Ocean Internal Wave Detection from SAR

2020 Third International Conference on Advances in Electronics, Computers and Communications (ICAECC) | 978-1-7281-9183-6/20/$31.00 ©2020 IEEE | DOI: 10.1109/ICAECC50550.2020.9339511

Images using Particle Swarm

Optimization
Chikati.Divya1 S. Vasavi 2 A.Sashikanth Sarma3
I/II M.Tech Professor Scientist
1,2
V R Siddhartha Engineering College, India
3
Space Application Center , ISRO, Ahmadabad, India
1
[email protected], [email protected], [email protected]

Abstract: The internal waves are the waves that occur in the Image processing made easy to take out the needed information
deep ocean. Many researchers found different methods for the from the images. By considering this principle we will
detection of internal waves. As we all know that machine learning
determine the internal waves based on pixel values or intensity
is the fastest growing technology to solve complicated problems.
So, here in this paper we took this advantage for the detection of values. Pixel values represent the important information in an
internal waves because this is also considered as a complicated image. In SAR image the internal waves viewed as vertical or
problem. Based on this a novel method is proposed for internal horizontal stripes. The canny edge detection is the famous
wave detection using neural networks which is a method that acts algorithm which it detect the edges while it suppress the speckle
as a human brain. A novel system using Particle swarm noise in SAR images. Fig 1(a) and (b) Shows the SAR image
optimization is proposed in this paper to automatically detect and the image that contains Edges of internal waves [6]. To
internal waves. Such system can differentiate with surface extract the features of these edges OpenCV, which is a library in
wavelets and slicks. This paper describes extraction of internal python programming, provides ORB detector. This paper can
wave parameters from SAR images. Initially ocean internal wave
use these features to construct the model that automatically
detection is modeled based on optimization algorithms. In the
proposed method, two models such as feature representation detects the internal waves using neural networks which is an
model and prediction model will be build. automatic learning system and it works as a human brain. The
main advantage of neural networks is that it tests the new object
Keyword: SAR, Particle Swarm Optimization, KDV , Feature which is new to the system or not trained object.
extraction (a)

I. INTRODUCTION

Physical oceanography has three types of ocean layers, those

are surface ocean, deep ocean and seafloor sediments. The
waves occur on the surface ocean are the surface waves
whereas internal waves occur in the deep ocean due to the
density or temperature changes. Usually the ocean surface is
warm compared to the deep ocean because of the atmosphere.
When the temperature is high then the density is low and vice
versa. So, when interface between these two layers is disturbed
because of density changes internal waves are formed. Finding (b)
these waves is very important and crucial task in physical
oceanography. Synthetic Aperture Radar (SAR) can create 2-D
or 3-D images of landscapes. The main advantage of SAR
images are it can capture the images through clouds and in
dark, mist conditions. So by using SAR images manually we
can determine the internal waves. But it is time taking and cost
effective process. It is indeed that we have to create an
automatic internal wave detection method. Many authors
found the methods using acoustic, frequency and amplitudes
signals. But these methods are difficult to understand. The
easy and effective way is using image processing integrated
with machine learning methods.
Fig1. (a)SAR image (b) after applying edge detection algorithm[6]

Authorized licensed use limited to: IEEE Xplore. Downloaded on May 15,2021 at 06:37:00 UTC from IEEE Xplore. Restrictions apply.
Internal Wave trains around Trinidad, as seen from space is automated methods for extracting target features and
given in Fig 2. classifying internal waves.

