Optimizing the Scale of Observation for Intertidal Habitat Classification through Multiscale Analysis

Abstract

Monitoring intertidal habitats, such as oyster reefs, salt marshes, and mudflats, is logistically challenging and often cost- and time-intensive. Remote sensing platforms, such as unoccupied aircraft systems (UASs), present an alternative to traditional approaches that can quickly and inexpensively monitor coastal areas. Despite the advantages offered by remote sensing systems, challenges remain concerning the best practices to collect imagery to study these ecosystems. One such challenge is the range of spatial resolutions for imagery that is best suited for intertidal habitat monitoring. Very fine imagery requires more collection and processing times. However, coarser imagery may not capture the fine-scale patterns necessary to understand relevant ecological processes. This study took UAS imagery captured along the Gulf of Mexico coastline in Florida, USA, and resampled the derived orthomosaic and digital surface model to resolutions ranging from 3 to 31 cm, which correspond to the spatial resolutions achievable by other means (e.g., aerial photography and certain commercial satellites). A geographic object-based image analysis (GEOBIA) workflow was then applied to datasets at each resolution to classify mudflats, salt marshes, oyster reefs, and water. The GEOBIA process was conducted within R, making the workflow open-source. Classification accuracies were largely consistent across the resolutions, with overall accuracies ranging from 78% to 82%. The results indicate that for habitat mapping applications, very fine resolutions may not provide information that increases the discriminative power of the classification algorithm. Multiscale classifications were also conducted and produced higher accuracies than single-scale workflows, as well as a measure of uncertainty between classifications.

Full text

Citation: Espriella, M.C.; Lecours, V.
Optimizing the Scale of Observation
for Intertidal Habitat Classification
through Multiscale Analysis. Drones
2022,6, 140. https://doi.org/
10.3390/drones6060140
Academic Editors: David R Green
and Brian S. Burnham
Received: 8 May 2022
Accepted: 4 June 2022
Published: 7 June 2022
Corrected: 23 September 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/).
drones
Article
Optimizing the Scale of Observation for Intertidal Habitat
Classification through Multiscale Analysis
Michael C. Espriella * and Vincent Lecours
School of Forest, Fisheries and Geomatics Sciences, University of Florida, Gainesville, FL 32653, USA;
vlecours@ufl.edu
*Correspondence: michaelespriella@ufl.edu
Abstract:
Monitoring intertidal habitats, such as oyster reefs, salt marshes, and mudflats, is logistically
challenging and often cost- and time-intensive. Remote sensing platforms, such as unoccupied aircraft
systems (UASs), present an alternative to traditional approaches that can quickly and inexpensively
monitor coastal areas. Despite the advantages offered by remote sensing systems, challenges remain
concerning the best practices to collect imagery to study these ecosystems. One such challenge is
the range of spatial resolutions for imagery that is best suited for intertidal habitat monitoring. Very
fine imagery requires more collection and processing times. However, coarser imagery may not
capture the fine-scale patterns necessary to understand relevant ecological processes. This study
took UAS imagery captured along the Gulf of Mexico coastline in Florida, USA, and resampled
the derived orthomosaic and digital surface model to resolutions ranging from 3 to 31 cm, which
correspond to the spatial resolutions achievable by other means (e.g., aerial photography and certain
commercial satellites). A geographic object-based image analysis (GEOBIA) workflow was then
applied to datasets at each resolution to classify mudflats, salt marshes, oyster reefs, and water.
The GEOBIA process was conducted within R, making the workflow open-source. Classification
accuracies were largely consistent across the resolutions, with overall accuracies ranging from 78%
to 82%. The results indicate that for habitat mapping applications, very fine resolutions may not
provide information that increases the discriminative power of the classification algorithm. Multiscale
classifications were also conducted and produced higher accuracies than single-scale workflows, as
well as a measure of uncertainty between classifications.
Keywords:
UAS; multiscale; geographic object-based image analysis; oyster; habitat mapping; scale;
drone; UAV; Florida; coastal
1. Introduction
Studying intertidal habitats as intricate mosaics forming a system rather than individ-
ual patches can allow for a better understanding of the complex and dynamic interactions
in these environments [
1
–
4
]. Ecosystem services provided by intertidal habitats, such as
oyster reefs, salt marshes, and mudflats, are often multifaceted as services may be amplified
or hindered by patch configuration. For example, the presence of the Crassostrea virginica
(eastern oyster) shell adjacent to a salt marsh can reduce erosion [
5
], the composition and
spatial configuration of intertidal habitats can affect the community structure [
6
,
7
], and
denitrification services provided by oyster reefs are dependent on habitat context [8].
Traditionally, sampling these environments has been very cost and time-intensive,
often relying on in situ approaches that do not provide a full picture of the ecosystem.
Remote sensing alternatives provide an opportunity to reduce costs while still capturing
valuable information about habitat coverage and configuration. Unoccupied aircraft sys-
tems (UASs) are among the least expensive and most accessible platforms to collect remote
sensing data on the environment. UASs have been used successfully to delineate intertidal
habitats with a high level of accuracy [
9
–
11
]. These habitat maps enable the analysis of
Drones 2022,6, 140. https://doi.org/10.3390/drones6060140 https://www.mdpi.com/journal/drones

Drones 2022,6, 140 2 of 19
spatial relationships between habitats and provide tools to produce more holistic, spatially
discrete analyses. Despite promising results, numerous questions remain regarding how to
efficiently and accurately monitor these systems using remote sensing. Understanding the
spatial resolution necessary to confidently draw conclusions about the ecosystem using
the collected imagery presents a significant knowledge gap. UASs can collect imagery
with sub-centimeter resolutions when flown at low altitudes, but collecting finer-resolution
data takes more time in the field and increases processing time significantly [
12
]. Targeting
ultra-high resolutions can also introduce additional monetary costs when acquiring sensors
with higher focal lengths capable of such fine observations.
In addition to data collection and processing efficiency, the information captured at a
given spatial resolution is an important consideration and an active research area [
13
]. Very-
high-resolution imagery (i.e., centimeter-scale) can introduce radiometric and geometric
distortions in an orthomosaic due to the variations in viewing geometry and angular
orientation of the vehicle [
14
]. Therefore, higher flight altitudes and, thereby, coarser
resolutions, are sometimes preferred to minimize perspective distortion, particularly when
the structure of the scene is of interest [
13
]. However, when using a broader scale, the
spectral response of a given pixel or object can be averaged out with that of neighboring
pixels or objects and may prevent capturing small objects in the scene, such as individual
oysters or oyster clusters. The ability to capture slight differences in spectral responses is
useful when delineating habitats in an intertidal environment with cover types, such as
oyster reefs and mudflats that have little spectral separability [
10
]. Studies have explored
the tradeoffs between parameters, such as image overlap and resolutions using UASs
(e.g., [
13
,
15
]). However, many of these studies are focused on the preservation of structure
(i.e., a digital surface model) rather than habitat mapping for monitoring purposes, and
most are terrestrial studies.
Multiscale approaches have become critical to ensure that the scales of observation
and analysis match the ecological scales relevant to a particular system [
16
,
17
] and can
help identify the most appropriate scales of observation and analysis for a given purpose
(e.g., [
18
,
19
]). In theory, using scales that best capture the patterns of interest improves
classification accuracy. In addition, finding which range of spatial resolution is most
appropriate will inform monitoring efforts by ensuring that imagery is collected in a
way that addresses objectives without hindering processing efforts with excessively large
datasets. While researchers often select an arbitrary scale they deem fit or use the native
resolution of a given dataset, multiscale approaches are becoming more common to ensure
the understanding of phenomena without introducing biases via arbitrary parametrization
(e.g., [
18
–
22
]). While UASs have been used to conduct multiscale studies in the intertidal
(e.g., [
23
,
24
]), these studies have often focused on identifying and mapping the distribution
of benthic fauna rather than habitat classification.
The objective of this study is to streamline habitat classifications in a way to optimize
survey efforts and minimize subjectivity. We developed a semi-automated classification
workflow to encourage consistent monitoring. The workflow will enable management
agencies to observe the temporal changes in habitat coverage, assess how to best address
habitat degradation, and pinpoint areas most prone to change where conservation efforts
should be targeted. A better understanding of trade-offs associated with the observation
scale will result in more informed management.
2. Materials and Methods
2.1. Study Site and UAS Survey
The study site is located at the mouth of Little Trout Creek (29
◦
15
0
34.98
00
N,
83◦4029.6900 W
),
within the Suwannee River estuary, on Florida’s Gulf of Mexico coast (Figure 1).
Little Trout Creek is a tidal creek with eastern oyster reef, salt marsh, and mudflat habi-
tats present. This area is known to have sustained major changes in recent years and is of
interest to local management agencies and researchers [
25
,
26
]. The Suwannee River estuary
is within Florida’s Big Bend—the coastline from Crystal River to Apalachee Bay—where,

