Automatic Seagrass Disturbance Pattern Identification on Sonar Images

Abstract

Natural and anthropogenic disturbances are causing degradation and loss of seagrass cover, often in the form of bare patches (potholes) and propeller-scaring from vessels. Degradation of seagrass habitat has increased significantly in recent years with losses totaling some 110 km2 per year. With seagrass habitat disappearing at historically unprecedented rates, development of new tools for mapping these disturbances is critical to understanding habitat distribution and seagrass abundance. Current methods for mapping seagrass coverage rely on appropriate meteorological conditions (satellite imagery), are high in cost (aerial photography), or lack resolution (in situ point surveys). All of these methods require low turbidity, and none is capable of automatically detecting bare patches (potholes) in seagrass habitat. Sonar-based methods for mapping seagrass can function in high turbidity, and are not affected by meteorological conditions. Here, we present an automatic method for detecting and quantifying potholes in sidescan sonar images collected in a very shallow, highly disturbed seagrass bed. Acoustic studies of shallow seagrass beds (<2 m) are scarce due to traditional approaches being limited by reduced horizontal swath in these depth ranges. The main challenges associated with these sidescan sonar images are random ambient noise and uneven backscatter intensity across the image. Our method combines adaptive histogram equalization and top-hat mathematical morphology transformation to remove image noises and irregularities. Then, boundaries of potholes are detected using optimum binarization as well as closing and erosion mathematical morphology filters. This method was applied to several sonar images taken from the Lower Laguna Madre in Texas at less than 2-m depth. Experimental results in comparison with ground-truthing demonstrated the effectiveness method by identifying potholes with 97% accuracy.

Full text

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
IEEE JOURNAL OF OCEANIC ENGINEERING 1
Peer-Reviewed Technical Communication
Automatic Seagrass Disturbance Pattern Identification on Sonar Images
Maryam Rahnemoonfar, Abdullah F. Rahman, Richard J. Kline, and Austin Greene
Abstract—Natural and anthropogenic disturbances are causing degrada-
tion and loss of seagrass cover, often in the form of bare patches (potholes)
and propeller-scaring from vessels. Degradation of seagrass habitat has
increased significantly in recent years with losses totaling some 110 km2
per year. With seagrass habitat disappearing at historically unprecedented
rates, development of new tools for mapping these disturbances is critical
to understanding habitat distribution and seagrass abundance. Current
methods for mapping seagrass coverage rely on appropriate meteorological
conditions (satellite imagery), are high in cost (aerial photography), or lack
resolution (in situ point surveys). All of these methods require low turbidity,
and none is capable of automatically detecting bare patches (potholes) in
seagrass habitat. Sonar-based methods for mapping seagrass can function
in high turbidity, and are not affected by meteorological conditions. Here,
we present an automatic method for detecting and quantifying potholes in
sidescan sonar images collected in a very shallow, highly disturbed seagrass
bed. Acoustic studies of shallow seagrass beds (<2 m) are scarce due to
traditional approaches being limited by reduced horizontal swath in these
depth ranges. The main challenges associated with these sidescan sonar
images are random ambient noise and uneven backscatter intensity across
the image. Our method combines adaptive histogram equalization and top-
hat mathematical morphology transformation to remove image noises and
irregularities. Then, boundaries of potholes are detected using optimum bi-
narization as well as closing and erosion mathematical morphology filters.
This method was applied to several sonar images taken from the Lower
Laguna Madre in Texas at less than 2-m depth. Experimental results in
comparison with ground-truthing demonstrated the effectiveness method
by identifying potholes with 97% accuracy.
Index Terms—Image analysis, image segmentation, morphological oper-
ations, sea floor, sonar.
I. INTRODUCTION
EXTENSIVE degradation of seagrass beds is taking place in
coastal areas around the globe because of natural and human-
induced disturbances. These negative impacts affect approximately
65% of the original seagrass communities, mainly in Europe, North
America, and Australia [1]. While large seagrass meadows can be ob-
served from satellite imagery and aerial methods these approaches can
be limited by high turbidity, poor meteorological conditions, or low
resolution. Mapping of seagrass degradation due to natural and human
disturbances such as potholes and propeller scars is essential to esti-
mating overall abundance, disturbance regimes, and the overall health
Manuscript received October 3, 2016; revised May 12, 2017 and September
7, 2017; accepted December 1, 2017. A short and early version of this paper
was previously published in Proc. SPIE, vol. 9844, Automatic Target Recogni-
tion XXVI, 98440C, May 12, 2016; doi: 10.1117/12.2224191. (Corresponding
author: Maryam Rahnemoonfar.)
Associate Editor: J. Cobb.
M. Rahnemoonfar is with the Computer Vision and Remote Sensing
Laboratory, School of Engineering and Computing Sciences, Texas A&M
University-Corpus Christi, Corpus Christi, TX 78412 USA (e-mail: maryam.
rahnemoonfar@tamucc.edu).
