especially in the Frasassi Gorge area, visible bands (green
and red, B3 and B4) is evident. These results are in line with previous studies [70,79,85–88],
which emphasized the importance of these bands for distinguishing and mapping tree
species, vegetation and habitats. The importance of visible bands is evident in the Frasassi
Gorge area, where, out of five indices selected through formula id #1, which achieved a
satisfactory Overall Accuracy, three are based exclusively on visible bands. This result is
significant for habitat mapping (Directive 92/43/EEC) because these indices, which are
often overlooked, can improve the accuracy of classification and offer the advantage of an
intuitive understanding of their variations [89].
In this study, the NIR had a lower contribution to classification accuracy than the other
bands mentioned above, despite the fact that its important role in vegetation mapping is
well known and proven [7,90]. NIR plays a key role in satellites with a higher spatial but
lower spectral resolution than Sentinel-2, such as IKONOS-2 and WorldView-2 [91,92].

4.4. Limits and Future Works

The first step in FDA is to transform raw data into functional objects by fitting discrete
observations with curves that approximate the underlying continuous process. Achieving
a balance between data fit and avoiding overfitting or neglecting essential aspects of the
estimated smooth function is a common goal in the smoothing process [36]. Developing
appropriate curves to describe the seasonal dynamics of vegetation across spectral bands
or indices is crucial for accurate supervised vegetation classifications. Although promising
results have been obtained in this and previous studies [29,30] using pixel-based functions
interpolated with GAM (with default parameters: Knots = 10 and cross-validation for
penalty value selection), future research could investigate how parameter variations and
alternative smoothing methods [93,94] can improve classification accuracy. However,
understanding the data-generating process and experimentation are fundamental tools in
spline smoothing [28,36].
Remote Sens. 2024, 16, 1224 17 of 26

The Ms models demonstrated a superior performance in both study areas. However,

the error matrices (Tables 3 and 4) revealed challenges in discriminating between some cate-
gories such as hornbeam and oak forests (e.g., 91AA* habitat). Incorporating topographical
variables [29,95] and more extensive reference data could enhance model performance. The
amount of reference data in our study, although well-distributed (see Figure 3), is relatively
small and this may negatively affect the performance of classification [96] and the selection
of time series. The main challenge, in fact, in mapping plant communities and habitats
lies in the time required for field data collection [6]. The activities of “drone truthing”,
obtaining reference data through drones [97–99], offers a cost-effective way for biologists to
verify satellite-derived maps, overcoming the limitations associated with ground-truthing
for habitat mapping [100]. The acquired RGB images allow for the recognition of plant
species [101], improving the efficiency of vegetation and habitat identification, even in
complex environments, by recognizing indicator species of plant communities [102]. We
are currently extending our analysis to other areas in the Central Apennines of Italy, where
we have obtained extensive reference data through both ‘ground-truthing’ and ‘drone-
truthing’. Preliminary results confirm the effectiveness of the Ms models in selecting a
minimal number of appropriate indices for the accurate classification of 16 different vegeta-
tion categories, demonstrating a significant level of discrimination for oak and hornbeam
forests in this context.
In addition to statistical validation, the robustness of the model can be qualitatively
assessed through the map generated by applying the model to all pixels [18,103]. In this
study, we chose not to perform mapping. This is because, even if feasible, it would have
been laborious given the numerous models developed. Our intention was to create a
standardized and easily adaptable methodology that could select the most suitable indices
for the study area.
Future developments will focus on evaluating other machine learning algorithms
besides RF. One intriguing option could be the use of Linear Discriminant Analysis (LDA),
which is also applicable in the functional context [104,105]. In the context of habitat and veg-
etation mapping, the adoption of a Hybrid statistical–functional model with LDA should
ensure good classification results and at the same time identify the seasonal discriminant
function that indicates the times when maximum differences between vegetation types
emerge. This approach would improve the interpretability of the results from an ecological
point of view, a crucial aspect for territorial entities engaged in habitat management and
conservation, as required by the Habitats Directive.

5. Conclusions
In this paper we studied different approaches to supporting the classification of
vegetation. These models combine machine learning, using RF, with the application of FDA
to dense satellite time series. Our main goal was to improve the accuracy of vegetation
and habitat classification in two different study areas. We achieved this by comparing
the performance of these models to that of the most common classification methods,
which apply machine learning directly to raw multi-temporal satellite data. Furthermore,
we analysed the effect of different formulas for calculating vegetation indices, using a
combinatorial approach. The goal was to identify the best approach and formula that
consistently generated the best classification accuracies in both study areas. Now, analysing
the results based on the research questions formulated at the beginning of this work, we
derive the following conclusions:
1. The Hybrid supervised classification approaches based on FDA produce higher accu-
racy than common machine learning methods applied directly to raw multi-temporal
data in both test areas.
2. Among the hybrid approaches examined, the Ms models achieve the highest accuracy
in both test sites. These models effectively combine FDA, by exploiting MFPCA that
compresses multiple time series based on different vegetation indices, with the use
of RF. Using a forward selection strategy, we identified a limited set of indices that
Remote Sens. 2024, 16, 1224 18 of 26

meaningfully represent crucial multispectral seasonal variations obtaining really good

results. Ms models are remarkably efficient, producing high accuracies with a low
number of input data.
3. Among the formulas explored for calculating vegetation indices, the formula id #15
proved to be the best performing one in both study areas. However, other formulas
have achieved good results (e.g., formula ids #17, #1), suggesting that further studies
could be conducted in different study areas and with more reference data. In general,
the use of indices rather than individual bands achieves better results.
4. This study demonstrated that Ms models can effectively identify a specific set of
indices for each study area, adapting to the ecological characteristics and vegetation
of the respective areas.
In conclusion, in scenarios characterized by an increasing availability of satellite data
(and then dense time-series), we believe that Ms models could play a role of significant
practical relevance in habitat monitoring and mapping. These models can identify the most
suitable indices, based on the specific characteristics of the study site and the ecological and
vegetation peculiarities of the analysed area, with the aim of maximizing the accuracy of
the classifications. Furthermore, the results obtained can be integrated with the field data
based on species recognition (for example, the Braun-Blanquet method), thus contributing
to the understanding and conservation of biodiversity in the study areas. These models
represent a promising contribution to overcoming the obstacle of transferability in remote
sensing for the conservation of Natura 2000 habitats [21]. The R code for these models is
available in [40] (repository habitatmapmfpca).