Drones 2022,6, 140 3 of 19
as of 2011, there has been an estimated 66% net loss of the oyster reef area since 1982 [
25
].
There has also been evidence of increased erosion of islands in the region, coinciding with
oyster reef decline and a decline in coastal forest coverage [
26
–
28
]. With about 100 km of
federally protected coastline, the Suwannee River estuary is largely undeveloped, and 25%
of the area’s economic activity is connected to natural
resources [29,30]
. Programs such as
the Oyster Integrated Mapping and Monitoring Program (OIMMP) represent increasing
efforts to map and monitor these resources [26].
Drones 2022, 6, x FOR PEER REVIEW 3 of 20
Figure 1. Study site for the UAS survey. The mouth of Little Trout Creek is located within the larger
tidal environment of the Suwannee Sound, fed by the Suwannee River in the Gulf of Mexico.
Little Trout Creek is a tidal creek with eastern oyster reef, salt marsh, and mudflat
habitats present. This area is known to have sustained major changes in recent years and
is of interest to local management agencies and researchers [25,26]. The Suwannee River
estuary is within Florida’s Big Bend—the coastline from Crystal River to Apalachee Bay—
where, as of 2011, there has been an estimated 66% net loss of the oyster reef area since
1982 [25]. There has also been evidence of increased erosion of islands in the region, coin-
ciding with oyster reef decline and a decline in coastal forest coverage [26–28]. With about
100 km of federally protected coastline, the Suwannee River estuary is largely undevel-
oped, and 25% of the area’s economic activity is connected to natural resources [29,30].
Programs such as the Oyster Integrated Mapping and Monitoring Program (OIMMP) rep-
resent increasing efforts to map and monitor these resources [26].
A DJI Inspire 2 was used to collect georeferenced imagery on 8 December 2018 at the
mouth of Little Trout Creek. The UAS was equipped with a Zenmuse X7 35 mm RGB
sensor. A flying height of 60 m above ground level and image overlaps of 80% along-track
and 75% across-track were set for the flight mission. Data collection was conducted to
coincide with the low tide at 8:30 a.m. local time to maximize coverage of the emerged
intertidal habitats. Four black-and-white checkered targets were deployed evenly across
the scene. The location of the center of each target was recorded using real-time kinematic
(RTK) positioning with a Trimble 5800 RTK system.
2.2. Imagery Processing
The survey produced 963 images and a ground sampling distance of 0.66 cm. Images
were imported into the Pix4D Mapper v. 4.2.26, and ground control targets were located
within the individual images and associated with coordinates recorded from RTK
Figure 1.
Study site for the UAS survey. The mouth of Little Trout Creek is located within the larger
tidal environment of the Suwannee Sound, fed by the Suwannee River in the Gulf of Mexico.
A DJI Inspire 2 was used to collect georeferenced imagery on 8 December 2018 at
the mouth of Little Trout Creek. The UAS was equipped with a Zenmuse X7 35 mm RGB
sensor. A flying height of 60 m above ground level and image overlaps of 80% along-track
and 75% across-track were set for the flight mission. Data collection was conducted to
coincide with the low tide at 8:30 a.m. local time to maximize coverage of the emerged
intertidal habitats. Four black-and-white checkered targets were deployed evenly across
the scene. The location of the center of each target was recorded using real-time kinematic
(RTK) positioning with a Trimble 5800 RTK system.
2.2. Imagery Processing
The survey produced 963 images and a ground sampling distance of 0.66 cm. Images
were imported into the Pix4D Mapper v. 4.2.26, and ground control targets were located
within the individual images and associated with coordinates recorded from RTK sampling

Drones 2022,6, 140 4 of 19
to ensure the spatial quality of outputs [
9
,
31
]. Pix4D produced estimates of the position,
angular orientation, and camera calibration parameters of images and generated an ortho-
mosaic and a digital surface model (DSM) using structure from motion photogrammetry
techniques (Figure 2). Ground control points were then digitized within the orthomosaic
using ArcGIS Pro v 2.4 to calculate the spatial accuracy by comparing the coordinates to
the known RTK observations [32].
Drones 2022, 6, x FOR PEER REVIEW 4 of 20
sampling to ensure the spatial quality of outputs [9,31]. Pix4D produced estimates of the
position, angular orientation, and camera calibration parameters of images and generated
an orthomosaic and a digital surface model (DSM) using structure from motion photo-
grammetry techniques (Figure 2). Ground control points were then digitized within the
orthomosaic using ArcGIS Pro v 2.4 to calculate the spatial accuracy by comparing the
coordinates to the known RTK observations [32].
Figure 2. Orthomosaic and DSM at the native imagery resolution of 0.66 cm.
The orthomosaic and DSM were then resampled to resolutions ranging from 3 to 31
cm by 2 cm increments. Resampling was conducted using the Resample tool with the
nearest neighbor as the resampling method in ArcGIS Pro v 2.4, resulting in 15 datasets of
different spatial resolutions. The lower limit of the range, 3 cm, was selected to be repre-
sentative of high-resolution UAS imagery. The upper limit of 31 cm was selected to corre-
spond with what can be acquired from other remote sensing sources (e.g., occupied air-
craft imagery, satellite imagery). This method of varying the resolution scale of input var-
iables (sometimes referred to as coarse-graining) has been used when studying marine
environments, often within the context of habitat selection models [33–36], and was iden-
tified as the most appropriate method to use when wanting to characterize specific fea-
tures or processes [17].
2.3. GEOBIA—Segmentation Optimization on Representative Subset Area
The analytical workflow is based on a geographic object-based image analysis
(GEOBIA), which first segments the imagery into relatively homogeneous groups of pix-
els, forming meaningful objects, before classifying them [37–39]. Segmentation parame-
ters are most often determined through trial and error, but here, we applied an optimiza-
tion procedure to remove subjectivity from the analysis and facilitate end uses from non-
experts. Figure 3 represents the workflow and the different tools used to accomplish each
step.
Figure 2. Orthomosaic and DSM at the native imagery resolution of 0.66 cm.
The orthomosaic and DSM were then resampled to resolutions ranging from 3 to 31 cm
by 2 cm increments. Resampling was conducted using the Resample tool with the nearest
neighbor as the resampling method in ArcGIS Pro v 2.4, resulting in 15 datasets of different
spatial resolutions. The lower limit of the range, 3 cm, was selected to be representative of
high-resolution UAS imagery. The upper limit of 31 cm was selected to correspond with
what can be acquired from other remote sensing sources (e.g., occupied aircraft imagery,
satellite imagery). This method of varying the resolution scale of input variables (sometimes
referred to as coarse-graining) has been used when studying marine environments, often
within the context of habitat selection models [
33
–
36
], and was identified as the most
appropriate method to use when wanting to characterize specific features or processes [
17
].
2.3. GEOBIA—Segmentation Optimization on Representative Subset Area
The analytical workflow is based on a geographic object-based image analysis (GEO-
BIA), which first segments the imagery into relatively homogeneous groups of pixels,
forming meaningful objects, before classifying them [
37
–
39
]. Segmentation parameters
are most often determined through trial and error, but here, we applied an optimization
procedure to remove subjectivity from the analysis and facilitate end uses from non-experts.
Figure 3represents the workflow and the different tools used to accomplish each step.
First, a representative subset of the scene was taken from each mosaic to inform the
optimization of the segmentation parameters (Figure 4). A subset was used for this proce-
dure because it is computationally expensive. The optimization for each resolution was
performed using the SegOptim package in R, providing a consistent and unbiased method
for segmentation parameter selection [
40
]. The SegOptim package combines the optimiza-
tion of segmentation parameters, image segmentation, and supervised classification in
a single workflow. SegOptim relies on third-party software to produce segmentations.
Here, the Orfeo ToolBox was selected, which provides open-source solutions for remote
sensing analysis and contains applications that can be used within command lines [
41
]. The

Drones 2022,6, 140 5 of 19
Orfeo ToolBox large-scale mean shift segmentation was selected to perform the segmenta-
tion, which has been shown to perform well in a variety of remote sensing applications
and from medium to very high resolutions [
42
]. Three parameters drive the large-scale
mean shift segmentation: spectral range radius (also referred to as range radius), spatial
radius, and the minimum object size in pixels. Spatial radius refers to the radius of the
neighborhood for averaging (higher spatial radius results in more smoothing). Spectral
range radius refers to the degree of spectral variability, measured by the Euclidean distance
(expressed in radiometry units) of the spectral signature that is allowed within an object
(higher values preserve edges less than lower values) [
43
]. While higher values for spatial
radius lead to more processing times, changing the spectral range radius does not affect
the processing time [
43
]. Minimum size refers to the minimum allowable object size in
pixels [
43
]. A range of appropriate values was set for each of the three parameters. Values
ranging from 50 to 125 for the spectral range radius and 20 to 80 for the spatial radius
were tested. The tested values for spectral and spatial radii remained constant for every
tested resolution. The minimum object size values remained constant in terms of area,
as the range in pixels was adjusted at every orthomosaic to allow for a minimum object
size ranging from 7 to 10 m
2
. We selected this object size to correspond with the smallest
habitat patch that we wanted to capture in the scene. As performed by Espriella et al.
(2020), a 3
×
3 edge-detecting Laplacian filter was applied to each orthomosaic to serve as
the input for the segmentation step, given its high performance in scenes where water is
present. The “gaOptimizeSegmentationParams” command was used within the SegOptim
package to test the quality of combinations within the defined parameter ranges. This
function optimizes segmentation parameters using genetic algorithms [
40
]. The inputs
of the genetic algorithm affect the computing time, as they direct functions, such as the
number of iterations. These parameters were kept constant for all resolutions and can be
found in Appendix A(Table A1).
Drones 2022, 6, x FOR PEER REVIEW 5 of 20
Figure 3. Diagram detailing the workflow from data acquisition to habitat classification.
First, a representative subset of the scene was taken from each mosaic to inform the
optimization of the segmentation parameters (Figure 4). A subset was used for this pro-
cedure because it is computationally expensive. The optimization for each resolution was
performed using the SegOptim package in R, providing a consistent and unbiased method
for segmentation parameter selection [40]. The SegOptim package combines the optimi-
zation of segmentation parameters, image segmentation, and supervised classification in
a single workflow. SegOptim relies on third-party software to produce segmentations.
Here, the Orfeo ToolBox was selected, which provides open-source solutions for remote
sensing analysis and contains applications that can be used within command lines [41].
The Orfeo ToolBox large-scale mean shift segmentation was selected to perform the seg-
mentation, which has been shown to perform well in a variety of remote sensing applica-
tions and from medium to very high resolutions [42]. Three parameters drive the large-
scale mean shift segmentation: spectral range radius (also referred to as range radius),
spatial radius, and the minimum object size in pixels. Spatial radius refers to the radius of
the neighborhood for averaging (higher spatial radius results in more smoothing). Spec-
tral range radius refers to the degree of spectral variability, measured by the Euclidean
distance (expressed in radiometry units) of the spectral signature that is allowed within
an object (higher values preserve edges less than lower values) [43]. While higher values
for spatial radius lead to more processing times, changing the spectral range radius does
not affect the processing time [43]. Minimum size refers to the minimum allowable object
size in pixels [43]. A range of appropriate values was set for each of the three parameters.
Values ranging from 50 to 125 for the spectral range radius and 20 to 80 for the spatial
radius were tested. The tested values for spectral and spatial radii remained constant for
every tested resolution. The minimum object size values remained constant in terms of
area, as the range in pixels was adjusted at every orthomosaic to allow for a minimum
object size ranging from 7 to 10 m
2
. We selected this object size to correspond with the
smallest habitat patch that we wanted to capture in the scene. As performed by Espriella
Figure 3. Diagram detailing the workflow from data acquisition to habitat classification.

