remote sensing

Article
Testing Textural Information Base on LiDAR and Hyperspectral
Data for Mapping Wetland Vegetation: A Case Study of Warta
River Mouth National Park (Poland)
Anna Jarocińska 1 , Jan Niedzielko 2 , Dominik Kopeć 2,3, * , Justyna Wylazłowska 2 , Bozhena Omelianska 1
and Jakub Charyton 2

1 Department of Geoinformatics, Cartography and Remote Sensing, Chair of Geomatics and Information
Systems, Faculty of Geography and Regional Studies, University of Warsaw, 00-927 Warsaw, Poland;
ajarocinska@uw.edu.pl (A.J.); b.omelianska@student.uw.edu.pl (B.O.)
2 MGGP Aero sp. z o.o., 33-100 Tarnów, Poland; jniedzielko@mggpaero.com (J.N.);
jwylazlowska@mggpaero.com (J.W.); jcharyton@mggpaero.com (J.C.)
3 Department of Biogeography, Paleoecology and Nature Conservation, Faculty of Biology and Environmental
Protection, University of Lodz, 90-237 Łódź, Poland
* Correspondence: dominik.kopec@biol.uni.lodz.pl

Abstract: One of the key issues in wetland monitoring is mapping vegetation. Remote sensing and
machine learning are used to acquire vegetation maps, which, due to the development of sensors
and data processing and analysis methods, have increasingly high accuracy. The objectives of this
study were to test: (i) which of the textural information (TI) features have the highest information
potential for identifying wetland communities; and (ii) whether the use of TI improves the accuracy
of wetland communities mapping using hyperspectral (HS) and Airborne Laser Scanning (ALS)
data. The analysis indicated that the mean and entropy features of the Gray Level Co-occurrence
Matrix had the highest potential to differentiate between various wetland communities. Adding these
features to the dataset resulted in a small increase (0.005) in average F1 accuracy based on HS data
and 0.011 for HS and ALS scenarios in wetland communities classification, and adding TI improved
Citation: Jarocińska, A.; Niedzielko,
the delineation of patch boundaries. A higher increase was noted for forest and scrub vegetation (by
J.; Kopeć, D.; Wylazłowska, J.;
0.019 for the HS scenario and 0.022 for the HS and ALS scenario) and rushes (only for the HS and
Omelianska, B.; Charyton, J. Testing
ALS scenario 0.017). It can be concluded that it is reasonable to use textural information for mapping
Textural Information Base on LiDAR
wetland communities, especially for areas with a high proportion of scrub and forest and rushes
and Hyperspectral Data for Mapping
Wetland Vegetation: A Case Study of
vegetation included in the analysis.
Warta River Mouth National Park
(Poland). Remote Sens. 2023, 15, 3055. Keywords: ALS; machine learning; classification; data fusion; floodplain; Permutation Importance;
https://doi.org/10.3390/rs15123055 CatBoost; HySpex; GLCM

Academic Editor: Melanie

Vanderhoof

Received: 22 April 2023 1. Introduction

Revised: 5 June 2023 Wetlands perform various important functions—for example, they regulate global
Accepted: 7 June 2023
and local climate [1], they are a location for rare species, and their proper functioning
Published: 11 June 2023
contributes to the protection of biodiversity [2]. Unfortunately, these areas are influenced
by climate change, especially the rise in temperature [3]. Any disturbances in the climate
lead to changes in hydrology and, as a consequence, changes in vegetation [1], so vegetation
Copyright: © 2023 by the authors.
monitoring can therefore detect changes in the functioning of wetlands [4].
Licensee MDPI, Basel, Switzerland. Vegetation mapping over large areas (such as national parks) with limited accessibility
This article is an open access article can be difficult using conventional methods. The use of remote sensing techniques can
distributed under the terms and be a good and recommended solution [5] because it allows for a significant reduction in
conditions of the Creative Commons field measurements and it enhances the objectivity and comparability of results [4]. In
Attribution (CC BY) license (https:// addition, the intensively developed machine learning (ML) algorithms currently allow
creativecommons.org/licenses/by/ different types of data, such as optical and ALS, to be combined and processed [6]. ML
4.0/). has high computational capabilities and adapts to the input data, which is why it can be a

Remote Sens. 2023, 15, 3055. https://doi.org/10.3390/rs15123055 https://www.mdpi.com/journal/remotesensing

Remote Sens. 2023, 15, 3055 2 of 26

good tool for processing remote sensing data. One of the newly introduced algorithms is
CatBoost, which was successfully used in the classification of remote-sensing images [7–9].
The combination of remote sensing data and ML algorithms gives good results in vegetation
classification and thus may be useful in classifying areas such as wetlands [10]. On the
other hand, the identification of diverse, natural vegetation based on remote sensing data
may not be precise enough. That is why there are studies to test the new techniques
that could improve ML results and better identify such vegetation using remote sensing
methods [4,11].
One of the main purposes of vegetation identification is its application in nature
conservation. However, the utility of this purpose is dependent on the high accuracy of
the acquired map, so it is important to test and implement methods that improve this. The
accuracy depends on many factors: the input dataset—types of data and used resolutions,
reference datasets, preprocessing procedures and classification procedure. One way to
improve the mapping results is to use relevant data as input; an example could be the
fusion of different spectral resolutions [11]. The spectral reflectance for vegetation is very
diverse, which makes identification by remote sensing techniques difficult. This problem
is particularly important in the case of diverse vegetation with high internal variability,
such as wetlands [4], so one of the solutions is to use data with a high spatial and spectral
resolution—for example, aerial hyperspectral data (HS) [12–14]. In addition, one of the
most successful data fusions used to improve mapping quality is a combination of HS
and Airborne Laser Scanning (ALS) data [15]. For Natura 2000 habitat mapping, using the
fusion of aerial HySpex images and ALS data, the F1 value for the Natura 2000 habitat
varied from 0.923 to 0.953 and was dependent on time and dataset. On the other hand,
even based on this combination of data with high resolutions, it is still not always possible
to correctly identify all habitats [10,16].
The natural vegetation is varied within the patches (has a high α-diversity), often
creating distinctive textures. On the other hand, the classifications using raster data are
mainly conducted on a pixel level, and each pixel is identified independently from its
neighbour. Based on the vegetation structure, it is worth using not only spectral features for
identification but also those that take into account the pixel’s neighbourhood. One solution
is to use textural information (TI) to estimate the internal complexity of the identified
classes, as well as differences in community patches. So far, TI has been most often used to
identify forest vegetation. Mangroves were classified based on WorldView-3 data, including
TI, with an overall accuracy (OA) of 0.94 [17]. In this case, the features of homogeneity,
contrast, entropy and correlation turned out to be significant. The Gray Level Co-occurrence
Matrix (GLCM) features were also used to classify forests, which allowed OA = 0.92 to
be achieved [18]. Based on data from the HyMap HS image, the TIs were calculated, and
tree species were classified [19]. Adding TI increased OA by 0.05 to 0.61. Adding TI also
improves the results of classification for crops [20]. However, other authors have not found
an increase in the accuracy of Phragmites australis identification when adding TI [21].
There are few studies evaluating the effectiveness of additional TIs for the identifica-
tion of non-forest natural and semi-natural communities. Based on data from WorldView-3,
TI was calculated, and this combination resulted in an OA of 0.91 for invasive weed classifi-
cation [22]. Moreover, the chosen TI features (mean, entropy, dissimilarity) calculated from
IKONOS images increased the accuracy of the classification of sub-Antarctic plant commu-
nities by 0.06 [23]. However, there are no studies that check whether the TI calculated on
HS and ALS data improved the classification results of wetland communities.
Very few studies were dedicated to the analysis of the usefulness of TI for identifying
natural and semi-natural non-forest communities, and none of these was dedicated to
wetland identification. TI brings new information, which is useful for class differentiation.
At the same time, adding more layers to the classification increases processing time. This is
especially important when using high-dimensional data such as HS. In this case, a signifi-
cant part of the analysis focuses on feature reduction. Different feature selection techniques
can be used to reduce the volume of data—for example, Permutation Importance [24] or
 At the same time, adding more layers to the classification increases processing time. This
is especially important when using high-dimensional data such as HS. In this case, a
significant part of the analysis focuses on feature reduction. Different feature selection
techniques can be used to reduce the volume of data—for example, Permutation
Remote Sens. 2023, 15, 3055 Importance [24] or Recursive Feature Elimination [25]. Therefore, adding additional layers 3 of 26

is only justified if they increase the accuracy of identification. To properly determine the
significance of layers, it is necessary to compare the mapping accuracy on the same dataset
Recursive Feature TI.
with and without Elimination [25]. Therefore, adding additional layers is only justified if
they increase
The aimthe of accuracy
this studyofwas
identification.
to test the To properly determine
applicability of selectedtheTI
significance of layers,
using Permutation
itImportance
is necessary to for
(PI) compare
mappingthe wetland
mappingcommunities
accuracy on using
the same dataset
HS and ALSwith
data.and
Thewithout
followingTI.
The aim
research of thiswere
questions study was to test the applicability of selected TI using Permutation
defined:
Importance (PI) TI
(i) Which for calculated
mapping wetland
from HS communities
and ALS has using
theHS and ALS
greatest data. The
potential forfollowing
wetland
research questions
communities mapping? were defined:
(i)
(ii)Which
Does theTITIcalculated
calculatedfrom
fromHSthe and ALS
HS, or thehas the ALS
HS and greatest
data,potential
improve for wetland
the accuracy
communities
of the wetland mapping?
communities mapping?
(ii)
(iii)Does
For the
whichTI calculated from the HS,does
plant communities or the HSuse
the andofALSTI data, improve
increase the accuracy
the classification
of the wetland
accuracy? communities mapping?
(iii)
TheFor which plant
analyses were communities
divided intodoestwothe use of TI increase
experiments: the classification
(1) identification accuracy?
of potentially
The analyses were divided into two experiments: (1) identification
influential TIs based on Permutation Importance and (2) comparison of classification of potentially
influential
results withTIs based
and withouton Permutation
TIs included in Importance and (2) comparison of classification
the input dataset.
results with and without TIs included in the input dataset.
2. Study Area and Object of Research
2. Study Area and Object of Research
The study was conducted in the Warta River Mouth National Park, located in western
The study was conducted in the Warta River Mouth National Park, located in western
Poland, in the mesoregions of the Gorzow Basin and Freienwalde Basin—a fragment of
Poland, in the mesoregions of the Gorzow Basin and Freienwalde Basin—a fragment of the
the Toruń-Eberswalde Ice Marginal Valley [26]. The survey covered the entire park and
Toruń-Eberswalde Ice Marginal Valley [26]. The survey covered the entire park and part of
part of its buffer zone—a total of 110.8 km2. Experiment 1 was carried out on a selected
2 . Experiment
its buffer zone—a total of 110.8 km 1 was carried out on a selected area of
area of 215.6 km2, representative of the park’s vegetation, while Experiment 2 was carried
15.6 km , representative of the park’s vegetation, while Experiment 2 was carried out on
out on the entire analysed area (Figure 1).
the entire analysed area (Figure 1).

