Can multiscale roughness help computer-assisted identification of coastal habitats in Florida?

Abstract

Coastal habitats are of natural, economic, and cultural importance in Florida, and there is a need for effective approaches to map and monitor them. Geographic Object-Based Image Analysis (GEOBIA) was previously applied to an orthomosaic and a Digital Surface Model (DSM) to automatically delineate oyster reef, salt marsh, and mudflat habitats in Little Trout Creek, Florida. Here we evaluated whether a multiscale measure of roughness has the potential to improve this GEOBIA workflow in this context where oysters are spectrally similar to the two other habitat types. Our results show that multiscale roughness can be used to distinguish the different coastal habitat types studied. The level of roughness of mudflats is usually higher at broader scales, and the magnitude of that roughness is relatively small. Marsh roughness was highest at finer scales, and its magnitude was higher compared to other habitat types likely due to marshes' vegetation cover, which is captured in the DSM. The highest magnitudes of roughness for oysters were smaller than, and found at slightly broader scales than, the highest roughness for marshes. Our results were strongly affected by edge effects because the studied habitats are discrete and discontinuous. Multiscale roughness has the potential to help delineate coastal habitats in Florida, but more work is needed to better understand the multiscale topographic patterns of different coastal habitats in Florida and elsewhere.

Full text

Can multiscale roughness help computer-assisted
identification of coastal habitats in Florida?
Vincent Lecours§ & Michael C. Espriella
School of Forest Resources & Conservation
University of Florida
7922 NW 71st Street, Gainesville, Florida, USA, 32606
§ vlecours@ufl.edu
Abstract—Coastal habitats are of natural, economic, and cultural
importance in Florida, and there is a need for effective approaches
to map and monitor them. Geographic Object-Based Image Analysis
(GEOBIA) was previously applied to an orthomosaic and a Digital
Surface Model (DSM) to automatically delineate oyster reef, salt
marsh, and mudflat habitats in Little Trout Creek, Florida. Here we
evaluated whether a multiscale measure of roughness has the
potential to improve this GEOBIA workflow in this context where
oysters are spectrally similar to the two other habitat types. Our
results show that multiscale roughness can be used to distinguish the
different coastal habitat types studied. The level of roughness of
mudflats is usually higher at broader scales, and the magnitude of
that roughness is relatively small. Marsh roughness was highest at
finer scales, and its magnitude was higher compared to other habitat
types likely due to marshes’ vegetation cover, which is captured in
the DSM. The highest magnitudes of roughness for oysters were
smaller than, and found at slightly broader scales than, the highest
roughness for marshes. Our results were strongly affected by edge
effects because the studied habitats are discrete and discontinuous.
Multiscale roughness has the potential to help delineate coastal
habitats in Florida, but more work is needed to better understand
the multiscale topographic patterns of different coastal habitats in
Florida and elsewhere.
I. INTRODUCTION
Coastal habitats like oyster reefs and salt marshes provide
economic opportunities as well as vital ecosystem services such as
shoreline erosion control, habitat and nursery for a variety of
species, and water filtration. However, many of these ecosystem
services are threatened by natural and anthropogenic factors (e.g.,
coastal development, sea-level rise, hurricanes). Mapping and
monitoring coastal habitats are critical to improving scientific
understanding of the complex dynamics of coastal ecosystems, to
better inform management, planning, and conservation efforts.
Florida’s coastal waters are the most economically valuable,
have the highest recreational use, and have one of the highest
concentrations of coastal communities in the United States. At a
2007 workshop, regional, state, and federal partners concluded that
although mapping coastal resources was a top priority, the lack of
a standard, reproducible approach was hindering broad-scale
efforts [1]. With the increased likelihood of extreme weather
events [2] that have the potential to impact coastal habitats in
Florida [3], there is a critical need to develop an effective and
reproducible mapping and monitoring workflow that can be used
to answer questions in a variety of contexts (e.g., sea-level rise,
community resilience, hurricane impact assessments).
In a recent article, Espriella et al. [4] proposed a reproducible
approach to detect and delineate three types of coastal habitats –
oyster reefs, salt marshes, and mudflats – in imagery collected with
Unoccupied Aircraft Systems (UAS). The approach is centered on
a two-level Geographic Object-Based Image Analysis (GEOBIA)
[5] that first identifies and extracts water areas from the data before
classifying the remaining objects into their respective habitats.
Both the RGB mosaic and the Digital Surface Model (DSM),
produced using structure-from-motion photogrammetry, were
used as inputs. However, with an overall classification accuracy of
79%, that GEOBIA alone may not be robust enough for accurate
temporal monitoring. Oysters had the lowest overall separability
from the other habitats, which is problematic from a management
perspective; oysters are one of the most important living coastal
resources actively managed in the Gulf of Mexico, are suffering
from major declines in the area, and thus are of particular interest.
In nature, oyster reefs are more structurally complex than
marsh and mudflats, at multiple scales. Therefore, we hypothesize
that geomorphometry can provide a means to help differentiate
these habitats from each other. While Espriella et al. [4] derived
local measures of terrain attributes (e.g., rugosity, relative
position) at multiple independent spatial scales [6] using relatively
few search neighborhoods [7], their feature-space optimization to
select the variables best fit to recognize the different habitats did
not identify any DSM-derived variables as being relevant. Here,
we evaluate the potential of a multiscale measure of roughness [8]
– as opposed to independent measures of roughness derived at
multiple scales – to help distinguish oyster reefs, salt marshes, and
mudflats from each other.
Vincent Lecours and Michael Espriella (2020) Can multiscale roughness help computer-assisted identification of coastal habitats in Florida?:
in Massimiliano Alvioli, Ivan Marchesini, Laura Melelli & Peter Guth, eds., Proceedings of the Geomorphometry 2020 Conference, doi:10.30437/GEOMORPHOMETRY2020_24.
111