A. F. Rahman, R. J. Kline, and A. Greene are with the Coastal Studies
Laboratory, School of Earth, Environmental, and Marine Sciences, Univer-
sity of Texas Rio Grande Valley, Brownsville, TX 78520 USA (e-mail: abdul-
lah.rahman@utrgv.edu; richard.kline@utrgv.edu; AustinLG@Hawaii.edu.
Digital Object Identifier 10.1109/JOE.2017.2780707
of related marine systems. Since the early 1990s, various remote sens-
ing techniques have been exploited for seagrass mapping [2]–[9] and
all have their limitations. For both aerial and satellite optical remote
sensing techniques, it is difficult to detect seagrass disturbances under
water. For optical remote sensing, light is attenuated as it passes through
the water column and reflects back from the benthos causing errors in
calculations. The attenuation is not only the function of depth of the wa-
ter column but also the sediment load, microalgae and organic matters
in the water column. As water depth increases, increased attenuation
makes optical imagery even more difficult to capture. Furthermore,
shallow coastal waters can be turbid due to wind and wave action, boat
traffic, coastal constructions, and other human activities—all of which
create limitations in seagrass and disturbance mapping effectiveness
using optical sensors. Methods typically used to categorize and map
terrestrial vegetation based on spectral reflectance of vegetation such as
normalized difference vegetation index do not function well in seagrass
habitats due to the reflective and refractive properties of the water col-
umn above the seagrass. These conditions limit most seagrass imagery
to simple visual analysis, where the effects of disturbance cannot be
accurately quantified. Only the visible bands of multispectral satellite
sensors are generally used to map seagrass occurrence and boundaries
in coastal shallow waters.
Underwater acoustic techniques have allowed many advances in the
field of remote sensing and these techniques can be used to produce
a high-definition, 2-D sonar image of seagrass ecosystems [10]. How-
ever,s everal studies have shown the inefficiency of operating traditional
acoustic instruments such as sidescan or multibeam sonar in shallow
conditions [11]–[15]. Recent work by Greene [16] has extended oper-
ational sidescan surveys of seagrass ecosystems into shallow seagrass
beds at depths of 2 m or less. Shallow habitats such as these have re-
mained largely understudied in acoustic surveys of submerged aquatic
vegetation. However, these advancements are largely due to a reduc-
tion in transducer beam angle and as such traditional techniques to
normalize backscatter-intensity [such as time variable gain (TVG)]
demonstrate limited effectiveness. The acoustic profile of this benthic
ecosystem is created when sound waves are emitted, reflected back,
and received by the transducer of a sonar device. The intensity and
contours of the image are determined by the position and amount of
time a sound wave takes to return to the transducer. In this paper,
we use sonar imagery as a tool to recognize disturbance patterns in
seagrass beds. Previous literature regarding sonar image pattern recog-
nition has mainly focused on searching for solid objects on a sandy
sea floor [17]–[24]. A Markovian segmentation algorithm was used by
Mignotte et al. [17] to segment the sonar image. Additionally, they used
an unsupervised hierarchical Markov random field model [18] to seg-
ment the image into two kinds of regions: 1) shadow and 2) sea-bottom
reverberation. Even though they obtained good results for simple con-
crete objects with regular shape on a sandy sea floor, their particular
method would be complicated and computationally expensive to apply
on seagrass images with irregular and complex patterns. Furthermore,
0364-9059 © 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
2IEEE JOURNAL OF OCEANIC ENGINEERING
another sonar segmentation method was described by Lianantonakis
et al. [19]. Their particular method focused on the binary segmentation
of high-resolution sonar images. The first step was to extract texture
features from a sidescan image containing two distinct regions. Then,
a region based active contour model was applied to the vector valued
images extracted from the original data. An automatic image analysis
program for the detection and identification of stationary targets, such
as meter-sized concrete artificial reefs on the sea floor was proposed
by Tian et al. [20]. In this algorithm, features were extracted using
grey level cooccurrence matrix and then classified by a Bayesian clas-
sifier. Another sonar image segmentation method was created by Ye
et al. [21]. Their method first involved the extraction of local texture
features of sonar images based on Gauss–Markov random field model
and integrated into the level-set energy functions. Even though the
sidescan sonar has been used for benthic mapping, there is no prevail-
ing method that automatically detects the extent of seagrass beds or
automatically identifies and maps its disturbance. To the best of our
knowledge, this research is the first of its kind on automated seagrass
disturbance identification using sidescan sonar imagery. Sonar images
are notorious for having random ambient noise and low signal-to-
noise ratio, which makes the segmentation of targets, such as seagrass,
difficult to accomplish. Additionally, brightness levels throughout the
image can be nonuniform, potentially making segmentation of natural
and man-made disturbances within seagrass difficult. Moreover, distur-
bance presents complex patterns in images causing most segmentation
techniques to fail. We collected sonar images in Lower Laguna Madre
of South Texas, which contains vast seagrass beds situated behind a
barrier island. To detect disturbances in seagrass structure, we describe
here, a novel technique based on mathematical morphology and adap-
tive histogram equalization for recognition of potholes within shallow
seagrass structure in sidescan sonar images.
The following are the contributions of this work.