Supplementary Materials: The following supporting information can be downloaded at: https://www.
mdpi.com/article/10.3390/rs16071224/s1.
Author Contributions: Conceptualization S.P., A.M., G.Q. and S.C.; Data curation S.P. and A.M.;
Formal analysis, S.P. and A.M.; Investigation, S.P., G.Q. and S.C.; Methodology, S.P., A.M. and G.Q.;
Software, S.P. and A.M.; Supervision, S.P. and S.C.; Writing—original draft, S.P., A.M., G.Q. and
S.C.; Writing—review and editing, S.P., A.M., G.Q. and S.C. All authors have read and agreed to the
published version of the manuscript.
Funding: This research received no external funding.
Data Availability Statement: The R code for these models is available at https://github.com/
geobotany/habitatmapmfpca (accessed on 27 March 2024).
Acknowledgments: The authors want to thank the Lorenzo Deplano, Riccardo Forconi and Cristian
Colavito at the Department of Information Engineering (DII) of Università Politecnica delle Marche
for their support to optimize the R code.
Conflicts of Interest: The authors declare no conflicts of interest.

Appendix A
Table A1. Selection of Sentinel-2 Images: All images were employed to represent spectral seasonal
variations as pixel-based functions, which were then used for Hybrid Statistical-Functional-Machine
Learning models with RF Models based on Functional Data Analysis. The * and ** scenes from
2019 were used for the baseline model (Pure Machine Learning Approach) with Random Forest
directly applied to raw time series for the Mount Conero and Frasassi Gorge areas, respectively.

Num Date Doy Week Month Num Date Doy Week Month
1 21 April 2017 111 16 4 48 13 October 2018 286 41 10
2 1 May 2017 121 18 5 49 12 November 2018 316 46 11
3 31 May 2017 151 22 5 50 7 December 2018 341 49 12
4 20 June 2017 171 25 6 51 12 December 2018 346 50 12
5 10 July 2017 191 28 7 52 27 December 2018 361 52 12
Remote Sens. 2024, 16, 1224 19 of 26

Table A1. Cont.

Num Date Doy Week Month Num Date Doy Week Month
6 20 July 2017 201 29 7 53 31 January 2019 31 5 1
7 30 July 2017 211 31 7 54 26 January 2019 26 4 1
8 9 August 2017 221 32 8 55 5 February 2019 36 6 2
9 19 August 2017 231 33 8 56 15 February 2019 ** 46 7 2
10 29 August 2017 241 35 8 57 20 February 2019 * 51 8 2
11 18 September 2017 261 38 9 58 25 February 2019 56 8 2
12 8 October 2017 281 41 10 59 2 March 2019 ** 61 9 3
13 18 October 2017 291 42 10 60 12 March 2019 71 11 3
14 28 October 2017 301 43 10 61 17 March 2019 76 11 3
15 27 November 2017 331 48 11 62 22 March 2019 *, ** 81 12 3
16 7 December 2017 341 49 12 63 1 April 2019 ** 91 13 4
17 22 December 2017 356 51 12 64 16 April 2019 * 106 16 4
18 6 January 2018 6 1 1 65 31 May 2019 151 22 5
19 15 February 2018 46 7 2 66 5 June 2019 *, ** 156 23 6
20 6 April 2018 96 14 4 67 15 June 2019 166 24 6
21 16 April 2018 106 16 4 68 25 June 2019 176 26 6
22 21 April 2018 111 16 4 69 30 June 2019 * 181 26 6
23 26 April 2018 116 17 4 70 5 July 2019 186 27 7
24 11 May 2018 131 19 5 71 20 July 2019 * 201 29 7
25 16 May 2018 136 20 5 72 25 July 2019 ** 206 30 7
26 21 May 2018 141 21 5 73 30 July 2019 211 31 7
27 31 May 2018 151 22 5 74 4 August 2019 * 216 31 8
28 10 June 2018 161 23 6 75 9 August 2019 221 32 8
29 15 June 2018 166 24 6 76 14 August 2019 226 33 8
30 20 June 2018 171 25 6 77 19 August 2019 ** 231 33 8
31 30 June 2018 181 26 6 78 24 August 2019 236 34 8
32 10 July 2018 191 28 7 79 29 August 2019 * 241 35 8
33 15 July 2018 196 28 7 80 8 September 2019 251 36 9
34 20 July 2018 201 29 7 81 13 September 2019 256 37 9
35 25 July 2018 206 30 7 82 18 September 2019 ** 261 38 9
36 30 July 2018 211 31 7 83 8 October 2019 * 281 41 10
37 4 August 2018 216 31 8 84 23 October 2019 ** 296 43 10
38 9 August 2018 221 32 8 85 7 November 2019 311 45 11
39 19 August 2018 231 33 8 86 1 January 2020 1 1 1
40 24 August 2018 236 34 8 87 6 January 2020 6 1 1
41 29 August 2018 241 35 8 88 5 February 2020 36 6 2
42 3 September 2018 246 36 9 89 15 February 2020 46 7 2
43 8 September 2018 251 36 9 90 20 February 2020 51 8 2
44 18 September 2018 261 38 9 91 11 March 2020 71 11 3
45 23 September 2018 266 38 9 92 16 March 2020 76 11 3
46 28 September 2018 271 39 9 93 21 March 2020 81 12 3
47 3 October 2018 276 40 10

Table A2. List of models for the Mount Conero area, displaying their accuracy (OA—Overall Accuracy
and sd—standard deviation; for c1–c4 vegetation types Producer’s Accuracy was reported) and model
complexity (pr—number of input predictors and Random Forest’s mtry value for tree splits).

Model Formula pr mtry OA sd c1 c2 c3 c4

B 0 38 4 0.818 0.095 0.768 0.893 0.477 0.837
M 0 6 1 0.812 0.076 0.692 0.901 0.538 0.844
M 1 2 1 0.838 0.085 0.730 0.887 0.754 0.867
M 2 18 1 0.768 0.082 0.757 0.887 0.015 0.800
M 3 6 1 0.825 0.076 0.654 0.941 0.462 0.878
M 4 34 1 0.790 0.081 0.714 0.930 0.031 0.841
M 5 36 2 0.793 0.075 0.768 0.899 0.015 0.859
M 6 30 5 0.675 0.092 0.400 0.893 0.138 0.704
Remote Sens. 2024, 16, 1224 20 of 26

Table A2. Cont.