Drones 2022,6, 140 6 of 19
Drones 2022, 6, x FOR PEER REVIEW 6 of 20
et al. (2020), a 3 × 3 edge-detecting Laplacian filter was applied to each orthomosaic to
serve as the input for the segmentation step, given its high performance in scenes where
water is present. The “gaOptimizeSegmentationParams” command was used within the
SegOptim package to test the quality of combinations within the defined parameter
ranges. This function optimizes segmentation parameters using genetic algorithms [40].
The inputs of the genetic algorithm affect the computing time, as they direct functions,
such as the number of iterations. These parameters were kept constant for all resolutions
and can be found in Appendix A (Table A1).
Figure 4. Training subset and delineated training areas for segmentation optimization.
To quantify the quality of the segmentation for each iteration, training data were pro-
vided in the form of rasters identifying samples for each of the four cover types (i.e.,
marsh, mud, oyster, and water) to conduct random forest classifications. Training areas
were first manually delineated with polygons in ArcGIS Pro in a way to depict the geom-
etry of each habitat patch (Figure 4). To be compatible with the SegOptim and Orfeo
ToolBox workflow, the polygons were then exported as rasters using the Polygon to Ras-
ter tool in ArcGIS Pro. The training raster resolution for each optimization corresponded
with the resolution of the orthomosaic being tested. These training areas were then used
to identify the objects included in the random forest classifier. In order for an object from
the segmentation to be included in the training dataset, 75% of its area must be within one
of the delineated classes. The means and standard deviations of the three bands (red,
green, blue) in the orthomosaic, calculated within segmented objects, were used to inform
the classifier. A 10-fold cross-validation was used to evaluate the quality of each classifi-
cation. The kappa coefficient of the classification was used as the metric to determine qual-
ity. The result of the optimization is the parameter values that produced the classification
with the highest kappa value.
2.4. GEOBIA-Segmentation and Classification of the Entire Scene
Following the optimization, the parameter values from the output were applied to
the entire scene. The implemented workflow follows the one introduced by Espriella et
al. (2020). First, as with the subset (cf. Section 2.3), a 3 × 3 Laplacian filter of each respective
resolution served as the layer for the large-scale mean shift segmentation. Following seg-
mentation, the water objects were removed from the scene by setting their object ID to
null within R. Given the homogeneity of the water, it is often segmented into just a few
objects, and removing these objects limits confusion between water and adjacent classes.
After removing the water objects, the remaining segmented objects were ready to be
classified. First, training areas were provided for the random forest classifier, as with the
subset for the optimization step (Figure 5). As with the optimization, 75% of an object in
the segmented layer must overlap with a training area in order to be used as training data
Figure 4. Training subset and delineated training areas for segmentation optimization.
To quantify the quality of the segmentation for each iteration, training data were
provided in the form of rasters identifying samples for each of the four cover types (i.e.,
marsh, mud, oyster, and water) to conduct random forest classifications. Training areas
were first manually delineated with polygons in ArcGIS Pro in a way to depict the geometry
of each habitat patch (Figure 4). To be compatible with the SegOptim and Orfeo ToolBox
workflow, the polygons were then exported as rasters using the Polygon to Raster tool
in ArcGIS Pro. The training raster resolution for each optimization corresponded with
the resolution of the orthomosaic being tested. These training areas were then used to
identify the objects included in the random forest classifier. In order for an object from the
segmentation to be included in the training dataset, 75% of its area must be within one of
the delineated classes. The means and standard deviations of the three bands (red, green,
blue) in the orthomosaic, calculated within segmented objects, were used to inform the
classifier. A 10-fold cross-validation was used to evaluate the quality of each classification.
The kappa coefficient of the classification was used as the metric to determine quality. The
result of the optimization is the parameter values that produced the classification with the
highest kappa value.
2.4. GEOBIA-Segmentation and Classification of the Entire Scene
Following the optimization, the parameter values from the output were applied to
the entire scene. The implemented workflow follows the one introduced by Espriella et al.
(2020). First, as with the subset (cf. Section 2.3), a 3
×
3 Laplacian filter of each respective
resolution served as the layer for the large-scale mean shift segmentation. Following
segmentation, the water objects were removed from the scene by setting their object ID to
null within R. Given the homogeneity of the water, it is often segmented into just a few
objects, and removing these objects limits confusion between water and adjacent classes.
After removing the water objects, the remaining segmented objects were ready to be
classified. First, training areas were provided for the random forest classifier, as with the
subset for the optimization step (Figure 5). As with the optimization, 75% of an object in
the segmented layer must overlap with a training area in order to be used as training data
for the classifier. The means and standard deviations of the red, green, and blue spectral
layers and the DSM were used to inform the classifier.
2.5. Accuracy Assessment
Classification accuracy was quantified using a stratified random sampling scheme
per class (i.e., marsh, mud, oyster, and water). Although the water class was not included
within the random forest classifier, it was included in the accuracy assessment to quantify
the accuracy of relying on the segmentation to delineate the water objects (cf. Section 2.4).
The sample size for the number of points was determined using an equation based on a
multinomial distribution with a 95% confidence level [
44
,
45
]. The sample size of 664 points

Drones 2022,6, 140 7 of 19
was evenly distributed in ArcGIS Pro among the four classes to ensure equal representa-
tion. A confusion matrix was developed for each of the 15 resolutions tested. Producer-,
user-, and overall accuracies, and a kappa coefficient, were calculated from each matrix.
Producer’s accuracy measures how often real features are represented in the generated clas-
sification, while the user’s accuracy measures how often an object is erroneously included
in a given class [
46
,
47
]. The Kappa coefficient quantifies accuracy as the actual agreement
minus chance agreement [
46
]. In addition to these performance metrics, a mean decrease in
Gini values from the random forest classifier was recorded to identify the most important
inputs to the models. The mean decrease in Gini is the total decrease in node impurity for a
predictor (when it forms a split in the forest), averaged across the trees [
48
,
49
]. The higher
the value, the more influential the variable was in the classification.
Drones 2022, 6, x FOR PEER REVIEW 7 of 20
for the classifier. The means and standard deviations of the red, green, and blue spectral
layers and the DSM were used to inform the classifier.
Figure 5. Delineated training areas for the classification of the entire scene.
2.5. Accuracy Assessment
Classification accuracy was quantified using a stratified random sampling scheme
per class (i.e., marsh, mud, oyster, and water). Although the water class was not included
within the random forest classifier, it was included in the accuracy assessment to quantify
the accuracy of relying on the segmentation to delineate the water objects (cf. Section 2.4).
The sample size for the number of points was determined using an equation based on a
multinomial distribution with a 95% confidence level [44,45]. The sample size of 664
points was evenly distributed in ArcGIS Pro among the four classes to ensure equal rep-
resentation. A confusion matrix was developed for each of the 15 resolutions tested. Pro-
ducer-, user-, and overall accuracies, and a kappa coefficient, were calculated from each
matrix. Producer’s accuracy measures how often real features are represented in the gen-
erated classification, while the user’s accuracy measures how often an object is errone-
ously included in a given class [46,47]. The Kappa coefficient quantifies accuracy as the
actual agreement minus chance agreement [46]. In addition to these performance metrics,
a mean decrease in Gini values from the random forest classifier was recorded to identify
the most important inputs to the models. The mean decrease in Gini is the total decrease
in node impurity for a predictor (when it forms a split in the forest), averaged across the
trees [48,49]. The higher the value, the more influential the variable was in the classifica-
tion.
2.6. Multiscale Analysis
To quantify agreement among the 15 classifications, each was resampled to 3 cm in
ArcGIS using the Resample tool with the nearest neighbor as the resampling method. This
allowed for the rasters to be stacked in R and the mode at each cell was recorded to deter-
mine the most agreed-upon class at that cell. A raster was generated from the mode values
as well as a raster recording how many rasters within the stack of 15 predicted the mode
class at a given cell. These representations allowed us to identify areas within the scene
Figure 5. Delineated training areas for the classification of the entire scene.
2.6. Multiscale Analysis
To quantify agreement among the 15 classifications, each was resampled to 3 cm
in ArcGIS using the Resample tool with the nearest neighbor as the resampling method.
This allowed for the rasters to be stacked in R and the mode at each cell was recorded to
determine the most agreed-upon class at that cell. A raster was generated from the mode
values as well as a raster recording how many rasters within the stack of 15 predicted
the mode class at a given cell. These representations allowed us to identify areas within
the scene that may have caused significant confusion as evidenced by weak agreement
between the classifications, as well as areas that performed well.
A classification that combined the most informative observation scales was also pro-
duced to quantify the accuracy of a multiscale classification. The resolutions that produced
the best user accuracies for each of the four classes were combined in one raster stack. To
stack the rasters, the orthomosaic and DSM were first resampled to the finest resolution
of the four selected rasters using the same method as above. All four orthomosaics and
DSMs were then stacked in R to generate one 16-band raster (3 spectral bands for each
orthomosaic and 1 band for each DSM). The Laplacian filter of the finest resolution was
used for segmentation using the parameters reported in the initial optimization procedure.