1) A novel approach for automated seagrass disturbance identifica-
tion is presented.
2) Our method is robust to noise, uneven backscatter intensity, and
complex seagrass pothole patterns.
3) Seagrass and pothole map are generated automatically for the
first time.
4) Testing and development of sidescan sonar for shallow water.
II. METHODOLOGY
Our approach consisted of two major steps, as shown in Fig. 1. In the
first step, we performed uneven backscatter intensity reduction, and the
image was enhanced based on adaptive histogram equalization, top-hat
transformation, and Gaussian adaptive thresholding. In the second step,
we identified seagrass blowouts (or potholes, a disturbance regime)
by applying, Otsu binarization, closing, and erosion morphological
operators [25], [26].
A. Image Enhancement
While capturing the data, the sidescan sonar transmits a high-
frequency acoustic signal in the water using two parallel transducers.
There were substantial variations in brightness found in the sides-
can sonar images because the objects closer to the transducer created
brighter reflections. Fig. 2 shows a sonar image in which the brightness
values vary across the image. In the central portion of the image, there
is a higher intensity of reflection, which makes the image look brighter;
whereas in both sides of the image the reflection is of lower intensity
and the image looks darker. The bright line in the middle of the image
is the first echo return of the sonar at the seafloor underneath the boat.
Fig. 1. Flowchart of overall methodology.
Potholes are visible on the image, scattered irregularly and shown as
depressions in the seagrass bed.
Uneven backscatter intensity at this shallow recording depth and
very low grazing angle creates a real challenge for automatic detection
of potholes in seagrass. Normalization of the images could not be
accomplished with commonly used methods such as TVG or beam
angle correction. Here, we explain that how we mitigated this issue.
From Fig. 2, it is clear that there are mainly three partitions that include
a bright partition in the middle and two relatively dark partitions in the
sides. We performed horizontal line by line scan from both directions.
The idea was to find two locations (b1and b2) within the line to divide
it into three different sections, namely two darker portions on the sides
and one bright portion in the middle. Locations b1and b2were selected
such that the bright portion is between b1and b2.Weusedhistogram
interpretation to select the threshold value (T)that we used for each line
to select b1and b2. In Fig. 3, we plotted a histogram of the highlighted
subset, where the expected break point lies. The peak of the histogram
in Fig. 3 was 0.4, so a threshold (T)of 0.4 was set to identify the
breakpoints b1and b2for each line.
The boundaries between darker and brighter partition is defined as a
set, B, of all the break points. Mathematically, Bcan be expressed as
B={(b1,b
2,i)|0≤i<n,b
1=F1(i, T ),b
2=F2(i, T )}(1)
where nis the number of horizontal pixel lines in the image, Tis
a predefined threshold parameter, and F1and F2are the functions
returning the length of the first run of pixels in ith line whose intensity

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
RAHNEMOONFAR et al.: AUTOMATIC SEAGRASS DISTURBANCE PATTERN IDENTIFICATION ON SONAR IMAGES 3
Fig. 2. Example of a sidescan sonar image with uneven backscatter intensity
across the image.
value is greater or equal to T.F1returns the length starting from first
pixel while scanning in the forward direction and F2returns the length
starting from the last pixel while scanning in a backward direction.
We then applied an adaptive histogram equalization separately on each
partition for adjusting image intensities to enhance contrast. Each row
of the image was partitioned into three sections using partition points
b1and b2, as described in (1). The histogram equalization [27] is an
effective image enhancing technique, summarized as follows: Let X=
{X(i, j)}be an image where {X(i, j )}denote gray value at location
(i, j). If the total number of pixels is N, and image intensity is digitized
to L, levels are then ∀X(i, j)∈{X0,X
1,...,X
L−1}. Suppose nkis
the total number of pixels with gray value Xkthen, the probability
density of the Xkwill be
p(Xk)=nk/N, k =0,1,...,L−1(2)
and its cumulative distribution function can be defined as
c(Xk)=
k

i=0
p(Xk),k=0,1,...,L−1.(3)
The transformation function can be defined as
f(Xk)=X0+(XL−1−X0)c(Xk),k=0,1,...,L−1.
(4)
If Y={Y(i, j)}is defined as an equalized image, then
Y=f(X)={f(X(i, j))|∀X(i, j )∈X}.(5)
After applying the adaptive histogram equalization, we used a
Top-hat mathematical morphology filter to further remove the un-
even backscatter intensity defects in the image. top-hat filtering
computes the morphological opening of the image and then sub-
tracts the result from the original image. It uses the structuring el-
ement. Top-hat transformation is defined as the function minus its
opening [26]
That(f)=f−(f◦b)(6)
where fis the original image, bis the structure element, and ◦is the
opening morphological operator. Opening of Aby Bis the erosion of
Aby Bfollowed by dilation of the result by B[28]. With Aand Bas
sets in Z2, the erosion of Aby B, denoted by ABis defined as
AB={z|(B)z∩Ac=∅} (7)
and the dilation of A by B can be defined as
A⊕B={z|
∧
(B)
z
∩A=∅}.(8)
Structure elements are small sets or subimages used to probe an
image under study. Here, the structure element used was a circle with
a radius of 15 pixels. The goal of top-hat transformation is to extract
the uniform background image. Therefore, all objects including pot-
holes must be removed during the erosion stage of top-hat. Since the
maximum size of a pothole is around 130 cm by 130 cm and the res-
olution of image is around 10 cm per pixel, a structure element of
size 15 pixels will remove all objects and only give us the background
image.