Figure 1.
Figure 1. The
The study
study area,
area, Warta
Warta River
River Mouth
MouthNational
NationalPark.
Park.

The Warta River floodplains are of key importance for wetland plant communities
and, associated with them, many animal species, especially birds. For this reason, the
analysed area has been placed under legal protection, first as the Słońsk nature reserve (in
1977) and, since 2001, as the Warta River Mouth National Park. In 1984, the research area
was included in the areas of the Ramsar Convention. In 2007, the analysed area became
part of the European Natura 2000 network as PLC080001 “Ujście Warty” [27]. The largest
 The Warta River floodplains are of key importance for wetland plant communities
and, associated with them, many animal species, especially birds. For this reason, the
analysed area has been placed under legal protection, first as the Słońsk nature reserve (in
Remote Sens. 2023, 15, 3055 1977) and, since 2001, as the Warta River Mouth National Park. In 1984, the research area 4 of 26
was included in the areas of the Ramsar Convention. In 2007, the analysed area became
part of the European Natura 2000 network as PLC080001 “Ujście Warty” [27]. The largest
area
area of of the
the park
park isis covered
covered by by Holocene
Holocene peats
peats up
up toto several
severalmetresmetresthick.
thick.Small
Smallareas
areasareare
also occupied by rivers, mud and sands [28,29]. The relief and true altitude are notvery
also occupied by rivers, mud and sands [28,29]. The relief and true altitude are not very
diverse. The
diverse. The lowest
lowest point
point of of the
the analysed
analysedarea areaisisatat9.6
9.6mmmsl,msl,the thehighest
highestatat20.620.6mmmsl,msl,
and the
and the average
average altitude
altitude is is 11.7
11.7 m m msl.
msl.
The park
The park has hasaacomplex
complex hydrological
hydrological network.
network. Its axis
Its is theisWarta
axis River, which
the Warta River, flows
which
through
flows the park’s
through territory
the park’s in a straight,
territory regulated
in a straight, channel,channel,
regulated and justand outside the park’s
just outside the
borders, it flows into the Oder River. The other important river
park’s borders, it flows into the Oder River. The other important river in the park is the in the park is the slow-
flowing, meandering
slow-flowing, meandering Postomia River.River.
Postomia The Warta and theand
The Warta flood
thecontrol dike running
flood control along
dike running
its right bank divide the park into two distinctly different areas.
along its right bank divide the park into two distinctly different areas. North of the Warta North of the Warta is theis
the higher-lying Polder Północny (North Polder), cut by numerous ditches and canalsfrom
higher-lying Polder Północny (North Polder), cut by numerous ditches and canals from
which water
which water is is pumped
pumped into into the
the Warta.
Warta. TheThe water
water level
levelthere
thereisisrelatively
relativelystable.
stable.South
Southofof
theWarta
the Wartariverbed
riverbedisisa aflat,
flat, extensive
extensive floodplain
floodplain (Figure
(Figure 2). 2).
ThisThis
area area
cancan remain
remain under-
underwater
water
for partforof part of the
the year, andyear,
water andlevel
water level fluctuations
fluctuations here canhere reach canupreach up to 4per
to 4 metres metres
yearper[30].
year [30]. It should be noted that the flooding of the park’s area
It should be noted that the flooding of the park’s area is caused by backwaters from the is caused by backwaters
from River
Oder the Oder andRiver
watersandcarried
watersby carried by the
the Warta Warta
River River
from from outside
outside the park. thePrecipitation
park. Precip-in
the area is less than the national average and amounts to 500–550 mm per year,per
itation in the area is less than the national average and amounts to 500–550 mm year,
of which
of which about half falls in May–August. The winter period is
about half falls in May–August. The winter period is the poorest for precipitation, and the the poorest for precipita-
tion, and of
duration thesnow
duration
cover of does
snownot cover does 25
exceed notdays.
exceed 25 days.annual
Average Average airannual air temper-
temperatures vary
atures
from vary
7 to 9 ◦from 7 to 9 °C [30,31].
C [30,31].

Figure 2. The photographs of study area.

Figure 2. The photographs of study area.

Untilthe
Until the18th
18thcentury,
century,the theWarta
Warta flowed
flowed inin many
many channels
channels andand branches
branches in the
in the ana-
analysed
lysed area, creating a complex hydrological network in a wide, swampy
area, creating a complex hydrological network in a wide, swampy valley overgrown with valley overgrown
with riparian
riparian forests.
forests. In the In thecentury,
18th 18th century,
the area the
wasarea was drained
drained and deforested.
and deforested. LivestockLivestock
grazing
and hay harvesting began in the resulting meadows and pastures. As a resultresult
grazing and hay harvesting began in the resulting meadows and pastures. As a of
of these
these activities, most of the park is now non-forested. The vegetation
activities, most of the park is now non-forested. The vegetation is dominated by reeds, is dominated by
reeds, rushes,
rushes, meadows meadows and pastures.
and pastures. On muddy On muddy
water water
banks,banks, communities
communities of annual
of annual plants
plants develop. The original vegetation is alluded to by willow thickets
develop. The original vegetation is alluded to by willow thickets and small-area forests and small-area
forests dominated
dominated by alderbyor alder
willow.or There
willow.areThere areNatura
several several2000
Natura 2000 in
habitats habitats
the park,in the park,
including
including some of European Union priority importance: 3130, 3150,
some of European Union priority importance: 3130, 3150, 3260, 3270, 6120 (priority),3260, 3270, 6120 (pri-
ority), 6210 (priority), 6430, 6440, 6510, 9170, 91E0 (priority),
6210 (priority), 6430, 6440, 6510, 9170, 91E0 (priority), 91F0 [27]. 91F0 [27].

3. Materials
3. Materials and
and Methods
Methods
3.1. Aerial Data
3.1.
The aerial dataset
The dataset was
wasacquired
acquiredonon2222 June 2020.
June TheThe
2020. flight was was
flight conducted with with
conducted the
use use
the of a of
multisensor platform
a multisensor consisting
platform of a RIEGL
consisting of a LiDAR
RIEGL VQ-780II scanner and
LiDAR VQ-780II HySpex
scanner and
hyperspectral
HySpex scanners:scanners:
hyperspectral VNIR-1800 (400–1000(400–1000
VNIR-1800 nm spectral
nmrange) andrange)
spectral SWIR-384and(930–2500
SWIR-384
nm) [32]. The
(930–2500 nm)platform
[32]. Thewas slightlywas
platform modified
slightlybymodified
adding abysecond
addingSWIR-384
a secondscanner
SWIR-384to
scanner to increase the spatial resolution of images in the SWIR range. Table 1 describes
all the sensors used and the basic settings of the flight parameters. For sensor specifi-
cation details, see the producer’s materials: Norsk Elektro Optikk AS, Skedsmokorset,
Norway, for HySpex, and RIEGL Laser Measurement Systems GmbH, Horn, Austria, for
VQ-780II [33,34]. Flights were conducted at an altitude of 1.3 km with a minimum sun
elevation angle of 30◦ .
Remote Sens. 2023, 15, 3055 5 of 26

Table 1. Characteristics of the sensors used in the study with the main parameters of data acquisition.

Sensor Producer Spatial Resolution FOV Side Overlap

HySpex VNIR-1800 NEO 1 m GSD 34 30
HySpex SWIR-384 x2 NEO 1 m GSD 2 × 16 (~1 overlap) ~30
7.6 point/m2
VQ-780II Riegl 60 50
(>15 with overlap)

3.2. Image Preprocessing

Hyperspectral data were parametrically geocoded in PARGE software (ReSe—Remote
Sensing Applications, Wil, Switzerland) [35] using flight navigation data and a camera
sensor model. Images were orthorectified on Digital Surface Model created from ALS data.
Raw data values were converted to at-sensor radiance (W·nm−1 ·sr−1 ·m−2 ) in the software
supplied by the equipment manufacturer (Norsk Elektro Optikk AS, Skedsmokorset, Nor-
way) [33]. Images from VNIR and the two SWIR scanners were combined geometrically into
a single hyperspectral data cube, setting split wavelength at 935 nm and trimming the rows
to a spatial range of SWIR imagery. Atmospheric compensation was made using ATCOR4
software (ReSe Applications GmbH, Wil, Switzerland) [36] using flat terrain topography,
variable water vapour, visibility estimation, and rural aerosol type. Bands of wavelengths
longer than 2.35 µm were removed due to a high noise level. A Savitzky–Golay filter with
a 6-band window side was applied to polish the spectra. The images were mosaicked to
the middle of overlapping areas between the flight lines.
The ALS point cloud orientation was made using RiProcess software (RIEGL Laser
Measurement Systems GmbH, Horn, Austria) [34]. Full waveform decomposition into
a point cloud was done using RiAnalyze software (RIEGL Laser Measurement Systems
GmbH, Horn, Austria) [37]. Point cloud classification was automatically pre-classified in
TerraSolid software (Terrasolid Ltd., Helsinki, Finland) [38]. After that, the point cloud was
manually classified into ASPRS standard classes.
As a part of preprocessing, the Minimum Noise Fraction (MNF) transformation was
calculated from the hyperspectral images, and 30 first bands were chosen for further
analysis. As additional information, seven spectral indices (SI) were calculated based on the
previous experiments and the literature [25]: Anthocyanin Reflectance Index 2, Carotenoid
Reflectance Index 1, Clay Minerals Ratio, Iron Oxide Ratio, Normalized Difference Nitrogen
Index, Red Green Ratio Index, WorldView Water Index. The indices were calculated using
ENVI software (Harris Geospatial Solutions, Broomfield, CO, USA) and the “Spectral
Indices” tool [39]. The reference to spectral indices can be found in the tool’s online manual
(ENVI Spectral Indices tool). A full list of SI is available in Table A1.
A classified point cloud was processed from ALS data to raster layers in order to be
used as features in classification. The digital terrain model (DTM) was calculated in the
TerraSolid software package [40] using the linear interpolation method from the points
classified as ground. Statistical metrics calculated in a 1 m spatial pixel cell are referred to
as ALS features (ALSF). These were calculated in lidR [41]. The lidR library can calculate
multiple statistical measures (mean, percentile, standard deviation, disparity, etc.) that
describe vegetation properties—for example, height, reflectance and density. Apart from
using the standard metrics delivered by the library, a set of custom metrics was defined
and calculated for the study area. A full list of ALSFs is available in Table A2. A group
of topographic indices (TOPO) was calculated based on the terrain model using SAGA
software [42] (Table A3).