II. METHODS
UAS imagery was collected on December 8th, 2018 at low tide,
at the mouth of Little Trout Creek (29° 15’ 34.98” N, 83° 4’ 29.68”
W), on the west coast of Florida (Fig. 1). The imagery was
collected at nadir using a DJI Inspire 2 equipped with a Zenmuse
X7 35 mm RGB sensor. The UAS was flown 60 m above ground
level, with an 80% along-track overlap and 75% across-track
overlap. Four checkered targets were evenly distributed across the
scene and located using a Trimble 5800 real-time kinematic
positioning system. In addition to the orthomosaic, a DSM was
produced using structure-from-motion photogrammetry in Pix4D
Mapper v. 4.2.27. The total area surveyed covered approximately
0.116 km2 and provided data with a 0.66 cm spatial resolution
(Fig. 2), with a root mean square error of 0.3 cm in longitude and
latitude, and 0.1 cm in elevation for the residuals of control points.
FIGURE 1. LOCATION OF THE STUDY AREA: LITTLE TROUT CREEK, FLORIDA, USA.
The GEOBIA ruleset of Espriella et al. [4] was applied to the
data. The resulting classification was used to extract 37 oyster
reefs, 22 salt marshes, and 33 mudflats areas from the DSM. The
areas were selected because they were fully encompassed within
the extent of the data (i.e., no areas from the boundaries of the
mosaic). Each extracted area was entered in the “Multiscale
Roughness” tool of WhiteboxTools v. 1.1.0 [9], with search
neighborhood radii ranging from 1 grid cell to 9,751 grid cells,
which corresponds to the length of the longest feature (a salt marsh
of about 64 m; cf. center of Fig. 2 and Fig. 3). The tool produced
two main outputs: a raster that indicates, for each pixel, the size of
the search neighborhood at which the measured roughness was the
highest, and a raster displaying the magnitude of the measured
roughness at that scale. Descriptive statistics of the two output
types were calculated for each habitat type.
FIGURE 2. UAS IMAGERY COLLECTED AT LOW TIDE ON DECEMBER 8TH, 2018.
III. RESULTS AND DISCUSSION
Figure 3 presents the GEOBIA classification results and the
scale and magnitude outputs for all the studied habitats. The
spatial distributions of the scale and magnitude values seem to be
influenced by the geometry of the features and the quality of the
DSM. For instance, high-magnitude values were found on long
and narrow features, and broader-scale values were found in areas
of interpolation artifacts where the presence of water affected
DSM production. In general, magnitude is the most promising
output to differentiate the three studied coastal habitats (Fig. 3);
patches of mud displayed a much lower magnitude than other
habitat types, which was expected considering the less complex
nature of mudflats, and salt marshes displayed a much higher
magnitude than other habitat types, likely because of the presence
of a vegetation cover captured in the DSM. Oyster reefs, which
are the most heterogeneous habitats, had intermediate magnitudes
of roughness at intermediate scales.
These observations are confirmed by the analysis of the
statistical distributions of scale and magnitude (Tab. 1). On
Geomorphometry 2020 Lecours and Espriella
112

