ArticlePDF Available

Seaweed Habitats on the Shore: Characterization through Hyperspectral UAV Imagery and Field Sampling

Authors:

Abstract

Intertidal macroalgal habitats are major components of temperate coastal ecosystems. Their distribution was studied using field sampling and hyperspectral remote mapping on a rocky shore of Porspoder (western Brittany, France). Covers of both dominating macroalgae and the sessile fauna were characterized in situ at low tide in 24 sampling spots, according to four bathymetric levels. A zone of ca. 17,000 m2 was characterized using a drone equipped with a hyperspectral camera. Macroalgae were identified by image processing using two classification methods to assess the representativeness of spectral classes. Finally, a comparison of the remote imaging data to the field sampling data was conducted. Seven seaweed classes were distinguished by hyperspectral pictures, including five different species of Fucales. The maximum likelihood (MLC) and spectral angle mapper (SAM) were both trained using image-derived spectra. MLC was more accurate to classify the main dominating species (Overall Accuracy (OA) 95.1%) than SAM (OA 87.9%) at a site scale. However, at sampling points scale, the results depend on the bathymetric level. This study evidenced the efficiency and accuracy of hyperspectral remote sensing to evaluate the distribution of dominating intertidal seaweed species and the potential for a combined field/remote approach to assess the ecological state of macroalgal communities.
Citation: Diruit, W.; Le Bris, A.;
Bajjouk, T.; Richier, S.; Helias, M.;
Burel, T.; Lennon, M.; Guyot, A.; Ar
Gall, E. Seaweed Habitats on the
Shore: Characterization through
Hyperspectral UAV Imagery and
Field Sampling. Remote Sens. 2022,14,
3124. https://doi.org/10.3390/
rs14133124
Academic Editor: Stuart Phinn
Received: 29 April 2022
Accepted: 26 June 2022
Published: 29 June 2022
Publisher’s Note: MDPI stays neutral
with regard to jurisdictional claims in
published maps and institutional affil-
iations.
Copyright: © 2022 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
remote sensing
Article
Seaweed Habitats on the Shore: Characterization through
Hyperspectral UAV Imagery and Field Sampling
Wendy Diruit 1,* , Anthony Le Bris 2, Touria Bajjouk 3, Sophie Richier 2, Mathieu Helias 1, Thomas Burel 1,
Marc Lennon 4, Alexandre Guyot 4and Erwan Ar Gall 1
1Univ Brest, CNRS, IRD, Ifremer, LEMAR, 29280 Plouzané, France;
mathieu.helias@etudiant.univ-brest.fr (M.H.); thomas.burel@univ-brest.fr (T.B.);
erwan.argall@univ-brest.fr (E.A.G.)
2
Centre d’Etude et de Valorisation des Algues (CEVA), 22195 Pleubian, France; anthony.lebris@ceva.fr (A.L.B.);
sophie.richier@ceva.fr (S.R.)
3Ifremer, Dynamiques des Ecosystèmes Côtiers (DYNECO)/Laboratoire d’Ecologie Benthique
Côtière (LEBCO), 29280 Plouzané, France; touria.bajjouk@ifremer.fr
4Hytech-Imaging, 115 Rue Claude Chappe, 29280 Plouzané, France; marc.lennon@hytech-imaging.fr (M.L.);
alexandre.guyot@hytech-imaging.fr (A.G.)
*Correspondence: wendy.diruit@univ-brest.fr
Abstract:
Intertidal macroalgal habitats are major components of temperate coastal ecosystems. Their
distribution was studied using field sampling and hyperspectral remote mapping on a rocky shore
of Porspoder (western Brittany, France). Covers of both dominating macroalgae and the sessile
fauna were characterized in situ at low tide in 24 sampling spots, according to four bathymetric
levels. A zone of ca. 17,000 m
2
was characterized using a drone equipped with a hyperspectral
camera. Macroalgae were identified by image processing using two classification methods to assess
the representativeness of spectral classes. Finally, a comparison of the remote imaging data to the
field sampling data was conducted. Seven seaweed classes were distinguished by hyperspectral
pictures, including five different species of Fucales. The maximum likelihood (MLC) and spectral
angle mapper (SAM) were both trained using image-derived spectra. MLC was more accurate to
classify the main dominating species (Overall Accuracy (OA) 95.1%) than SAM (OA 87.9%) at a site
scale. However, at sampling points scale, the results depend on the bathymetric level. This study
evidenced the efficiency and accuracy of hyperspectral remote sensing to evaluate the distribution of
dominating intertidal seaweed species and the potential for a combined field/remote approach to
assess the ecological state of macroalgal communities.
Keywords:
seaweeds; hyperspectral; UAVs; intertidal ecology; rocky shores; supervised classification;
vegetation cover
1. Introduction
The intertidal zone hosts considerable diversity together with a great abundance of
benthic organisms [
1
,
2
] and has long been monitored as a control ecosystem in ecological
processes. Seaweeds are the major component of flora on temperate rocky shores, where
they can commonly form extensive canopies, structuring macroalgal communities compa-
rable to terrestrial forest systems in their arrangement [
3
,
4
]. Seaweed species are vertically
distributed on the shore according to several abiotic factors such as desiccation, hydrody-
namics, light and salinity, themselves largely influenced by tide oscillations [
5
,
6
]. Temperate
rocky shores are globally dominated by fucoids (i.e., large Phaeophyceae from the order
Fucales), from high to low levels of the shore and by kelps (i.e., large Phaeophyceae from
the order Laminariales sensu lato) in the lower intertidal fringe and the subtidal area [
7
].
Along the north east Atlantic coastline, up to six successive macroalgal communities may
Remote Sens. 2022,14, 3124. https://doi.org/10.3390/rs14133124 https://www.mdpi.com/journal/remotesensing
Remote Sens. 2022,14, 3124 2 of 26
be found [
8
,
9
], which can be reduced to 2–5 depending on the geographical area, the
substratum or the hydrodynamic conditions [10].
Brittany is a long-term monitored area for macroalgal diversity (approximatively
650 species [11]
) and resources (e.g., Benthic Network research program since 2005). These
characteristics are examples of a prime area to fully describe seaweed-dominated habitats
through remote sensing. Therefore, remote sensing for macroalgal covers has undergone
early development since the 1960s [1217].