3.3. Reference Data

Together with the acquisition of aerial data, ground reference data were acquired.
The time of data acquisition (both aerial and terrestrial) was designed to capture the full
development of plant communities located in the study area. The field measurements
were carried out in the period 8 June 2020–27 August 2020. The main goal in acquiring the
Remote Sens. 2023, 15, 3055 6 of 26

ground reference data was to collect a set representing the variability of all the communities
in the study area.
A total of 1608 reference polygons were collected, including 1475 representing plant
communities and 133 polygons for other land cover elements—for example, areas without
vegetation and surface water (Table 2). Reference polygons were circles with a defined
radius and phytosociological affiliation. The radius ranged from 1 to 5 m, depending on
the type of plant community—aquatic and non-forest vegetation 1 to 3 m, shrub and forest
vegetation 3 to 5 m. Phytosociological affiliation was determined by the species composition
and plant proportions in the patch. The coordinates of the centre of the circle were recorded
in the field using a Trimble Catalyst DA1 GNSS system [43], with a measurement accuracy
of 1 m, and a mobile application for GIS data collection—Mapit [44].

Table 2. Plant communities are represented in the reference botanical ground data.

Experiment 1 Experiment 2

Ref. Class Name— Vertical Structure Ref.

Class Name Syntaxonomic Class Description
Polygons 1 Units (Plant Dominants) Polygons 1

Underwater plants
Aquatic Lemnetea and Aquatic macrophyte vegetation from
vegetation 10/232 Potametea and plants on the 90/1971
Cl. Lemnetea minoris and Cl. Potametea
water surface
Phalaridetum Marsh vegetation dominated by
arundinaceae high perennials 144/3885
73/2084 Phalaris arundinaceae
Rushes Magnocaricion Marsh vegetation from All. Magnocaricion high perennials 96/2840
Phragmition Marsh vegetation from All. Phragmition high perennials 170/4603
Isoëto- Amphibious short annual pioneer
Nanojuncetea low annuals 42/613
vegetation from Cl. Isoeto-Nanojuncetea
Annuals 16/277 Annual pioneer nitrophilous vegetation
Bidentetea low annuals 78/1867
from Cl. Bidentetea tripartiti
Pastures vegetation, periodically covered
Trifolio- with flood water and vegetation of
Agrostietalia and trodden surfaces from O. Trifolio low perennials 110/3520
Plantaginetalia fragiferae-Agrostietalia stoloniferae and
O. Plantaginetalia majoris
Meadows, Wet meadows and nitrophilous
Molinietalia low perennials 190/6398
grasslands 33/1073 perennials from O. Molinietalia caeruleae
and pastures Lowland hay meadows from
Arrhenatheretalia low perennials 54/1647
O. Arrhenatheretalia
Koelerio- Xeric sand semi-dry calcareous
Corynephoretea grasslands from Cl. Koelerio
and low perennials 107/3430
glaucae-Corynephoretea canescentis and
Festuco-Brometea Cl. Festuco-Brometea

Nitrophilous Nitrophilous perennials and shrubs from

23/392 Artemisietea and high perennials 190/4725
Epilobietea the Cl. Artemisietea vulgaris and
perennials Cl. Epilobietea angustifolii
Swamp forests and shrubs from
Salicetea purpureae shrubs and trees 53/4208
Cl. Salicetea purpureae
Ribeso Swamp forests from Ass. Ribeso
nigri-Alnetum and trees 56/4472
Alno-Ulmion nigri-Alnetum and All. Alno-Ulmion
Forests and 29/1880 Salicetum Shrub communities from Ass. Salicetum
shrubs pentandro-cinereae shrubs 18/1055
pentandro-cinereae
Vaccinio-Piceetea Pine forests from Cl. Vaccinio-Piceetea and trees 20/880
others communities with pine
Others wooded Wooded communities without
communities syntaxonomic assignment shrubs and trees 82/2683

Areas
without 10/171 Areas without Land areas without vegetation no plants 57/1591
vegetation vegetation

Surface Surface water without aquatic

water 10/825 Surface water no plants 51/3553
macrophyte vegetation
1 Number of reference polygons/Reference areas in m2 .
Remote Sens. 2023, 15, x FOR PEER REVIEW 7 of 27

Remote Sens. 2023, 15, 3055 7 of 26

After completion of the field measurements, the collected reference polygons were
combined into classesofusing
After completion syntaxonomic
the field measurements,criteria, including
the collected vertical
reference structure,
polygons wereresulting
combined
mainly fromintothe
classes
typeusing syntaxonomic
of plant dominants criteria,
(Tableincluding vertical
2) [45,46]. structure,
According resulting
to the criteria used,
mainly fromunits
individual the type
mayof plant
belong dominants (Table
to different 2) [45,46].
levels According
of the to the criteria
phytosociological used,
hierarchy. The
individual units may belong to different levels of the phytosociological hierarchy. The
proposed division of classes reflects the actual diversity of the communities of the study
proposed division of classes reflects the actual diversity of the communities of the study
area well. Reference polygons were distributed as evenly as possible, and the abundance
area well. Reference polygons were distributed as evenly as possible, and the abundance of
of each
each class
class corresponded
corresponded tofrequency
to the the frequency
of eachof plant
each community
plant community in the analysed
in the analysed area area
(Figure 3).
(Figure 3).

3.The
Figure 3.
Figure Thedistribution
distributionof of
reference polygons.
reference polygons.
3.4. Data Analysis
3.4. Data Analysis
The analyses were divided into two steps: identification of potentially influential
Thecomparison
TIs and analyses were divided into
of classification two
with andsteps: identification
without TIs includedofin
potentially influential TIs
the input dataset
and comparison of classification with and without TIs included in the input
(Figure 4). To acquire information about the utility of TI, two experiments were performed. dataset (Fig-
ure 4). To
The first oneacquire
aimed toinformation
optimize theabout the
analysis andutility of TI,Permutation
find, using two experiments werewhat
Importance, performed.
TI could
The firstbe
oneinfluential
aimed toinoptimize
wetland communities mapping.
the analysis and The second
find, using experiment
Permutation focused what
Importance,
on could
TI the determination of the
be influential differences
in wetland in classification
communities accuracyThe
mapping. between datasets
second without focused
experiment
TI and with TI, separately, for HS images only and for HS and ALS data.
on the determination of the differences in classification accuracy between datasets without
The results of both experiments were validated using reference data. The polygons
TI and with TI, separately, for HS images only and for HS and ALS data.
were divided into training and validation data using stratified random sampling. Based on
the validation data, an accuracy assessment was performed, where an F1 score for each
class was acquired based on user and producer accuracy. Then, the average F1 for all
classes (F1 macro score) was calculated, and this parameter was the basis of the analysis.
Remote Sens.
Remote Sens.2023,
2023,15,
15,x3055
FOR PEER REVIEW 8 of826of 27

Figure
Figure 4. Research scheme
4. Research schemeofofexperiments
experiments1 and
1 and
2; 2; HS—hyperspectral
HS—hyperspectral data,
data, ALS—Airborne
ALS—Airborne Laser
Laser
Scanning data, MNF—Minimum Noise Fraction, SI—spectral indices, TOPO—topographic indices,
Scanning data, MNF—Minimum Noise Fraction, SI—spectral indices, TOPO—topographic indices,
ALSF—ALS features, REF—reference polygons, TI—textural information, TIHS—influential TI fea-
ALSF—ALS features, REF—reference polygons, TI—textural information, TIHS —influential TI fea-
tures based on HS data, TIHS+ALS—influential TI features based on HS and ALS data, F1—accuracy
tures based on HS data, TIHS+ALS —influential TI features based on HS and ALS data, F1—accuracy
calculated for classes in classification process.
calculated for classes in classification process.

3.4.1.The
Theresults of both of
Determination experiments were
Influential TI validated1)using reference data. The polygons
(Experiment
were The
divided into training
first experiment wasand validation
performed on adata
part using
of the stratified
study area.random
The areasampling.
covers more Based
on
than 15.7 km (Figure 1). Additionally, image resizing reduced the amount of data andeach
the validation
2 data, an accuracy assessment was performed, where an F1 score for
class wastime.
analysis acquired based
The map on user
legend wasand producer
simplified accuracy. Then,
by combining the average
individual F1 for
classes into all clas-
higher
ses (F1 macro score)
syntaxonomic was calculated,
units (Table 2) to obtainand this parameter
a minimum was thefor
of 10 polygons basis of the analysis.
individual classes.

3.4.1. The Determination of Influential TI (Experiment 1)

The first experiment was performed on a part of the study area. The area covers more
than 15.7 km2 (Figure 1). Additionally, image resizing reduced the amount of data and
Remote Sens. 2023, 15, 3055 9 of 26

The first step was to calculate TI GLCM (Table A4) [47]. Eight different texture features
were calculated for each ALS and HS layer, with a kernel size of 3 per 3 pixels: mean,
variance, homogeneity, contrast, dissimilarity, entropy, correlation and second moment
(Table A5). Next, original HS and ALS bands were stacked with the calculated TI into one
raster file with 1521 bands: 30 MNF bands, 240 TI based on MNF; 7 SI, 56 TI calculated
using SI; 93 ALSF and 744 TI based on ALSF products; and 9 TOPO with 72 TI based on
TOPO. For more information, see Section 3.2. Image Preprocessing.
To determine the TI influential in wetland communities mapping, a Permutation
Importance (PI) analysis was conducted. The PI procedure is described as the decrease
in a model’s accuracy when one layer is randomly shuffled [48]. The drop in accuracy
shows how much the model depends on the feature. The analysis was based on a mean
decrease in the F1 macro score for each layer. The PI was performed using cross-validation
(validation performed 10 times with the stratified sampling, with 90% calibration and
10% validation polygons). As a result, the value of PI was analysed for each layer for the
10 iterations. It was assumed that if the value was above 0, the layer potentially improved
the classification accuracy. The TI was considered significant if the PI value was above 0
in at least 5 out of 10 iterations. A list of influential TI features was subsequently created:
TIHS based on HS data and TIHS+ALS based on HS and ALS data.