To further enhance the image, we applied the Gaussian adaptive
thresholding technique. In this method, a variable threshold was cal-
culated at every point, (x,y) based on the properties computed in a
neighborhood of (x,y). In Gaussian adaptive thresholding, the thresh-
old value is a weighted sum of the small neighborhood around each
pixel. The neighborhood windows were chosen small enough so that
the backscatter intensity of each is approximately uniform. A threshold
is calculated for each pixel based on the convolution of the image with
Gaussian function as follows [26]:
T(x, y)=
a

k
b

l=−b
G(s, t)f(x−k, y −l)(9)
where Gis the Gaussian function of two variables and has the basic
form of
G(x, y)= 1
2πσ2e−x2+y2
2σ2(10)
where σis the standard deviation.
B. Seagrass Pattern Identification
After enhancing the image and removing the effect of nonuni-
form backscatter intensity, the process of extracting potholes from
the sonar images was conducted. The image was binarized using an
optimum Otsu threshold [29]. If the threshold to binarize the image
is t, then optimal threshold can be defined as maximum of σ2
B(t)
as follows:
σ2
B(t∗)= max
0≤t≤L−1{σ2
B(t)}(11)
where σ2
B(t)is class variance and Lis total number of gray levels in
the image. σ2
B(t)can be defined as follows:
σ2
B(t)=ω0(t)[μ0(t)−μT]2+ω1(t)[μ1(t)−μT]2(12)
where μT=L−1
i=0 iPi,ω0(t)=l
i=0 Pi,ω1(t)=1−ω0(t),μ(t)=
l
i=0 iPi,μ0(t)=(μ(t))/(ω0(t)),μ1(t)=(μT−μ(t))/(1−ω0(t)),
Pi=(ni)/(N),N=L−1
i=0 ni,andniis the number of pixels with
gray level i.
After binarization, there were a lot of small holes remaining which
were, in fact, part of seagrass texture. To eliminate these small holes,
we applied a closing morphological filter. If Ais the image and Bis
structure element, then the Closing of Aby Bcan be defined as the
dilation of Aby Bfollowed by erosion of the result of the dilation

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
4IEEE JOURNAL OF OCEANIC ENGINEERING
Fig. 3. Histogram of the highlighted subset of the seagrass image.
Fig. 4. Location of the study site with the six sonar transects.
by B[26]
AB=(A⊕B)B. (13)
For our analysis, we used a circle as the structural element with
diameter of 11 pixels since we were interested in detecting pothole
with a diameter greater than 11 pixels. The closing operation smoothed
out sections of contours by eliminating small holes and filling gaps in
the contour. Subsequently, the boundary of objects in the image was
calculated using the following formula [26]:
β(A)= A−(AB)(14)
where Ais the original image and is the erosion mathematical mor-
phology operator.
III. EXPERIMENTAL RESULTS
A. Data Description
Data were collected from the seagrass beds of the Lower Laguna
Madre in Southern Texas on May 23, 2016 from an average depth of
75 cm. A specialized sidescan sonar unit was constructed consisting of
a towfish with two Lowrance Structure Scan HD LSS-2 transducers,
a dual-beam 200-kHz down-imaging transducer connected to a Hum-
minbird 998C HD SI control unit [30]. The sidescan unit was operated
at 800 kHz, and the transducers were offset at 25°from the horizon-
tal to allow improved horizontal swath in shallow environments [16].
Vessel position was measured with a heading-sensor equipped GPS
receiver mounted directly over the transducer. Water depth was mea-
sured via the 200-kHz down-imaging transducer. All navigation and

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
RAHNEMOONFAR et al.: AUTOMATIC SEAGRASS DISTURBANCE PATTERN IDENTIFICATION ON SONAR IMAGES 5
Fig. 5. Transect images with a zoomed subset of final result.
acoustic signals were recorded continuously by the Humminbird 998C
HD SI control unit on a secure digital (SD) card for later process-
ing. Combining each line of reflected signal with its position, time,
and depth produced an image of the seafloor (see Fig. 2). A total
of six transects approximately 5000 ×60 000 pixels and overlapping
by 50% were processed individually to image an area of approxi-
mately 88 000 m2. Fig. 4 shows the location of the study site on
GeoEye imagery with the six transects of sonar images used in this
experiment.
B. Seagrass Pattern Recognition
The images used in this experiment were large transects. Each tran-
sect was 600 m long, and spaced 20 m apart with a horizontal swath
of approximately 50 m. Total area covered between all six transects
was 88 000 m2(150 m ×600 m). We applied the seagrass detection
algorithm on entire transect images, however, for a better display, only
a subset is shown here. Fig. 5 shows the original transect images along
with the subset chosen for display and part of the result superimposed
on the subset.
