Exploring Seafloor Stretching in Mariana Trench
Arc via the Squeeze and Excitation network with
High-resolution Multibeam Bathymetric Survey
Shasha Liu, Jie Wang, Lina Zang, RuiNian*,
Xiaoyu Li, Bo He
School of Information Science and Engineering
No. 238 Songling Road
Ocean University of China, Qingdao
E-Mails: liushasha117@163.com;
2896090535@qq.com; zanglina0103@163.com;
nianrui_80@163.com; 982502041@qq.com;
bhe@ouc.edu.cn
Abstract—Multibeam bathymetry data could represent nearly
continuous coverage depth measurements of the seafloor and
reveal geomorphological regions. Recent studies have utilized
multibeam bathymetry data to provide geological maps, but their
delineations were done manually. Manual classification and
delineation are inherently subjective and therefore can be
inaccurate. In this paper, we try to develop one strategy to
explore seafloor stretching in Mariana trench arc via squeeze
and excitation network, combining data clustering, slope and
gradient. In our experiments, we use the high-resolution
multibeam bathymetric data collected by NOAA Office of Ocean
Exploration and Research (OER). The geomorphological seabed
in the Mariana region is automatically classified into different
classes. The experimental results demonstrate that
geomorphological seabed classification strategy achieves a robust,
automated delineation approach.
Keywords—Multibeam Bathymetry; Geomorphological Seabed;
Classification; Convolutional Neural Networks
I. INTRODUCTION
There are so many exciting geological features within the
Mariana region. It’s because tectonic plates are colliding in
some areas and spreading apart in others, producing a wide
variety of environments that make the Mariana region such a
fascinating place to explore. Recent developments in
multibeam sonar, however, have resulted in detailed acoustic
surveys that provide an unprecedented view of the seafloor at a
broad range of spatial scales. With the advent of multibeam
technology [1], marine benthic habitat mapping has entered a
new era. Ship borne multibeam bathymetry surveys, have
provided spectacular detail of deep sea terrain revealing
numerous previously unrecognized features. Multibeam
surveys provide the detailed bathymetry data necessary for the
production of submarine Digital Terrain Models (DTM)
which, unlike traditional charting formats, lend themselves
well to terrain analysis. Multibeam data have proven their
value for habitat mapping and studies of the distribution of
benthic fauna [2].
Authorized licensed use limited to: Academia Navala Mircea cel Batran Constanta. Downloaded on October 05,2023 at 08:00:35 UTC from IEEE Xplore. Restrictions apply.
978-1-7281-1450-7/19/$31.00 ©2019 IEEE
Amaury Lendasse
Department of Information and Logistics Technology,
The University of Houston, Houston City, Texas
77204(713)743-2255, USA;
Arcada University of Applied Sciences, 00550 Helsinki,
Finland;
E-Mails: alendass@central.uh.edu;
lendasse@gmail.com
However, recent studies have utilized multibeam
bathymetry data to provide geological maps, but their
delineations were done manually. Manual classification (by
visual appearance) and delineation are inherently subjective
and therefore can be inaccurate[3].
One possible approach to automating the delineation of
seafloor regions involves classification analysis of multibeam
bathymetry data representing seafloor topography data.
Among them, we try to develop one strategy to explore
seafloor stretching in Mariana trench arc via squeeze and
excitation network, combining data clustering, slope and
gradient. We try to first cluster the high-resolution multibeam
bathymetric data. Cluster analysis is aimed at classifying
elements into categories on the basis of their similarity. One
kind of clustering by fast search and find of density peaks is
used for data clustering. Multibeam surveys can provide
detailed bathymetry data for geomorphological seabed from
which quantitative descriptor slope of the seabed terrain may
be obtained. Bathymetry data are parameterised using
quadratic surfaces. It is used a pixel window around a
particular point to approximate this surface. Different
geomorphological features can be represented by gradient
information extracted from the image. According to the slope
and gradient, we can make a more detailed geomorphological
seabed classification. In order to explore seafloor stretching in
Mariana trench arc, one kind of deep learning framework
called “Squeeze-and-Excitation” will be further introduced for
those more detailed geomorphological seabed classification
process. This mechanism that allows the network to perform
feature recalibration, through which it can learn to use global
information to selectively emphasise informative features and
suppress less useful ones [4].