3.4.2. Mapping of Wetland Communities with the Use of TI (Experiment 2)

The second experiment aimed to determine whether textured features significantly
improve the classification accuracy of wetland communities mapping. The results from
experiment 1 were used to determine classification data for the whole study area. Only
products that were found to be influential in experiment 1 were calculated. To assess the
usefulness of the TI based on the accuracy for 50 classification iterations, four scenarios
were created: two for HS data and two for the combination of HS and ALS data. Two
pairs were compared: for HS and HS + TIHS scenarios—to check if the texture layers
improve classification results using only hyperspectral images, and for HS + ALS and
HS + ALS + TIHS+ALS scenarios—to analyse whether the TI calculated for HS and ALS data
improve classification results.
In the comparison CatBoost, classification was used. The classification process was
conducted using the CatBoost library machine learning algorithms [49]. CatBoost is based
on a gradient boosting technique that builds strong predictors by combining base (weaker)
predictors through a greedy iterative procedure of fitting weak predictors with gradient
descent. In the case of this library, base predictors are decision trees. The library uses
symmetric trees with the same depth applied to each leaf node of the tree. The usage of
symmetric trees reduces the overfitting of base predictors. Base predictors are iteratively
added to the ensemble model.
New trees are built to approximate the gradients (error values) of the created
model [50,51]. The target of this process is to find the optimal split points of the trees
that correspond to the minimum loss function. CatBoost uses a biased pairwise gradient
estimation technique to improve the speed of calculating gradient values, which are used
to optimize the selected loss function. The CatBoost validation was performed 50 times
on reference polygons with 50%/50% training/validation stratified random sampling on
reference polygons.
For the accuracy assessment, F1 scores were calculated for each class and as an average
from 50 iterations. The acquired results for four scenarios were analysed based on F1
values for each class and also for the mean F1 score. The mean values of F1 calculated
from 50 classification iterations were compared in pairs: for HS and HS + TIHS scenarios
and for HS + ALS and HS + ALS + TIHS+ALS scenarios. The data were tested for normal
distribution, and based on the information, the t-test was performed to determine if there
was a significant difference between the means of the two groups. The test was calculated
for two datasets: the F1 values separately for each class and the average from all classes.
 and based on the information, the t-test was performed to determine if there was a signif-
icant difference between the means of the two groups. The test was calculated for two
datasets: the F1 values separately for each class and the average from all classes.
To compare the results, prediction maps were generated for each scenario. Due to the
Remote Sens. 2023, 15, 3055 10 of 26
very large amount of data, a representative fragment of the study area was selected for
visualization (Figure 1). Each of the 50 fitted models was used to produce a map, and the
final class for each pixel was determined by majority voting. The results are presented in
mapsTo forcompare the results, prediction maps were generated for each scenario. Due to the
each scenario.
very large amount of data, a representative fragment of the study area was selected for
visualization (Figure 1). Each of the 50 fitted models was used to produce a map, and the
4. Results
final class for each pixel was determined by majority voting. The results are presented in
4.1.
mapsA Selection of TI Features Influential for Communities Identification (Experiment 1)
for each scenario.
In PI, 1112 TI were analysed: 240 were calculated from MNF, 65 from SI, 72 from
4. Results
TOPO and 744 from ALSF. The goal of experiment 1 was to determine what TI could be
4.1. A Selection of TI Features Influential for Communities Identification (Experiment 1)
considered influential in mapping wetland communities. This information was based on
In PI, 1112 TI were analysed: 240 were calculated from MNF, 65 from SI, 72 from
10 iterations of PI. The results showed that 91 TI features were determined as potentially
TOPO and 744 from ALSF. The goal of experiment 1 was to determine what TI could be
influential at least once: 34 layers based on ALSF, 18 based on TOPO, 38 based on MNF
considered influential in mapping wetland communities. This information was based on
and 1 based on
10 iterations SI (Figures
of PI. The results5 and 6). Among
showed that 91 the eight different
TI features types of TI
were determined as (mean, variance,
potentially
homogeneity, contrast,
influential at least dissimilarity,
once: 34 layers based entropy,
on ALSF, correlation and second
18 based on TOPO, moment),
38 based on MNFthe TI
mean layers computed 16 from the HS and 26 from ALS layers, which were
and 1 based on SI (Figures 5 and 6). Among the eight different types of TI (mean, variance, indicated as
influential
homogeneity, at contrast,
least once by the PIentropy,
dissimilarity, algorithm, while and
correlation the second
dissimilarity
moment),layer was
the TI the least
mean
influential: four 16
layers computed from
fromHStheand eleven
HS and fromALS
26 from ALS. Of the
layers, large
which number
were of as
indicated TIinfluential
features calcu-
at least
lated oncethe
from by ALSF,
the PI algorithm,
only 34 TIswhile the dissimilarity
calculated from the layer was the
24 ALSF least influential:
layers four
were determined as
from HS
influential.and eleven from ALS. Of the large number of TI features calculated from the ALSF,
only 34 TIs calculated from the 24 ALSF layers were determined as influential.

Figure
Figure 5. The number
numberofofinfluential
influentialTITI acquired
acquired forfor
HS HS
andand
ALS.ALS.
The The figure
figure showsshows how many
how many times times
each
each type TI was
type of TI wasindicated
indicatedatatleast
least once
once as influential
as influential (the(the PI value
PI value was above
was above 0) divided
0) divided into HSinto HS
and
andALS
ALS data.
data.

For experiment 2, it was decided to analyse features that were indicated as influential
five times or more out of 10 PI iterations. The HS and ALS data were analysed separately.
The mean layer calculated for 14 different ALS or HS layers was influential: five from ALSF,
five from TOPO and four from MNF. Moreover, only one entropy layer, calculated based
on MNF, was found to be influential in five PI iterations. As a result, in experiment 2, the
mean and entropy were calculated for HS data and the mean for ALS data.
In the case of ALS data, only mean TI was calculated for all TOPO and chosen ALSF.
In the case of ALSF, most of the layers were not influential in the classification based on
PI results. To reduce the amount of data, only 24 out of 93 ALSF (those ALSF for which
calculated TI were indicated as influential at least one time) were chosen to calculate mean
TI. From nine TOPO bands, nine mean TI were calculated. Experiment 2 resulted in four
scenarios: HS, HS + TIHS , HS + ALS and HS + ALS + TIHS+ALS (Tables 3 and A4).
Remote
RemoteSens.
Sens.2023,
2023,15,
15,x 3055
FOR PEER REVIEW 11 11
of of2726

Figure
Figure6.6.Information
Informationon
onhow
howmany
manytimes
timesout
out of
of 10
10 iterations
iterations aa TI
TI feature
feature was
was indicated
indicatedas
asinfluential
influen-
tial by the PI algorithm. Usability is defined by the PI algorithm as a mean decrease in the
by the PI algorithm. Usability is defined by the PI algorithm as a mean decrease in the F1 macro F1 macro
score
score (after a random shuffle of the feature value). For values above 0, the layer was indicated as
(after a random shuffle of the feature value). For values above 0, the layer was indicated as influential
influential in the iteration. On the figure are presented only layers that were indicated as influential
in the iteration. On the figure are presented only layers that were indicated as influential at least once.
at least once.
Remote Sens. 2023, 15, 3055 12 of 26

Table 3. The four scenarios used for the classification of the whole study area (Experiment 2).

Nb Scenario Products Layers

1 HS HS data: MNF, SI 37
HS data: MNF, SI, TI (mean and entropy) calculated
2 HS + TIHS 111
based on HS data
3 HS + ALS HS and ALS data: MNF, SI, ALSF and TOPO 139
HS and ALS data: MNF, SI, TI (mean and entropy)
4 HS + ALS + TIHS+ALS calculated based on HS data; ALSF and TOPO, and 246
mean TI calculated based on ALS data

4.2. The Accuracy of Wetland Communities Mapping Depending on the Use of TI (Experiment 2)
4.2.1. The Influence of TI on Classification Accuracy Using HS Data
The result of the second experiment was classifications performed for the entire study
area (Figure 7, Table 4).
Table 4. The average F1 values calculated based on 50 iterations of four scenarios.

HS + HS + HS + ALS +
F1 Value (Mean) HS
TIHS ALS TIHS+ALS
average F1 0.730 * 0.735 * 0.764 * 0.775 *
Lemnetea and Potametea 0.800 0.799 0.834 0.837
Phalaridetum arundinaceae 0.681 0.684 0.707 * 0.735 *
Magnocaricion 0.793 0.796 0.784 0.791
Phragmition 0.681 0.688 0.724 * 0.740 *
Isoëto-Nanojuncetea 0.346 0.353 0.381 0.393
Bidentetea 0.650 0.643 0.643 0.643
Trifolio-Agrostietalia and Plantaginetalia 0.665 0.671 0.777 0.780
Molinietalia 0.674 0.676 0.705 0.710
Arrhenatheretalia 0.462 0.436 0.459 0.446
Koelerio-Corynephoretea and Festuco-Brometea 0.787 0.786 0.803 0.812
Artemisietea and Epilobietea 0.510 * 0.523 * 0.618 0.632
Salicetea purpureae 0.867 * 0.904 * 0.894 * 0.925 *
Ribeso nigri-Alnetum and Alno-Ulmion 0.928 0.928 0.943 * 0.960 *
Salicetum pentandro-cinereae 0.843 * 0.913 * 0.885 * 0.925 *
Vaccinio-Piceetea 0.885 0.875 0.854 0.860
Others wooded communities 0.684 0.684 0.820 * 0.836 *
Areas without vegetation 0.936 * 0.921 * 0.963 0.966
Surface water 0.944 0.944 0.960 0.959
* The differences in the average are statistically significant.

The mean and distribution of F1 values for each of the 18 classes were used to assess
differences between the HS and HS + TIHS scenarios. Based on the average accuracy of
F1 for all classes, it can be concluded that the results for the HS scenario (mean F1 for
18 classes was equal to 0.730) are lower than those of HS + TIHS (F1 = 0.735), and the
difference is statistically significant based on the t-test.
For the HS dataset, the average accuracy of F1 for a single class varied from 0.346 (Isoëto-
Nanojuncetea), which is a very low value and the class was not properly classified, to 0.928
(Ribeso nigri-Alnetum and Alno-Ulmion), which indicate correct identification. For seven
classes, including five communities classes (Lemnetea and Potametea, Salicetea purpureae,
Ribeso nigri-Alnetum and Alno-Ulmion, Salicetum pentandro-cinereae, Vaccinio-Piceetea), the
F1 accuracy was above 0.800, which indicates the correct classification of these units. An
average F1 value was below 0.500 for two classes only (Isoëto-Nanojuncetea, Arrhenatheretalia).
Similar conclusions can be drawn from the results calculated based on the HS + TIHS
scenario. For the same two classes (Isoëto-Nanojuncetea, Arrhenatheretalia), the accuracies
were below 0.500; and for four classes, the F1 were above 0.800 (Salicetea purpureae, Ribeso
nigri-Alnetum and Alno-Ulmion, Salicetum pentandro-cinereae, Vaccinio-Piceetea).
Remote
RemoteSens.
Sens.2023,
2023,15,15,
3055
x FOR PEER REVIEW 13
13 ofof2627

Figure7.7.The
Figure Thedistribution
distributionofofF1F1values
valuescalculated
calculatedinin5050iterations
iterationsofofthe
theCatBoost
CatBoostalgorithm
algorithmfor
forHS
HS
andHS
and HS+TI HS scenarios.
+ TIHS scenarios.Differences
Differencesforforclasses
classesmarked
markedwith with**were
werestatistically
statistically significant.
significant.
Remote Sens. 2023, 15, 3055 14 of 26

Based on the accuracies acquired for the two scenarios, HS and HS + TIHS , it can
be concluded that the differences are not high. For nine classes, adding TI to the HS
data increased the F1 accuracy: Phalaridetum arundinaceae, Magnocaricion, Phragmition,
Isoëto-Nanojuncetea, Trifolio-Agrostietalia and Plantaginetalia, Molinietalia, Artemisietea and
Epilobietea, Salicetea purpureae and Salicetum pentandro-cinereae, but the differences were
statistically not significant and do not exceed 0.007. The F1 values for other wooded
communities and surface water were almost the same for both scenarios.
In the case of four classes, the differences in average F1 were statistically signifi-
cant: Artemisietea and Epilobietea, Salicetea purpureae, Salicetum pentandro-cinereae and areas
without vegetation. The largest differences (0.070) were noted for the Salicetum pentan-
dro-cinereae, and slightly lower (0.036) for Salicetea purpureae. For only one class—areas
without vegetation—the F1 value was higher (by 0.015) for the HS scenario compared to the
HS + TIHS .
For the forests and bushes classes (Salicetea purpureae, Ribeso nigri-Alnetum and Alno-
Ulmion, Salicetum pentandro-cinereae, Vaccinio-Piceetea and other wooded communities), the
accuracy is improved by adding TI—the average F1 for the HS scenario was equal to 0.841
and for HS + TIHS , 0.861. Moreover, better results for scenarios with TI were noted for
rushes (Phalaridetum arundinaceae, Magnocaricion, Phragmition): 0.718 for HS and 0.723 for
HS + TIHS .