Fig. 6 shows the original image and all the intermediate results while
identifying the potholes along with the ground-truth. We present here,
a subset of one of the six transect images.
The uneven, nonlinear backscatter intensity gradients are likely the
product of either the modified sidescan array’s reduced beam pattern,
sound attenuation by seagrass aerenchym, the very shallow environ-
ment in which the sidescan was designed to operate, or a combination
of these. These factors did not permit backscatter-intensity normaliza-
tion with a TVG without a heavy loss in detail along the center track.
However, automatic gain correction (AGC) appeared to approximate

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
6IEEE JOURNAL OF OCEANIC ENGINEERING
Fig. 6. (a) Subset of the original image. (b) Result after applying histogram equalization. (c) Result after applying the top-hat filter. (d) Result after applying
the Gaussian smoothing filter. (e) Result after applying binarization. (f) Result after removing small holes from the binary image. (g) Image obtained after
superimposing the boundaries of the potholes over the original image. (h) Ground-truth.
our own adaptive gain equalization. A traditional sidescan array op-
erating over deeper seagrass meadows may be able to simplify image
processing and pattern recognition steps by utilizing these automated
techniques such as TVG correction. Fig. 6(a) represents the subset of
the original image. In Fig. 6(b), it can be observed that after applying
adaptive histogram equalization (5), the darker sides of the original
image were brighter but still the middle section of the image was still
brighter, as compared to sides. To mitigate this effect, we applied Top-
hat filter (6) on the image in Fig. 6(b) and (c). This filter was effective
at balancing the backscatter intensity. The image is too sharp to be
binarized at this point of time because of the seagrass texture. Due to
the extreme sharpness of the seagrass texture in the image, a Gaussian
filter was used before binarization to prevent artifacts being identified
as potholes that were only the texture of the seagrass. After apply-
ing Gaussian filter (9) with σ=2,a smoother image was observed in
Fig. 6(d) with reduced sharpness of the seagrass texture. After bina-
rizing of the image (11), we can see many small potholes in Fig. 6(e).
A closing morphological filter (13) was applied, revealing the actual
potholes in Fig. 6(f). Finally, the boundaries of potholes were detected
using (14). The final polygons were superimposed on the original im-
age, shown in Fig. 6(g). Fig. 6(h) shows the ground-truth image. This
image was obtained by manual inspection of the image and annotation
by an expert with detailed experience in the seagrass area that was
imaged.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
RAHNEMOONFAR et al.: AUTOMATIC SEAGRASS DISTURBANCE PATTERN IDENTIFICATION ON SONAR IMAGES 7
TAB L E I
EXAMPLE OF AN ERROR MATR I X
TAB L E I I
OVERALL AND CLASSWISE ACCURACY ASSESSMENT
Average overall na¨
ıve accuracy (%) 97.73
Average overall kappa accuracy (%) 85.96
Average producer’s Potholes 83.33
kappa Accuracy Seagrass 88.84
Average user’s Potholes 89.00
kappa Accuracy Seagrass 83.03
TABLE III
OVERALL AND CLASSWISE ERROR ASSESSMENT
Average overall kappa error (10–5)5.58
Average producer’s Potholes 7.83
kappa Accuracy Seagrass 6.76
Average user’s Potholes 7.84
kappa Accuracy Seagrass 6.76
C. Performance Evaluation and Accuracy Metrics
Visual interpretation of the results suggests that the algorithm was
able to detect potholes efficiently and an accuracy assessment con-
firms this same result in Table II. We used several matrices for the
accuracy assessment including Na¨
ıve and kappa accuracy measures of
both, overall and classwise comparisons. na¨
ıve overall accuracy was
computed as follows [31]:
na
¨
lve overall accuracy =k
i=1 nii
n.(15)
Assuming that nsamples are distributed into k2cells, where each
sample is assigned to one of kcategories in the remotely sensed clas-
sification (rows) and, independently, to one of the same kcategories in
the reference data set (columns). As shown in Table I, let nij denote
the number of samples classified into category i(i=1,2,...,k)in
the remotely sensed classification and category j(j=1,2,...,k)in
the reference data set (ground-truth).
Generally, na¨
ıve accuracy measure is used, but in this case, it is not
a reliable measure because it does not take into account the random
agreement for two classes. Since we had only two classes (seagrass and
potholes), there was a 50% chance for a pixel to belong to any one of
the classes, we used kappa statistics [15] to compute both overall and
classwise accuracies.