The structure of the paper is as follows: Section II
describes our automated geomorphological seabed
classification model. Section III introduces the clustering
algorithms. Section IV introduces slope and gradient
information. In Section V, automated geomorphological
 seabed classification is introduced. Section VI lists our
experimental results and analysis. Conclusions are drawn in
Section VII.
II. GENERAL FRAMEWORK
In this paper, we try to develop one strategy to explore
seafloor stretching in Mariana trench arc via squeeze and
excitation network, combining data clustering, slope and
gradient. A brief flow chart of our proposed strategy is shown
in Fig. 1, which is made up of several steps, including
clustering algorithms, slope and gradient, squeeze and
excitation network. In order to explore seafloor stretching in
Mariana trench arc, we try to first cluster the high-resolution
multibeam bathymetric data. Cluster analysis is aimed at
classifying elements into categories on the basis of their
similarity. One kind of clustering by fast search and find of
density peaks is used for data clustering. Clustering
algorithms based on the idea that cluster centers are
characterized by a higher density than their neighbors and by
a relatively large distance from points with higher densities.
This idea forms the basis of a clustering procedure in which
the number of clusters arises intuitively, outliers are
automatically spotted and excluded from the analysis, and
clusters are recognized regardless of their shape and of the
dimensionality of the space in which they are embedded.
Multibeam surveys can provide detailed bathymetry data for
geomorphological seabed from which quantitative descriptor
slope of the seabed terrain may be obtained. Slope is thought
to be an important factor in determining benthic habitat and
colonization in the deep sea at a variety of scales. Flat areas
tend to exhibit different features that are different from those
on steeply sloping areas. Bathymetry data are parameterised
using quadratic surfaces. It is used a pixel window around a
particular point to approximate this surface since local slope is
a function of nearest-neighbour pixels. Different
geomorphological features can be represented by gradient
information extracted from the image. We adopt the gradient
operator which is essentially expressed by convolution masks.
Authorized licensed use limited to: Academia Navala Mircea cel Batran Constanta. Downloaded on October 05,2023 at 08:00:35 UTC from IEEE Xplore. Restrictions apply.
Gradient magnitude can be obtained through the partial
derivatives of the image along horizontal and vertical
directions using the gradient operator. In order to explore
seafloor stretching in Mariana trench arc, one kind of deep
learning framework called “Squeeze-and-Excitation” (SE)
will be further introduced for slope, gradient and original
depth information for those more detailed geomorphological
seabed classification process, in the following inter-connected
steps. For any given transformation (e.g. a residual block), SE
block could perform feature recalibration. a) Squeeze: The
features are first passed through a squeeze operation, which
aggregates the feature maps across spatial dimensions to
produce a channel descriptor. This descriptor embeds the
global distribution of channel-wise feature responses,
enabling information from the global receptive field of the
network to be leveraged by its lower layers. b) Excitation:
This is followed by an excitation operation, in which
sample-specific activations, learned for each channel by a
self-gating mechanism based on channel dependence, govern
the excitation of each channel. The feature maps are then
reweighted to generate the output of the SE block which can
then be fed directly into subsequent layers. SENets could be
developed by integrating SE blocks into popular network
architectures Residual Network (ResNet). This mechanism
that allows the network to perform feature recalibration,
through which it can learn to use global information to
selectively emphasise informative features and suppress less
useful ones.
III. ROUGH CLASSIFICATION
The classification of trenchs and island arcs in multibeam
bathymetry images is preliminarily determined by CFDP
(Clustering by fast search and find of density peaks), thereby
improving the classification accuracy of trenches and island
arcs
 Fig. 1. The flow chart of system
Cluster analysis is the formal study of methods and
algorithms for grouping, or clustering, objects according to
measured or perceived intrinsic characteristics or similarity.