4.2.2. The Influence of TI on Classification Accuracy Using HS and ALS Data

The results of the second comparison of the two scenarios, HS + ALS and
HS + ALS + TIHS+ALS , generally indicate that the accuracy for the scenario with TI is
slightly higher (Figure 8, Table 4). The mean value of F1 for all 18 classes is higher for
HS + ALS + TIHS+ALS (F1 = 0.775) compared to the HS + ALS scenario (F1 = 0.764), and this
small difference is statistically significant based on the t-test.
For the HS + ALS scenario, the average F1 for a single class varied from Isoëto-
Nanojuncetea (0.381), which indicates poor quality of identification, to 0.943 for Ribeso
nigri-Alnetum and Alno-Ulmion, and even higher for two non-vegetation classes: areas with-
out vegetation (0.963) and surface water (0.960). For nine classes, including communities
classes (Lemnetea and Potametea, Koelerio-Corynephoretea and Festuco-Brometea, Salicetea pur-
pureae, Ribeso nigri-Alnetum and Alno-Ulmion, Salicetum pentandro-cinereae, Vaccinio-Piceetea
and other wooded communities) the F1 values exceed 0.800, so these areas are properly
classified. Two classes (Isoëto-Nanojuncetea, Arrhenatheretalia) were not correctly classified,
with the average class accuracy below 0.500.
The general results for the HS + ALS + TIHS+ALS scenario are similar to HS + ALS—the
F1 values varied from 0.393 for Isoëto-Nanojuncetea to 0.960 for Ribeso nigri-Alnetum and
Alno-Ulmion and higher for areas without vegetation (0.966). The same two classes as for
HS + ALS were poorly classified (F1 less than 0.500): Isoëto-Nanojuncetea and Arrhenathere-
talia. An average F1 higher than 0.800 and good classification quality were noted for the
same classes as for HS + ALS: Lemnetea and Potametea, Koelerio-Corynephoretea and Festuco-
Brometea, Salicetea purpureae, Ribeso nigri-Alnetum and Alno-Ulmion, Salicetum pentandro-
cinereae, Vaccinio-Piceetea, other wooded communities, areas without vegetation, and
surface water.
For almost all of the classes, the F1 value was higher for the HS + ALS + TIHS+ALS
scenario than for HS + ALS, although the differences were not high. The highest differences
were noted for Salicetum pentandro-cinereae (+0.040), Salicetea purpureae (+0.030) and Phalar-
idetum arundinaceae (+0.027). Only for the Arrhenatheretalia class was the F1 accuracy for
HS + ALS higher (0.446) compared to HS + ALS + TIHS+ALS (0.459). No difference in F1
between the two scenarios was noted for Bidentetea and surface water. For six classes
(Phalaridetum arundinaceae, Phragmition, Salicetea purpureae, Ribeso nigri-Alnetum and Alno-
Ulmion, Salicetum pentandro-cinereae and other wooded communities), the differences were
statistically significant and average accuracy was higher for the HS + ALS + TIHS+ALS
Remote Sens. 2023, 15, 3055 15 of 26

compared to the HS + ALS scenario for these classes. Within these five classes, the smallest
difference was noted for Phragmition (0.015).
For the forests and bushes classes, the accuracy improved by adding TI,
and the differences were higher compared to previous scenarios: for HS + ALS and
HS + ALS + TIHS+ALS , the F1 values were equal to 0.879 and 0.901, respectively. The
same situation was noted for rushes—the average F1 value for HS + ALS + TIHS+ALS was
higher by 0.017 than for the HS + ALS scenario.

4.2.3. Influence of TI on the Effectiveness of Patches Delineation Based on Acquired

Communities Maps
A comparative analysis of the prediction maps showed clear differences between
the two types of scenarios (without TI − HS and HS + ALS and with TI – HS + TIHS
and HS + ALS + TIHS+ALS ) (Figure 9). The main difference was the weaker “salt and
pepper” effect using TI. In addition, the patches of each class were less fragmented, and
the boundaries between patches were clearer and easier to distinguish. As a result of these
differences, maps with the application of TI better represented the spatial differentiation
of plant communities occurring in the analysis area. Adding TI to both scenarios (HS and
HS + ALS) resulted in a clear reduction in the visibility of the mosaic line, which resulted
in an improvement in the course of the boundary between patches.
To compare the results for single classes, four areas of interest were analysed A, B, C,
and D (see Figure 9), and the differences between scenarios were observed. For area A in
the HS scenario, Phalaridetum arundinaceae patches were weakly isolated. Adding TI to both
scenarios (HS and HS + ALS) resulted in an improvement in the course of the boundary
between patches: for HS + TIHS, the improvement was noted for Phalaridetum arundinaceae,
whereas for HS + ALS + TIHS+ALS , the differences were also visible for the Molinietalia,
Trifolio-Agrostietalia and Plantaginetalia, and Artemisietea and Epilobietea classes.
In area B, in the HS and HS + ALS scenarios, patches of annual Bidentetea were classified
in a shallow depression, and areas of Phragmition and Artemisietea and Epilobietea were
highly fragmented and surrounded by patches of Trifolio-Agrostietalia and Plantaginetalia
pastures. Adding TI to the classifications resulted in clearly separating patches of Bidentetea
and Artemisietea, and Epilobietea from patches of Trifolio-Agrostietalia and Plantaginetalia,
which is consistent with the actual co-occurrence of patches of these plant communities.
The C area maps produced with the HS and HS + ALS scenarios were dominated by
medium-sized patches of Artemisietea and Epilobietea and highly fragmented patches of
Phragmition. Using the HS + TIHS and HS + ALS + TIHS+ALS scenarios resulted in a decrease
in the Artemisietea and Epilobietea area, in favour of rush classes (Phalaridetum arundinaceae
and Phragmition), which better correspond to the ground situation.
For area D, in both scenarios without TI, a very strong fragmentation of the patches of
each class was visible. Moreover, not very clear boundaries between patches were noted:
for the HS scenario in Molinietalia, Phalaridetum arundinaceae and Trifolio-Agrostietalia and
Plantaginetalia classes; for the HS + ALS scenario in Molinietalia and Trifolio-Agrostietalia and
Plantaginetalia. On the maps produced by the HS + TIHS and HS + ALS + TIHS+ALS scenarios,
the course of the boundary between patches of the mentioned classes was improved.
 Remote Sens. 2023, 15, x FOR PEER REVIEW 15 of 27
Remote Sens. 2023, 15, 3055 16 of 26

Figure 8. The distribution of F1 values calculated in 50 iterations of CatBoost algorithm for HS and
Figure 8. The distribution of F1 values calculated in 50 iterations of CatBoost algorithm for HS and
HS + ALS + TIHS+ALS scenarios. Differences for classes marked with * were statistically significant.
HS + ALS + TIHS+ALS scenarios. Differences for classes marked with * were statistically significant.
Remote
Remote Sens. 2023, 15,
Sens. 2023, 15, 3055
x FOR PEER REVIEW 1717of
of 27
26

Figure 9. The
Figure 9. communities map
The communities map calculated
calculated based
based on
on 50 CatBoost classifications
50 CatBoost classifications for
for four scenarios
four scenarios
performed on a part of the Park. The area is marked in Figure 1. Areas marked with A, B, C and D
performed on a part of the Park. The area is marked in Figure 1. Areas marked with A, B, C and D
are described in detail in the text.
are described in detail in the text.
To compare the results for single classes, four areas of interest were analysed A, B, C,
and D (see Figure 9), and the differences between scenarios were observed. For area A in
Remote Sens. 2023, 15, 3055 18 of 26

5. Discussion
5.1. The Utility of TI in Wetland Communities Mapping
The research results indicate that the mean TI, calculated based on HS and ALS
data, and entropy TI, calculated using HS images, are useful for differentiation
(Figures 5 and 6). Similar information was acquired in other studies. In the case of
bamboo forest classification, mean TI was defined as the most useful [52]. For tree species
identification, the mean was also one of the important TI, together with contract and dis-
similarity, or mean and correlation, dependent on the analysis [53]. In other studies, in rice
leaf blast classification, one useful TI, among others, was entropy [54]. To sum up, mean
and entropy were also useful in other studies.
Adding TI to the dataset, whether HS only or HS and ALS, improved the classification
results for most of the classes by about 0.01–0.02 based on the F1 value (Figures 7 and 8).
Such results can also be found in other studies, mainly concerning trees. For forest species
classification based on the multispectral images, overall accuracy improved by about 0.12
for one study area and 0.05 for a second one by adding textural information [55]. For the
classification of Australian woodlands, the OA was improved by 0.05 by adding TI to the
hyperspectral HyMap images [19]. The differences for forests classification are quite clear,
whereas the differences in this study for F1 values are not high (the maximum difference
was 0.07 for Salicetum pentandro-cinereae between HS and HS + TIHS ) but visible on acquired
prediction maps (Figure 9). The patch borders are more clearly recognized, and the “salt
and pepper” effect is less intensive compared to the maps without TI used as input layers.
The differences are strongly dependent on the analysed class. For the forests and
bushes classes, the accuracy is improved by adding TI: for these classes, Salicetea purpureae,
Ribeso nigri-Alnetum and Alno-Ulmion, Salicetum pentandro-cinereae, Vaccinio-Piceetea and
other wooded communities, the F1 value was higher. The average F1 for these communities
for the HS scenario was equal to 0.841 and for HS + TIHS, 0.861. The differences were even
higher for the HS + ALS and HS + ALS + TIHS+ALS scenarios—the F1 values were equal to
0.879 and 0.901, respectively. Probably the better accuracy is related to the size of individual
elements of the habitat. A single tree or bush is big enough to be highlighted by the TI
layers. In the case of rushes (Phalaridetum arundinaceae, Magnocaricion and Phragmition),
the use of TI also increased the F1 accuracy. These are communities with one dominant
species, similar in structure to arable fields. In this case, the use of TI mean, and entropy
emphasizes the internal homogeneity of the community. Similar conclusions can be drawn
for arable land [54]. The TI did not significantly change the accuracy within classes that
were very heterogeneous, where it was not possible to distinguish individual elements due
to the limited spatial resolution of the data. Moreover, individual objects (plants) in these
communities are smaller than the GSD size—smaller than 1 m2 .
There is a less visible “salt and pepper” effect on maps classified with TI. Moreover, the
borders of flight lines are less visible. This is probably due to the use of the TI mean, which
averages the pixel values. Additionally, on maps classified using TI, patch borders are better
defined. This is probably related to the better identification of individual objects within
classes. This difference in the map is not reflected in F1 accuracy because the reference
polygons were located inside the communities patches, not close to the border. In order to
be able to assess the accuracy of the entire map and the patch borders, it is necessary to
collect reference data at the borders.