The kappa index of agreement (KIA) reveals how much better, or
worse, the classifier is than would be expected by random chance. The
overall kappa index is defined as follows [31]:
κ=θ1−θ2
1−θ2
(16)
where κis kappa, θ1=(1/n)k
i=1 nii and θ2=(1/n2)
k
i=1 ni+n+i,n+iis Marginal sum of column iand ni+Marginal
sum of row i
kappa user’s accuracy =nnii −ni+n+i
nni+−ni+n+i
(17)
kappa producer’s accuracy =nnjj −n+jnj+
nn+j−n+jnj+
.(18)
We computed the average of all the six images for both the over-
all na¨
ıve and classwise kappa accuracies. The average overall na¨
ıve
accuracy was 97.73%. Despite the high overall na¨
ıve accuracy, we
also computed kappa accuracies because na¨
ıve accuracy is not free
from random agreement and may not be a reliable measure for two
classes. For classwise accuracies, we computed both producer’s and
user’s accuracy. Producer’s accuracy, also known as precision, corre-
sponds to error of omission (exclusion). It accounts for the samples,
which are not classified in a class that they actually belong to, while
user’s accuracy, also known as recall, presents the reliability of classes
in the classified image. Along with the accuracies, we also computed
the standard error of each computed accuracy. Tables II and III show
that the pothole detection accuracies are high and errors are very low
(on the order of 10–5) and that potholes have more user’s accuracy
as compared to seagrass. Pothole identification was more reliable as
compared to seagrass identification. However, some of the potholes
could not be identified correctly and classified as seagrass, leading to
a lower producer’s accuracy. The average overall and classwise ac-
curacies and errors are presented in Tables II and III. Figs. 7 and
8 show accuracies and error, respectively, for all the six images. It
can be observed from Fig. 7 that the overall kappa for six images
ranged from 82% to 90% accuracy with an average kappa accuracy
of 85.96%.
Figs. 9–11 show some subset of the seagrass images, where big,
intermediate, and small size potholes were accurately identified.
In this study, potholes were identified with the high accuracy regard-
less of their size (see Figs. 9–11). Moreover, our algorithm was able to
detect potholes in any portion of the seagrass image despite the uneven
backscatter intensity in the sonar image. In Fig. 11(a) and (b), potholes
are in the brighter portion of the seagrass image (middle of the image)
and all other subsets from Figs. 9–11 potholes are in darker portion of
the image.
There were few false negative and false positives in our pothole iden-
tifications; however, some false positives occur when there is similar
texture inside and outside of potholes. In Fig. 12, it can be observed that
the pothole boundary is not identified accurately. Moreover, a close ob-
servation of the Fig. 12 reveals that the area within the pothole, which
was not detected as pothole (highlighted by upper red box), has a similar
texture to that of seagrass (highlighted by lower red box). Two possible
reasons for the texture similarity are: 1) presence of macroalgae inside
the pothole or 2) due to the new growth of seagrass inside the pothole.
Algae within the potholes may reflect sonar signals similar to the sea-

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
8IEEE JOURNAL OF OCEANIC ENGINEERING
Fig. 7. Overall and classwise kappa accuracies for six transect images.
Fig. 8. Overall and classwise standard error of kappa for six transect images.
Fig. 9. Accurately identified big size potholes over different seagrass images (left: original image; right: detected polygons with our proposed method).

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
RAHNEMOONFAR et al.: AUTOMATIC SEAGRASS DISTURBANCE PATTERN IDENTIFICATION ON SONAR IMAGES 9
Fig. 10. Accurately identified intermediate size potholes over different seagrass images.
Fig. 11. Accurately identified small size potholes over different seagrass images.
Fig. 12. (a) Subset of the original image having a large pothole and (b) subset
of the image containing boundaries of the potholes identified by the algorithm
along with the zoomed subsets of the areas within the pothole and outside the
pothole.
grass, leading to misidentification of potholes as seagrass as seen in
Figs. 7 and 8 in terms of user’s and producer’s accuracy and errors.
Another problem in disturbance identification was with boat pro-
peller scars. Currently, the algorithm presented here is not able to
detect propeller scars accurately. Since the scars are very narrow linear
features, they dissolve with the seagrass while smoothing the image and
are left unidentified. Future research will investigate separate narrow
linear feature detection based algorithms to identify the propeller scars.
IV. CONCLUSION
In this study, we demonstrated an automated method to detect sea-
grass potholes using sidescan sonar imagery from a modified sides-
can array functioning in a shallow environment. Presence of uneven
backscatter intensity and noise in sidescan sonar images, in addition to
the complex pothole patterns, created several challenges in recognizing
seagrass disturbance pattern in the sonar images. We approached these
challenges in two steps, namely, 1) image enhancement and 2) seagrass
pattern recognition. For image enhancement, our automated method
combined adaptive histogram equalization and top-hat mathematical
transform to remove image noises and irregularities. Commonly used
sonar gain normalization methods such as TVG or AGC may accom-
plish similar results as adaptive histogram equalization when operating
in deeper conditions via traditional sidescan arrays. Surveying a shallow
and highly disturbed seagrass meadow, we designed a wholly automatic
technique to detect the boundary of seagrass bed potholes using mathe-
matical morphology filters. We applied our algorithm to sidescan sonar
images collected from Lower Laguna Madre in Texas and experimental
results in comparison with the ground-truthing show the high accuracy
(∼97%) of the proposed technique in detecting the potholes. Paired
with a sidescan array modified for use in very shallow depths, these
results demonstrate an efficient and accurate method of automatically
identifying disturbance in shallow seagrass meadows rarely attempted
using acoustic instruments. In the future, we plan to extend this work
to real-time identification of seagrass disturbance patterns.