Cluster analysis does not use category labels that tag objects
with prior identifiers, i.e., class labels. The aim of clustering is
to find structure in data and is therefore exploratory in nature
[5].
The clustering algorithm we used in the article is
density-based methods: CFDP (Clustering by fast search and
find of density peaks) [6]. The core idea of the algorithm is
that cluster centers are characterized by a higher density than
their neighbors and by a relatively large distance from points
with higher densities. Compared with DBSCAN
(Density-Based Spatial Clustering of Application with Noise)
algorithm, the algorithm process does not need to determine
the density threshold [7]. This approach only requires
measuring (or computing) the distance between all the pairs of
data points and does not require parameterizing a probability
distribution or a multidimensional density function. In this
method, the number of clusters arises intuitively, outliers are
automatically spotted and excluded from the analysis, and
clusters are recognized regardless of their shape and of the
dimensionality of the space in which they are embedded.
In this paper, we use clustering algorithm: CFDP
(Clustering by fast search and find of density peaks). The
algorithm has its basis in the assumptions that cluster centers
are surrounded by neighbors with lower local density and that
they are at a relatively large distance from any points with a
higher local density. For each data point i , we compute two
Authorized licensed use limited to: Academia Navala Mircea cel Batran Constanta. Downloaded on October 05,2023 at 08:00:35 UTC from IEEE Xplore. Restrictions apply.
quantities: its local density ρi and its distance δi from points
of higher density. This approach only requires measuring (or
computing) the distance between all the pairs of data points.
The algorithm flow can roughly be divided into the following
steps:
Computing the local density ρi . The local density
ρi of data point embedded image is defined as:
ρi =  x(dij − d c ) (1)
j
where x ( x ) = 1 if x < 0 and x ( x ) = 0 otherwise, and dc is a
cutoff distance. Basically, ρi is equal to the number of
points that are closer than dc .
Computing the distance δi . δi is measured by
computing the minimum distance d ij between the point i
and any other point with higher density: δi = min dij . For
j:ρ j > ρi
( )
the point with highest density, we conventionally take
δ i = max j ( d ij ) .
Find cluster centers. The only points of high δi and
relatively high ρi are the cluster centers. Other points have a
relatively high δi and a low ρi can be considered as
clusters composed of outliers. After the cluster centers have
been found, each remaining point is assigned to the same
cluster as its nearest neighbor of higher density.
 Find a border region for each cluster. Firstly, find a border To calculate slope from a regularly gridded digital
region, defined as the set of points assigned to that cluster but elevation models (DEMs), an assumption must be made about
being within a distance dc from data points belonging to the way the grid points represent the land surface. It is
other clusters. Then find for each cluster the point of highest common to use a 3 × 3 pixel window around a particular
point to approximate this surface since local slope is a
density within its border region. We denote its density by ρ b .
function of nearest-neighbour pixels. Different assumptions of
The points of the cluster whose density is higher than ρ b are shape within this window will provide different results.