5.2. Applicability of the Results

The conducted analysis was performed to identify the possibility of better vegetation
identification. In natural ecosystems, such as wetlands, the vegetation is diverse and correct
delineation can be difficult. Based on the results, it can be concluded that TI mean and
entropy are useful for the classification of forest and shrub areas as well as communities
with one dominant species. The differences in accuracy metrics are not high, but the use
of TI improves the delineation of patches. The better identification of patch boundaries is
the basis for maps to be useful in environmental protection management. So, in the case of
Remote Sens. 2023, 15, 3055 19 of 26

forests, shrubs and communities with one dominant species, it is recommended to add TI
to the dataset.
On the other hand, in the case of grassland and meadow communities, TIs with the
applied spatial resolution will probably not increase accuracy. In this case, it was not
possible to capture the high internal differentiation of the vegetation in TI.
In future studies, it is necessary to test the TI calculated using data with a higher than
HS data spatial resolution—for example, an orthophoto map. In this case, it would be
possible to recognize individual objects of meadow and grassland communities.
Based on the analyses, it can be concluded that the method of including TI mean,
and entropy in order to increase the F1 accuracy of the classification can be implemented
without prior Permutation Importance analysis. However, the study area has to be geo-
graphically comparable with similar community classes. In addition, it is also necessary to
use comparable input data.

6. Conclusions
Research conducted in the Warta River Mouth National Park (Poland) using HS and
ALS data indicates that:
1. The textural information with the highest information potential for identifying wet-
land communities are mean for HS and ALS data and entropy for HS data;
2. The addition of textural information in the dataset leads to an increase in mean F1
accuracy of 0.005 when using HS data and 0.011 when using a fusion of HS and
ALS data;
3. The resulting maps from the scenarios using TI allow for better delineation of the
patch boundaries of individual community units and eliminate the “salt and pepper”
effect and the visibility of mosaic lines. In order to analyse this change in quality in
the maps, it is necessary to have verification polygons located as close as possible to
the real patch boundary. Since there was a small proportion of polygons close to the
patch boundary in the reference dataset, the changes in accuracy measures were also
smaller than the visual differences;
4. A comparison of the classification with TI and without TI shows the greatest increase
in accuracy after the application of TI for scrub and forest communities (by 0.019 for
the scenarios with HS and 0.022 for the scenarios with HS + ALS).
In summary, it can be concluded that the use of textural information allows for more
precise mapping of wetland communities. Further research is needed to test textural
information derived from other data types, such as high-resolution RGB orthophotos.

Author Contributions: Conceptualization, A.J., D.K. and J.N.; methodology, A.J.; software, J.C.;
validation, D.K.; formal analysis, A.J., J.N. and J.C.; investigation, A.J., J.W., B.O. and D.K.; resources,
J.N.; data curation, J.N.; writing—original draft preparation, A.J., D.K., J.W., J.N., B.O. and J.C.;
writing—review and editing, A.J. and D.K.; visualization, A.J. and B.O.; supervision, D.K.; project
administration, D.K.; funding acquisition, D.K. All authors have read and agreed to the published
version of the manuscript.
Funding: Remote sensing and reference data were made as part of the project entitled: “Assessment of
the state of natural resources of the “Ujście Warty” National Park and valuable fragments of its buffer
zone using modern remote sensing techniques combined with the development of the interoperable
Park Spatial Information System”. The project is financed by the European Union—Operational
Programme Infrastructure and Environment 2014–2020 under the program 2.4.4d—assessment of the
state of natural resources in national parks using modern remote sensing techniques, competition call
POIS.02.04.00-IW.02-004D1/17. The publication was co-financed by the University of Warsaw.
Data Availability Statement: The airborne images were acquired by the MGGP Aero company and
delivered to the National Park “Ujście Warty”, which is the owner of the data. Reference polygons
were acquired during field mapping.
Acknowledgments: The authors are grateful to the National Park “Ujście Warty” for the possibility
of using airborne remote sensing data and permission to conduct field research in the park.
Remote Sens. 2023, 15, 3055 20 of 26

Conflicts of Interest: The authors declare no conflict of interest.

Appendix A

Table A1. Spectral indices calculated based on hyperspectral bands.

Index Formula Source

Anthocyanin Reflectance Index 2 ARI2 = R800(1/R550 − 1/R700) [56]
Carotenoid Reflectance Index 1 CRI1 = 1/R510 − 1/R550 [57]
Clay Minerals CM = R1650/R2215 [58]
Iron Oxide IO =R660/R485 [59]
Normalized Difference Nitrogen Index NDNI = (log(1/R1510) − log(1/R1680))/(log(1/R1510) + log(1/R1680)) [60]
Red Green Ratio Index RGRI = (∑(i = 600)ˆ699 Ri)/(∑(j = 500)ˆ599 Rj) [61]
WorldView Water Index WWI = (R427 − R950)/(R427 + R950) [62]

Table A2. The list of ALS features (ALSF).

ALSF Description Source

All returns above mean divided by (total first
ARAMean [63]
returns) × 100
All returns above mode divided by (total first
ARAMOde [63]
returns) × 100
1st decile of height 10th percentile of height values [41]
2nd decile of height 20th percentile of height values [41]
3nd decile of height 30th percentile of height values [41]
4nd decile of height 40th percentile of height values [41]
5nd decile of height 50th percentile of height values [41]
6nd decile of height 60th percentile of height values [41]
7nd decile of height 70th percentile of height values [41]
8nd decile of height 80th percentile of height values [41]
9nd decile of height 90th percentile of height values [41]
Maximum value of deviation from pulse shape in
Deviation max [41]
the grid cell
Mean value of deviation from pulse shape in the
Deviation mean [41]
grid cell
Median value of deviation from pulse shape in the
Deviation median [41]
grid cell
Minimum value of deviation from pulse shape in the
Deviation min [41]
grid cell
Range of deviation values from pulse shape in the
Deviation range [41]
grid cell
Squared mean value of deviation from pulse shape
Deviation rms [41]
in the grid cell
Variance of deviation from pulse shape in the grid
Deviation var [41]
cell
25th percentile of deviation from pulse shape in the
25th percentile of dev [41]
grid cell
75th percentile of deviation from pulse shape in the
75th percentile of dev [41]
grid cell
Largest eigenvalue of the covariance matrix of the
Largest eigen of the cov matrix [41]
points 3D position in the grid cell
Largest eigenvalue of the covariance matrix of the
Medium eigen of the cov matrix [41]
points 3D position in the grid cell
Largest eigenvalue of the covariance matrix of the
Smallest eigen of the cov matrix [41]
points 3D position in the grid cell
Fraction of first return Fraction of first return pulses intercepted by tree [64]
Mean value of number of points defined in 3D fixed
First_Echo_Ratio_Mean neighborhood divided by number of points defined [65]
in fixed 2D neighborhood in the grid cell
Remote Sens. 2023, 15, 3055 21 of 26

Table A2. Cont.

ALSF Description Source

Min value of number of points defined in 3D fixed
First_Echo_Ratio_Min neighborhood divided by number of points defined [65]
in fixed 2D neighborhood in the grid cell
Range of values of number of points defined in 3D
First_Echo_Ratio_Range fixed neighborhood divided by number of points [65]
defined in fixed 2D neighborhood in the grid cell
Root mean square of values of number of points
defined in 3D fixed neighborhood divided by
First_Echo_Ratio_Rms [65]
number of points defined in fixed 2D neighborhood
in the grid cell
Variance of values of number of points defined in 3D
First_Echo_Ratio_Var fixed neighborhood divided by number of points [65]
defined in fixed 2D neighborhood in the grid cell
Fraction of all returns Fraction of all returns classified as tree [64]
Max height above gr first
Maximum height above ground of all first returns [66]
returns
90th–25th perc 90th percentile–25th percentile of height values [67]
90th–50th perc 90th percentile–50th percentile of height values [67]
99th–25th perc 99th percentile–25th percentile of height values [67]
99th–50th perc 99th percentile–50th percentile of height values [67]
The coefficient of variation of all height points
Var coeff all height points [68]
within each pixel
Coefficient of variation percentage of heights of all
Var coeff first return % [66]
first returns relative to all returns
L-moment 1 1st L-moment of height values [63]
L-moment 2 2st L-moment of height values [63]
L-moment 3 3st L-moment of height values [63]
L-moment 4 4st L-moment of height values [63]
L-moment kurtosis L-moment kurtosis of height values [63]
L-moment skewness L-moment skewness of height values [63]
The Median Absolute Deviation from Median
Height value (HMAD) of all height points within
MADev from Median Height [68]
each pixel, where HMAD = 1.4826 × median
(|height − median height|)
Median of the absolute deviations from the overall
MADev from overall mode [63]
mode
Measure of horizontality of points based on
Horizontality eigenvalues of the covariance matrix of the points [41]
3D position in the grid cell
25th Percentile intensity 25th Percentile of intensity [67]
50th Percentile intensity 50th Percentile of intensity [67]
75th Percentile intensity 75th Percentile of intensity [67]
99th Percentile intensity 99th Percentile of intensity [67]
Kurtosis of intensity Kurtosis of Intensity [69]
Kurtosis of reflectance Kurtosis of Reflectance [69]
Maximum of intensity Maximum of Intensity [69]
Maximum of reflectance Maximum of Reflectance [69]
Mean of intensity Mean of Intensity [69]
Mean of reflectance Mean of Reflectance [69]
Median of intensity Median of Intensity [69]
Median of reflectance Median of Reflectance [69]
Minimum of intensity Minimum of Intensity [69]
Minimum of reflectance Minimum of Reflectance [69]
Percentage of intensity values for heights below the
% intens 10percentile height [41]
10th percentile of heights
Percentage of reflectance values for heights below
% reflect 10percentile height [41]
the 10th percentile of heights
Remote Sens. 2023, 15, 3055 22 of 26

Table A2. Cont.