Seaweed communities have been recognized as a quality element for the classifica-
tion of coastal water bodies as part of the European Water Framework Directory (WFD,
2000/60/EC; E.C., 2000 [
18
]) and several metrics based on the good ecological state of
macroalgal communities have been developed along the European coasts [
6
,
9
,
19
24
]. On
rocky shores, the occurrence and abundance of vegetation can easily be estimated visually
through the cover-abundance scale, or percentage-cover indices, without damaging the
habitat [
5
]. Even if these estimations are easy to implement, they may be time consuming
and some locations remain difficult to reach. In this context, using remote sensing imagery
for spatialization is an interesting alternative to a site-specific scale [
25
,
26
], and could help
survey shifting ecosystems [27].
Both multispectral and hyperspectral imagery are routinely used on terrestrial vegeta-
tion, for instance, to estimate crop yields [
28
31
]. By contrast with other plants, seaweeds
have a larger phylum-specific diversity of pigments, which can be discriminated by an-
alyzing spectral characteristics at different wavelengths [
32
]. Pigment diversity in algae
contributed to the early development of seaweed detection through airborne remote sens-
ing [
33
]. Later, mapping of macroalgal communities was processed using satellite imagery
(IKONOS, SPOT, Sentinel-2), with scale refining depending on the sharpness of the sen-
sors aboard [
34
37
], and promoted combined airborne/ground spectra acquisition for
macroalgal mapping. Another powerful tool to study coastal environments is the use of
free-access satellite images, which could help to produce extensive habitat mapping, in
order to observe natural variations in habitats overtime [38].
These methods enable the collection of homogeneous data over broad spatial scales but
are inaccurate when applied to heterogeneous habitats, varying at a centimeter in scale [
39
].
Such approaches are complexified in coastal areas due to tidal variations and highly mosaic
environments [
40
]. Furthermore, data acquisition is generally altered by the occurrence of a
water layer [
41
] and often disturbed by atmospheric conditions (noticeably, cloud cover and
light reflection). The development and easy access to both unmanned aerial vehicles (UAVs)
and hyperspectral sensors further promoted the remote mapping and characterization
of intertidal habitats [
42
44
]. Since the 1970s, automated methods (i.e., classification
algorithms) have been developed to classify multi-/hyper-spectral images [
45
47
]. The
present work focuses on an easy habitat classification through high spatial resolution
pictures, obtained by a UAV and by applying commonly used algorithms. To characterize
seaweed-dominated habitats, maximum likelihood (MLC) is currently the most widely
used method of supervised classifications [
48
], along with the spectral angle mapper
(SAM) [4952].
To date, there are still few studies comparing both mapping intertidal seaweed us-
ing multispectral [
38
,
42
,
53
] or hyperspectral sensors on UAVs, and an accurate spatial
resolution (less than 5 cm). Indeed, the majority of studies focus on kelp beds at lower
resolution (spatial and/or spectral) [
54
57
]. Rossiter et al. (2020) [
49
,
58
] successfully clas-
sified shores using both multispectral and hyperspectral sensors focusing on the Fucales
Ascophyllum nodosum, but did not compare sensor data with macroalgal in situ covers.
To fill these gaps existing between remote sensing and field sampling, a two-way
approach was conducted: on the one hand, in situ sampling of macroalgal communities, and
on the other hand, hyperspectral UAV imagery acquisition and automated classifications.
In that prospect, several objectives were defined:
1.
Distinguishing macroalgae from seawater, substratum and associated non-algal or-
ganisms based on classification results from hyperspectral imagery.
Remote Sens. 2022,14, 3124 3 of 26
2.
Using hyperspectral data to discriminate the main species of fucoids from green and
red macroalgae.
3. Testing the accuracy of supervised classification algorithms.
4. Comparing field and remotely estimated cover-abundance data.
The working hypothesis of this study is that the two classifications, obtained from
hyperspectral images, would yield similar results between the two, successfully differ-
entiating macroalgae and would correspond to those obtained in the field. The present
experiment was performed on a seaweed-dominated shore of western Brittany to test
the complementarity between both approaches and to study the distribution of seaweed
habitats. The aim of the study is to evaluate the correspondence between the distribu-
tion of species in macroalgal habitats obtained by both in situ sampling on the shore and
hyperspectral imagery by a UAV.
2. Materials and Methods
2.1. Studied Site and Communities
The study was performed on the coasts of north-west Brittany, on the site of Porspoder
(48
28.88
0
N/4
46.29
0
W) (Figure 1). The site is about 230 m long and 100 m wide, with
a maximal tidal range of 8.15 m and mainly exhibiting dense macroalgal canopies with
some pools, boulder fields and bedrock. The six macroalgal communities typically found
in the north-east Atlantic were vertically distributed on the shore [
8
] and were grouped
into 4 bathymetric levels for the study. These levels correspond to either single or mixed
communities named by the dominating Fucales and Laminariales: (1) Pelvetia canaliculata
plus Fucus spiralis communities, (2) Ascophyllum nodosum/Fucus vesiculosus community,
(3) Fucus serratus community and (4) Himanthalia elongata/Bifurcaria bifurcata plus Laminaria
digitata communities. These levels are referred to, respectively, as Pc-Fspi, An, Fser and
He-Ld hereafter.
Remote Sens. 2022, 14, x FOR PEER REVIEW 3 of 26
2. Using hyperspectral data to discriminate the main species of fucoids from green and
red macroalgae.
3. Testing the accuracy of supervised classification algorithms.
4. Comparing field and remotely estimated cover-abundance data.
The working hypothesis of this study is that the two classifications, obtained from
hyperspectral images, would yield similar results between the two, successfully differen-
tiating macroalgae and would correspond to those obtained in the field. The present ex-
periment was performed on a seaweed-dominated shore of western Brittany to test the
complementarity between both approaches and to study the distribution of seaweed hab-
itats. The aim of the study is to evaluate the correspondence between the distribution of
species in macroalgal habitats obtained by both in situ sampling on the shore and hyper-
spectral imagery by a UAV.