Model Formula pr mtry OA sd c1 c2 c3 c4

M 7 26 5 0.802 0.082 0.703 0.927 0.385 0.807
M 8 34 5 0.663 0.120 0.454 0.930 0.185 0.567
M 9 36 3 0.797 0.091 0.703 0.935 0.262 0.807
M 10 14 1 0.778 0.087 0.768 0.918 0.000 0.789
M 11 10 2 0.790 0.088 0.686 0.868 0.508 0.830
M 12 22 3 0.732 0.088 0.708 0.882 0.000 0.730
M 13 10 1 0.827 0.074 0.719 0.955 0.354 0.844
M 14 34 4 0.671 0.098 0.562 0.859 0.385 0.570
M 15 10 3 0.856 0.072 0.762 0.938 0.615 0.870
M 16 6 2 0.814 0.076 0.659 0.904 0.615 0.848
M 17 6 2 0.838 0.081 0.730 0.899 0.585 0.893
M 18 10 2 0.829 0.081 0.714 0.913 0.492 0.878
M 19 18 4 0.794 0.081 0.735 0.904 0.354 0.796
M 20 36 6 0.793 0.090 0.703 0.921 0.215 0.826
mF 0 46 2 0.826 0.084 0.751 0.910 0.477 0.852
mF 1 258 11 0.816 0.082 0.714 0.893 0.492 0.863
mF 2 274 13 0.844 0.079 0.773 0.921 0.554 0.859
mF 3 290 7 0.849 0.074 0.719 0.944 0.615 0.870
mF 4 290 7 0.857 0.070 0.746 0.938 0.631 0.881
mF 5 630 15 0.857 0.070 0.751 0.955 0.585 0.867
mF 6 674 21 0.841 0.074 0.741 0.930 0.585 0.856
mF 7 294 15 0.854 0.066 0.724 0.924 0.738 0.878
mF 8 954 22 0.831 0.084 0.757 0.930 0.508 0.830
mF 9 818 19 0.856 0.071 0.730 0.972 0.523 0.870
mF 10 518 21 0.835 0.077 0.757 0.907 0.554 0.863
mF 11 658 20 0.860 0.072 0.751 0.963 0.631 0.856
mF 12 910 28 0.842 0.078 0.746 0.941 0.477 0.867
mF 13 118 2 0.828 0.084 0.703 0.907 0.631 0.859
mF 14 710 20 0.845 0.071 0.762 0.938 0.615 0.833
mF 15 634 24 0.845 0.072 0.730 0.927 0.662 0.863
mF 16 674 11 0.833 0.088 0.724 0.907 0.600 0.867
mF 17 610 7 0.847 0.070 0.730 0.932 0.646 0.863
mF 18 610 3 0.850 0.066 0.730 0.949 0.631 0.856
mF 19 250 6 0.852 0.077 0.735 0.944 0.600 0.870
mF 20 122 3 0.818 0.094 0.708 0.882 0.692 0.841
Ms 0 14 3 0.812 0.086 0.730 0.913 0.354 0.848
Ms 1 6 1 0.835 0.075 0.719 0.938 0.431 0.878
Ms 2 19 3 0.839 0.073 0.762 0.944 0.369 0.867
Ms 3 10 2 0.849 0.082 0.746 0.932 0.615 0.867
Ms 4 8 3 0.779 0.086 0.719 0.834 0.338 0.856
Ms 5 10 2 0.859 0.072 0.751 0.941 0.677 0.870
Ms 6 14 2 0.842 0.071 0.697 0.961 0.585 0.848
Ms 7 6 1 0.860 0.082 0.697 0.958 0.769 0.863
Ms 8 10 2 0.838 0.082 0.778 0.927 0.431 0.859
Ms 9 10 3 0.848 0.072 0.708 0.966 0.523 0.870
Ms 10 24 4 0.840 0.081 0.751 0.944 0.631 0.815
Ms 11 14 2 0.860 0.074 0.757 0.938 0.646 0.881
Ms 12 10 1 0.851 0.074 0.686 0.972 0.662 0.852
Ms 13 10 3 0.844 0.084 0.751 0.932 0.523 0.870
Ms 14 10 2 0.838 0.077 0.762 0.913 0.554 0.859
Ms 15 10 2 0.872 0.078 0.800 0.966 0.585 0.867
Ms 16 10 3 0.844 0.079 0.795 0.938 0.446 0.848
Ms 17 6 1 0.847 0.079 0.757 0.938 0.554 0.856
Ms 18 10 3 0.857 0.073 0.773 0.921 0.677 0.874
Ms 19 10 2 0.850 0.075 0.741 0.930 0.646 0.867
Ms 20 6 2 0.851 0.075 0.697 0.941 0.754 0.863
Remote Sens. 2024, 16, 1224 21 of 26

Table A3. List of models for the Frasassi Gorge area, displaying their accuracy (OA—Overall Accuracy
and sd—standard deviation; for v1–v8 vegetation types Producer’s Accuracy was reported) and
model complexity (pr—number of input predictors and Random Forest’s mtry value for tree splits).