ALSF Description Source

Percentage of intensity values for heights below the
% intens 30percentile height [41]
30th percentile of heights
Percentage of reflectance values for heights below
% reflect 30percentile height [41]
the 30th percentile of heights
Percentage of intensity values for heights below the
% intens 50percentile height [41]
50th percentile of heights
Percentage of reflectance values for heights below
% reflect 50percentile height [41]
the 50th percentile of heights
Percentage of intensity values for heights below the
% intens 70percentile height [41]
70th percentile of heights
Percentage of reflectance values for heights below
% reflect 70percentile height [41]
the 70th percentile of heights
Percentage of intensity values for heights below the
% intens 90percentile height [41]
90th percentile of heights
Percentage of reflectance values for heights below
% reflect 90percentile height [41]
the 90th percentile of heights
Interquartile range (P75–P25) of deviation from
Interquartile range of dev [41]
pulse shape in the grid cell
Interquartile range (P75–P25) of deviation from
Interquartile range of dev [41]
pulse shape in the grid cell
Range of reflectance Range of reflectance [67]
Values values [67]
St dev of intensity Standard deviation of intensity [69]
St dev of reflectance Standard deviation of reflectance [69]
Skewn of intensity Skewness of intensity [69]
Skewn of reflectance Skewness of reflectance [69]
Measure of linearity of points based on eigenvalues
Linearity of the covariance matrix of the points 3D position in [41]
the grid cell
Median absolute deviation = median (|height −
Median abs dev [64]
median height|) of tree returns Meters MAD
The total number of all the points within each pixel
Nb of points below GT that are below the specified Ground Threshold [68]
value (GT)
Nb of modes Number of Modes [70]
Standard deviation of heights for points between
St dev non ground [41]
0 and 1 m
The total number of all the points within each pixel
Nb of points above CT that are above the specified Crown Threshold [68]
value (CT)
% returns above mean Percentage all returns above mean/total all returns [63]

Table A3. The list of Topographic indices list (TOPO).

Feature Index Full Name

direct insolation Direct Insolation
duration of insolation Duration of Insolation
modified catchment area Modified Catchment Area
multi-resolution ridge top flatness Multi-resolution index of the Ridge Top Flatness
multi-resolution valley bottom flatness Multi-resolution Index of Valley Bottom Flatness
total insolation Total Insolation
topographic position index Topographic Position Index
topographic wetness index Topographic Wetness Index
diffuse insolation Diffuse Insolation
Remote Sens. 2023, 15, 3055 23 of 26

Table A4. Textural Feature formulas, where P(i)—the probability of each pixel value and the variable
µ represents the mean of P.

Feature Formula
Ng Ng
Mean ∑i=1 ∑ j=1 i × P(i, j)
Ng Ng 2
Variance ∑i=1 ∑ j=1 (i − µ) × P(i, j)
Ng Ng
Homogeneity ∑i=1 ∑ j=1 1+(i1− j 2 × P(i, j)
)
Ng Ng 2
Contrast ∑i=1 ∑ j=1 P(i, j)(i − j)
Ng Ng
Dissimilarity ∑i=1 ∑ j=1 P(i, j)|i − j|
Ng Ng
Entropy ∑i=1 ∑ j=1 P(i, j) log( P(i, j))
Ng Ng 2
Second Moment ∑i=1 ∑ j=1 { P(i, j)}
∑i ∑ j (i, j) P(i, j)−µ x µy
Correlation
σx σy

Table A5. The four scenarios layers used for classification in experiment 2.

Nb Scenario Products
1 HS HS data: 30 MNF, 7 SI
HS data: 30 MNF, 7 SI, texture features (mean and entropy)
2 HS + TIHS
calculated based on HS bands
3 HS + ALS HS and ALS data: 30 MNF, 7 SI, 93 ALSF and 9 TOPO
HS and ALS data: 30 MNF, 7 SI, mean and entropy calculated
based on HS data; 93 ALSF, 9 TOPO, 24 mean texture features
calculated based on chosen ALSF(ARAMean, ARAMOde,
deviation mean, deviation range, deviation rms, deviation var,
duration of insolation, First_Echo_Ratio_Mean,
First_Echo_Ratio_Min, First_Echo_Ratio_Rms,
4 HS + ALS + TIHS+ALS
First_Echo_Ratio_Var, var coeff all height points, var coeff
first return %, L-moment 3, L-moment 4, L-moment kurtosis,
L-moment skewness, maximum of intensity, mean of intensity,
median of intensity, % reflect 10percentile height, % reflect
30percentile height, st dev of intensity, st dev non ground)
and 9 mean texture features calculated using TOPO products

References
1. Amesbury, M.J.; Gallego-Sala, A.; Loisel, J. Peatlands as Prolific Carbon Sinks. Nat. Geosci. 2019, 12, 880–881. [CrossRef]
2. Kimmel, K.; Mander, Ü. Ecosystem Services of Peatlands: Implications for Restoration. Prog. Phys. Geogr. Earth Environ. 2010, 34,
491–514. [CrossRef]
3. Nichols, J.E.; Peteet, D.M. Rapid Expansion of Northern Peatlands and Doubled Estimate of Carbon Storage. Nat. Geosci. 2019, 12,
917–921. [CrossRef]
4. Guo, M.; Li, J.; Sheng, C.; Xu, J.; Wu, L. A Review of Wetland Remote Sensing. Sensors 2017, 17, 777. [CrossRef] [PubMed]
5. Food and Agricultire Organization of the United Nations. Peatlands Mapping and Monitoring. Food & Agriculture ORG: S.l.,
2020. Available online: https://www.fao.org/documents/card/en/c/ca8200en/ (accessed on 10 April 2023).
6. Kopeć, D.; Michalska-Hejduk, D.; Sławik, Ł.; Berezowski, T.; Borowski, M.; Rosadziński, S.; Chormański, J. Application of
Multisensoral Remote Sensing Data in the Mapping of Alkaline Fens Natura 2000 Habitat. Ecol. Indic. 2016, 70, 196–208.
[CrossRef]
7. Huang, H.; Wu, D.; Fang, L.; Zheng, X. Comparison of Multiple Machine Learning Models for Estimating the Forest Growing
Stock in Large-Scale Forests Using Multi-Source Data. Forests 2022, 13, 1471. [CrossRef]
8. Ozturk, M.Y.; Colkesen, I. Evaluation of Effectiveness of Patch Based Image Classification Technique Using High Resolution
Worldview-2 Image. In The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences; Copernicus
GmbH: Safranbolu, Turkey, 2021; Volume XLVI-4-W5-2021, pp. 417–423. [CrossRef]
9. Samat, A.; Li, E.; Du, P.; Liu, S.; Miao, Z.; Zhang, W. CatBoost for RS Image Classification With Pseudo Label Support From
Neighbor Patches-Based Clustering. IEEE Geosci. Remote Sens. Lett. 2022, 19, 8004105. [CrossRef]
10. Zhang, C. Combining Hyperspectral and Lidar Data for Vegetation Mapping in the Florida Everglades. Photogramm. Eng. Remote
Sens. 2014, 80, 733–743. [CrossRef]
Remote Sens. 2023, 15, 3055 24 of 26

11. Rapinel, S.; Hubert-Moy, L.; Clément, B. Combined Use of LiDAR Data and Multispectral Earth Observation Imagery for Wetland
Habitat Mapping. Int. J. Appl. Earth Obs. Geoinf. 2015, 37, 56–64. [CrossRef]
12. 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]
13. Jollineau, M.Y.; Howarth, P.J. Mapping an Inland Wetland Complex Using Hyperspectral Imagery. Int. J. Remote Sens. 2008, 29,
3609–3631. [CrossRef]
14. Rosso, P.H.; Ustin, S.L.; Hastings, A. Mapping Marshland Vegetation of San Francisco Bay, California, Using Hyperspectral Data.
Int. J. Remote Sens. 2005, 26, 5169–5191. [CrossRef]
15. Szporak-Wasilewska, S.; Piórkowski, H.; Ci˛eżkowski, W.; Jarzombkowski, F.; Sławik, Ł.; Kopeć, D. Mapping Alkaline Fens,
Transition Mires and Quaking Bogs Using Airborne Hyperspectral and Laser Scanning Data. Remote Sens. 2021, 13, 1504.
[CrossRef]
16. 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]
17. Wang, T.; Zhang, H.; Lin, H.; Fang, C. Textural–Spectral Feature-Based Species Classification of Mangroves in Mai Po Nature
Reserve from Worldview-3 Imagery. Remote Sens. 2016, 8, 24. [CrossRef]
18. Mohammadpour, P.; Viegas, D.X.; Viegas, C. Vegetation Mapping with Random Forest Using Sentinel 2 and GLCM Texture
Feature—A Case Study for Lousã Region, Portugal. Remote Sens. 2022, 14, 4585. [CrossRef]
19. Bunting, P.; He, W.; Zwiggelaar, R.; Lucas, R. Combining Texture and Hyperspectral Information for the Classification of Tree
Species in Australian Savanna Woodlands. In Innovations in Remote Sensing and Photogrammetry; Lecture Notes in Geoinformation
and, Cartography; Jones, S., Reinke, K., Eds.; Springer: Berlin, Heidelberg, 2009; pp. 19–26. [CrossRef]
20. Yalçın, H. Phenology Monitoring Of Agricultural Plants Using Texture Analysis. In Proceedings of the 2015 Fourth International
Conference on Agro-Geoinformatics (Agro-geoinformatics), Istanbul, Turkey, 20–24 July 2015. [CrossRef]
21. Abeysinghe, T.; Simic Milas, A.; Arend, K.; Hohman, B.; Reil, P.; Gregory, A.; Vázquez-Ortega, A. Mapping Invasive Phragmites
Australis in the Old Woman Creek Estuary Using UAV Remote Sensing and Machine Learning Classifiers. Remote Sens. 2019,
11, 1380. [CrossRef]
22. Shendryk, Y.; Rossiter-Rachor, N.A.; Setterfield, S.A.; Levick, S.R. Leveraging High-Resolution Satellite Imagery and Gradient
Boosting for Invasive Weed Mapping. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2020, 13, 4443–4450. [CrossRef]
23. Murray, H.; Lucieer, A.; Williams, R. Texture-Based Classification of Sub-Antarctic Vegetation Communities on Heard Island. Int.
J. Appl. Earth Obs. Geoinf. 2010, 12, 138–149. [CrossRef]
24. Xu, Z.; Chen, J.; Xia, J.; Du, P.; Zheng, H.; Gan, L. Multisource Earth Observation Data for Land-Cover Classification Using
Random Forest. IEEE Geosci. Remote Sens. Lett. 2018, 15, 789–793. [CrossRef]
25. Demarchi, L.; Kania, A.; Ci˛eżkowski, W.; Piórkowski, H.; Oświecimska-Piasko, Z.; Chormański, J. Recursive Feature Elimination
and Random Forest Classification of Natura 2000 Grasslands in Lowland River Valleys of Poland Based on Airborne Hyperspectral
and LiDAR Data Fusion. Remote Sens. 2020, 12, 1842. [CrossRef]
26. Solon, J.; Borzyszkowski, J.; Bidłasik, M.; Richling, A.; Badora, K.; Balon, J.; Brzezińska-Wójcik, T.; Chabudziński, Ł.;
Dobrowolski, R.; Grzegorczyk, I.; et al. Physico-Geographical Mesoregions of Poland: Verification and Adjustment of Boundaries
on the Basis of Contemporary Spatial Data. Geogr. Pol. 2018, 91, 143–170. [CrossRef]
27. Warta River Mouth National Park. Available online: https://zpppn.pl/warta-river-mouth-national-park-en/park (accessed on
17 April 2023).
28. Jaworski, M. Mapa Hydrogeologiczna Polski; PIG: Słubice, Poland, 1986.
29. Skompski, S. Mapa Geologiczna Polski, Mapa Utworów Powierzchniowych; Wyd. Geologiczne: Słubice, Poland, 1976.
30. Choiński, A.; Ławniczak, A.E.; Ptak, M. Park Narodowy “Ujście Warty”. In Wody w Parkach Narodowych Polski; Bogdanowicz, R.,
Jokiel, P., Pociask-Karteczka, J., Eds.; IGiGP Uniwersytet Jagielloński: Kraków, Poland, 2012.
31. Climate of Poland in 2021. Available online: https://www.imgw.pl/sites/default/files/2022-06/imgw-pib-klimat-polski-2021
-eng-final.pdf (accessed on 18 April 2023).
32. Sławik, Ł.; Niedzielko, J.; Kania, A.; Piórkowski, H.; Kopeć, D. Multiple Flights or Single Flight Instrument Fusion of Hyperspectral
and ALS Data? A Comparison of Their Performance for Vegetation Mapping. Remote Sens. 2019, 11, 970. [CrossRef]
33. HySpex. Available online: https://www.hyspex.com/ (accessed on 7 July 2021).
34. RiProcess Data Sheet for RIEGL Scan Data. Available online: http://www.riegl.com/uploads/tx_pxpriegldownloads/RiProcess_
Datasheet_2020-08-20_01.pdf (accessed on 17 April 2023).
35. PARGE Airborne Image Rectification. Available online: http://www.rese-apps.com/software/parge/index.html (accessed on
7 July 2021).
36. ATCOR for Airborne Remote Sensing. Available online: https://www.rese-apps.com/software/atcor-4-airborne/index.html
(accessed on 17 April 2023).
37. RiAnalyze Data Sheet for Automated Resolution of Range Ambiguities. Available online: https://www.rieglusa.com/pdf/als/
rianalyze-datasheet.pdf (accessed on 17 April 2023).
Remote Sens. 2023, 15, 3055 25 of 26