2. Materials and Methods
2.1. Studied Site and Communities
The study was performed on the coasts of north-west Brittany, on the site of Por-
spoder (48°28.88 N/4°46.29 W) (Figure 1). The site is about 230 m long and 100 m wide,
with a maximal tidal range of 8.15 m and mainly exhibiting dense macroalgal canopies
with some pools, boulder fields and bedrock. The six macroalgal communities typically
found in the north-east Atlantic were vertically distributed on the shore [8] and were
grouped into 4 bathymetric levels for the study. These levels correspond to either single
or mixed communities named by the dominating Fucales and Laminariales: (1) Pelvetia
canaliculata plus Fucus spiralis communities, (2) Ascophyllum nodosum/Fucus vesiculosus
community, (3) Fucus serratus community and (4) Himanthalia elongata/Bifurcaria bifurcata
plus Laminaria digitata communities. These levels are referred to, respectively, as Pc-Fspi,
An, Fser and He-Ld hereafter.
Figure 1.
Study site of Porspoder (Brittany, France: 48
28.88
0
N/4
46.29
0
W) showing the 24 in situ
sampling spots surveyed during the study. The color of the circles indicates the intertidal level
considered: red circles, P. canaliculataF. spiralis; yellow circles, A. nodosum; black circles, F. serratus;
blue circles, H. elongata. The dotted lines correspond to the UAV flight lines.
Remote Sens. 2022,14, 3124 4 of 26
2.2. Sampling Method
Field sampling was conducted in Spring 2021 (28 April to 1 June). A total of
24 sampling
spots (i.e., 6 for each of the 4 levels) were monitored at low tide. The sampling spots were
referenced using pictures and GPS positioning (Garmin GPS 73,
±
3 m). The sampling
protocol followed the methodology described in Burel et al. (2019b) [
59
]. A
mobile
plastic
grid structure of 1.65 m
×
1.65 m divided into 25 quadrats of
33 cm ×33 cm
was used to
delimit each sampling spot. Covers of benthic fauna, flora and bare rock were estimated
visually on the entire surface delimited by the plastic structure from 0 to 100 percent, with a
5 percent
pace. That approach, known as ‘undisturbed sampling’, describes the distribution
of the main groups of benthic organisms plus the substratum during emersion.
2.3. Remote Sensing Acquisition
Acquisitions were made by Hytech-Imaging (Plouzané, Brittany, France) using a
NEO HysPex Mjolnir V-1240 sensor (Oslo, Norway) (Table 1). The sensor was set on an
octocopter UAV based on Gryphon Dynamics X8 architecture (Figure 2), with a gStabi H16
stabilization, containing an Applanix APX15 inertial unit with an L1/L2 GPS receiver and
a GPS L1/L2 Tallysman enabling geolocation. The UAV and the central acquisition unit of
the sensor were remotely controlled by a radio link.
Table 1.
Characteristics of the hyperspectral visible near infrared (VNIR) Mjolnir_V-1240 sensor.
FOV = field of view.
Spectral
Range
Spatial
Pixels
Spectral
Resolution
Spectral
Sampling
Number of
Bands
FOV Across
Track
iFOV Across/
3Along Track Coding
0.4–1 µm 1240 4.5 nm 3 nm 200 200.27/0.27 mrad 12 bits
Remote Sens. 2022, 14, x FOR PEER REVIEW 4 of 26
Figure 1. Study site of Porspoder (Brittany, France: 48°28.88 N/4°46.29 W) showing the 24 in situ
sampling spots surveyed during the study. The color of the circles indicates the intertidal level con-
sidered: red circles, P. canaliculataF. spiralis; yellow circles, A. nodosum; black circles, F. serratus;
blue circles, H. elongata. The dotted lines correspond to the UAV flight lines.
2.2. Sampling Method
Field sampling was conducted in Spring 2021 (28 April to 1 June). A total of 24 sam-
pling spots (i.e., 6 for each of the 4 levels) were monitored at low tide. The sampling spots
were referenced using pictures and GPS positioning (Garmin GPS 73, ±3 m). The sampling
protocol followed the methodology described in Burel et al. (2019b) [59]. A mobile plastic
grid structure of 1.65 m × 1.65 m divided into 25 quadrats of 33 cm × 33 cm was used to
delimit each sampling spot. Covers of benthic fauna, flora and bare rock were estimated
visually on the entire surface delimited by the plastic structure from 0 to 100 percent, with
a 5 percent pace. That approach, known as ‘undisturbed sampling, describes the distri-
bution of the main groups of benthic organisms plus the substratum during emersion.
2.3. Remote Sensing Acquisition
Acquisitions were made by Hytech-Imaging (Plouzané, Brittany, France) using a
NEO HysPex Mjolnir V-1240 sensor (Oslo, Norway) (Table 1). The sensor was set on an
octocopter UAV based on Gryphon Dynamics X8 architecture (Figure 2), with a gStabi
H16 stabilization, containing an Applanix APX15 inertial unit with an L1/L2 GPS receiver
and a GPS L1/L2 Tallysman enabling geolocation. The UAV and the central acquisition
unit of the sensor were remotely controlled by a radio link.
Table 1. Characteristics of the hyperspectral visible near infrared (VNIR) Mjolnir_V-1240 sensor.
FOV = field of view.
Spectral
Range
Spatial
Pixels
Spectral
Resolution
Spectral
Sampling
Number of
Bands
FOV Across
Track
iFOV Across/
Along Track Coding
0.41 µm 1240 4.5 nm 3 nm 200 20° 0.27/0.27 mrad 12 bits
Figure 2. UAV octocopter used for the acquisitions.
Acquisitions were performed on the 24 June 2021 at a 64 m height to obtain a resolu-
tion of 2 cm (Table 2). To perform the image acquisition, two technicians were involved
to pilot the UAV and to operate the hyperspectral sensor. The acquisition lasted about 30
Figure 2. UAV octocopter used for the acquisitions.
Acquisitions were performed on the 24 June 2021 at a 64 m height to obtain a resolution
of 2 cm (Table 2). To perform the image acquisition, two technicians were involved to pilot
the UAV and to operate the hyperspectral sensor. The acquisition lasted about 30 min. The
flight plan was designed to cover a subsection of the site of Porspoder, including all of the
field sampling spots (Figure 1).
Remote Sens. 2022,14, 3124 5 of 26
Table 2. Parameters of the aerial survey.