Drones 2022,6, 140 8 of 19
The procedure then followed the steps from the individual resolution classification (cf.
Section 2.4) and the accuracy assessment steps outlined above (cf. Section 2.5).
3. Results
3.1. Imagery Processing
From the digitized ground control points, a root mean square error of 2.1 cm in latitude
and 1.6 cm in longitude was derived for the orthomosaic. Figure 6shows the levels of
detail allowed by the selected resolutions. The three-band orthomosaics ranged in size
from
1.5 GB
at 3 cm resolution to 15.5 MB at 31 cm. The DSMs range in size from 750 MB at
3 cm resolution to 7.7 MB at 31 cm.
Drones 2022, 6, x FOR PEER REVIEW 8 of 20
that may have caused significant confusion as evidenced by weak agreement between the
classifications, as well as areas that performed well.
A classification that combined the most informative observation scales was also pro-
duced to quantify the accuracy of a multiscale classification. The resolutions that pro-
duced the best user accuracies for each of the four classes were combined in one raster
stack. To stack the rasters, the orthomosaic and DSM were first resampled to the finest
resolution of the four selected rasters using the same method as above. All four orthomo-
saics and DSMs were then stacked in R to generate one 16-band raster (3 spectral bands
for each orthomosaic and 1 band for each DSM). The Laplacian filter of the finest resolu-
tion was used for segmentation using the parameters reported in the initial optimization
procedure. The procedure then followed the steps from the individual resolution classifi-
cation (cf. Section 2.4) and the accuracy assessment steps outlined above (cf. Section 2.5).
3. Results
3.1. Imagery Processing
From the digitized ground control points, a root mean square error of 2.1 cm in lati-
tude and 1.6 cm in longitude was derived for the orthomosaic. Figure 6 shows the levels
of detail allowed by the selected resolutions. The three-band orthomosaics ranged in size
from 1.5 GB at 3 cm resolution to 15.5 MB at 31 cm. The DSMs range in size from 750 MB
at 3 cm resolution to 7.7 MB at 31 cm.
Figure 6. Area of about 10 × 10 m with marsh, oyster, mud, and water at varying resolutions. A
checkered ground control target is visible in some of the resolutions. The difference in details for
each resolution highlights how even an expert/manual delineation of these habitats by photointer-
pretation would lead to different results.
3.2. Segmentation Optimization
Table 1 represents the spatial radius, spectral range radius, and minimum size pa-
rameter values that the SegOptim package reported as optimal for each resolution. The
spatial radius was 58.5 on average across the resolutions, while the spectral range radius
was 70.93 on average. The spatial radius was more variable than the spectral range radius,
with standard deviations of 12.5 and 9.72, respectively. The average minimum object size
was 7.57 m
2
with a standard deviation of 0.47 m
2
.
Figure 6.
Area of about 10
×
10 m with marsh, oyster, mud, and water at varying resolutions. A
checkered ground control target is visible in some of the resolutions. The difference in details for each
resolution highlights how even an expert/manual delineation of these habitats by photointerpretation
would lead to different results.
3.2. Segmentation Optimization
Table 1represents the spatial radius, spectral range radius, and minimum size pa-
rameter values that the SegOptim package reported as optimal for each resolution. The
spatial radius was 58.5 on average across the resolutions, while the spectral range radius
was 70.93 on average. The spatial radius was more variable than the spectral range radius,
with standard deviations of 12.5 and 9.72, respectively. The average minimum object size
was 7.57 m2with a standard deviation of 0.47 m2.
3.3. Segmentation and Classification
Segmentations captured boundaries between habitats well, particularly between water
and intertidal habitats (Figure A1). Given the relative homogeneity of the water surface,
water was most often successfully grouped into a single object because adjacent objects did
not exceed the spectral threshold parameter. However, up to three “water” objects were
removed for some resolutions.
The average object for the mud class was the largest at 17.0 m
2
(standard
deviation = 1.8 m2
).
The average object in the oyster class was 16.3 m
2
(standard deviation = 1.7 m
2
), and the marsh
class had the smallest average object area of 15.7 m
2
(standard deviation = 3.4 m
2
). Figure 7
displays the percent difference from the mean area over the 15 classifications per habitat type.

Drones 2022,6, 140 9 of 19
The resolutions were largely consistent, with many displaying minimal differences from the
mean value. However, some resolutions saw large variations. The 15 cm resolution produced
the instance with the largest difference from the mean value with the marsh habitat representing
63% less area than the average marsh area among all classifications.
Table 1. Results for each parameter at the respective resolution.
Resolution Spectral Radius Spatial Radius Min Size
(Pixels) Min Size (m2)
3 cm 56 56 8286 7.46
5 cm 63 61 2932 7.33
7 cm 65 71 1626 7.97
9 cm 71 39 931 7.54
11 cm 62 70 740 8.96
13 cm 69 55 455 7.69
15 cm 63 45 329 7.40
17 cm 67 55 261 7.54
19 cm 86 59 207 7.47
21 cm 76 77 161 7.10
23 cm 86 73 133 7.04
25 cm 74 39 113 7.06
27 cm 66 60 107 7.80
29 cm 89 72 92 7.73
31 cm 71 45 77 7.40
Drones 2022, 6, x FOR PEER REVIEW 9 of 20
Table 1. Results for each parameter at the respective resolution.
Resolution Spectral Radius Spatial Radius Min Size (Pixels) Min Size (m2)
3 cm 56 56 8286 7.46
5 cm 63 61 2932 7.33
7 cm 65 71 1626 7.97
9 cm 71 39 931 7.54
11 cm 62 70 740 8.96
13 cm 69 55 455 7.69
15 cm 63 45 329 7.40
17 cm 67 55 261 7.54
19 cm 86 59 207 7.47
21 cm 76 77 161 7.10
23 cm 86 73 133 7.04
25 cm 74 39 113 7.06
27 cm 66 60 107 7.80
29 cm 89 72 92 7.73
31 cm 71 45 77 7.40
3.3. Segmentation and Classification
Segmentations captured boundaries between habitats well, particularly between wa-
ter and intertidal habitats (Figure A1). Given the relative homogeneity of the water sur-
face, water was most often successfully grouped into a single object because adjacent ob-
jects did not exceed the spectral threshold parameter. However, up to three “water” ob-
jects were removed for some resolutions.
The average object for the mud class was the largest at 17.0 m2 (standard deviation =
1.8 m2). The average object in the oyster class was 16.3 m2 (standard deviation = 1.7 m2),
and the marsh class had the smallest average object area of 15.7 m2 (standard deviation =
3.4 m2). Figure 7 displays the percent difference from the mean area over the 15 classifica-
tions per habitat type. The resolutions were largely consistent, with many displaying min-
imal differences from the mean value. However, some resolutions saw large variations.
The 15 cm resolution produced the instance with the largest difference from the mean
value with the marsh habitat representing 63% less area than the average marsh area
among all classifications.
Figure 7. The percent difference from the mean area across the 15 classifications for each class at the
respective resolution. This visualization allows for initial observations about what classes are being
Figure 7.
The percent difference from the mean area across the 15 classifications for each class at the
respective resolution. This visualization allows for initial observations about what classes are being
confused for each other (e.g., when there is a large difference in the mud class, there is also often a
large difference in the water class that offsets).
3.4. Accuracy Assessment
The classifications all performed well, with kappa coefficients ranging from 0.708
(
29 cm
resolution) to 0.764 (5 cm resolution). Overall accuracies ranged from 78% at 29 cm
to 82% at 5 cm. The classifications with the highest and lowest performances are displayed
in Figure 8.

Drones 2022,6, 140 10 of 19
Drones 2022, 6, x FOR PEER REVIEW 10 of 20
confused for each other (e.g., when there is a large difference in the mud class, there is also often a
large difference in the water class that offsets).
3.4. Accuracy Assessment
The classifications all performed well, with kappa coefficients ranging from 0.708 (29
cm resolution) to 0.764 (5 cm resolution). Overall accuracies ranged from 78% at 29 cm to
82% at 5 cm. The classifications with the highest and lowest performances are displayed
in Figure 8.
Figure 8. Classification corresponding with the highest kappa coefficient at 5 cm (left) compared to
the 29 cm resolution (right), which had the lowest kappa coefficient. The comparison allows visual-
izing areas within the scene that pose challenges, such as the southwest corner where mud is mis-
taken for water in the 29 cm classification.
Figure 9 represents the trend of kappa values over the range of resolutions. Perfor-
mance was largely consistent over the span of the resolutions. The lowest-performing res-
olutions, 19 and 29 cm, struggled with differentiating mud from other habitat types, re-
sulting in 56% and 60% producer accuracies, respectively. Most misclassifications within
the mud class were with water for both the 19 cm and the 29 cm classifications. Very coarse
resolutions performed similarly, and sometimes better than finer resolutions that took sig-
nificantly more processing times.
Figure 9. Change in kappa coefficient as a function of input resolution.
Figure 8.
Classification corresponding with the highest kappa coefficient at 5 cm (
left
) compared
to the 29 cm resolution (
right
), which had the lowest kappa coefficient. The comparison allows
visualizing areas within the scene that pose challenges, such as the southwest corner where mud is
mistaken for water in the 29 cm classification.
Figure 9represents the trend of kappa values over the range of resolutions. Perfor-
mance was largely consistent over the span of the resolutions. The lowest-performing
resolutions, 19 and 29 cm, struggled with differentiating mud from other habitat types,
resulting in 56% and 60% producer accuracies, respectively. Most misclassifications within
the mud class were with water for both the 19 cm and the 29 cm classifications. Very coarse
resolutions performed similarly, and sometimes better than finer resolutions that took
significantly more processing times.
Drones 2022, 6, x FOR PEER REVIEW 10 of 20
confused for each other (e.g., when there is a large difference in the mud class, there is also often a
large difference in the water class that offsets).
3.4. Accuracy Assessment
The classifications all performed well, with kappa coefficients ranging from 0.708 (29
cm resolution) to 0.764 (5 cm resolution). Overall accuracies ranged from 78% at 29 cm to
82% at 5 cm. The classifications with the highest and lowest performances are displayed
in Figure 8.
Figure 8. Classification corresponding with the highest kappa coefficient at 5 cm (left) compared to
the 29 cm resolution (right), which had the lowest kappa coefficient. The comparison allows visual-
izing areas within the scene that pose challenges, such as the southwest corner where mud is mis-
taken for water in the 29 cm classification.
Figure 9 represents the trend of kappa values over the range of resolutions. Perfor-
mance was largely consistent over the span of the resolutions. The lowest-performing res-
olutions, 19 and 29 cm, struggled with differentiating mud from other habitat types, re-
sulting in 56% and 60% producer accuracies, respectively. Most misclassifications within
the mud class were with water for both the 19 cm and the 29 cm classifications. Very coarse
resolutions performed similarly, and sometimes better than finer resolutions that took sig-
nificantly more processing times.
Figure 9. Change in kappa coefficient as a function of input resolution.
Figure 9. Change in kappa coefficient as a function of input resolution.
While the kappa coefficient represents the classification’s performance across all
classes, the user’s accuracy is of particular interest in this context, given the focus on
monitoring and the value of the classifications to end users. The four cover types performed
better in terms of user accuracies at different resolutions. Marsh and oyster performed
better at coarser resolutions. The marsh habitat had a 96% user ’s accuracy at 23 cm and
the oyster habitat had an 84% user’s accuracy at 29 cm. Conversely, water and mud had

Drones 2022,6, 140 11 of 19
the highest user accuracies at finer resolutions. Water performed best at 5 cm with a user’s
accuracy of 86% and mud performed best at 3 cm with a user’s accuracy of 86%.
3.5. Variable Importance
The DSM mean was consistently the most influential metric included in the random
forest classification, as measured by the mean decrease in Gini (Figure 10). The standard
deviation of the DSM and standard deviation of the blue spectral band were the next most
influential variables. While the standard deviation of the blue band and the standard
deviation of the DSM produced similar mean decreases in Gini values for finer resolutions,
beginning at the 21 cm resolution, the standard deviation of the DSM began indicating
higher variable importance. This change may indicate that coarser resolutions are better
able to characterize variations in habitat covers’ topography without the influence of noise,
while the influence of spectral variation remains largely constant.
1
Figure 10.
Boxplots representing the mean decrease in Gini for each input variable over the range of
resolutions. Black dots represent outliers.
3.6. Mode of Classifications
The water class had the highest agreement among the classifications, averaging an
agreement of 13.83 out of 15 rasters. Marsh had the next highest agreement at an average
of 13.72, with oyster and mud displaying averages of 12.47 and 12.03, respectively. Areas of
disagreement often corresponded with the edges of habitats, as evidenced by the outlines
depicted in Figure 11. Areas of mud partially inundated with water (e.g., the southwest
corner of the scene) also coincided with areas of disagreement between the classification
rasters. Across the scene, 52% of the area had 100% agreement between the 15 rasters.
When tested using the accuracy assessment points, the mode raster produced a kappa
coefficient of 0.821 and overall accuracy of 87%. These high-performance metrics indicate
that not only did classifications often agree, but also the agreed-upon class was frequently
the correct class.
3.7. Multiscale Classification
The multiscale classification performed better than any individual classification in
terms of the kappa coefficient and overall accuracy, producing results of 0.778 and 83%,
respectively. However, there was a high level of omission for the mud class, resulting in a
57% producer’s accuracy. The mud class was most often misclassified as water. Despite
the underperformance in terms of the producer’s accuracy, the user’s accuracy of 88% for
mud was the highest among any of the classifications. The oyster class produced a user’s
accuracy of 86%, which was also higher than any individual classification, while marsh
and water had user accuracies of 89% and 75%, respectively. Table 2shows the comparison