38. TerraSolid Terrascan User Guide. Available online: http://www.terrasolid.com/guides/tscan/index.html (accessed on 17

April 2023).
39. Harris Geospatial Solutions, Broomfield, CO, USA. Available online: https://www.l3harrisgeospatial.com/Software-Technology/
ENVI (accessed on 17 April 2023).
40. Terrascan, Terrasolid’s Software for LiDAR Data Processing and 3D Vector Data Creation. Available online: https://terrasolid.
com/products/terrascan/ (accessed on 17 April 2023).
41. Roussel, J.-R.; Auty, D.; Coops, N.C.; Tompalski, P.; Goodbody, T.R.H.; Meador, A.S.; Bourdon, J.-F.; de Boissieu, F.; Achim, A.
LidR: An R Package for Analysis of Airborne Laser Scanning (ALS) Data. Remote Sens. Environ. 2020, 251, 112061. [CrossRef]
42. Conrad, O.; Bechtel, B.; Bock, M.; Dietrich, H.; Fischer, E.; Gerlitz, L.; Wehberg, J.; Wichmann, V.; Böhner, J. System for Automated
Geoscientific Analyses (SAGA) v. 2.1.4. Geosci. Model Dev. 2015, 8, 1991–2007. [CrossRef]
43. Trimble Catalyst | Catalyst GNSS Systems|Trimble Geospatial. Available online: https://geospatial.trimble.com/products-and-
solutions/trimble-catalyst (accessed on 18 April 2023).
44. Mapit GIS LTD—Aplikacje na Androida w Google Play. Available online: https://play.google.com/store/apps/dev?id=921411
8068832022925&hl=pl&gl=US (accessed on 18 April 2023).
45. Kacki,
˛ Z.; Czarniecka, M.; Swacha, G. Statistical Determination of Diagnostic, Constant and Dominant Species of the Higher
Vegetation Units of Poland. Monogr. Bot. 2014, 103, 1–267. [CrossRef]
46. Matuszkiewicz, W. Przewodnik Do Oznaczania Zbiorowisk Roślinnych Polski; Wydawnictwo Naukowe PWN: Warszawa,
Poland, 2008.
47. Haralick, R.M.; Shanmugam, K.; Dinstein, I. Textural Features for Image Classification. IEEE Trans. Syst. Man Cybern. 1973,
SMC-3, 610–621. [CrossRef]
48. Breiman, L. Random Forests. Mach. Learn. 2001, 45, 5–32. [CrossRef]
49. CatBoost. Available online: https://catboost.ai/en/docs/ (accessed on 18 April 2023).
50. Dorogush, A.V.; Ershov, V.; Gulin, A. CatBoost: Gradient Boosting with Categorical Features Support. arxiv 2018. [CrossRef]
51. Prokhorenkova, L.; Gusev, G.; Vorobev, A.; Dorogush, A.V.; Gulin, A. CatBoost: Unbiased Boosting with Categorical Features.
arxiv 2019. [CrossRef]
52. Zhou, G.; Ni, Z.; Zhao, Y.; Luan, J. Identification of Bamboo Species Based on Extreme Gradient Boosting (XGBoost) Using
Zhuhai-1 Orbita Hyperspectral Remote Sensing Imagery. Sensors 2022, 22, 5434. [CrossRef]
53. Tang, J.; Liang, J.; Yang, Y.; Zhang, S.; Hou, H.; Zhu, X. Revealing the Structure and Composition of the Restored Vegetation Cover
in Semi-Arid Mine Dumps Based on LiDAR and Hyperspectral Images. Remote Sens. 2022, 14, 978. [CrossRef]
54. Feng, S.; Cao, Y.; Xu, T.; Yu, F.; Zhao, D.; Zhang, G. Rice Leaf Blast Classification Method Based on Fused Features and
One-Dimensional Deep Convolutional Neural Network. Remote Sens. 2021, 13, 3207. [CrossRef]
55. Franklin, S.E.; Hall, R.J.; Moskal, L.M.; Maudie, A.J.; Lavigne, M.B. Incorporating Texture into Classification of Forest Species
Composition from Airborne Multispectral Images. Int. J. Remote Sens. 2000, 21, 61–79. [CrossRef]
56. Gitelson, A.A.; Merzlyak, M.N.; Chivkunova, O.B. Optical Properties and Nondestructive Estimation of Anthocyanin Content in
Plant Leaves. Photochem. Photobiol. 2001, 74, 38–45. [CrossRef]
57. Gitelson, A.; Zur, Y.; Chivkunova, O.; Merzlyak, M. Assessing Carotenoid Content in Plant Leaves with Reflectance Spectroscopy.
Photochem. Photobiol. 2002, 75, 272–281. [CrossRef]
58. Drury, S.A. Image Interpretation in Geology. Geocarto Int. 1987, 2, 48. [CrossRef]
59. Segal, D. Theoretical Basis for Differentiation of Ferric-Iron Bearing Minerals, Using Landsat MSS Data. In Proceedings of
the International Symposium on Remote Sensing of Environment, 2nd Thematic Conference, Remote Sensing for Exploration
Geology, Fort Worth, TX, USA, 6–10 December 1982; pp. 949–951.
60. Fourty, T.; Baret, F.; Jacquemoud, S.; Schmuck, G.; Verdebout, J. Leaf Optical Properties with Explicit Description of Its Biochemical
Composition: Direct and Inverse Problems. Remote Sens. Environ. 1996, 56, 104–117. [CrossRef]
61. Gamon, J.A.; Surfus, J.S. Assessing Leaf Pigment Content and Activity with a Reflectometer. New Phytol. 1999, 143, 105–117.
[CrossRef]
62. Wolf, A.F. Using WorldView-2 Vis-NIR Multispectral Imagery to Support Land Mapping and Feature Extraction Using Normalized
Difference Index Ratios. In Proceedings of the SPIE 8390, Algorithms and Technologies for Multispectral, Hyperspectral, and
Ultraspectral Imagery, Baltimore, MD, USA, 14 May 2012; p. 83900. [CrossRef]
63. Botequim, B.; Fernandes, P.M.; Borges, J.G.; González-Ferreiro, E.; Guerra-Hernández, J. Improving Silvicultural Practices for
Mediterranean Forests through Fire Behaviour Modelling Using LiDAR-Derived Canopy Fuel Characteristics. Int. J. Wildland Fire
2019, 28, 823. [CrossRef]
64. Marrs, J.; Ni-Meister, W. Machine Learning Techniques for Tree Species Classification Using Co-Registered LiDAR and Hyper-
spectral Data. Remote Sens. 2019, 11, 819. [CrossRef]
65. Pfeifer, N.; Mandlburger, G.; Otepka, J.; Karel, W. OPALS—A Framework for Airborne Laser Scanning Data Analysis. Comput.
Environ. Urban Syst. 2014, 45, 125–136. [CrossRef]
66. Shen, X.; Cao, L. Tree-Species Classification in Subtropical Forests Using Airborne Hyperspectral and LiDAR Data. Remote Sens.
2017, 9, 1180. [CrossRef]
67. García, M.; Saatchi, S.; Casas, A.; Koltunov, A.; Ustin, S.L.; Ramirez, C.; Balzter, H. Extrapolating Forest Canopy Fuel Properties in
the California Rim Fire by Combining Airborne LiDAR and Landsat OLI Data. Remote Sens. 2017, 9, 394. [CrossRef]
Remote Sens. 2023, 15, 3055 26 of 26

68. Li, A.; Dhakal, S.; Glenn, N.F.; Spaete, L.P.; Shinneman, D.J.; Pilliod, D.S.; Arkle, R.S.; McIlroy, S.K. Lidar Aboveground Vegetation
Biomass Estimates in Shrublands: Prediction, Uncertainties and Application to Coarser Scales. Remote Sens. 2017, 9, 903.
[CrossRef]
69. Ørka, H.O.; Næsset, E.; Bollandsås, O.M. Classifying Species of Individual Trees by Intensity and Structure Features Derived
from Airborne Laser Scanner Data. Remote Sens. Environ. 2009, 113, 1163–1174. [CrossRef]
70. Falkowski, M.J.; Evans, J.S.; Martinuzzi, S.; Gessler, P.E.; Hudak, A.T. Characterizing Forest Succession with Lidar Data: An
Evaluation for the Inland Northwest, USA. Remote Sens. Environ. 2009, 113, 946–956. [CrossRef]

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.