Flight
Altitude
Ground Sampling
Distance Swath Mapped
Area
Viewing
Angle
Flight
Lines
64 m 2 cm 23 m 1.76 ha 204
Images (Figure 3) were collected between 09h28 and 09h47 UTC at low tide (tidal
coefficient 92 corresponding to tidal range of 6.1 m). During the acquisitions, light was
diffused due to cloud cover.
Remote Sens. 2022, 14, x FOR PEER REVIEW 5 of 26
min. The flight plan was designed to cover a subsection of the site of Porspoder, including
all of the field sampling spots (Figure 1).
Table 2. Parameters of the aerial survey.
Flight
Altitude
Ground Sampling
Distance Swath Mapped
Area
Viewing
Angle
Flight
Lines
64 m 2 cm 23 m 1.76 ha 20° 4
Images (Figure 3) were collected between 09h28 and 09h47 UTC at low tide (tidal
coefficient 92 corresponding to tidal range of 6.1 m). During the acquisitions, light was
diffused due to cloud cover.
Figure 3. Porspoder orthophoto (RGB) obtained during the flight on the 24 June 2021. Detailed sec-
tions of the color image illustrating the different bathymetric levels on the shore are represented:
Pc-Fspi (red square), An (yellow square), Fser (black square) and He-Ld (blue square).
2.4. Pre-Processing
To obtain a georeferenced image in spectral radiance (W·m−2·sr−1·µm−1), the hyper-
spectral image was processed from raw data (level 0) to a radiometrically and geometri-
cally calibrated image (level 1c) using the HYPIP (HYPperspectral Image Preprocessing)
Figure 3.
Porspoder orthophoto (RGB) obtained during the flight on the 24 June 2021. Detailed
sections of the color image illustrating the different bathymetric levels on the shore are represented:
Pc-Fspi (red square), An (yellow square), Fser (black square) and He-Ld (blue square).
2.4. Pre-Processing
To obtain a georeferenced image in spectral radiance (W
·
m
2·
sr
1·µ
m
1
), the hyper-
spectral image was processed from raw data (level 0) to a radiometrically and geometrically
calibrated image (level 1c) using the HYPIP (HYPperspectral Image Preprocessing) chain
of Hytech-imaging that includes ATCOR/PARGE software applications (ReSe Applica-
tions, Wil, Switzerland). To calculate the surface reflectance, atmospheric corrections were
Remote Sens. 2022,14, 3124 6 of 26
performed in a two-step process: first, using the ATCOR-4 software, and then empirically
adjusting each spectrum. To adjust each spectrum, coefficients of gain and bias were cal-
culated per spectral band, by linear regression between surface reflectance data and the
reflectance signature. This reflectance signature was obtained by positioning pre-calibrated
targets (tarps) near the area of interest overflown during the survey.
2.5. Data Classification
For this study, supervised classifications were performed, where categories (classes)
correspond to spectral signatures defined by the user. A class contains a characteristic
spectral signature for each dominating fucoid species, macroalgal group or an abiotic
component and corresponds to homogeneous regions delineated on the UAV image. The
software then assigns each pixel of the image into a cover type to which its signature is
most comparable [
60
]. The supervised classifications were performed after defining regions
of interests (ROIs) which are training data. ROIs were created for each class using ‘ROI
tool’ in ENVI version 5.6.1 (Exelis Visual Information Solutions, Boulder, CO, USA) by
manually circling pixel areas on the image. More than one training ROI were usually
used to represent a particular class (ROIs = multiple polygons) (Table 3). The number of
polygons and pixels per class depend on the surface occupied by each species. For example,
covers of P. canaliculata and F. spiralis are low compared to those of F. serratus or H. elongata,
which represent more homogeneous and larger classes. Classes were selected in agreement
with the hyperspectral image and pictures taken during field sampling.
Table 3. Number of ROIs and pixels for each class.
Class Number of ROIs Number of Pixels
P. canaliculata 76 29,899
F. spiralis 10 551
A. nodosum 233 334,002
F. serratus 227 894,910
H. elongata 145 353,825
Green 482 73,592
Red 509 41,808
Substratum 408 1,834,496
Water 235 1,073,044
Nine classes were thus defined for the site of Porspoder (Figure 4), including five
classes of dominating Fucales (‘Pelvetia canaliculata’, Fucus spiralis’, Ascophyllum nodosum’,
Fucus serratus and Himanthalia elongata’), and two classes related to green and red seaweeds
(respectively, ‘Green’ and ‘Red’) were created. A ‘Substratum’ class was defined grouping
bedrock, boulders, gravel and sand, and, finally, a ’Water’ class was also created, gathering
immerged parts of the shore (pools or subtidal zone). The ‘Substratum’ and ‘Water’ classes
were classified in the same way as the other classes and have subsequently been removed
from the maps to improve their clarity and interpretation. Due to the complexity of
accurately identifying benthic fauna on the UAV image, no appropriate class was created,
and these data were grouped together as the ‘Substratum’ class.
Training data (i.e., ROIs mean spectra) were checked for class separability using the
Jeffries–Matusita distance [
61
]. The values of the resulting output between each pair of
classes ranged between 0 and 2, with values greater than 1.9 indicating almost perfect
separability between them [
46
]. A large class separability indicates that accurate training
areas have been selected, whereas values approaching zero suggest either the need for
more training areas or classes that are inherently similar in their spectral properties.
Remote Sens. 2022,14, 3124 7 of 26
Remote Sens. 2022, 14, x FOR PEER REVIEW 7 of 26
Figure 4. Hierarchical tree of decision to make classes, inspired from Congalton et al. (1999) [62].
Two supervised classification methods were performed to test the representativeness
of the spectral classes running the software ENVI version 5.6.1 (Exelis Visual Information
Solutions, Boulder, CO, USA), i.e., the algorithms maximum likelihood classification
(MLC) and spectral angle mapper (SAM).