Drones 2022,6, 140 12 of 19
of all classifications in terms of producer accuracies, user accuracies, overall accuracies, and
kappa coefficients.
Drones 2022, 6, x FOR PEER REVIEW 12 of 20
that not only did classifications often agree, but also the agreed-upon class was frequently
the correct class.
Figure 11. The class most often represented in each raster cell (left) and how often that class was
represented out of the 15 classifications (right).
3.7. Multiscale Classification
The multiscale classification performed better than any individual classification in
terms of the kappa coefficient and overall accuracy, producing results of 0.778 and 83%,
respectively. However, there was a high level of omission for the mud class, resulting in
a 57% producer’s accuracy. The mud class was most often misclassified as water. Despite
the underperformance in terms of the producer’s accuracy, the user’s accuracy of 88% for
mud was the highest among any of the classifications. The oyster class produced a user’s
accuracy of 86%, which was also higher than any individual classification, while marsh
and water had user accuracies of 89% and 75%, respectively. Table 2 shows the compari-
son of all classifications in terms of producer accuracies, user accuracies, overall accura-
cies, and kappa coefficients.
Table 2. Performance metrics for each of the individual scale classifications, and the two multiscale
methods. PA refers to producer’s accuracy and UA refers to user’s accuracy.
Marsh Mud Oyster Water
PA UA PA UA PA UA PA UA Overall Kappa
3 cm 92% 90% 49% 86% 87% 74% 96% 75% 80% 0.736
5 cm 92% 86% 73% 77% 78% 80% 88% 86% 82% 0.764
7 cm 88% 89% 73% 71% 72% 84% 87% 76% 79% 0.727
9 cm 91% 92% 69% 83% 75% 77% 90% 74% 81% 0.743
11 cm 93% 92% 80% 74% 75% 78% 77% 83% 81% 0.749
13 cm 89% 93% 72% 74% 74% 76% 83% 77% 79% 0.723
15 cm 88% 95% 74% 74% 81% 71% 77% 82% 80% 0.728
17 cm 91% 89% 78% 75% 73% 84% 86% 80% 82% 0.757
19 cm 85% 95% 56% 80% 80% 76% 93% 69% 78% 0.713
21 cm 93% 91% 72% 78% 78% 82% 87% 78% 82% 0.759
23 cm 88% 96% 67% 74% 78% 81% 87% 72% 80% 0.732
Figure 11.
The class most often represented in each raster cell (
left
) and how often that class was
represented out of the 15 classifications (right).
Table 2.
Performance metrics for each of the individual scale classifications, and the two multiscale
methods. PA refers to producer’s accuracy and UA refers to user’s accuracy.
Marsh Mud Oyster Water
PA UA PA UA PA UA PA UA Overall Kappa
3 cm 92% 90% 49% 86% 87% 74% 96% 75% 80% 0.736
5 cm 92% 86% 73% 77% 78% 80% 88% 86% 82% 0.764
7 cm 88% 89% 73% 71% 72% 84% 87% 76% 79% 0.727
9 cm 91% 92% 69% 83% 75% 77% 90% 74% 81% 0.743
11 cm 93% 92% 80% 74% 75% 78% 77% 83% 81% 0.749
13 cm 89% 93% 72% 74% 74% 76% 83% 77% 79% 0.723
15 cm 88% 95% 74% 74% 81% 71% 77% 82% 80% 0.728
17 cm 91% 89% 78% 75% 73% 84% 86% 80% 82% 0.757
19 cm 85% 95% 56% 80% 80% 76% 93% 69% 78% 0.713
21 cm 93% 91% 72% 78% 78% 82% 87% 78% 82% 0.759
23 cm 88% 96% 67% 74% 78% 81% 87% 72% 80% 0.732
25 cm 88% 90% 72% 78% 78% 81% 89% 77% 81% 0.751
27 cm 95% 89% 67% 83% 82% 71% 82% 83% 81% 0.745
29 cm 85% 90% 60% 82% 73% 84% 97% 64% 78% 0.708
31 cm 87% 91% 75% 79% 79% 83% 89% 77% 82% 0.763
Mode 95% 93% 76% 87% 86% 84% 91% 83% 87% 0.821
Multiscale
92% 89% 58% 88% 86% 86% 96% 75% 83% 0.778
Variable importance was similar to that of the individual resolutions, with the DSM
means as the most influential variable. However, 8 of 12 spectral standard deviations
carried more importance than the DSM standard deviations. Figure A2 displays the mean
decrease in Gini for the multiscale classification.
4. Discussion
4.1. Classification Performance
The GEOBIA classifications performed well at every resolution tested. The range
of resolutions tested represents high-resolution UAS imagery (3 cm) up to what can be
accessed via commercial satellite imagery (31 cm). Espriella et al. (2020) used GEOBIA
to classify intertidal habitats with high accuracy. However, the workflow relied solely on
the native resolution of the imagery, 6 mm. This ultra-high-resolution imagery resulted in
lengthy processing times. The current study is unique in its exploration of different input

Drones 2022,6, 140 13 of 19
resolutions in the GEOBIA workflow. While Espriella et al. (2020) used a multitude of layers,
including spectral indices and geometric properties to develop a classification workflow,
this study relies only on the orthomosaic and DSM to further encourage straightforward
and consistent monitoring.
The high performance of the classification workflow over the range of resolutions
indicates that targeting excessively fine resolutions on the scale of millimeters or 1 cm
may not be necessary for intertidal habitat mapping purposes along Florida’s Big Bend
coastline. Coarser resolutions correspond with higher flying heights for UASs, meaning
more area can be covered in a single mission. The ability to yield greater coverage without
sacrificing classification accuracy has significant implications for monitoring agencies that
have limited resources. The 19 and 29 cm resolutions performed more modestly than the
other resolutions, but still produced high-performance metrics for habitat mapping applica-
tions. Despite their lower accuracies, kappa coefficients of 0.713 (19 cm) and 0.708 (
29 cm
)
are very promising metrics. Kappa values between 0.6 and 0.8 are considered to have a
‘substantial’ agreement, which is the case for all tested resolutions [
50
]. Additionally, there
was substantial agreement between the classifications at multiple scales (Figure 11). This
agreement provides further evidence that for habitat mapping, very fine resolutions may
not be necessary for this environment. Despite the high performances of the classifications
at resolutions that coincide with occupied aerial or satellite imagery, it is important to note
that such imagery is rarely captured at the ideal conditions (e.g., lowest tides and adequate
scene illumination) allowable by the flexibility of UAS surveys.
To increase the accuracy of the classifications, steps should be explored to better
differentiate mud, as it was the class that performed the worst in terms of producer’s
accuracy in 12 of the 15 resolutions tested. The mud class was most often confused with
water, followed by the oyster. From a classification perspective, mud has the disadvantage
of having qualities that may be consistent with water when examining the DSM (i.e.,
relatively low elevation and homogenous surface), while having qualities that may be
consistent with oysters in terms of the spectral response characterized in the orthomosaic.
Although classification accuracy is useful to quantify the reliability of a habitat map,
there is often limited understanding of the location of errors or areas of low accuracy [
51
].
As evidenced by Figure 11, classification errors are frequently related to habitat patch
boundaries and transitionary zones [
51
–
53
]. The map displaying the percent agreement
highlights how the edges of habitats coincided with lower levels of agreement among classi-
fications. The southwest corner of the study area was also a challenging area. This area was
primarily mud that was partially inundated with water at the time of the survey, explaining
why there was disagreement among the classifications at the different resolutions. Figure 7
also provides insight into challenges with boundaries. While most classifications produced
similar habitat areas, the 15 cm resolution produced significantly less marsh habitat, as the
edges of the habitat were often segmented into water objects. The inclusion of structural
features beyond the DSM has the potential to improve the segmentation in these areas.
Additionally, classifications based on fuzzy logic principles would potentially help improve
results in these challenging environments as they can better represent natural transitional
areas between habitat types; they do not impose discrete boundaries between adjacent
habitat types [
54
]. Fuzzy logic and other soft classification principles also allow objects
to belong to more than one class. This could improve areas such as mud inundated with
water and fringing oysters that integrate with the salt marsh habitat.
While the classifications performed well at each resolution, there were notable findings
regarding which scale best characterized each cover type. Water and mud performed best
in terms of user accuracies at finer scales of 5 and 3 cm, respectively. Conversely, marsh
and oyster habitats produced the highest user accuracies at coarser resolutions of 23 and
29 cm
, respectively. Selecting a proper observational scale is an ongoing area of research as
no one scale is appropriate for the study of all ecological processes [
19
,
55
–
57
]. Multiscale
workflows can help address the limitations of single-scale workflows by highlighting those
that are most informative, a consideration that is as important as variable selection [19].