REFERENCES
[1] H. K. Lotze et al., “Depletion, degradation, and recovery potential of
estuaries and coastal seas,” Science, vol. 312, pp. 1806–1809, 2006.
[2] M. Hossain, J. Bujang, M. Zakaria, and M. Hashim, “Assessment of
landsat 7 scan line corrector-off data gap-filling methods for seagrass
distribution mapping,” Int. J. Remote Sens., vol. 36, pp. 1188–1215, 2015.
[3] A. G. Dekker, V. E. Brando, and J. M. Anstee, “Retrospective seagrass
change detection in a shallow coastal tidal Australian lake,” Remote Sens.
Environ., vol. 97, pp. 415–433, 2005.
[4] H. Dierssen, A. Chlus, and B. Russell, “Hyperspectral discrimination of
floating mats of seagrass wrack and the macroalgae sargassum in coastal
waters of greater Florida Bay using airborne remote sensing,” Remote
Sens. Environ., vol. 167, pp. 247–258, 2015.
[5] V. Pasqualini et al., “Use of SPOT 5 for mapping seagrasses: An appli-
cation to Posidonia Oceanica,” Remote Sens. Environ., vol. 94, pp. 39–45,
2005.
[6] S. Phinn, C. Roelfsema, A. Dekker, V. Brando, and J. Anstee, “Mapping
seagrass species, cover and biomass in shallow waters: An assessment
of satellite multi-spectral and airborne hyper-spectral imaging systems in
moreton bay (Australia),” Remote Sens. Environ., vol. 112, pp. 3413–3425,
2008.
[7] C. M. Roelfsema et al., “Multi-temporal mapping of seagrass cover,
species and biomass: A semi-automated object based image analysis ap-
proach,” Remote Sens. Environ., vol. 150, pp. 172–187, 2014.
[8] C. C. Wabnitz, S. Andr´
efou¨
et, D. Torres-Pulliza, F. E. M¨
uller-Karger,
and P. A. Kramer, “Regional-scale seagrass habitat mapping in the Wider
Caribbean region using landsat sensors: Applications to conservation and
ecology,” Remote Sens. Environ., vol. 112, pp. 3455–3467, 2008.
[9] J. Hedley, B. Russell, K. Randolph, and H. Dierssen, “A physics-based
method for the remote sensing of seagrasses,” Remote Sens. Environ.,
vol. 174, pp. 134–147, 2016.
[10] M. Hossain, J. S. Bujang, M. Zakaria, and M. Hashim, “The application of
remote sensing to seagrass ecosystems: an overview and future research
prospects,” Int. J. Remote Sens., vol. 36, pp. 61–114, 2015.
[11] B. M. Sabol, R. E. Melton, R. Chamberlain, P. Doering, and K. Haunert,
“Evaluation of a digital echo sounder system for detection of submersed
aquatic vegetation,” Estuaries Coasts, vol. 25, pp. 133–141, 2002.
[12] T. Sagawa et al., “Technical Note. Mapping seagrass beds using IKONOS
satellite image and side scan sonar measurements: A Japanese case study,”
Int. J. Remote Sens., vol. 29, pp. 281–291, 2008.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
10 IEEE JOURNAL OF OCEANIC ENGINEERING
[13] M. Paul, A. Lefebvre, E. Manca, and C. L. Amos, “An acoustic method
for the remote measurement of seagrass metrics,” Estuarine, Coastal Shelf
Sci., vol. 93, pp. 68–79, 2011.
[14] A. Lefebvre, C. Thompson, K. Collins, and C. Amos, “Use of a high-
resolution profiling sonar and a towed video camera to map a Zostera
marina bed, Solent, UK,” Estuarine, Coastal Shelf Sci., vol. 82, pp. 323–
334, 2009.
[15] G. De Falco, R. Tonielli, G. Di Martino, S. Innangi, S. Simeone, and I. M.
Parnum, “Relationships between multibeam backscatter, sediment grain
size and Posidonia Oceanica seagrass distribution,”Continental Shelf Res.,
vol. 30, pp. 1941–1950, 2010.
[16] A. Greene, “Applications of side scan and parametric echosounders for
mapping shallow seagrass habitats and their associated organic carbon,”
M.S. thesis, Dept. Bio. Sci., School of Earth, Environ., Marine Sci., Univ.
Texas Rio Grande Valley, Brownsville, TX, USA, 2017.
[17] M. Mignotte, C. Collet, P. Perez, and P. Bouthemy, “Three-class Marko-
vian segmentation of high-resolution sonar images,” Comput. Vis. Image
Understanding, vol. 76, no. 3, pp. 191–204, 1999.
[18] M. Mignotte, C. Collet, P. Perez, and P. Bouthemy, “Sonar image seg-
mentation using an unsupervised hierarchical MRF model,” IEEE Trans.
Image Process., vol. 9, no. 7, pp. 1216–1231, Jul. 2000.
[19] M. Lianantonakis and Y. R. Petillot, “Sidescan sonar segmentation using
active contours and level set methods,” presented at the OCEANS Eur.