considered part of the cluster core (robust assignation). The In this paper we use the following method to calculate the
others are considered part of the cluster halo (suitable to be slope:
considered as noise). As shown in the Fig. 2, points 4 and 2
are the cluster center, and after first three steps, point 3 and
point 1 are divided into clusters of point 2. Suppose that the
Slope = tan ( 2
Slopewe + Slopesn2 ) (2)
distance between point 3 and point 1 to point 4 is less than
dc , then point 3 and point 1 are both border region points, but Where the Slopewe and Slopesn can calculated by the
the density of point 1 is greater than point 3, so point 3 is following operators:
classified as the cluster of points 4, and density of points in
point 2 clusters that are less than point 1 density are ( e8 + 2e1 + e5 ) − ( e7 + 2e3 + e6 )
reassigned to other nearest clusters. Slopewe =
8 × Cellsize
(3)
( e7 + 2e4 + e8 ) − ( e6 + 2e2 + e5 )
Slopesn =
8 × Cellsize
Cellsize is the Grid DEM interval length, and e1 - e8 is the
3 × 3 pixel window:

e5 e2 e6

e1 e e3

e8 e4 e7
(4)
Different geomorphological features can be represented by
gradient information extracted from the image. In this paper,
we comprehensively compare the processing effects of Sobel
operator, Laplace operator (the second derivative test), Prewitt
Fig. 2. The clustering algorithm in two dimensions operator, Scharr operator, Roberts operator (derivative), and
finally we adopt the Scharr gradient operator, which is
IV. THE TERRAIN VARIABLE
essentially expressed by convolution masks as follow:
Multibeam surveys can provide detailed bathymetry data
for geomorphological seabed from which quantitative  −3 0 3   − 3 −10 −3
 −10 0 10  0 0 
descriptor slope of the seabed terrain may be obtained. Slope
is thought to be an important factor in determining benthic    0 (5)
habitat and colonization in the deep sea at a variety of scales.  −3 0 3   3 10 3 
The slope of the seabed is a key variable in a wide selection of ∇x ∇y
seabed analysis. It affects the speed of water currents and (a) horizontal direction (b) vertical direction
consequent erosion, slumping potential, and distribution of
benthic fauna [8]. The methods for estimating these The gradient magnitude (GM) is given by:
directional gradients from a 3 × 3 set of elevation grid cells
or points exist, in general slope is defined as the seabed G = ∇ 2x + ∇ 2y (6)
gradient in the direction of maximum inclination [9].
V. DETAILED CLASSIFICATION
Slope is thought to be an important factor in determining
benthic habitat and colonization in the deep sea at a variety of One possible approach to automating the delineation of
scales. Flat areas tend to exhibit different seabed facies and seafloor regions involves classification analysis of multibeam
support communities that are different from those on steeply bathymetry data representing seafloor topography data.
sloping areas [10]. Slope may also contribute to current flow Classification models estimated by various methods such as
amplification, which has consequences for the supply of food machine learning. For example, Artificial Neural Networks
to the benthic fauna. In relation to anthropogenic impact, we (ANNs) have been playing the dominant roles because of its
also note that slope may be a limiting factor in the use of benefits on generalization, flexibility, non-linearity, fault
particular fishing gears [11]. tolerance, selforganization, adaptive learning, and
computation in parallel. Except for ANNs, there are a lot of

Authorized licensed use limited to: Academia Navala Mircea cel Batran Constanta. Downloaded on October 05,2023 at 08:00:35 UTC from IEEE Xplore. Restrictions apply.
 machine learning variations, such as Support Vector Machine
(SVM), Deep Belief Networks (DBN) [12], the Restricted
Boltzmann machine (RBM) [13], Convolutional neural
networks (CNNs), etc.
In this paper, we use the squeeze and excitation network
for detailed classification of the obtained terrain variable slope
and gradient information, combined with the original
bathymetry data. A different aspect of architectural design
—— the channel relationship is investigated, by introducing a
new architectural unit, the “Squeeze-and-Excitation” (SE)
block. The convolutional neural network is based on the
convolution operation and extracts the information features by
blending the spatial information and channel information in
the local receptive field. In order to improve the presentation
capabilities of the network, many existing work has shown the
benefits of enhanced spatial coding. The
"Squeeze-and-Excitation" (SE) block focuses on the channel
and uses a novel architectural unit that learns to use global
information to selectively emphasize information features and
suppress less useful features. By explicitly modeling the
interdependencies between channels, adaptively recalibrating
the channel's feature response, improving the representation
of the network by explicitly modeling the interdependencies
between convolution feature channels. Moreover, the SE
block produces significant performance improvements for the
most advanced deep architectures available at a small
computational cost.