Drones 2022,6, 140 14 of 19
Lecours and Espriella (2020) suggest that multiscale roughness metrics can help de-
lineate intertidal habitats [
57
]. The study finds that mudflats present a comparatively
lower magnitude of roughness, with the highest roughness values captured at coarse scales.
Conversely, the marsh habitat displayed the highest magnitude of roughness among the
habitats and did so at a fine scale. Oyster reefs displayed an intermediate magnitude of
roughness at an intermediate scale when compared to mudflat and marsh habitats. These
findings within the context of the current study highlight the significance of variable im-
portance, relative to the scale (Figure A2). The mean of the DSM as the most influential
parameter across the tested resolutions is logical, given the varying elevations across the
cover types. The standard deviation of the DSM and the standard deviation of the blue
spectral band were the next most influential parameters, following the DSM mean. The
standard deviation of the DSM exhibited a higher level of importance at coarser scales
when compared to the blue spectral band standard deviation. The standard deviation of
the DSM highlights topographic heterogeneity, and with the assumption that marsh and
oyster habitats are more structurally complex than mud and water, should contribute to
the classification accordingly. However, it is notable that in terms of classification accuracy,
the habitats performed counter to these assumptions. Marsh and oyster habitats had the
highest performances in terms of user accuracies at coarser scales, while mud and water
had their highest performance at finer scales. Despite these differences, it is also important
to consider that classifications performed well across the resolutions (e.g., the marsh class
displayed its highest user accuracy of 96% at 23 cm, but also had a user accuracy of 90% at
3 cm). These differences again highlight that scale selection should be context-dependent.
The multiscale classification conducted further demonstrates the relevancy of using
multiscale workflows. The multiscale classification performed better than any individual
resolution in terms of the kappa coefficient and provided further insight into which vari-
ables are most important to consider (Figure A2). As with the single-scale classifications,
the DSM means were the most influential. However, unlike the single-scale classifications,
the standard deviations of the spectral bands, highlighting spectral heterogeneity, were
often more influential than the DSM standard deviations (Figure A2). Additionally, the
spectral variation was most influential at finer scales, suggesting the signatures of the
habitat types may be lost at coarser scales. The classification generated by the mode of each
individual resolution produced a very high kappa coefficient of 0.821, demonstrating the
utility of quantifying a classification agreement (Figure 11). However, this method was
used primarily to understand areas of uncertainty and would be computationally intensive
to rely on for consistent monitoring.
4.2. Considerations, Challenges, and Future Directions
While the Orfeo ToolBox large-scale mean shift algorithm produced quality segmen-
tations, there were initial challenges with arriving at an appropriate segmentation for the
scene. The segmentation developed artifacts across flight paths, likely caused by a slight
variation in atmospheric conditions and tide levels that generated artificial changes in spec-
tral response. As a result, objects would display boundaries in a north–south orientation,
given the east–west flight pattern of the UAS. The SegOptim package supports multiple
segmentation algorithms from third-party software. The RSGISLib Shepherd’s k-means
segmentation was tested before the Orfeo ToolBox large-scale mean shift segmentation.
However, after exploring the data and segmentations, there was significant difficulty in
reconciling the flight path artifacts. The large-scale mean shift better handled any artifacts
introduced by flight paths. In future studies, this can be partially addressed by including a
light sensor on the UAS that normalizes spectral responses by measuring incoming light.
Another difficulty was posed during the testing of classifications in the optimization step.
In order for the optimization to run completely, the validation must produce a specified
minimum number of train cases and test cases; regarding the optimizations conducted
for this study, five of each were required (Table A1). For this dataset, this meant finding a
middle ground that generated representative objects that did not compromise their shapes

Drones 2022,6, 140 15 of 19
or sizes, while maintaining enough objects to meet the testing parameters. For example,
setting the spectral range radius too high would result in all water objects in the subset
for optimization being merged. As a result, there would not be sufficient objects to run
validation steps. Another processing consideration was the inclusion of the DSM. The
DSM was not included in the optimization procedure, as the optimization is very com-
putationally intensive. Additionally, non-expert end users may not have the ability or
resources to develop quality digital surface models. Therefore, by excluding it from the
optimization, it ensures end users can rely on the optimized segmentation parameters
when only using an orthomosaic (or potentially a single) image in the case of an occupied
aircraft or satellite imagery.
Despite the near-equal performances across the different resolutions for classification,
results for studies that target other objectives, such as analyzing habitat geometries, could
prove different as the resolution changes. For example, should one be interested in cal-
culating the perimeter of oyster reefs for spatial analysis, the resolution is an important
consideration. To illustrate this fact, a classified oyster reef at the 5 cm resolution had an
area of 2106 m
2
and a perimeter of 1879 m, while the same reef had an area of 2005 m
2
and a perimeter of 949 m at the 31 cm resolution. While the area was largely preserved,
seeing only a 4.9% difference between the resolutions, the perimeters represent a 65.8%
difference. This difference highlights how researchers must consider the scale-dependency
of metrics, as the area proved to be less scale-dependent than the perimeter. Ultimately,
it is essential to consider objectives when selecting an appropriate target resolution [
16
].
No one scale of observation is suitable for all objectives, and when using imagery to study
ecological processes, the scale should be considered within the context of the biological and
environmental processes being studied [16,19,58,59].
While this study used only three spectral bands, given the accessibility of RGB sensors,
spectral resolution is another important consideration in habitat mapping workflows. For
example, Chand and Bollard (2021) suggest that spectral resolution is more critical to classify
oyster reefs than spatial resolution, as an RGB dataset with a higher spatial resolution did
not perform as well as a coarser dataset that included red-edge and near-infrared bands [
60
].
The inclusion of a near infrared band has the potential to better differentiate water and
mud, where most misclassifications occurred, given its ability to differentiate water and
non-water bodies [
61
]. Optimizing spectral resolution along with spatial resolution will
further improve habitat classifications and our understanding of what metrics to target in
data collection.
The GEOBIA workflow detailed in this study was conducted within R using open-
source solutions. An example of the workflow is made available in the Supplemen-
tary Materials. This study was able to achieve a similar level of accuracy as the one
by
Espriella et al. (2020)
, which relied on proprietary software. This accessibility allows for
consistent, holistic monitoring of intertidal habitats along Florida’s Gulf of Mexico coastline.
Future surveys can use the results of this study to decide what is an appropriate resolution,
or resolutions, for the monitoring objectives. This workflow streamlines data collection and
processing for future studies in the intertidal. While altitude (and indirectly, resolution) has
the strongest effect on flight time, other flight parameters, such as image overlaps, can also
be studied to quantify trade-offs and further streamline the process [13].
5. Conclusions
Florida’s intertidal habitats along the Gulf of Mexico coast are dynamic ecosystems
experiencing significant changes. Generating reliable methods to monitor these resources
allows management to appropriately respond to their potential deterioration. Remote sensing
platforms such as UASs provide resources to conduct consistent temporal monitoring. Here,
classifications were conducted using GEOBIA to determine how scale affects the accuracy of
habitat maps produced at different resolutions. In general, results showed that classification
accuracy did not change dramatically when coarsening data resolution, which has direct
implications for data collection and processing. Results also (1) confirmed that one single scale

Drones 2022,6, 140 16 of 19
is not optimal for the study of all habitat types; (2) showed how promising the implementation
of multiscale analysis is for coastal habitat classification; and (3) highlighted that fuzzy
classifications should be explored to capture the areas of transition among habitat types,
which were often the sources of misclassification. Coupling information obtained from sensors
with a repeatable workflow that is freely available provides managers with the necessary
tools to conduct consistent monitoring and identifies scales and variables of interest for three
intertidal habitat types. The inclusion of GEOBIA within the workflow also creates many other
possibilities, such as quantifying spatial relationships among habitats or tracking geometries,
as commonly performed in seascape ecology. Habitat maps are valuable tools for quantifying
change; this study contributes to the present knowledge gaps regarding best practices to
collect and process high-quality data.
Supplementary Materials:
The following supporting information can be downloaded at: https:
//www.mdpi.com/article/10.3390/drones6060140/s1, R code example.
Author Contributions: Conceptualization, M.C.E. and V.L.; methodology, M.C.E.; software, M.C.E.;
validation, M.C.E.; formal analysis, M.C.E.; investigation, M.C.E.; resources, M.C.E. and V.L.; data
curation, M.C.E.; writing—original draft preparation, M.C.E.; writing—review and editing, M.C.E.
and V.L.; visualization, M.C.E.; supervision, V.L.; project administration, V.L.; funding acquisition,
V.L. All authors have read and agreed to the published version of the manuscript.
Funding:
This work was supported by the USDA National Institute of Food and Agriculture (Hatch
project FLA-FOR-005981; multiscale analysis of oyster resources using geomorphometry and remote
sensing technologies), by funds allocated to V.L. by the University of Florida Senior Vice President
for Agriculture and Natural Resources, and by an Early-Career Research Fellowship to V.L. from the
Gulf Research Program of the National Academies of Sciences, Engineering, and Medicine.
Informed Consent Statement: Not applicable.
Data Availability Statement: Data can be made available upon reasonable request.
Acknowledgments:
Thanks are due to Peter Frederick for providing the imagery and Steve Beck,
Sean Denney, Lindsey Garner, and Andrew Ortega who led the imagery and RTK data collection.
Additional thanks go out to João Gonçalves for providing support with the SegOptim package.
Conflicts of Interest: The authors declare no conflict of interest.
Appendix A
Table A1. Parameters used in the SegOptim optimization step.
Parameter Value Used
Training threshold 0.75
Balance method “ubOver” (over-sampling)
Evaluation metric “Kappa” (kappa coefficient)
Evaluation method “10FCV” (10-fold cross-validation)
Min train cases 20
Min cases by class train 5
Min cases by class test 5
Population size 20
Probability of crossover 0.8
Probability of mutation 0.2
Max number of iterations 5
Run 20
Keep best TRUE