average, the scale of maximum roughness was broader for
mudflats than for oysters and marshes. However, averages are
likely biased by outliers caused by edge effects: the skewness of
the distributions for scale shows that they are highly skewed for
marsh and oysters, and moderately skewed for mudflats.
Distributions of scale for marshes and oysters are leptokurtic, and
that of mudflats is platykurtic. The distributions of magnitudes for
mudflats and oysters are heavily skewed, with a high and sharp
peak and long and fat tails caused by many outliers. The
distribution of magnitudes for marshes is relatively symmetrical
but platykurtic, with a short and thin tail. Given these results, we
do not expect that the averages presented in Tab. 1 are fully
representative. We adjusted them by manually removing outliers
from the distribution and obtained revised averages for scale of
1,231.66 (≈9.4 m) for marshes, 1,980.06 (≈19.6 m) for mudflats,
and 1879.68 (≈15.0 m) for oysters. However, the median values
suggest that patterns of highest roughness can be found at about
1.5 m for marshes, 12.9 m for mudflats, and 5.3 m for oysters,
which is more consistent with what can be observed in the field
in terms of habitat complexity and habitat patch size. Setting these
results into the natural context can thus serve as additional
evidence that edge effects influenced some of the statistics (e.g.,
average, standard deviation). In fact, the cumulative distributions
of scale for all habitat types showed a stabilization in slope
between 125 and 400, which corresponds to 0.8 to 2.6 m,
indicating that most of the high measured roughness would be
found at scales finer than 3 m.
It is noteworthy that all habitat types reached a local peak of
maximum roughness at search radii of 30 cells (oysters and mud)
or 33 cells (marsh), which corresponds to 21±1 cm. While this is
an interesting result, it should be interpreted with caution: given
the different natures of the habitat types, it is improbable that they
display local roughness at almost the same exact scale (with a
precision of 6 mm). A possible explanation is that intrinsic noise
is present in the DSM at this specific scale range and captured by
the analysis. In terms of magnitude, the statistical distributions
confirm that magnitudes of roughness are generally smaller for
mudflats and higher for marshes.
TABLE 1. DESCRIPTIVE STATISTICS OF THE SCALE AND MAGNITUDE OF THE
MULTISCALE ROUGHNESS MEASURE FOR EACH HABITAT TYPE.
This work is an initial exploration of the potential of measures
of multiscale topographic patterns to help identify coastal
habitats. However, limitations include the use of the results of an
imperfect spectral-based GEOBIA classification to guide the
selection of habitat features for this analysis. For instance, Fig. 4
shows that one of the objects identified as marsh is partly
misclassified: only the central section of this object is a vegetated
salt marsh – the surrounding habitat is oysters. However, the
entirety of this area was considered as marsh for the analysis
because it was based on the objects defined and classified by the
GEOBIA workflow. Both the scale and magnitude of the
multiscale roughness captured that difference, with the marsh
having a finer-scale roughness of higher magnitude than the
surrounding oysters. This directly highlights the potential of these
Marsh Mud Oyster Marsh Mud Oyster
Number of Cells 154,945,348 238,774,206 438,236,795 154,945,348 238,774,206 438,236,795
Minimum 1 1 1 3.90 2.18 3.31
Maximum 9,750 9,750 9,750 108.65 120.51 101.21
Average 1,428.40 2,973.09 2,276.70 21.34 7.90 10.77
Median 229 1,948 804
25.76 6.30 9.82
Standard Deviation 2,287.13 3,127.44 2,943.04 11.01 5.64 5.47
Variance 5,230,967.94 9,780,854.37 8,661,511.55 121.33 31.76 29.88
Skewness 2.00 0.87 1.30 -0.10 2.91 3.15
Kurtosis 6.19 2.51 3.48 1.77 14.55 17.43
Scale
Magnitude
FIGURE 3. GEOBIA CLASSIFICATION, AND SCALE AND MAGNITUDE OF THE MULTISCALE ROUGHNESS MEASURE APPLIED TO THE EXTRACTED FEATURES.
Geomorphometry 2020 Lecours and Espriella
113