MLC calculates the probability that an individual pixel belongs to a specific class and
is based on an estimated probability density function derived from the defined reference
classes [63]. MLC is a popular classifier [64]. The use of spectral profiles by this method
requires ROIs based on multiple pixels. Following this method, the classification is based
on the selection of the most representative spectral profiles in ROIs of the same class upon
different flight lines. The MLC classifier assumes a Gaussian distribution for each input
training class [65] and it can be expressed by the following equation:
(
)
=
ln
(
)
1
2
ln
|
|
1
2
(
)
(
)
(1)
where is a given spectral class, equals n-dimensional data, () is the probability
that class occurs in the image and it is assumed the same for all classes, ||
is the
determinant of the covariance matrix of the data in class ,
is the inverse matrix and
is the mean vector. The advantage of MLC as a parametric classifier is that it considers
the variance-covariance within the class distributions and, for normally distributed data,
MLC performs better than the other known parametric classifiers [66]. However, for data
with a non-normal distribution, the results may be unsatisfactory.
SAM identifies the spectral similarity between two spectra collected from an image
or distributed from a spectral library [67]. The resulting classification is rather based on
the angular orientations of spectral vectors [29]. Similarities within pairs of spectra (refer-
ence and classification) can be compared regardless of differences in brightness, and the
pairs are treated as vectors in an n-dimensional space [68]. SAM is expressed by the fol-
lowing equation, taken from Kruse et al. 1993 [67]:
α =

󰇩
(
)
(
)
󰇪
(2)
where t is the spectra for a pixel, r is for the reference spectrum pixel, α is the spectral
angle between t an r (measured in radians or degrees) and n is the number of bands.
The average spectral reflectance curves from the ROIs were extracted since SAM re-
quires endmember spectra. If two ROIs were identical, they were averaged in order to
obtain one curve with the maximum possible data. The use of spectra derived directly
from the image is usually better than using ground or library spectra due to better
Figure 4. Hierarchical tree of decision to make classes, inspired from Congalton et al. (1999) [62].
Two supervised classification methods were performed to test the representativeness
of the spectral classes running the software ENVI version 5.6.1 (Exelis Visual Information
Solutions, Boulder, CO, USA), i.e., the algorithms maximum likelihood classification (MLC)
and spectral angle mapper (SAM).
MLC calculates the probability that an individual pixel belongs to a specific class and
is based on an estimated probability density function derived from the defined reference
classes [
63
]. MLC is a popular classifier [
64
]. The use of spectral profiles by this method
requires ROIs based on multiple pixels. Following this method, the classification is based
on the selection of the most representative spectral profiles in ROIs of the same class upon
different flight lines. The MLC classifier assumes a Gaussian distribution for each input
training class [65] and it can be expressed by the following equation:
gi(x) = ln p(ωi)1
2ln |
i
| 1
2(xmi)t
1
i
(xmi)(1)
where
i
is a given spectral class,
x
equals n-dimensional data,
p
(
ωi
) is the probability
that class
ωi
occurs in the image and it is assumed the same for all classes,
|
i
|
is the
determinant of the covariance matrix of the data in class
ωi
,
1
i
is the inverse matrix and
mi
is the mean vector. The advantage of MLC as a parametric classifier is that it considers the
variance-covariance within the class distributions and, for normally distributed data, MLC
performs better than the other known parametric classifiers [
66
]. However, for data with a
non-normal distribution, the results may be unsatisfactory.
SAM identifies the spectral similarity between two spectra collected from an image or
distributed from a spectral library [
67
]. The resulting classification is rather based on the
angular orientations of spectral vectors [
29
]. Similarities within pairs of spectra (reference
and classification) can be compared regardless of differences in brightness, and the pairs
are treated as vectors in an n-dimensional space [
68
]. SAM is expressed by the following
equation, taken from Kruse et al. 1993 [67]:
α=cos1
nb
i=1tiri
(nb
i=1tiri)
1
2(nb
i=1tiri)
1
2
(2)
where tis the spectra for a pixel, ris for the reference spectrum pixel,
α
is the spectral angle
between tan r(measured in radians or degrees) and nis the number of bands.
The average spectral reflectance curves from the ROIs were extracted since SAM
requires endmember spectra. If two ROIs were identical, they were averaged in order to
obtain one curve with the maximum possible data. The use of spectra derived directly from
Remote Sens. 2022,14, 3124 8 of 26
the image is usually better than using ground or library spectra due to better inclusions of
errors related to atmospheric corrections, calibration and effects of sensor responses [29].
For both classifications (SAM and MLC), no detection threshold was selected, so that
all pixels could be classified.
2.6. Data Analysis
Accuracy assessment for classification was checked using ground truth (or reference)
ROIs based on the same method as the training data [
63
,
69
]. These polygons were indepen-
dent of the training ROIs and their number represented one third of training ROIs. The
accuracy assessment tool was used to create the confusion matrix and derive quantitative
measures of accuracy (i.e., kappa coefficient, overall accuracy, user/producer accuracy,
errors of commission/omission) using ENVI version 5.6.1 (Exelis Visual Information So-
lutions, Boulder, CO, USA). User accuracy is the probability of correct class assignment,
calculated by dividing the number of correctly classified pixels by the total number of pixels
in the class, and producer accuracy is the correctly classified reference pixels, calculated by
dividing the number of correctly classified pixels by the total number of pixels that should
be in a class.
Each grid structure was replaced using ‘Advanced Digitizing toolbar’ (‘Move Feature’
and ‘Rotate Feature’ options) on Qgis. Corresponding polygons were accurately positioned
using pictures taken during the field sampling, in order to decrease the potential GPS error
and to compare the exact same position.
To compare in situ data and classification data, vectors of the grid structure were
replaced on the Porspoder image using the ‘Vector to ROI’ tool, and the percentage of pixels
for each class in ROIs was extracted with the ROI statistics tool on ENVI.
Statistical analyses were conducted using the R environment [
70
]. Normality and
homoscedasticity were first tested on each biological and classification variable, corre-
sponding to seaweed species and substratum covers, with Shapiro–Wilk and F Test/Levene
tests, respectively. These tests then determined what analyses were the most suitable
(parametric or not). In order to represent the distribution of the replicates described by
the three approaches, a distance-based redundancy analysis (db-RDA) was constructed,
based on the method described by Escobar-Briones et al. (2008) [
71
]. Values for each class
(apart from ‘Water’) were first converted into a distance matrix by calculating the Hellinger
distance for each class in the whole dataset. Then, a principal coordinates analysis (PCoA)
was performed on this matrix. The PCoA allows to convert the distance between items
(distance matrix) into a map-based visualization (each item is assigned a location in a
low-dimensional space, materialized by its eigenvector) in order to better understand the
relation between each object. All the PCoA eigenvectors were used as input into an RDA
in order to build the db-RDA. The db-RDA represents on a single plot the position of the
different replicates using the PCoA eigenvalues, as well as the species (classes) and the
explanatory variables (level and method).