The Squeeze-and-Excitation block can be divided into
four steps:
The first step is transformation, using the symbol Ftr to
indicate. The purpose of this step is to match the size of the
input with the final output, and add SE block without
changing the original network structure. For simplicity of
exposition, in the notation that follows we take Ftr to be a
standard convolutional operator. Let V = [ v1 , v 2 , … , v C ]
denote the learned set of filter kernels, where vc refers to the
parameters of the c-th filter. We can then write the outputs of
Ftr as U = [ u1 , u 2 , … , u C ] where
C′
u c = vc ∗ X =  vcs ∗ x s
s =1
Here ∗ denotes convolution, v c = [v1c , v c2 ,… , v Cc ′ ] and
X = [x1 , x2 ,…, xC′ ] (to simplify the notation, bias terms are
s
omitted). Here v c is a 2D spatial kernel, and therefore
represents a single channel of vc which acts on the
corresponding channel of X. Since the output is produced by a
summation through all channels, the channel dependencies are
implicitly embedded in vc , but these dependencies are
entangled with the spatial correlation captured by the filters.
Our goal is to ensure that the network is able to increase its
sensitivity to informative features so that they can be
exploited by subsequent transformations, and to suppress less
useful ones. We propose to achieve this by explicitly
modelling channel interdependencies to recalibrate filter
responses in two steps, squeeze and excitation, before they are
fed into next transformation.
Authorized licensed use limited to: Academia Navala Mircea cel Batran Constanta. Downloaded on October 05,2023 at 08:00:35 UTC from IEEE Xplore. Restrictions apply.
The second step is squeeze. In order to tackle the issue of
exploiting channel dependencies, we first consider the signal
to each channel in the output features. Each of the learned
filters operate with a local receptive field and consequently
each unit of the transformation output U is unable to exploit
contextual information outside of this region. This is an issue
that becomes more severe in the lower layers of the network
whose receptive field sizes are small.
To mitigate this problem, we propose to squeeze global
spatial information into a channel descriptor. This is achieved
by using global average pooling to generate channel-wise
statistics. Formally, a statistic z ∈ R C is generated by
shrinking U through spatial dimensions W × H , where the
c-th element of z is calculated by:
W H
1

z c = Fsq ( u c ) =  uc ( i , j )
W × H i =1 j =1
(8)
The transformation output U can be interpreted as a
collection of the local descriptors whose statistics are
expressive for the whole image. Exploiting such information
is prevalent in feature engineering work. Using the simplest
method, global average pooling, while more sophisticated
aggregation strategies could be employed here as well.
The third step is excitation. To make use of the
information aggregated in the squeeze operation, we follow it
with a third operation which aims to fully capture
channel-wise dependencies. To fulfil this objective, the
function must meet two criteria: first, it must be flexible (in
particular, it must be capable of learning a nonlinear
interaction between channels) and second, it must learn a
non-mutually-exclusive relationship as multiple channels are
allowed to be emphasised opposed to one-hot activation. To
meet these criteria, SE block uses a simple gating mechanism
with a sigmoid activation:
s = Fex ( z, W ) = σ ( g ( z, W ) ) = σ ( W2δ ( W1z ) ) (9)
C
×C
where δ refers to the ReLU [14] function, W1 ∈ R r
C
C×
and W2 ∈ R r . To limit model complexity and aid
generalisation, we parameterise the gating mechanism by
(7) forming a bottleneck with two fully-connected (FC) layers
around the non-linearity, i.e. a dimensionality-reduction layer
with parameters W1 with reduction ratio r, a ReLU and then
a dimensionality-increasing layer with parameters W2 .
The last step is scale. The output of excitation step is a
tensor s, which is used to characterize the weight of C feature
maps in tensor U. The final output of the block is obtained by
rescaling the transformation output U with the activations:
xc = Fscale ( u c ,s c ) = s c ⋅ u c (10)
where X  = [ x , x , … , x ] and F ( u , s ) refers to
1 2 C scale c c
channel-wise multiplication between the feature map
u c ∈ R W × H and the scalar sc .
The activations act as channel weights adapted to the
input-specific descriptor z. In this regard, SE blocks
 intrinsically introduce dynamics conditioned on the input,
helping to boost feature discriminability.
An SENet is constructed by stacking a set of SE blocks. In
practice, it is generated by replacing each original block (i.e.
residual block) with its corresponding SE counterpart (i.e.