Model Formula pr mtry OA sd v1 v2 v3 v4 v5 v6 v7 v8

B 0 62 3 0.770 0.071 0.829 0.507 0.804 0.910 0.313 0.707 0.813 0.922
M 0 26 4 0.785 0.070 0.771 0.571 0.857 0.813 0.350 0.707 0.850 0.974
M 1 26 5 0.778 0.076 0.882 0.436 0.864 0.871 0.287 0.653 0.825 0.935
M 2 30 5 0.675 0.083 0.676 0.371 0.696 0.794 0.550 0.600 0.775 0.791
M 3 30 5 0.817 0.063 0.853 0.600 0.839 0.877 0.400 0.827 0.875 0.974
M 4 34 5 0.733 0.080 0.682 0.421 0.718 0.903 0.550 0.653 0.813 0.930
M 5 34 5 0.731 0.079 0.700 0.529 0.721 0.826 0.525 0.707 0.800 0.883
M 6 34 5 0.646 0.078 0.682 0.314 0.857 0.587 0.137 0.293 0.675 0.887
M 7 34 5 0.823 0.080 0.924 0.500 0.879 0.865 0.413 0.893 0.800 0.978
M 8 36 6 0.644 0.081 0.618 0.300 0.893 0.613 0.187 0.093 0.588 0.948
M 9 22 2 0.754 0.077 0.747 0.386 0.843 0.761 0.425 0.747 0.850 0.957
M 10 36 4 0.667 0.081 0.506 0.393 0.718 0.710 0.512 0.640 0.750 0.900
M 11 36 6 0.708 0.075 0.712 0.164 0.768 0.839 0.463 0.680 0.775 0.952
M 12 36 4 0.668 0.090 0.588 0.407 0.743 0.626 0.375 0.613 0.763 0.913
M 13 30 3 0.764 0.074 0.835 0.307 0.896 0.787 0.300 0.747 0.775 0.974
M 14 34 3 0.634 0.072 0.659 0.179 0.882 0.690 0.050 0.053 0.838 0.870
M 15 30 4 0.798 0.069 0.841 0.543 0.829 0.897 0.375 0.773 0.788 0.978
M 16 18 3 0.788 0.071 0.771 0.536 0.904 0.787 0.400 0.693 0.875 0.948
M 17 18 2 0.810 0.071 0.812 0.521 0.843 0.839 0.562 0.893 0.825 0.978
M 18 18 4 0.813 0.078 0.924 0.586 0.807 0.852 0.463 0.787 0.813 0.978
M 19 30 4 0.764 0.071 0.841 0.486 0.825 0.761 0.338 0.773 0.813 0.935
M 20 34 4 0.786 0.070 0.771 0.493 0.861 0.890 0.325 0.760 0.800 0.978
mF 0 58 2 0.773 0.073 0.735 0.614 0.839 0.839 0.312 0.493 0.938 0.970
mF 1 250 3 0.773 0.066 0.806 0.550 0.850 0.819 0.350 0.640 0.863 0.922
mF 2 275 16 0.816 0.067 0.924 0.571 0.814 0.903 0.613 0.680 0.813 0.948
mF 3 202 7 0.829 0.062 0.912 0.579 0.807 0.942 0.588 0.707 0.875 0.978
mF 4 550 22 0.811 0.065 0.924 0.550 0.821 0.890 0.488 0.733 0.813 0.961
mF 5 550 9 0.819 0.073 0.935 0.600 0.821 0.890 0.525 0.720 0.813 0.952
mF 6 202 12 0.808 0.072 0.947 0.564 0.768 0.903 0.450 0.760 0.913 0.943
mF 7 606 15 0.818 0.066 0.953 0.579 0.789 0.897 0.563 0.693 0.813 0.978
mF 8 530 2 0.816 0.065 0.894 0.500 0.893 0.884 0.350 0.707 0.938 0.970
mF 9 998 17 0.816 0.062 0.853 0.529 0.900 0.871 0.475 0.720 0.813 0.978
mF 10 470 21 0.792 0.065 0.894 0.536 0.782 0.865 0.550 0.707 0.850 0.935
mF 11 606 12 0.825 0.065 0.912 0.536 0.882 0.890 0.500 0.707 0.813 0.978
mF 12 886 20 0.825 0.065 0.935 0.529 0.879 0.923 0.550 0.707 0.813 0.935
mF 13 498 1 0.785 0.064 0.853 0.493 0.879 0.832 0.375 0.627 0.863 0.935
mF 14 782 26 0.798 0.071 0.947 0.571 0.786 0.871 0.350 0.720 0.875 0.948
mF 15 646 25 0.806 0.067 0.935 0.557 0.761 0.910 0.475 0.707 0.850 0.978
mF 16 470 1 0.784 0.068 0.835 0.464 0.868 0.839 0.413 0.653 0.863 0.943
mF 17 510 10 0.813 0.066 0.906 0.600 0.786 0.903 0.500 0.693 0.875 0.978
mF 18 438 12 0.805 0.066 0.912 0.571 0.779 0.871 0.488 0.707 0.875 0.978
mF 19 202 4 0.820 0.063 0.906 0.607 0.814 0.890 0.575 0.693 0.813 0.978
mF 20 474 6 0.789 0.067 0.788 0.557 0.846 0.858 0.338 0.707 0.925 0.952
Ms 0 22 3 0.812 0.076 0.865 0.586 0.821 0.839 0.625 0.733 0.813 0.970
Ms 1 22 4 0.845 0.065 0.929 0.550 0.839 0.923 0.663 0.827 0.825 0.987
Ms 2 26 1 0.824 0.073 0.924 0.471 0.893 0.884 0.475 0.840 0.925 0.922
Ms 3 18 2 0.829 0.075 0.953 0.543 0.839 0.942 0.713 0.693 0.762 0.930
Ms 4 22 1 0.832 0.081 0.912 0.579 0.871 0.910 0.425 0.800 0.863 0.970
Ms 5 14 2 0.842 0.070 0.924 0.714 0.846 0.813 0.587 0.773 0.863 0.978
Ms 6 22 4 0.815 0.065 0.941 0.464 0.814 0.903 0.437 0.813 0.925 0.970
Ms 7 18 3 0.836 0.065 0.971 0.514 0.868 0.890 0.525 0.867 0.775 0.974
Ms 8 34 3 0.840 0.075 0.900 0.679 0.879 0.897 0.437 0.733 0.925 0.952
Ms 9 18 1 0.840 0.060 0.906 0.486 0.936 0.806 0.437 0.973 0.875 1.000
Ms 10 18 2 0.811 0.071 0.959 0.579 0.846 0.852 0.400 0.747 0.825 0.930
Remote Sens. 2024, 16, 1224 22 of 26

Table A3. Cont.

Model Formula pr mtry OA sd v1 v2 v3 v4 v5 v6 v7 v8

Ms 11 22 2 0.828 0.075 0.853 0.450 0.893 0.935 0.625 0.787 0.813 0.978
Ms 12 18 2 0.814 0.087 0.900 0.529 0.843 0.890 0.437 0.907 0.813 0.939
Ms 13 22 4 0.833 0.074 0.924 0.600 0.864 0.865 0.538 0.787 0.813 0.970
Ms 14 22 3 0.840 0.074 0.865 0.664 0.868 0.897 0.525 0.840 0.888 0.948
Ms 15 22 3 0.865 0.070 0.953 0.586 0.921 0.935 0.575 0.773 0.888 0.978
Ms 16 22 4 0.828 0.064 0.906 0.493 0.896 0.845 0.425 0.880 0.900 0.974
Ms 17 22 4 0.856 0.055 1.000 0.550 0.893 0.910 0.550 0.827 0.838 0.978
Ms 18 22 2 0.835 0.055 0.953 0.457 0.889 0.910 0.613 0.827 0.838 0.943
Ms 19 26 4 0.811 0.063 0.894 0.493 0.864 0.890 0.425 0.773 0.863 0.957
Ms 20 22 2 0.822 0.064 0.747 0.621 0.929 0.865 0.550 0.707 0.763 1.000