To compare more precisely the cover of each of the classes for the three methods and
for each bathymetric level, Kruskal–Wallis tests (non-parametric) were performed followed
by a post-hoc Dunn test to identify variables that were statistically different.
3. Results
3.1. In Situ Vegetation Cover
Large discrepancies were observed in the covers between bathymetric levels, the
lowest levels being characterized by a dominance of seaweeds, whereas bare rock showed a
large occurrence in the upper level. Covers of macroalgal classes differed between the four
levels (Figure 5). The Pc-Fspi level, corresponding to the upper shore (5.2–6.1 m above chart
datum (CD)), was lightly vegetalized, with bare rock occupying 55% of the surface. The
level was dominated by the Fucales P. canaliculata for about 32.5%. The remaining covers
were well distributed between F. spiralis and red seaweeds (5% each), whereas benthic
fauna (barnacles and limpets) corresponded to a cover of 2.5%.
Remote Sens. 2022,14, 3124 9 of 26
Remote Sens. 2022, 14, x FOR PEER REVIEW 9 of 26
covers were well distributed between F. spiralis and red seaweeds (5% each), whereas ben-
thic fauna (barnacles and limpets) corresponded to a cover of 2.5%.
Figure 5. Average covers of macroalgal groups and sessile fauna, and percentage of bare rock ob-
served in situ at each bathymetric level. Covers are given in percentages. Fucoids and other brown
species are grouped in the Brown class, and erect and crustose red algae are grouped in the Red
class.
The An level (middle shore, 3.44.4 m above CD) was largely dominated by the Fu-
cales A. nodosum (60%) and F. serratus (5.9%). Red seaweeds then covered about 25% of
the surface (22.5% erect and 2.5% crustose). Bare rock and limpets completed the remain-
ing surface (6.7% and 2.5%, respectively).
In the Fser level (lower shore, 3.12.3 m above CD), macroalgal covers became con-
spicuously dominant compared to bare rock and sessile fauna. Indeed, the cover of F. ser-
ratus was close to 100% (94.2%), while the rest corresponded to erect red algae (5.8%).
In the He-Ld level (2.81.6 m ab. CD), the distribution between macroalgal groups
was equilibrated, with a co-dominance of H. elongata (39.2%) and erect red seaweeds
(36.7%). The Laminariales L. digitata also presented large covers (17.5%), and in addition,
there were little covers of crustose red and green seaweeds (2.5% and 4.2%, respectively).
Thus, An, Fser and He-Ld had a higher cover of Phaeophyceae (more than one half)
compared to the other macroalgal groups of species (65.8%, 94.2% and 56.7% of cover,
respectively). By contrast, Pc-Fspi showed only a bit more than one third of cover by Phae-
ophyceae (37.5%).
3.2. Classifications Results
3.2.1. MLC Results
The results from the class separability test of image-derived spectra showed that all
of the class pairs had values greater than 1.90, indicating globally a good class separation
(Figure 6) [72].
Figure 5.
Average covers of macroalgal groups and sessile fauna, and percentage of bare rock
observed in situ at each bathymetric level. Covers are given in percentages. Fucoids and other
brown species are grouped in the ‘Brown’ class, and erect and crustose red algae are grouped in
the ‘Red’ class.
The An level (middle shore, 3.4–4.4 m above CD) was largely dominated by the
Fucales A. nodosum (60%) and F. serratus (5.9%). Red seaweeds then covered about 25% of
the surface (22.5% erect and 2.5% crustose). Bare rock and limpets completed the remaining
surface (6.7% and 2.5%, respectively).
In the Fser level (lower shore, 3.1–2.3 m above CD), macroalgal covers became conspic-
uously dominant compared to bare rock and sessile fauna. Indeed, the cover of F. serratus
was close to 100% (94.2%), while the rest corresponded to erect red algae (5.8%).
In the He-Ld level (2.8–1.6 m ab. CD), the distribution between macroalgal groups was
equilibrated, with a co-dominance of H. elongata (39.2%) and erect red seaweeds (36.7%).
The Laminariales L. digitata also presented large covers (17.5%), and in addition, there were
little covers of crustose red and green seaweeds (2.5% and 4.2%, respectively).
Thus, An, Fser and He-Ld had a higher cover of Phaeophyceae (more than one half)
compared to the other macroalgal groups of species (65.8%, 94.2% and 56.7% of cover,
respectively). By contrast, Pc-Fspi showed only a bit more than one third of cover by
Phaeophyceae (37.5%).
3.2. Classifications Results
3.2.1. MLC Results
The results from the class separability test of image-derived spectra showed that all
of the class pairs had values greater than 1.90, indicating globally a good class separation
(Figure 6) [72].
Remote Sens. 2022,14, 3124 10 of 26
Remote Sens. 2022, 14, x FOR PEER REVIEW 10 of 26
(a) (b)
Figure 6. Average reflectance of the different spectral classes between 450 and 900 nm: (a) macroal-
gal groups, water and substratum mean reflectance; (b) detailed spectra of each Fucales.
The MLC classifier, trained using image-derived spectra, revealed a dense cover of
intertidal Fucales (26.9% of the site) (Figure 7). The overall classification accuracy for the
MLC was 95.1% and the kappa coefficient was 0.93. The four bathymetric/vegetation lev-
els appeared clearly, forming four distinctive bands. The Pc-Fspi level (upper shore) was
dominated by a thin band of both P. canaliculata and F. spiralis (1.3% and 0.1% of total
pixels, respectively). The An and Fser levels (mid-shore) were dominated by a large band
of A. nodosum and of F. serratus (5.6% and 6.8% of total pixels, respectively). The He-Ld
level (lower shore) was characterized by the important development of H. elongata (9.7%
of total pixels) and a cover of red macroalgae greater than in higher levels (1.7% for all of
the site). Green algae were mainly present in the lower shore (1.7% of total pixels). The
‘Substratumand ‘Waterclasses represented the majority of the site (51.8% and 21.3% of
total pixels for the site, respectively). The macroalgal classes A. nodosum’, F. serratusand
H. elongatashowed the highest producer/user accuracies (Table 4). There were some mis-
classifications between the Fucales A. nodosumand F. serratus (1.56%), and between A.
nodosumand P. canaliculata(7.04%). The lowest producer/user accuracy was for F. spi-
ralis,with some misclassifications between F. spiralis and P. canaliculata(44.13%) and
between F. spiralisand A. nodosum(14.81%). ‘Greenand ‘Redalgae were also well clas-
sified (96.92% and 90.54%, respectively) but there were some misclassifications between
‘Redalgae and the Fucales P. canaliculata(3.04%) and H. elongata(2.06%).