measures to augment the GEOBIA and differentiate coastal
habitats in Florida.
Another limitation of this work is that each of the 92 features
studied was analyzed independently, which artificially increased
edge effects. The complex dynamics of this coastal ecosystem
mean that oysters can be directly adjacent to mudflats and
marshes. As such, analyzing a patch of multiple habitats as one,
then separating it into different habitats post-analysis before
computing statistics could have reduced the influence of edge
effects. However, this would be very computationally intensive
given the size of the DSM (24 GB).
IV. CONCLUSIONS
Coastal geomorphometry has recently been identified as a
future application of geomorphometry that will present challenges
due to the presence of features both over and under the waterline
[10]. Here we presented such an application, and these challenges
were highlighted by a strong influence of edge effects and feature
geometry that artificially increased the average scale at which
maximum roughness was observed and the average magnitude of
that roughness. However, we concluded that mudflats display
relatively smaller amplitudes of roughness over broader scales
and that salt marshes display the highest roughness over relatively
finer scales. Oyster reefs showed intermediate patterns of
roughness, with both amplitudes and scales between those of the
two other habitat types. While we hypothesized that oyster reefs
would show the highest roughness at the finest scales, the finer-
scale patterns of salt marshes may be explained by the presence
of characteristic vegetation on the marshes, which creates
relatively high roughness patterns in the DSM. Future work
should repeat the analyses using a Digital Terrain Model (DTM)
instead of a DSM. In theory, the DTM would preserve the
complex fine-scale structures of oyster reefs while omitting the
vegetation over salt marshes that created local roughness.
Denoising algorithms could also be applied to the models to
ensure that the multiscale analyses capture the scales at which
patterns are observed rather than the noise in the data. Finally,
because we demonstrated that multiscale roughness shows
potential to help differentiate coastal habitat types from each
other, we recommend evaluating the suitability of other
multiscale geomorphometric measures, such as multiscale
topographic position [11], multiscale maximum spherical
standard deviation [8], multiscale maximum difference from
mean elevation [11], and multiscale topographic anisotropy [12]
for the identification of coastal habitats.
REFERENCES
[1] Robbins, L., S. Wolfe, E. Raabe, 2008. “Mapping of Florida’s coastal
and marine resources: setting priorities workshop”. USGS Open-File
Report, 1157, 32 p.
[2] Gao, Y., J.S. Fu, J.B. Drake, Y. Liu, J.-F. Lamarque, 2012. “Projected
changes of extreme weather events in the eastern United States based on a
high resolution climate modeling system”. Environ. Res. Lett. 7, 1-12.
[3] Unmenhofer, C.C., G.A. Meehl, 2017. “Extreme weather and climate events
with ecological relevance: a review”. Philos. T. R. Soc. B. 372, 1-13.
[4] Espriella, M., V. Lecours, P.C. Frederick, E.V. Camp, B. Wilkinson, 2020.
“Quantifying intertidal habitat relative coverage in a Florida estuary using
UAS imagery and GEOBIA”. Remote Sens. 12(4), 677, 1-17.
[5] Blaschke, T., 2010. “Object based image analysis for remote sensing”.
ISPRS J. Photogramm. Remote Sens. 65, 2-16.
[6] Grohmann C.H., C. Riccomini, 2009. “Comparison of roving-window and
search-window techniques for characterizing landscape morphometry”.
Comput. Geosci. 35, 2164-2169.
[7] Deng, Y., J.P. Wilson, 2008. “Multi-scale and multi-criteria mapping of
mountain peaks as fuzzy entities”. Int. J. Geogr. Inf. Sci. 22, 205-218.
[8] Lindsay, J.B., D.R. Newman, A. Francioni, 2019. “Scale-optimized surface
roughness for topographic analysis”. Geosci. 9(7), 322.
[9] WhiteboxTools, https://jblindsay.github.io/ghrg/WhiteboxTools/index.html
[10] Lecours, V., M.F. Dolan, A. Micallef, V.L. Lucieer, 2016. “A review of
marine geomorphometry, the quantitative study of the seafloor”. Hydrol.
Earth Syst. Sci. 20, 3207-3244.
[11] Newman, D.R., J.B. Lindsay, J.M.H. Cockburn, 2018. “Evaluating metrics
of local topographic position for multiscale geomorphometric analysis”.
Geomorphology 312, 40-50.
[12] Newman, D.R., J.B. Lindsay, J.M.H. Cockburn, 2018. “Measuring
hyperscale topographic anisotropy as a continuous landscape property”.
Geosci. 8, 278.
FIGURE 4. EXAMPLE OF ORTHOMOSAIC, DSM, AND THE SCALE AND MAGNITUDE
COMPONENTS OF THE MULTISCALE ROUGHNESS MEASURE FOR ONE OF THE 22 SALT
MARSHES, ONE OF THE 33 MUDFLATS, AND ONE OF THE 37 OYSTER REEFS STUDIED.
Geomorphometry 2020 Lecours and Espriella
114