References
1. The Habitats Directive. Council Directive 92/43/EEC of 21 May 1992 on the Conservation of Natural Habitats and of Wild Fauna
and Flora. Off. J. L 1992, 206, 7–50.
2. Evans, D. The Habitats of the European Union Habitats Directive. Biol. Environ. Proc. R. Irish Acad. 2006, 106B, 167–173. [CrossRef]
3. Corbane, C.; Lang, S.; Pipkins, K.; Alleaume, S.; Deshayes, M.; García Millán, V.E.; Strasser, T.; Vanden Borre, J.; Toon, S.; Michael, F.
Remote Sensing for Mapping Natural Habitats and Their Conservation Status—New Opportunities and Challenges. Int. J. Appl.
Earth Obs. Geoinf. 2015, 37, 7–16. [CrossRef]
4. Vanden Borre, J.; Paelinckx, D.; Mücher, C.A.; Kooistra, L.; Haest, B.; De Blust, G.; Schmidt, A.M. Integrating Remote Sensing in
Natura 2000 Habitat Monitoring: Prospects on the Way Forward. J. Nat. Conserv. 2011, 19, 116–125. [CrossRef]
5. Schmidt, T.; Schuster, C.; Kleinschmit, B.; Förster, M. Evaluating an Intra-Annual Time Series for Grassland Classification—How
Many Acquisitions and What Seasonal Origin Are Optimal? IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2014, 7, 3428–3439.
[CrossRef]
6. Rapinel, S.; Rozo, C.; Delbosc, P.; Bioret, F.; Bouzillé, J.B.; Hubert-Moy, L. Contribution of Free Satellite Time-Series Images to
Mapping Plant Communities in the Mediterranean Natura 2000 Site: The Example of Biguglia Pond in Corse (France). Mediterr.
Bot. 2020, 41, 181–191. [CrossRef]
7. Marzialetti, F.; Giulio, S.; Malavasi, M.; Sperandii, M.G.; Acosta, A.T.R.; Carranza, M.L. Capturing Coastal Dune Natural
Vegetation Types Using a Phenology-Based Mapping Approach: The Potential of Sentinel-2. Remote Sens. 2019, 11, 1506.
[CrossRef]
8. Bajocco, S.; Ferrara, C.; Alivernini, A.; Bascietto, M.; Ricotta, C. Remotely-Sensed Phenology of Italian Forests: Going beyond the
Species. Int. J. Appl. Earth Obs. Geoinf. 2019, 74, 314–321. [CrossRef]
9. Grignetti, A.; Salvatori, R.; Casacchia, R.; Manes, F. Mediterranean Vegetation Analysis by Multi-Temporal Satellite Sensor Data.
Int. J. Remote Sens. 1997, 18, 1307–1318. [CrossRef]
10. Marzialetti, F.; Di Febbraro, M.; Malavasi, M.; Giulio, S.; Acosta, A.T.R.; Carranza, M.L. Mapping Coastal Dune Landscape
through Spectral Rao’s Q Temporal Diversity. Remote Sens. 2020, 12, 2315. [CrossRef]
11. Sittaro, F.; Hutengs, C.; Semella, S.; Vohland, M. A Machine Learning Framework for the Classification of Natura 2000 Habitat
Types at Large Spatial Scales Using MODIS Surface Reflectance Data. Remote Sens. 2022, 14, 823. [CrossRef]
12. Mahmud, S.; Redowan, M.; Ahmed, R.; Khan, A.A.; Rahman, M.M. Phenology-Based Classification of Sentinel-2 Data to Detect
Coastal Mangroves. Geocarto Int. 2022, 37, 14335–14354. [CrossRef]
13. Raab, C.; Stroh, H.G.; Tonn, B.; Meißner, M.; Rohwer, N.; Balkenhol, N.; Isselstein, J. Mapping Semi-Natural Grassland
Communities Using Multi-Temporal RapidEye Remote Sensing Data. Int. J. Remote Sens. 2018, 39, 5638–5659. [CrossRef]
14. Hubert-Moy, L.; Fabre, E.; Rapinel, S. Contribution of SPOT-7 Multi-Temporal Imagery for Mapping Wetland Vegetation. Eur.
J. Remote Sens. 2020, 53, 201–210. [CrossRef]
15. Jarocińska, A.; Kopeć, D.; Niedzielko, J.; Wylazłowska, J.; Halladin-Dabrowska,
˛ A.; Charyton, J.; Piernik, A.; Kamiński, D.
The Utility of Airborne Hyperspectral and Satellite Multispectral Images in Identifying Natura 2000 Non-Forest Habitats for
Conservation Purposes. Sci. Rep. 2023, 13, 4549. [CrossRef] [PubMed]
16. Tarantino, C.; Forte, L.; Blonda, P.; Vicario, S.; Tomaselli, V.; Beierkuhnlein, C.; Adamo, M. Intra-Annual Sentinel-2 Time-Series
Supporting Grassland Habitat Discrimination. Remote Sens. 2021, 13, 277. [CrossRef]
17. Buck, O.; Millán, V.E.G.; Klink, A.; Pakzad, K. Using Information Layers for Mapping Grassland Habitat Distribution at Local to
Regional Scales. Int. J. Appl. Earth Obs. Geoinf. 2015, 37, 83–89. [CrossRef]
18. Rapinel, S.; Mony, C.; Lecoq, L.; Clément, B.; Thomas, A.; Hubert-Moy, L. Evaluation of Sentinel-2 Time-Series for Mapping
Floodplain Grassland Plant Communities. Remote Sens. Environ. 2019, 223, 115–129. [CrossRef]
19. Durell, L.; Scott, J.T.; Hering, A.S. Hybrid Forecasting for Functional Time Series of Dissolved Oxygen Profiles. Data Sci. Sci. 2023,
2, 2152401. [CrossRef]
Remote Sens. 2024, 16, 1224 23 of 26