Figure 6.
Average reflectance of the different spectral classes between 450 and 900 nm: (
a
) macroalgal
groups, water and substratum mean reflectance; (b) detailed spectra of each Fucales.
The MLC classifier, trained using image-derived spectra, revealed a dense cover of
intertidal Fucales (26.9% of the site) (Figure 7). The overall classification accuracy for the
MLC was 95.1% and the kappa coefficient was 0.93. The four bathymetric/vegetation
levels appeared clearly, forming four distinctive bands. The Pc-Fspi level (upper shore)
was dominated by a thin band of both P. canaliculata and F. spiralis (1.3% and 0.1% of total
pixels, respectively). The An and Fser levels (mid-shore) were dominated by a large band of
A. nodosum and of F. serratus (5.6% and 6.8% of total pixels, respectively). The He-Ld level
(lower shore) was characterized by the important development of H. elongata (9.7% of total
pixels) and a cover of red macroalgae greater than in higher levels (1.7% for all of the site).
Green algae were mainly present in the lower shore (1.7% of total pixels). The ‘Substratum’
and ‘Water classes represented the majority of the site (51.8% and 21.3% of total pixels
for the site, respectively). The macroalgal classes
A. nodosum
, F. serratus and H. elongata
showed the highest producer/user accuracies (Table 4). There were some misclassifications
between the Fucales A. nodosum and F. serratus (1.56%), and between A. nodosum and
P. canaliculata (7.04%). The lowest producer/user accuracy was for F. spiralis, with some
misclassifications between F. spiralis and P. canaliculata (44.13%) and between F. spiralis
and A. nodosum (14.81%). ‘Green’ and ‘Red’ algae were also well classified (96.92% and
90.54%, respectively) but there were some misclassifications between ‘Red’ algae and the
Fucales P. canaliculata (3.04%) and H. elongata (2.06%).
Remote Sens. 2022,14, 3124 11 of 26
Remote Sens. 2022, 14, x FOR PEER REVIEW 11 of 26
Figure 7. Maximum likelihood classification (MLC), trained using image-derived spectra resulting
from the hyperspectral UAV survey at Porspoder. Seven macroalgal cover classes are displayed
over the UAV RGB imagery and an orthophotography (Mégalis Bretagne et collectivités territoriales
bretonnes2015), giving an overview of the site. The Substratum and Water classes are not rep-
resented on the map. Class codes Green and Red represent grouped green and red macroalgal
species, respectively.
Table 4. Maximum likelihood classification (MLC) confusion matrix, calculated, using ENVI 5.6.1,
by comparing pixels of known class locations to those predicted by the classification workflow for
each of the nine cover classes. Results are displayed as percentages of pixels assigned, correctly or
incorrectly, to each class. User/producer accuracies (User Acc. and Prod. Acc., respectively) are also
presented.
Class P. canaliculata
F. spiralis
A. nodosum
F. serratus
H. elongata
Green
Red
Substratum
Water
Total
User Acc.
Unclassified 0 0 0 0 0 0 0 0 0 0 -
P. canaliculata
97.82 44.13 7.04 0.01 0.02 0 3.04
0.90 0.03 1.28
26.04
F. spiralis
0.13 39.00 0.23 0 0.07 0 0.85
0.09 0.24 0.14
6.23
A. nodosum
0.79 14.81 89.32 5.59 0.03 0.27 0.78
0.04 0.02 5.63
91.93
Figure 7.
Maximum likelihood classification (MLC), trained using image-derived spectra resulting
from the hyperspectral UAV survey at Porspoder. Seven macroalgal cover classes are displayed
over the UAV RGB imagery and an orthophotography (Mégalis Bretagne et collectivités territoriales
bretonnes—2015), giving an overview of the site. The ‘Substratum’ and ‘Water classes are not
represented on the map. Class codes ‘Green’ and ‘Red’ represent grouped green and red macroalgal
species, respectively.
Remote Sens. 2022,14, 3124 12 of 26
Table 4.
Maximum likelihood classification (MLC) confusion matrix, calculated, using ENVI 5.6.1,
by comparing pixels of known class locations to those predicted by the classification workflow for
each of the nine cover classes. Results are displayed as percentages of pixels assigned, correctly or
incorrectly, to each class. User/producer accuracies (User Acc. and Prod. Acc., respectively) are
also presented.
Class P. canaliculata F. spiralis A. nodosum F. serratus H. elongata Green Red Substratum Water Total User Acc.