Drones 2022,6, 140 17 of 19
Drones 2022, 6, x FOR PEER REVIEW 17 of 20
Acknowledgments: Thanks are due to Peter Frederick for providing the imagery and Steve Beck,
Sean Denney, Lindsey Garner, and Andrew Ortega who led the imagery and RTK data collection.
Additional thanks go out to João Gonçalves for providing support with the SegOptim package.
Conflicts of Interest: The authors declare no conflict of interest.
Appendix A
Table A1. Parameters used in the SegOptim optimization step.
Parameter Value Used
Training threshold 0.75
Balance method “ubOver” (over-sampling)
Evaluation metric “Kappa” (kappa coefficient)
Evaluation method “10FCV” (10-fold cross-validation)
Min train cases 20
Min cases by class train 5
Min cases by class test 5
Population size 20
Probability of crossover 0.8
Probability of mutation 0.2
Max number of iterations 5
Run 20
Keep best TRUE
Figure A1. The 5 cm segmentation with (left) and without (right) the water object. In the case of the
5 cm resolution, water was segmented into a single object. Given the spectral homogeneity of the
water, multiple objects were merged into one because adjacent objects did not exceed the spectral
threshold outlined by the spectral range radius parameter.
Figure A1.
The 5 cm segmentation with (
left
) and without (
right
) the water object. In the case of the
5 cm resolution, water was segmented into a single object. Given the spectral homogeneity of the
water, multiple objects were merged into one because adjacent objects did not exceed the spectral
threshold outlined by the spectral range radius parameter.
Drones 2022, 6, x FOR PEER REVIEW 18 of 20
Figure A2. The mean decrease in Gini for each metric included in the multiscale classification.
References
1. Wiens, J.A. Landscape Mosaics and Ecological Theory. In Mosaic Landscapes and Ecological Processes; Hansson, L., Fahrig, L.,
Merriam, G., Eds.; Springer: Dordrecht, The Netherlands, 1995; pp. 1–26.
2. Wedding, L.M.; Lepczyk, C.A.; Pittman, S.J.; Friedlander, A.M.; Jorgensen, S. Quantifying Seascape Structure: Extending Ter-
restrial Spatial Pattern Metrics to the Marine Realm. Mar. Ecol. Prog. Ser. 2011, 427, 219–232.
3. Olds, A.D.; Nagelkerken, I.; Huijbers, C.M.; Gilby, B.L.; Pittman, S.J.; Schlacher, T.A. Connectivity in Coastal Seascapes. In Sea-
scape Ecology, 1st ed.; Pittman, S.J., Ed.; John Wiley & Sons Ltd.: Hoboken, NJ, USA, 2018; pp. 261–364.
4. Pittman, S.J. Introducing Seascape Ecology. In Seascape Ecology, 1st ed.; Pittman, S.J., Ed.; John Wiley & Sons Ltd.: Hoboken, NJ,
USA, 2018; pp. 1–25.
5. Meyer, D.L.; Townsend, E.C.; Thayer, G.W. Stabilization and Erosion Control Value of Oyster Cultch for Intertidal Marsh. Res-
tor. Ecol. 1997, 5, 93–99.
6. Grabowski, J.H.; Hughes, A.R.; Kimbro, D.L.; Dolan, M.A. How Habitat Setting Influences Restored Oyster Reef Communities.
Ecology 2005, 86, 1926–1935.
7. Hosack, G.R.; Dumbauld, B.R.; Ruesink, J.L.; Armstrong, D.A. Habitat Associations of Estuarine Species: Comparisons of Inter-
tidal Mudflat, Seagrass (Zostera Marina), and Oyster (Crassostrea Gigas) Habitats. Estuaries Coasts 2006, 29, 1150–1160.
8. Smyth, A.R.; Piehler, M.F.; Grabowski, J.H. Habitat Context Influences Nitrogen Removal by Restored Oyster Reefs. J. Appl.
Ecol. 2015, 52, 716–725.
9. Windle, A.E.; Poulin, S.K.; Johnston, D.W.; Ridge, J.T. Rapid and Accurate Monitoring of Intertidal Oyster Reef Habitat Using
Unoccupied Aircraft Systems and Structure from Motion. Remote Sens. 2019, 11, 2394.
10. Espriella, M.C.; Lecours, V.; Frederick, P.C.; Camp, E.V.; Wilkinson, B. Quantifying Intertidal Habitat Relative Coverage in a
Florida Estuary Using UAS Imagery and GEOBIA. Remote Sens. 2020, 12, 677.
11. Ridge, J.T.; Gray, P.C.; Windle, A.E.; Johnston, D.W. Deep Learning for Coastal Resource Conservation: Automating Detection
of Shellfish Reefs. Remote Sens. Ecol. Conserv. 2020, 6, 431–440.
12. Fraser, B.T.; Congalton, R.G. Issues in Unmanned Aerial Systems (UAS) Data Collection of Complex Forest Environments. Re-
mote Sens. 2018, 10, 908.
13. Seifert, E.; Seifert, S.; Vogt, H.; Drew, D.; van Aardt, J.; Kunneke, A.; Seifert, T. Influence of Drone Altitude, Image Overlap, and
Optical Sensor Resolution on Multi-View Reconstruction of Forest Images. Remote Sens. 2019, 11, 1252.
14. Whitehead, K.; Hugenholtz, C.H. Remote Sensing of the Environment with Small Unmanned Aircraft Systems (UASs), Part 1:
A Review of Progress and Challenges. J. Unmanned Veh. System 2014, 2, 69–85.
15. Torres-Sánchez, J.; López-Granados, F.; Borra-Serrano, I.; Peña, J.M. Assessing UAV-Collected Image Overlap Influence on
Computation Time and Digital Surface Model Accuracy in Olive Orchards. Precis. Agric. 2018, 19, 115–133.
16. Lecours, V.; Devillers, R.; Schneider, D.; Lucieer, V.; Brown, C.; Edinger, E. Spatial Scale and Geographic Context in Benthic
Habitat Mapping: Review and Future Directions. Mar. Ecol. Prog. Ser. 2015, 535, 259–284.
17. Misiuk, B.; Lecours, V.; Dolan, M.F.J.; Robert, K. Evaluating the Suitability of Multi-Scale Terrain Attribute Calculation Ap-
proaches for Seabed Mapping Applications. Mar. Geod. 2021, 44, 327–385.
Figure A2. The mean decrease in Gini for each metric included in the multiscale classification.
References
1.
Wiens, J.A. Landscape Mosaics and Ecological Theory. In Mosaic Landscapes and Ecological Processes; Hansson, L., Fahrig, L.,
Merriam, G., Eds.; Springer: Dordrecht, The Netherlands, 1995; pp. 1–26.
2.
Wedding, L.M.; Lepczyk, C.A.; Pittman, S.J.; Friedlander, A.M.; Jorgensen, S. Quantifying Seascape Structure: Extending
Terrestrial Spatial Pattern Metrics to the Marine Realm. Mar. Ecol. Prog. Ser. 2011,427, 219–232. [CrossRef]
3.
Olds, A.D.; Nagelkerken, I.; Huijbers, C.M.; Gilby, B.L.; Pittman, S.J.; Schlacher, T.A. Connectivity in Coastal Seascapes. In Seascape
Ecology, 1st ed.; Pittman, S.J., Ed.; John Wiley & Sons Ltd.: Hoboken, NJ, USA, 2018; pp. 261–364.
4.
Pittman, S.J. Introducing Seascape Ecology. In Seascape Ecology, 1st ed.; Pittman, S.J., Ed.; John Wiley & Sons Ltd.: Hoboken, NJ,
USA, 2018; pp. 1–25.

Drones 2022,6, 140 18 of 19
5.
Meyer, D.L.; Townsend, E.C.; Thayer, G.W. Stabilization and Erosion Control Value of Oyster Cultch for Intertidal Marsh. Restor.
Ecol. 1997,5, 93–99. [CrossRef]
6.
Grabowski, J.H.; Hughes, A.R.; Kimbro, D.L.; Dolan, M.A. How Habitat Setting Influences Restored Oyster Reef Communities.
Ecology 2005,86, 1926–1935. [CrossRef]
7.
Hosack, G.R.; Dumbauld, B.R.; Ruesink, J.L.; Armstrong, D.A. Habitat Associations of Estuarine Species: Comparisons of
Intertidal Mudflat, Seagrass (Zostera Marina), and Oyster (Crassostrea Gigas) Habitats. Estuaries Coasts
2006
,29, 1150–1160.
[CrossRef]
8.
Smyth, A.R.; Piehler, M.F.; Grabowski, J.H. Habitat Context Influences Nitrogen Removal by Restored Oyster Reefs. J. Appl. Ecol.
2015,52, 716–725. [CrossRef]
9.
Windle, A.E.; Poulin, S.K.; Johnston, D.W.; Ridge, J.T. Rapid and Accurate Monitoring of Intertidal Oyster Reef Habitat Using
Unoccupied Aircraft Systems and Structure from Motion. Remote Sens. 2019,11, 2394. [CrossRef]
10.
Espriella, M.C.; Lecours, V.; Frederick, P.C.; Camp, E.V.; Wilkinson, B. Quantifying Intertidal Habitat Relative Coverage in a
Florida Estuary Using UAS Imagery and GEOBIA. Remote Sens. 2020,12, 677. [CrossRef]
11.
Ridge, J.T.; Gray, P.C.; Windle, A.E.; Johnston, D.W. Deep Learning for Coastal Resource Conservation: Automating Detection of
Shellfish Reefs. Remote Sens. Ecol. Conserv. 2020,6, 431–440. [CrossRef]
12.
Fraser, B.T.; Congalton, R.G. Issues in Unmanned Aerial Systems (UAS) Data Collection of Complex Forest Environments. Remote
Sens. 2018,10, 908. [CrossRef]
13.
Seifert, E.; Seifert, S.; Vogt, H.; Drew, D.; van Aardt, J.; Kunneke, A.; Seifert, T. Influence of Drone Altitude, Image Overlap, and
Optical Sensor Resolution on Multi-View Reconstruction of Forest Images. Remote Sens. 2019,11, 1252. [CrossRef]
14.
Whitehead, K.; Hugenholtz, C.H. Remote Sensing of the Environment with Small Unmanned Aircraft Systems (UASs), Part 1: A
Review of Progress and Challenges. J. Unmanned Veh. System 2014,2, 69–85. [CrossRef]
15.
Torres-Sánchez, J.; López-Granados, F.; Borra-Serrano, I.; Peña, J.M. Assessing UAV-Collected Image Overlap Influence on
Computation Time and Digital Surface Model Accuracy in Olive Orchards. Precis. Agric. 2018,19, 115–133. [CrossRef]
16.
Lecours, V.; Devillers, R.; Schneider, D.; Lucieer, V.; Brown, C.; Edinger, E. Spatial Scale and Geographic Context in Benthic
Habitat Mapping: Review and Future Directions. Mar. Ecol. Prog. Ser. 2015,535, 259–284. [CrossRef]
17.
Misiuk, B.; Lecours, V.; Dolan, M.F.J.; Robert, K. Evaluating the Suitability of Multi-Scale Terrain Attribute Calculation Approaches
for Seabed Mapping Applications. Mar. Geod. 2021,44, 327–385. [CrossRef]
18.
Miyamoto, M.; Kiyota, M.; Murase, H.; Nakamura, T.; Hayashibara, T. Effects of Bathymetric Grid-Cell Sizes on Habitat Suitability
Analysis of Cold-Water Gorgonian Corals on Seamounts. Mar. Geod. 2017,40, 205–223. [CrossRef]
19. Misiuk, B.; Lecours, V.; Bell, T. A Multiscale Approach to Mapping Seabed Sediments. PLoS ONE 2018,13, e0193647.
20.
Florinsky, I.V.; Kuryakova, G.A. Determination of Grid Size for Digital Terrain Modelling in Landscape Investigations—Exemplified
by Soil Moisture Distribution at a Micro-Scale. Int. J. Geogr. Inf. Sci. 2000,14, 815–832. [CrossRef]
21.
Gottschalk, T.K.; Aue, B.; Hotes, S.; Ekschmitt, K. Influence of Grain Size on Species–Habitat Models. Ecol. Model.
2011
,222,
3403–3412. [CrossRef]
22.
Dolan, M.F.J.; Van Lancker, V.; Guinan, J.; Al-Hamdani, Z.; Leth, J.; Thorsnes, T. Terrain Characterization from Bathymetry Data at
Various Resolutions in European Waters—Experiences and Recommendations; Geological Survey of Norway Report No. 2012.045;
Geological Survey of Norway: Trondheim, Norway, 2012.
23.
Blanchet, H.; Gouillieux, B.; Alizier, S.; Amouroux, J.-M.; Bachelet, G.; Barillé, A.-L.; Dauvin, J.-C.; de Montaudouin, X.;
Derolez, V.
;
Desroy, N.; et al. Multiscale Patterns in the Diversity and Organization of Benthic Intertidal Fauna among French Atlantic
Estuaries. J. Sea Res. 2014,90, 95–110. [CrossRef]
24.
Azhar, M.; Schenone, S.; Anderson, A.; Gee, T.; Cooper, J.; van der Mark, W.; Hillman, J.R.; Yang, K.; Thrush, S.F.; Delmas, P. A
Framework for Multiscale Intertidal Sandflat Mapping: A Case Study in the Whangateau Estuary. ISPRS J. Photogramm. Remote
Sens. 2020,169, 242–252. [CrossRef]
25.
Seavey, J.R.; Pine III, W.E.; Frederick, P.; Sturmer, L.; Berrigan, M. Decadal Changes in Oyster Reefs in the Big Bend of Florida’s
Gulf Coast. Ecosphere 2011,2, 1–14. [CrossRef]
26.
Radabaugh, K.R.; Geiger, S.P.; Moyer, P.P. Oyster Integrated Mapping and Monitoring Program Report for the State of Florida; FWRI
Technical Report No. 22; Fish and Wildlife Research Institute, Florida Fish and Wildlife Conservation Commission: St. Petersburg,
FL, USA, 2019.
27. McCarthy, M.J.; Dimmitt, B.; Muller-Karger, F.E. Rapid Coastal Forest Decline in Florida’s Big Bend. Remote Sens. 2018,10, 1721.
[CrossRef]
28.
Vitale, N.; Brush, J.; Powell, A. Loss of Coastal Islands Along Florida’s Big Bend Region: Implications for Breeding American
Oystercatchers. Estuaries Coasts 2021,44, 1173–1182. [CrossRef]
29.
Main, M.B.; Allen, G.M. Florida’s Environment: North Central Region; Wildlife Ecology and Conservation Department, Florida
Cooperative Extension Service, Institute of Food and Agricultural Sciences, University of Florida: Gainesville, FL, USA, 2007.
30.
Moore, J.F.; Pine, W.E., III; Frederick, P.C.; Beck, S.; Moreno, M.; Dodrill, M.J.; Boone, M.; Sturmer, L.; Yurek, S. Trends in Oyster
Populations in the Northeastern Gulf of Mexico: An Assessment of River Discharge and Fishing Effects over Time and Space.
Mar. Coast. Fish. 2020,12, 191–204. [CrossRef]
31.
Pix4D Mapper [Computer Software]. Available online: https://www.pix4d.com/product/pix4dmapperphotogrammetry-
software (accessed on 1 March 2022).