20. Huang, S.; Tang, L.; Hupy, J.P.; Wang, Y.; Shao, G. A Commentary Review on the Use of Normalized Difference Vegetation Index
(NDVI) in the Era of Popular Remote Sensing. J. For. Res. 2021, 32, 1–6. [CrossRef]
21. Vanden Borre, J.; Spanhove, T.; Haest, B. Towards a Mature Age of Remote Sensing for Natura 2000 Habitat Conservation:
Poor Method Transferability as a Prime Obstacle. In The Roles of Remote Sensing in Nature Conservation; Springer International
Publishing: Cham, Switzerland, 2017; pp. 11–37.
22. Xue, J.; Su, B. Significant Remote Sensing Vegetation Indices: A Review of Developments and Applications. J. Sens. 2017, 2017,
1353691. [CrossRef]
23. Fatima, N.; Javed, A. Assessment of Land Use Land Cover Change Detection Using Geospatial Techniques in Southeast Rajasthan.
J. Geosci. Environ. Prot. 2021, 9, 299–319. [CrossRef]
24. Barrett, B.; Raab, C.; Cawkwell, F.; Green, S. Upland Vegetation Mapping Using Random Forests with Optical and Radar Satellite
Data. Remote Sens. Ecol. Conserv. 2016, 2, 212–231. [CrossRef] [PubMed]
25. Nagendra, H.; Lucas, R.; Honrado, J.P.; Jongman, R.H.G.; Tarantino, C.; Adamo, M.; Mairota, P. Remote Sensing for Conservation
Monitoring: Assessing Protected Areas, Habitat Extent, Habitat Condition, Species Diversity, and Threats. Ecol. Indic. 2013,
33, 45–59. [CrossRef]
26. Pasquarella, V.J.; Holden, C.E.; Kaufman, L.; Woodcock, C.E. From Imagery to Ecology: Leveraging Time Series of All Available
Landsat Observations to Map and Monitor Ecosystem State and Dynamics. Remote Sens. Ecol. Conserv. 2016, 2, 152–170. [CrossRef]
27. Gillanders, S.N.; Coops, N.C.; Wulder, M.A.; Gergel, S.E.; Nelson, T. Multitemporal Remote Sensing of Landscape Dynamics and
Pattern Change: Describing Natural and Anthropogenic Trends. Prog. Phys. Geogr. Earth Environ. 2008, 32, 503–528. [CrossRef]
28. Ramsay, J.O.; Silverman, B.W. Functional Data Analysis; Ramsay, R., Silverman, B., Eds.; Springer Series in Statistics; Springer:
New York, NY, USA, 2005; ISBN 978-0-387-40080-8.
29. Pesaresi, S.; Mancini, A.; Quattrini, G.; Casavecchia, S. Functional Analysis for Habitat Mapping in a Special Area of Conservation
Using Sentinel-2 Time-Series Data. Remote Sens. 2022, 14, 1179. [CrossRef]
30. Pesaresi, S.; Mancini, A.; Quattrini, G.; Casavecchia, S. Mapping Mediterranean Forest Plant Associations and Habitats with
Functional Principal Component Analysis Using Landsat 8 NDVI Time Series. Remote Sens. 2020, 12, 1132. [CrossRef]
31. Coviello, L.; Martini, F.M.; Cesaretti, L.; Pesaresi, S.; Solfanelli, F.; Mancini, A. Clustering of Remotely Sensed Time Series Using
Functional Principal Component Analysis to Monitor Crops. In Proceedings of the 2022 IEEE Workshop on Metrology for
Agriculture and Forestry (MetroAgriFor), Perugia, Italy, 3–5 November 2022; pp. 141–145.
32. Hurley, M.A.; Hebblewhite, M.; Gaillard, J.; Dray, S.; Taylor, K.A.; Smith, W.K.; Zager, P.; Bonenfant, C. Functional Analysis of
Normalized Difference Vegetation Index Curves Reveals Overwinter Mule Deer Survival Is Driven by Both Spring and Autumn
Phenology. Philos. Trans. R. Soc. Lond. B Biol. Sci. 2014, 369, 20130196. [CrossRef]
33. Pesaresi, S.; Mancini, A.; Casavecchia, S. Recognition and Characterization of Forest Plant Communities through Remote-Sensing
NDVI Time Series. Diversity 2020, 12, 313. [CrossRef]
34. Ramsay, J.O. When the Data Are Functions. Psychometrika 1982, 47, 379–396. [CrossRef]
35. Kennedy, R.E.; Andréfouët, S.; Cohen, W.B.; Gómez, C.; Griffiths, P.; Hais, M.; Healey, S.P.; Helmer, E.H.; Hostert, P.; Lyons, M.B.; et al.
Bringing an Ecological View of Change to Landsat-Based Remote Sensing. Front. Ecol. Environ. 2014, 12, 339–346. [CrossRef]
[PubMed]
36. Levitin, D.J.; Nuzzo, R.L.; Vines, B.; Ramsay, J.O. Introduction to Functional Data Analysis. Can. Psychol. 2007, 48, 135–155.
[CrossRef]
37. Ramsay, J.O.; Dalzell, C.J. Some Tools for Functional Data Analysis. J. R. Stat. Soc. Ser. B 1991, 53, 539–572. [CrossRef]
38. Happ, C.; Greven, S. Multivariate Functional Principal Component Analysis for Data Observed on Different (Dimensional)
Domains. J. Am. Stat. Assoc. 2018, 113, 649–659. [CrossRef]
39. Wang, J.-L.; Chiou, J.-M.; Müller, H.-G. Functional Data Analysis. Annu. Rev. Stat. Its Appl. 2016, 3, 257–295. [CrossRef]
40. Geobotanic Group at Università Politecnica delle Marche. Dataset and R Code Related to the Habitat Mapping with Functional
Hybrid Machine Learning. Available online: https://github.com/geobotany (accessed on 15 January 2024).
41. Rivas-Martínez, S.; Sáenz, S.R.; Penas, A. Worldwide Bioclimatic Classification System. Glob. Geobot. 2011, 1, 1–634.
42. Pesaresi, S.; Biondi, E.; Casavecchia, S. Bioclimates of Italy. J. Maps 2017, 13, 955–960. [CrossRef]
43. Biondi, E.; Casavecchia, S.; Gigante, D. Contribution to the Syntaxonomic Knowledge of the Quercus Ilex L. Woods of the Central
European Mediterranean Basin. Fitosociologia 2003, 40, 129–156.
44. Biondi, E.; Gubellini, L.; Pinzi, M.; Casavecchia, S. The Vascular Flora of Conero Regional Nature Park (Marche, Central Italy).
Flora Mediterr. 2012, 22, 67–167. [CrossRef]
45. Biondi, E. L’ostrya Carpinifolia Scop. Sul Litorale Delle Marche (Italia Centrale). Stud. Geobot. 1982, 2, 141–147.
46. Baiocco, M.; Casavecchia, S.; Biondi, E.; Pietracapina, A. Indagini Geobotaniche per Il Recupero Del Rimboschimento Del Monte
Conero (Italia Centrale). Doc. Phytosociol. 1996, 16, 387–425.
47. Blasi, C.; Di Pietro, R.; Filesi, L. Syntaxonomical Revision of Quercetalia Pubescenti-Petraeae in the Italian Peninsula. Fitosociologia
2004, 41, 87–164.
48. Blasi, C.; Feoli, E.; Avena, G.C. Due Nuove Associazioni Dei Quercetalia Pubescentis Dell’Appennino Centrale. Stud. Geobot. 1982,
2, 155–167.
Remote Sens. 2024, 16, 1224 24 of 26