Unclassified 0 0 0 0 0 0 0 0 0 0 -
P. canaliculata 97.82 44.13 7.04 0.01 0.02 0 3.04 0.90 0.03 1.28 26.04
F. spiralis 0.13 39.00 0.23 0 0.07 0 0.85 0.09 0.24 0.14 6.23
A. nodosum 0.79 14.81 89.32 5.59 0.03 0.27 0.78 0.04 0.02 5.63 91.93
F. serratus 0.01 0 1.56 91.71 0.03 0.01 0.04 0 0 6.80 98.61
H. elongata 0.09 0.15 0.08 0.38 93.53 0.18 2.60 0 4.00 9.69 90.72
Green 0.17 0.15 0.73 0.35 0.07 96.92 0.70 0.11 0.21 1.66 88.85
Red 0.01 1.76 0.75 1.78 3.14 1.64 90.54 0.02 0.22 1.68 67.03
Substratum 0.67 0 0.14 0.03 0 0.22 0.25 96.59 0.15 51.78 99.90
Water 0.30 0 0.15 0.15 3.10 0.75 1.20 2.25 95.15 21.34 92.76
Total 100 100 100 100 100 100 100 100 100 100 -
Prod. Acc. 97.82 39.00 89.32 91.71 93.53 96.92 90.54 96.59 95.15 - -
3.2.2. SAM Results
The SAM classifier, trained using image-derived spectra, revealed a similar cover of
intertidal Fucales as MLC (27.6% of the site) (Figure 8). The overall classification accuracy
for the SAM was 87.9% and the kappa coefficient was 0.82. By contrast with MLC, the four
bathymetric levels appeared less distinct. The Pc-Fspi level was dominated by a thin band
of both P. canaliculata and F. spiralis (1.4% and 1.3% of total pixels, respectively) with a better
cover of F. spiralis than for MLC. The An and Fser levels (mid-shore) were dominated by a
large band of A. nodosum and of F. serratus (5.5% and 3.5% of total pixels, respectively), but
the cover of F. serratus was less important than for MLC. The He-Ld level (lower shore) was
characterized by the important development of H. elongata (11.8% of total pixels). The cover
of red macroalgae was distributed on all of the site (2.2% of total pixels) and was more
present in the Fser level compared to the MLC results. Green algae were mainly present in
the lower shore (1.9% of total pixels). The ‘Substratum’ and ‘Water’ classes represented the
majority of the site (54.8% and 17.5% of the site, respectively). The macroalgal classes H.
elongata and ‘Green’ showed the highest producer/user accuracies (Table 5). As for MLC,
there was some misclassification. First, 18% of P. canaliculata pixels had been classified
as F. spiralis (9.44%) and H. elongata (9.49%). The lowest producer/user accuracy was
for F. spiralis’, with the largest misclassification (37.54%) in A. nodosum’ and 14.37% of
pixels in P. canaliculata’. Of the A. nodosum’ pixels, 18.76% were misclassified as F. spiralis’,
but also 7.35% and 4.62% of A. nodosum pixels were misclassified as F. serratus and P.
canaliculata’, respectively. Of the F. serratus pixels, 18.86% were misclassified as ‘Red’, and
17.37% of pixels should have been classified as F. serratus’, when they were in fact classified
as A. nodosum’. ‘Green’ and ‘Red’ algae were globally well classified, but there was some
misclassification between ‘Red’ algae and some Fucales, such as A. nodosum’. ’Substratum’
and ‘Water classes had the highest producer/user accuracy and so were well classified on
the entire image.
Remote Sens. 2022,14, 3124 13 of 26
Remote Sens. 2022, 14, x FOR PEER REVIEW 13 of 26
Figure 8. Spectral angle mapper (SAM), trained using image-derived spectra resulting from the hy-
perspectral UAV survey at Porspoder. Seven macroalgal cover classes are displayed over the UAV
RGB imagery and an orthophotography (Mégalis Bretagne et collectivités territoriales bretonnes
2015), giving an overview of the site. The Substratum and Water classes are not represented on
the map. Class codes Green and Red represent grouped green and red macroalgal species, respec-
tively.
3.3. Comparison of Field Sampling and Hyperspectral Classification
Covers determined by field sampling are compared here to the classification results
by both MLC and SAM. A visual representation of the comparison between in situ sam-
pling, an infrared picture and the two methods is provided in Figure 9 and Appendix A
Figures A1A3.
Figure 8.
Spectral angle mapper (SAM), trained using image-derived spectra resulting from the
hyperspectral UAV survey at Porspoder. Seven macroalgal cover classes are displayed over the UAV
RGB imagery and an orthophotography (Mégalis Bretagne et collectivités territoriales bretonnes—
2015), giving an overview of the site. The ‘Substratum’ and ‘Water’ classes are not represented on the
map. Class codes ‘Green’ and ‘Red’ represent grouped green and red macroalgal species, respectively.
Table 5.
Spectral angle mapper (SAM) confusion matrix, calculated, using ENVI 5.6.1, by comparing
pixels of known class locations to those predicted by the classification workflow, for each of the nine
cover classes. Results are displayed as percentages of pixels assigned, correctly or incorrectly, to each
class. User/producer accuracies (User Acc. and Prod. Acc., respectively) are also presented.
Class P. canaliculata F. spiralis A. nodosum F. serratus H. elongata Green Red Substratum Water Total User Acc.
Unclassified 0 0 0 0 0.06 0 0 0 0.24 0.06 -
P. canaliculata 65.22 14.37 4.62 9.604 0.72 6.32 3.97 0.03 0.04 1.44 15.44
F. spiralis 9.44 25.95 18.76 2.78 0 0.33 0.70 0.01 0 1.35 0.43
A. nodosum 3.02 37.54 67.06 17.37 0.04 0.32 22.84 0.01 0.01 5.48 70.97
F. serratus 1.97 8.80 7.35 38.80 1.36 1.44 8.21 0.01 0.04 3.54 80.26
H. elongata 9.49 1.47 0.44 11.78 95.01 10.76 7.26 0.01 7.92 11.76 75.96
Green 7.72 6.45 0.77 0.75 0.86 77.33 1.42 0.64 0.73 1.89 62.07
Red 2.56 5.13 0.93 18.86 0.60 0.23 54.82 0 0.02 2.19 31.09
Substratum 0.29 0 0.05 0.01 0 0.43 0.06 99.04 8.41 54.81 96.79
Water 0.30 0.29 0.02 0.02 1.35 2.82 0.72 0.24 82.60 17.49 98.23
Total 100 100 100 100 100 100 100 100 100 100 -
Prod. Acc. 65.22 25.95 67.06 38.8 95.01 77.33 54.82 99.04 82.6 - -
Remote Sens. 2022,14, 3124 14 of 26
3.3. Comparison of Field Sampling and Hyperspectral Classification
Covers determined by field sampling are compared here to the classification re-
sults by both MLC and SAM. A visual representation of the comparison between in
situ sampling, an infrared picture and the two methods is provided in Figure 9and
Appendix AFigures A1A3.
Remote Sens. 2022, 14, x FOR PEER REVIEW 14 of 26
(a)
(b) (c) (d)
Figure 9. (a) Picture of a sampling spot on the Pc-Fspi level at Porspoder taken during field sampling
in June 2021. (b) NIR-G-B image of the same s