SE-residual block). The flexibility of the SE block means that
it can be directly applied to transformations beyond standard
convolutions. We develop SENets by integrating SE blocks
into ResNet. Residual networks and their variants have shown
to be highly effective at learning deep representations. We use
the Resnet-50 with SE (called SE-ResNet) block for
classification. Squeeze and excitation both act before
summation with the identity branch.
During training, we follow standard practice and perform
data augmentation with random-size cropping to 224 × 224
pixels and random horizontal flipping. By introducing this
deep learning framework called "Squeeze-and-Excitation", a
more detailed classification of seabed topography can be
achieved.
VI. SIMULATION EXPERIMENT
In our simulation experiments, we use the
high-resolution multibeam bathymetric data collected by
NOAA Office of Ocean Exploration and Research (OER).
Mariana region was surveyed using a Kongsberg EM302
multibeam sonar. Mariana trench arc region stands on the
western Pacific Ocean, and it located at 144º0'0''E~149º0'0''E,
13º0'0' N~23º0'0''N. The total time of exploration is 572.0812
hours, the total track length is 9727.3393 kilometers, and the
average speed is 17.003 km/ hour (9.1809 knots). An
automated delineation strategy has been set up, making it
possible to exploring seafloor stretching in Mariana Trench
Arc.
The classification of the geomorphology is based on
information such as the bathymetry data clustering, the slope
and the gradient zonal statistics. Figure 3 shows the
classification of trenches and island arcs initially obtained by
clustering by fast search and find of density peaks. Figure 3
shows the overall clustering of the entire exploration area,
Authorized licensed use limited to: Academia Navala Mircea cel Batran Constanta. Downloaded on October 05,2023 at 08:00:35 UTC from IEEE Xplore. Restrictions apply.
which is divided into two categories: trenches and island arcs.
Figure 4 shows the slope map obtained by correlation
calculation. Gradient could affect the accumulation of organic
sediment (and biomass). Different geomorphological features
can be represented by gradient information extracted from the
image. Figure 5 shows the gradient map obtained by
correlation calculation. Figures show the slope and gradient of
the local area. The accuracy of the classification can be
improved by combining the slope map and the gradient map
with the original sounding data. The squeeze and excitation
network adaptively recalibrates channel-wise feature
responses by explicitly modelling interdependencies
between channels to improve classification accuracy. SE
blocks produce significant performance improvements at
slight computational cost. To illustrate the cost of the
module, we take the comparison between ResNet-50 and
SE-ResNet-50, the computational overhead of
SE-ResNet-50 is increased by 0.25% compared to
ResNet-50. The accuracy of SENet is 84.9%, and the
accuracy of ResNet is 83.1%. The accuracy of SENet is
obviously superior to ResNet. The small additional
computational overhead required by the SE block is
justified by its contribution to model performance. We use
squeeze and excitation network to classify the multibeam
bathymetry datasets, and divide the trenches and island arcs
into more detailed landforms, and finally divide them into 8
categories including trench edge slope, trench seamount group,
trench bottom basin, slope fault basin, island slope ridge,
island slope deep water terrace, island platform, and island
slope. The classification results in the Mariana region are
shown in Figure 6. After comparing with the original map, we
can find that our geomorphological seabed classification
strategy can well realize the geomorphic delineation of the
Mariana region. Experiments demonstrate the favorable
performances of the proposed strategy in both efficiency and
accuracy.
 Figure 3 Clustering mapping

Figure 4 Slope mapping

Figure 5 Gradient mapping

Authorized licensed use limited to: Academia Navala Mircea cel Batran Constanta. Downloaded on October 05,2023 at 08:00:35 UTC from IEEE Xplore. Restrictions apply.