49. Pedrotti, F.; Ballelli, S.; Biondi, E.; Cortini Pedrotti, C.; Orsomando, E. Resoconto Dell’escursione Della Società Italiana Di
Fitosociologia Nelle Marche Ed in Umbria (11–14 Giugno 1979). Not. Fitosociologico 1980, 16, 73–75.
50. Allegrezza, M.; Pesaresi, S.; Ballelli, S.; Tesei, G.; Ottaviani, C. Influences of Mature Pinus Nigra Plantations on the Floristic-
Vegetational Composition along an Altitudinal Gradient in the Central Apennines, Italy. iForest 2020, 13, 279–285. [CrossRef]
51. Biondi, E.; Casavecchia, S. Inquadramento Fitosociologico Della Vegetazione Arbustiva Di Un Settore Dell’Appennino Settentri-
onale. Fitosociologia 2002, 39, 65–73.
52. Biondi, E.; Allegrezza, M.; Zuccarello, V. Syntaxonomic Revision of the Apennine Grasslands Belonging to Brometalia Erecti, and
an Analysis of Their Relationships with the Xerophilous Vegetation of Rosmarinetea Officinalis (Italy). Phytocoenologia 2005, 35,
129–164. [CrossRef]
53. Allegrezza, M.; Biondi, E.; Ballelli, S.; Formica, E. La Vegetazione Dei Settori Rupestri Calcarei Dell’Italia Centrale. Fitosociologia
1997, 32, 91–120.
54. Ranghetti, L.; Boschetti, M.; Nutini, F.; Busetto, L. “Sen2r”: An R Toolbox for Automatically Downloading and Preprocessing
Sentinel-2 Satellite Data. Comput. Geosci. 2020, 139, 104473. [CrossRef]
55. Zeng, Y.; Hao, D.; Huete, A.; Dechant, B.; Berry, J.; Chen, J.M.; Joiner, J.; Frankenberg, C.; Bond-Lamberty, B.; Ryu, Y.; et al. Optical
Vegetation Indices for Monitoring Terrestrial Ecosystems Globally. Nat. Rev. Earth Environ. 2022, 3, 477–493. [CrossRef]
56. ESA. Sentinel-2 User Handbook. Available online: https://sentinel.esa.int/documents/247904/685211/sentinel-2_user_
handbook (accessed on 15 January 2024).
57. Fisher, J.I.; Mustard, J.F.; Vadeboncoeur, M.A. Green Leaf Phenology at Landsat Resolution: Scaling from the Field to the Satellite.
Remote Sens. Environ. 2006, 100, 265–279. [CrossRef]
58. Schuster, C.; Schmidt, T.; Conrad, C.; Kleinschmit, B.; Förster, M. Grassland habitat mapping by intra-annual time series
analysis—Comparison of RapidEye and TerraSAR-X satellite data. Int. J. Appl. Earth Obs. Geoinf. 2015, 34, 25–34. [CrossRef]
59. Lambert, J.; Drenou, C.; Denux, J.; Balent, G.; Cheret, V. Monitoring Forest Decline through Remote Sensing Time Series Analysis.
GISci. Remote Sens. 2013, 50, 437–457. [CrossRef]
60. Hyndman, R.; Athanasopoulos, G.; Bergmeir, C.; Caceres, G.; Chhay, L.; O’Hara-Wild, M.; Petropoulos, F.; Razbash, S.; Wang, E.;
Yasmeen, F. Forecast: Forecasting Functions for Time Series and Linear Models. R Package Version 8.6. Available online:
https://cran.r-project.org/package=forecast (accessed on 3 August 2020).
61. Hyndman, R.J.; Khandakar, Y. Automatic Time Series Forecasting: The Forecast Package for R. J. Stat. Softw. 2008, 27, 1–22.
[CrossRef]
62. Wood, S.N. Generalized Additive Models: An Introduction with R; Chapman and Hall/CRC: New York, NY, USA, 2017;
ISBN 9781315370279.
63. Younes, N.; Joyce, K.E.; Maier, S.W. All Models of Satellite-Derived Phenology Are Wrong, but Some Are Useful: A Case Study
from Northern Australia. Int. J. Appl. Earth Obs. Geoinf. 2021, 97, 102285. [CrossRef]
64. Di Salvo, F.; Ruggieri, M.; Plaia, A. Functional Principal Component Analysis for Multivariate Multidimensional Environmental
Data. Environ. Ecol. Stat. 2015, 22, 739–757. [CrossRef]
65. Dai, X.; Hadjipantelis, P.Z.; Han, K.; Ji, H. Fdapace: Functional Data Analysis and Empirical Dynamics. R Package Version 0.5.5.
Available online: https://cran.r-project.org/package=fdapace (accessed on 3 August 2020).
66. Happ-Kurz, C. MFPCA: Multivariate Functional Principal Component Analysis for Data Observed on Different Dimensional
Domains. R Package Version 1.3-6. Available online: https://cran.r-project.org/web/packages/MFPCA/index.html (accessed
on 22 March 2022).
67. Breiman, L. Random Forests. Mach. Learn. 2001, 45, 5–32. [CrossRef]
68. Belgiu, M.; Drăguţ, L. Random Forest in Remote Sensing: A Review of Applications and Future Directions. ISPRS J. Photogramm.
Remote Sens. 2016, 114, 24–31. [CrossRef]
69. Evans, J.S.; Cushman, S.A. Gradient Modeling of Conifer Species Using Random Forests. Landsc. Ecol. 2009, 24, 673–683. [CrossRef]
70. Le Dez, M.; Robin, M.; Launeau, P. Contribution of Sentinel-2 Satellite Images for Habitat Mapping of the Natura 2000 Site
‘Estuaire de La Loire’ (France). Remote Sens. Appl. Soc. Environ. 2021, 24, 100637. [CrossRef]
71. Marcinkowska-Ochtyra, A.; Ochtyra, A.; Raczko, E.; Kopeć, D. Natura 2000 Grassland Habitats Mapping Based on Spectro-
Temporal Dimension of Sentinel-2 Images with Machine Learning. Remote Sens. 2023, 15, 1388. [CrossRef]
72. Wakulińska, M.; Marcinkowska-Ochtyra, A. Multi-Temporal Sentinel-2 Data in Classification of Mountain Vegetation. Remote
Sens. 2020, 12, 2696. [CrossRef]
73. Congalton, R.G. A Review of Assessing the Accuracy of Classifications of Remotely Sensed Data. Remote Sens. Environ. 1991,
37, 35–46. [CrossRef]
74. Cohen, J. A Coefficient of Agreement for Nominal Scales. Educ. Psychol. Meas. 1960, 20, 37–46. [CrossRef]
75. Kuhn, M. Building Predictive Models in R Using the Caret Package. J. Stat. Softw. 2008, 28, 1–26. [CrossRef]
76. Pham-Duc, B.; Nguyen, H.; Phan, H.; Tran-Anh, Q. Trends and Applications of Google Earth Engine in Remote Sensing and Earth
Science Research: A Bibliometric Analysis Using Scopus Database. Earth Sci. Inform. 2023, 16, 2355–2371. [CrossRef]
77. Gorelick, N.; Hancher, M.; Dixon, M.; Ilyushchenko, S.; Thau, D.; Moore, R. Google Earth Engine: Planetary-Scale Geospatial
Analysis for Everyone. Remote Sens. Environ. 2017, 202, 18–27. [CrossRef]
Remote Sens. 2024, 16, 1224 25 of 26