Drones 2022,6, 140 19 of 19
32. ESRI ArcGIS Pro v 2.4 [Computer Software]. Available online: https://pro.arcgis.com/es/pro-app (accessed on 1 March 2022).
33.
Alvarez-Berastegui, D.; Ciannelli, L.; Aparicio-Gonzalez, A.; Reglero, P.; Hidalgo, M.; López-Jurado, J.L.; Tintoré, J.; Alemany, F.
Spatial Scale, Means and Gradients of Hydrographic Variables Define Pelagic Seascapes of Bluefin and Bullet Tuna Spawning
Distribution. PLoS ONE 2014,9, e109338. [CrossRef]
34.
Scales, K.L.; Hazen, E.L.; Jacox, M.G.; Edwards, C.A.; Boustany, A.M.; Oliver, M.J.; Bograd, S.J. Scale of Inference: On the
Sensitivity of Habitat Models for Wide-Ranging Marine Predators to the Resolution of Environmental Data. Ecography
2017
,40,
210–220. [CrossRef]
35.
Pittman, S.J.; Brown, K.A. Multi-Scale Approach for Predicting Fish Species Distributions across Coral Reef Seascapes. PLoS ONE
2011,6, e20583. [CrossRef]
36.
Schneider, D.C. Scale and Scaling in Seascape Ecology. In Seascape Ecology, 1st ed.; Pittman, S.J., Ed.; John Wiley & Sons Ltd.:
Hoboken, NJ, USA, 2018; pp. 89–117.
37.
Gibbes, C.; Adhikari, S.; Rostant, L.; Southworth, J.; Qiu, Y. Application of Object Based Classification and High Resolution
Satellite Imagery for Savanna Ecosystem Analysis. Remote Sens. 2010,2, 2748–2772. [CrossRef]
38. Blaschke, T. Object Based Image Analysis for Remote Sensing. ISPRS J. Photogramm. Remote Sens. 2010,65, 2–16. [CrossRef]
39.
Diesing, M.; Green, S.L.; Stephens, D.; Lark, R.M.; Stewart, H.A.; Dove, D. Mapping Seabed Sediments: Comparison of Manual,
Geostatistical, Object-Based Image Analysis and Machine Learning Approaches. Cont. Shelf Res. 2014,84, 107–119. [CrossRef]
40.
Gonçalves, J.; Pôças, I.; Marcos, B.; Mücher, C.A.; Honrado, J.P. SegOptim—A New R Package for Optimizing Object-Based Image
Analyses of High-Spatial Resolution Remotely-Sensed Data. Int. J. Appl. Earth Obs. Geoinf. 2019,76, 218–230. [CrossRef]
41. Grizonnet, M.; Michel, J.; Poughon, V.; Inglada, J.; Savinaud, M.; Cresson, R. Orfeo ToolBox: Open Source Processing of Remote
Sensing Images. Open Geospat. Data Softw. Stand. 2017,2, 15. [CrossRef]
42.
Michel, J.; Youssefi, D.; Grizonnet, M. Stable Mean-Shift Algorithm and Its Application to the Segmentation of Arbitrarily Large
Remote Sensing Images. IEEE Trans. Geosci. Remote Sens. 2015,53, 952–964. [CrossRef]
43. OTB Development Team. OTB CookBook Documentation; CNES: Paris, France, 2018.
44.
Jensen, J.R. Introductory Digital Image Processing: A Remote Sensing Perspective, 3rd ed.; Pearson Education: Upper Saddle River, NJ,
USA, 2005.
45.
Congalton, R.G. A review of assessing the accuracy of classifications in remotely sensed data. Remote Sens. Environ.
1991
,37,
35–46. [CrossRef]
46.
Fung, T.; LeDrew, E. The determination of optimal threshold levels for change detection using various accuracy indices.
Photogramm. Eng. Remote Sens. 1988,54, 1449–1454.
47. Story, M.; Congalton, R.G. Accuracy assessment: A user’s perspective. Photogramm. Eng. Remote Sens. 1986,52, 397–399.
48.
Luz Calle, M.; Urrea, V. Letter to the Editor: Stability of Random Forest Importance Measures. Brief. Bioinform.
2011
,12, 86–89.
[CrossRef]
49.
Han, H.; Guo, X.; Yu, H. Variable Selection Using Mean Decrease Accuracy and Mean Decrease Gini Based on Random Forest. In
Proceedings of the 2016 7th IEEE International Conference on Software Engineering and Service Science (ICSESS), Beijing, China,
26–28 August 2016; pp. 219–224.
50.
Landis, J.R.; Kock, G.G. The measurement of observer agreement for categorical data. Biometrics
1977
,33, 159–174. [CrossRef]
[PubMed]
51.
Langford, W.T.; Gergel, S.E.; Dietterich, T.G.; Cohen, W. Map Misclassification Can Cause Large Errors in Landscape Pattern
Indices: Examples from Habitat Fragmentation. Ecosystems 2006,9, 474–488. [CrossRef]
52.
Edwards, G.; Lowell, K.E. Modeling Uncertainty in Photointerpreted Boundaries. Photogramm. Eng. Remote Sens.
1996
,15,
377–391.
53.
Plourde, L.; Congalton, R. Sampling Method and Sample Placement: How Do They Affect the Accuracy of Remotely Sensed
Maps? Photogramm. Eng. Remote Sens. 2003,69, 289–297. [CrossRef]
54.
Fiorentino, D.; Lecours, V.; Brey, T. On the Art of Classification in Spatial Ecology: Fuzziness as an Alternative for Mapping
Uncertainty. Front. Ecol. Evol. 2018,6, 231. [CrossRef]
55. Wiens, J.A. Spatial Scaling in Ecology. Funct. Ecol. 1989,3, 385–397. [CrossRef]
56. Willis, K.J.; Whittaker, R.J. Species Diversity—Scale Matters. Science 2002,295, 1245–1248. [CrossRef] [PubMed]
57.
Lecours, V.; Espriella, M. Can Multiscale Roughness Help Computer-Assisted Identification of Coastal Habitats in Florida? In
Proceedings of the Geomorphometry 2020 Conference, Perugia, Italy, 22–26 June 2020; IRPI CNR: Perugia, Italy, 2020.
58. Goodchild, M.F. Scale in GIS: An Overview. Geomorphology 2011,130, 5–9. [CrossRef]
59.
Bradter, U.; Kunin, W.E.; Altringham, J.D.; Thom, T.J.; Benton, T.G. Identifying Appropriate Spatial Scales of Predictors in Species
Distribution Models with the Random Forest Algorithm. Methods Ecol. Evol. 2013,4, 167–174. [CrossRef]
60.
Chand, S.; Bollard, B. Low Altitude Spatial Assessment and Monitoring of Intertidal Seagrass Meadows beyond the Visible
Spectrum Using a Remotely Piloted Aircraft System. Estuar. Coast. Shelf Sci. 2021,255, 107299. [CrossRef]
61.
Mondejar, J.P.; Tongco, A.F. Near Infrared Band of Landsat 8 as Water Index: A Case Study around Cordova and Lapu-Lapu City,
Cebu, Philippines. Sustain Environ. Res. 2019,29, 16. [CrossRef]