 (a) The original map
Figure 6 Automated geomorphological seabed classification
VII. CONCLUSION
In this paper, we try to develop one strategy to automated
classify geomorphological seabed via very deep convolutional
networks, combining data clustering, slope and gradient. In
our simulation experiments, we use the multibeam data
collected by NOAA Office of Ocean Exploration and
Research (OER). After comparing with the original map, we
can find that our geomorphological seabed classification
strategy can well realize the geomorphic division of the
Mariana region. Experiments demonstrate the favorable
performances of the proposed strategy in both efficiency and
accuracy.
ACKNOWLEDGEMENTS
This work is partially supported by the National
High-Tech R&D 863 Program (2014AA093410), the Natural
Science Foundation of P. R. China (31202036), the National
Program of International S&T Cooperation (2015DFG32180),
the National Science & Technology Pillar Program
(2012BAD28B05), and the Natural Science Foundation of P.
R. China (41376140).
REFERENCES
[1] Kenny A J, Cato I, Desprez M, Desprez, M., Fader, G., Schüttenhelm,
R. T. E., Side, J. An overview of seabed-mapping technologies in the
context of marine habitat classification. ICES Journal of Marine
Science, 2003, 60(2): 411-418.
[2] Wilson M F J, O’Connell B, Brown C, Guinan, J. C., Grehan, A. J.
Multiscale terrain analysis of multibeam bathymetry data for habitat
mapping on the continental slope. Marine Geodesy, 2007, 30(1-2): 3-35.
[3] Cutter Jr G R, Rzhanov Y, Mayer L A. Automated segmentation of
seafloor bathymetry from multibeam echosounder data using local
Authorized licensed use limited to: Academia Navala Mircea cel Batran Constanta. Downloaded on October 05,2023 at 08:00:35 UTC from IEEE Xplore. Restrictions apply.
(b) The classification results
Fourier histogram texture features. Journal of Experimental Marine
Biology and Ecology, 2003, 285: 355-370.
[4] Hu J, Shen L, Sun G. Squeeze-and-excitation networks. Proceedings of
the IEEE conference on computer vision and pattern recognition. 2018:
7132-7141.
[5] Jain A K. Data clustering: 50 years beyond K-means. Pattern
recognition letters, 2010, 31(8): 651-666.
[6] Rodriguez A, Laio A. Clustering by fast search and find of density
peaks. Science, 2014, 344(6191): 1492-1496.
[7] Ester M, Kriegel H P, Sander J, Xiaowei, X. A density-based algorithm
for discovering clusters in large spatial databases with noise. Kdd. 1996,
96(34): 226-231.
[8] Kostylev V E, Courtney R C, Robert G, Todd, B. J. Stock evaluation of
giant scallop (Placopecten magellanicus) using high-resolution acoustics
for seabed mapping. Fisheries Research, 2003, 60(2-3): 479-492.
[9] Jerosch K, Kuhn G, Krajnik I, Scharf, F. K., Dorschel, B. A
geomorphological seabed classification for the Weddell Sea, Antarctica.
Marine Geophysical Research, 2016, 37(2): 127-141.
[10] Ichino M C, Clark M R, Drazen J C, Jamieson A, Jones. D O.B, Adrian
P.Martin, Rowden A A, Shank T M, Yancey P H., Ruhl H A. The
distribution of benthic biomass in hadal trenches: a modelling approach
to investigate the effect of vertical and lateral organic matter transport to
the seafloor. Deep Sea Research Part I: Oceanographic Research Papers,
2015, 100: 21-33.
[11] Wilson M F J, O’Connell B, Brown C, Guinan, J. C., Grehan, A. J.
Multiscale terrain analysis of multibeam bathymetry data for habitat
mapping on the continental slope. Marine Geodesy, 2007, 30(1-2): 3-35.
[12] Hinton G E, Osindero S, Teh Y W. A fast learning algorithm for deep
belief nets. Neural computation, 2006, 18(7): 1527-1554.
[13] Hinton G E. Training products of experts by minimizing contrastive
divergence. Neural computation, 2002, 14(8): 1771-1800.
[14] Nair V, Hinton G E. Rectified linear units improve restricted boltzmann
machines. Proceedings of the 27th international conference on machine
learning (ICML-10). 2010: 807-814.