78. Pettorelli, N.; Vik, J.O.; Mysterud, A.; Gaillard, J.-M.; Tucker, C.J.; Stenseth, N.C. Using the Satellite-Derived NDVI to Assess
Ecological Responses to Environmental Change. Trends Ecol. Evol. 2005, 20, 503–510. [CrossRef]
79. Grabska, E.; Hostert, P.; Pflugmacher, D.; Ostapowicz, K. Forest Stand Species Mapping Using the Sentinel-2 Time Series. Remote
Sens. 2019, 11, 1197. [CrossRef]
80. Vrieling, A.; Meroni, M.; Darvishzadeh, R.; Skidmore, A.K.; Wang, T.; Zurita-Milla, R.; Oosterbeek, K.; O’Connor, B.; Paganini, M.
Vegetation Phenology from Sentinel-2 and Field Cameras for a Dutch Barrier Island. Remote Sens. Environ. 2018, 215, 517–529.
[CrossRef]
81. Pasquarella, V.J.; Holden, C.E.; Woodcock, C.E. Improved Mapping of Forest Type Using Spectral-Temporal Landsat Features.
Remote Sens. Environ. 2018, 210, 193–207. [CrossRef]
82. Alvera-Azcárate, A.; Sirjacobs, D.; Barth, A.; Beckers, J.-M. Outlier Detection in Satellite Data Using Spatial Coherence. Remote
Sens. Environ. 2012, 119, 84–91. [CrossRef]
83. Balestra, M.; Pierdicca, R.; Cesaretti, L.; Quattrini, G.; Mancini, A.; Galli, A.; Malinverni, E.S.; Casavecchia, S.; Pesaresi, S.
A comparison of pre-processing approaches for remotely sensed time series classification based on functional analysis. ISPRS
Ann. Photogramm. Remote Sens. Spat. Inf. Sci. 2023. [CrossRef]
84. Liu, C.; Ray, S.; Hooker, G.; Friedl, M. Functional Factor Analysis for Periodic Remote Sensing Data. Ann. Appl. Stat. 2012,
6, 601–624. [CrossRef]
85. Fassnacht, F.E.; Neumann, C.; Forster, M.; Buddenbaum, H.; Ghosh, A.; Clasen, A.; Joshi, P.K.; Koch, B. Comparison of Feature
Reduction Algorithms for Classifying Tree Species with Hyperspectral Data on Three Central European Test Sites. IEEE J. Sel. Top.
Appl. Earth Obs. Remote Sens. 2014, 7, 2547–2561. [CrossRef]
86. Saini, R.; Ghosh, S.K. Analyzing the Impact of Red-Edge Band on Land Use Land Cover Classification Using Multispectral
RapidEye Imagery and Machine Learning Techniques. J. Appl. Remote Sens. 2019, 13, 044511. [CrossRef]
87. Schuster, C.; Förster, M.; Kleinschmit, B. Testing the Red Edge Channel for Improving Land-Use Classifications Based on
High-Resolution Multi-Spectral Satellite Data. Int. J. Remote Sens. 2012, 33, 5583–5599. [CrossRef]
88. Immitzer, M.; Vuolo, F.; Atzberger, C. First Experience with Sentinel-2 Data for Crop and Tree Species Classifications in Central
Europe. Remote Sens. 2016, 8, 166. [CrossRef]
89. Meyer, G.E.; Neto, J.C. Verification of Color Vegetation Indices for Automated Crop Imaging Applications. Comput. Electron.
Agric. 2008, 63, 282–293. [CrossRef]
90. Alcaraz-Segura, D.; Cabello, J.; Paruelo, J. Baseline Characterization of Major Iberian Vegetation Types Based on the NDVI
Dynamics. Plant Ecol. 2009, 202, 13–29. [CrossRef]
91. Saini, R. Integrating Vegetation Indices and Spectral Features for Vegetation Mapping from Multispectral Satellite Imagery Using
AdaBoost and Random Forest Machine Learning Classifiers. Geomat. Environ. Eng. 2022, 17, 57–74. [CrossRef]
92. Illarionova, S.; Shadrin, D.; Trekin, A.; Ignatiev, V.; Oseledets, I. Generation of the NIR Spectral Band for Satellite Images with
Convolutional Neural Networks. Sensors 2021, 21, 5646. [CrossRef] [PubMed]
93. Chen, J.; Jo, P. A Simple Method for Reconstructing a High-Quality NDVI Time-Series Data Set Based on the Savitzky–Golay
Filter. Remote Sens. Environ. 2004, 91, 332–344. [CrossRef]
94. Li, S.; Xu, L.; Jing, Y.; Yin, H.; Li, X.; Guan, X. High-Quality Vegetation Index Product Generation: A Review of NDVI Time Series
Reconstruction Techniques. Int. J. Appl. Earth Obs. Geoinf. 2021, 105, 102640. [CrossRef]
95. Marcinkowska-Ochtyra, A.; Gryguc, K.; Ochtyra, A.; Kopeć, D.; Jarocińska, A.; Sławik, Ł. Multitemporal Hyperspectral Data
Fusion with Topographic Indices—Improving Classification of Natura 2000 Grassland Habitats. Remote Sens. 2019, 11, 2264.
[CrossRef]
96. Tuia, D.; Persello, C.; Bruzzone, L. Domain Adaptation for the Classification of Remote Sensing Data: An Overview of Recent
Advances. IEEE Geosci. Remote Sens. Mag. 2016, 4, 41–57. [CrossRef]
97. Piel, A.K.; Crunchant, A.; Knot, I.E.; Chalmers, C.; Fergus, P.; Mulero-Pázmány, M.; Wich, S.A. Noninvasive Technologies for
Primate Conservation in the 21st Century. Int. J. Primatol. 2022, 43, 133–167. [CrossRef]
98. Suir, G.; Saltus, C.; Sasser, C.; Harris, J.; Reif, M.; Diaz, R.; Giffin, G. Evaluating Drone Truthing as an Alternative to Ground Truthing:
An Example with Wetland Plant Identification; Engineer Research and Development Center (U.S.): Vicksburg, MS, USA, 2021.
99. Szantoi, Z.; Smith, S.E.; Strona, G.; Koh, L.P.; Wich, S.A.; Szantoi, Z.; Smith, S.E.; Strona, G.; Koh, L.P.; Serge, A. Mapping
Orangutan Habitat and Agricultural Areas Using Landsat OLI Imagery Augmented with Unmanned Aircraft System Aerial
Photography. Int. J. Remote Sens. 2017, 38, 2231–2245. [CrossRef]
100. Wich, S.A.; Koh, L.P. Conservation Drones: Mapping and Monitoring Biodiversity; Oxford University Press: Oxford, UK, 2018;
pp. 51–54.
101. Onishi, M.; Ise, T. Explainable Identification and Mapping of Trees Using UAV RGB Image and Deep Learning. Sci. Rep. 2021,
11, 903. [CrossRef]
102. Gigante, D.; Attorre, F.; Venanzoni, R.; Acosta, A.T.R.; Agrillo, E.; Aleffi, M.; Alessi, N.; Allegrezza, M.; Angelini, P.; Angiolini, C.; et al.
A Methodological Protocol for Annex I Habitats Monitoring: The Contribution of Vegetation Science. Plant Sociol. 2016, 53, 77–87.
[CrossRef]
Remote Sens. 2024, 16, 1224 26 of 26

103. Correll, M.D.; Hantson, W.; Hodgman, T.P.; Cline, B.B.; Elphick, C.S.; Gregory Shriver, W.; Tymkiw, E.L.; Olsen, B.J. Fine-Scale
Mapping of Coastal Plant Communities in the Northeastern USA. Wetlands 2019, 39, 17–28. [CrossRef]
104. Epifanio, I.; Ventura-Campos, N. Hippocampal Shape Analysis in Alzheimer’s Disease Using Functional Data Analysis. Stat.
Med. 2014, 33, 867–880. [CrossRef] [PubMed]
105. Ramsay, J.O.; Silverman, B.W. Applied Functional Data Analysis: Methods and Case Studies; Ramsay, J.O., Silverman, B.W., Eds.;
Springer Series in Statistics; Springer: New York, NY, USA, 2002; Volume 45, ISBN 978-0-387-95414-1.

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.