Conf., Brest, France, 2005.
[20] W. Tian, “Automatic target detection and analyses in side-scan sonar
imagery,” in Proc. WRI Global Congr. Intell. Syst., 2009, pp. 397–403.
[21] X.-F. Ye, Z.-H. Zhang, P. X. Liu, and H.-L. Guan, “Sonar image seg-
mentation based on GMRF and level-set models,” Ocean Eng., vol. 37,
pp. 891–901, 2010.
[22] J. T. Cobb, K. C. Slatton, and G. J. Dobeck, “A parametric model for
characterizing seabed textures in synthetic aperture sonar images,” IEEE
J. Ocean. Eng., vol. 35, no. 2, pp. 250–266, Apr. 2010.
[23] F. A. Sadjadi, “Studies in adaptive automated underwater sonar mine
detection and classification—Part 1: Exploitation methods,” Proc. SPIE,
vol. 9476, 2015, Art. no. 94760K.
[24] J. C. Isaacs, “Laplace–Beltrami eigenfunction metrics and geodesic shape
distance features for shape matching in synthetic aperture sonar,” in Proc.
IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recogn. Workshops, 2011,
pp. 14–20.
[25] Z. Qu and L. Zhang, “Research on image segmentation based on the
improved Otsu algorithm,” in Proc. 2nd Int. Conf. Intell. Human–Mach.
Syst. Cybern., 2010, pp. 228–231.
[26] R. C. Gonzalez and R. E. Woods, Digital Image Processing. Upper Saddle
River, NJ, USA: Prentice-Hall, 2008.
[27] Y. Wang, Q. Chen, and B. Zhang, “Image enhancement based on equal
area dualistic sub-image histogram equalization method,” IEEE Trans.
Consum. Electron., vol. 45, no. 1, pp. 68–75, Feb. 1999.
[28] R. M. Haralick, S. R. Sternberg, and X. Zhuang, “Image analysis us-
ing mathematical morphology,” IEEE Trans. Pattern Anal. Mach, Intell.,
vol. PAMI-9, no. 4, pp. 532–550, Jul. 1987.
[29] N. Otsu, “A threshold selection method from gray-level histograms,” Au-
tomatica, vol. 11, pp. 23–27, 1975.
[30] M. L. Quammen and C. P. Onuf, “Laguna madre: seagrass changes con-
tinue decades after salinity reduction,” Estuaries, vol. 16, pp. 302–310,
1993.
[31] R. G. Congalton and K. Green, Assessing the Accuracy of Remotely Sensed
Data: Principles and Practices. Boca Raton, FL, USA: CRC Press, 2008.
Maryam Rahnemoonfar received the Ph.D. degree
in computer science from the University of Salford,
Manchester, U.K., in 2010 and the M.Sc. degree in
remote sensing engineering from the University of
Tehran, Tehran, Iran, in 2005.
She is currently an Assistant Professor in Com-
puter Science and the Director of the Computer Vi-
sion and Remote Sensing Laboratory (Bina lab),
Texas A&M University Corpus Christi, Corpus
Christi, TX, USA. Her research interests include im-
age processing, computer vision, machine learning,
remote sensing, and synthetic aperture radar.
Abdullah F. Rahman received the Ph.D. degree in
hydrology and remote sensing from the University of
Arizona, Tucson, AZ, USA, in 1996.
He is currently a Professor at the School of Earth,
Environmental, and Marine Sciences, University of
Texas Rio Grande Valley (UTRGV), Brownsville,
TX, USA. He is a member of the Blue Carbon Sci-
entific Working Group, an international group of re-
searchers studying carbon in coastal ecosystems of
seagrasses, mangroves, and tidal salt marshes. Be-
fore joining UTRGV, he was an Associate Professor
at Indiana University, Bloomington, IN, USA. His expertise is on the use of
remote sensing for studying carbon stocks and fluxes of ecosystems.
Richard J. Kline received the Ph.D. degree in ma-
rine science from the University of Texas at Austin,
Austin, Texas, USA, in 2010 and the M.Sc. degree in
fisheries and aquatic sciences from the University of
Florida, Gainesville, Florida, USA, in 2004.
He is currently an Associate Professor at the
School of Earth, Environmental, and Marine Sci-
ences, University of Texas Rio Grande Valley
(UTRGV), Brownsville, TX, USA. His research in-
terests include the fields of conservation and ap-
plied ecology and physiology, especially applica-
tion of new technology to address research questions in coastal and marine
environments.
Austin Greene received the B.S. degree in evolution,
ecology, and biodiversity from the University of Cali-
fornia, Davis, CA, USA, in 2014 and the M.S. degree
in biological sciences from the University of Texas
Rio Grande Valley, Brownsville, TX, USA, in 2017.
He is currently working toward the Ph.D. degree at
the University of Hawaii at Manoa, Honolulu, HI,
USA, studying the drivers and spatial distribution of
disease on coral reefs.
His research interests include anthropogenic
drivers of environmental change and the development
of low-cost sensors to encourage their study.