II. LITERATURE SURVEY

Authors of [1] explained An effective internal wave detection

method is using PCANet. In this they collect the images by
using DJI drone. First they labeled each image as internal
wave or non internal wave. Finally they apply the
PCANet(principle component analysis network) for training
and testing new images. To locate and detection of internal
waves the authors of [2] use wavelet transform which gives
the time localization of the frequency components. Wavelets
are poor for giving the information about orientation. Random
transform is used to orientate the captured internal waves. For
the characterization of IW the author uses continuous wavelet
transform. The image from the sentinel-3 satellite is used in
[3] to detect interface waves. Author integrate the
OLCI(Ocean and Land Color Instrument), which has spatial
resolution about 300m, and SRAL provides 100% SAR
altimetry data. These two instruments are boarded on sentinel-
Fig 2: Image of internal waves off northern Trinidad, taken
3A satellite. Author used the Ku and C bands of SRAL to find
from International Space Station (ISS) on 18 January
the frequency of the signal.[4] is based on previous work the
2013[21]
author uses OLCI and SRAL for the detection of IW packets.
Apart from that it also uses a multimodal deep learning
Internal waves at tidal frequencies are produced by tidal
method SONet which is a library that build on the top on
flow over topography/bathymetry, and are known as internal
tensor flow. If the amplitude of internal waves is high then
tides as in Fig 3.
these waves are called as internal solitary waves. To detect
these waves the authors of [5] presented four methods those
are random forest (RT), conventional neural network (CNN)
and multilayer perception to construct the model for training
the internal waves. SVM (support vector machine) is used to
classification. Grid search SVM is best fit for the detection of
ISW. In this the total data is divided into ¾ for training and ¼
for testing.This work gave the best results with minimum root
mean square errors and average relative errors. [6] explains
that ENVISAT is a satellite that consist radar and optical
instruments. ASAR(advanced synthetic aperture radar) is
boarded on ENVISAT and it can detect any surface changes
on the earth. It automatically detects the internal waves and
also generate automatic maps of the internal waves over the
UK continental shelf. The author uses the canny edge
detection algorithm to detect the internal waves. The
algorithm modulated using ASAR images to improve the
accuracy and performance. Acoustic sensors are used in [7]
for the detection of internal waves. When the ISW is absent
then the acoustic data is remains unchanged also referred to as
reference data. Whereas when ISW pass through the sensor
node the acoustic data is changed. Then author determines the
Fig 3: A satellite synthetic aperture radar image of internal correlation between received data and reference data.
waves east of Honshu. The figure is adapted from a figure in Miniaturized microfiber optic salinity sensor is used in [8] for
Jackson’s internal wave atlas [22]. the detection of ISW. This sensor is worked in harsh
environments and it has a good stability in water current. The
The generation of internal waves in Indian Scenario may be author performs an experiment in laboratory and CCD
because of barotropic tides at Andaman and Nicobar cameras were used to capture the propagation of ISW. A MZI
Islands, non linear internal waves in the Andaman Sea, sensor is placed in a tank that consists of different densities of
semidiurnal tidal internal waves in the Indian Coast. water. When the ISW is passing through the sensor then
salinity is low. The authors of [9] represent Empirical Mode
A Motivation Decomposition (EMD) method for the internal wave
In SAR images we can find Ocean internal waves with detection. First, the signal is divided into n IMFs. Wave
irregular bright or dark stripes. Detecting these stripes with energy theory says that which has more diversion that
manual methods may take time and hence requires contains the more energy. Internal wave contains the energy

Authorized licensed use limited to: IEEE Xplore. Downloaded on May 15,2021 at 06:37:00 UTC from IEEE Xplore. Restrictions apply.
in the form of signal. So, signal which consists of higher
energy that is internal wave.
To determine the amplitude of an internal wave the authors of
[10] used three deep learning models and results are shown
that the amplitude produced by deep learning models are close
to the actual values. In model 3, 15 types of texture
characteristics are used as inputs which shows best results
compare with the measured values. This paper concludes that
usage of texture characteristic parameters of optical remote
sensing images are best to determine the amplitude of internal
waves. AUV (Autonomous underwater vehicle) are also used
in [11] for detection of internal waves. Using of multiple
AUV’s will give best results. Here two AUV’s are used named
as Unicorn and Harpo. They can communicate and transfer the
data through acoustic propagation. Two AUV’s consists of
CT, pressure sensors and a computer to process the data
automatically. Here they used MOOS software system. This
software boarded on both AUV’s to allow communication
with each other. To detect nonlinear internal waves [12] uses
2-Dimensional wavelet transform. Inherently SAR images
have speckle noise that may be effect the detection process.
So, first we have to remove the noise in the SAR images.
Fig 4:SAR image showing internal waves in Andaman sea[16]
DWT is used to detect the internal waves and CWT is used to
detect the wavelength of IW packet. They conclude that Himawari-8 is used in [17] to continuously tack the nonlinear
wavelet analysis is very useful tool when analyzing nonlinear internal waves. Author uses four cases of nonlinear internal
internal waves. MODIS (Moderate Resolution Imaging waves of Indonesia Sea. Based on this data, the top nine
Spectra radiometer) images have high resolution, swath leading nonlinear internal waves were calculated. Himawari-8
area[13]. A survey based on MODIS imagery is conducted is geostationary satellite that has high temporal and spatial
from august 2002 to May 2004 and captured 3581 internal resolution sensors. This provides the special power to trace
wave events in different areas. MODIS imagery can be used the nonlinear internal waves. [18] uses SAR images taken
to combine with other sensors to give the best results. Pairs of from the ERS-1 SAR and RADARSAT. To detect the internal
images that are different in time can be used to calculate the soliton amplitudes here two methods are used. By using curve
propagation speed of internal waves. It gives accurate results. fitting method the author obtain the characteristic of an
SAR images of the gulf of California have appeared internal soliton from SAR image. Peak-to-Peak method is
propagation of IWs in the north part of gulf[14]. By taking used to determine the high and low peaks of the radar
series of nine repeated shows that the internal waves observed backscatter. They implement methods on Portuguese
are caused by tides. The wave packets reach its maximum in continental shelf and South China Sea. Table 1 compares
spring tides and minimum in neap tides. The author represents different methods to detect internal waves
that the observed waves are nonlinear waves with peak-to-
peak amplitude. The authors of [15] uses two methods for Table 1: Different methods to detect internal waves
linear and nonlinear internal waves. Linear internal waves are
detected using fast Fourier transform (FFT). It converts the Method Advantage Disadvantage
original data to frequency signal for plain results. For the
detection of nonlinear signals Fourier transform is not OLCI Higher Higher Cost
relevant. So, the author uses the CWT for the detection of Resolution
nonlinear internal signals. The author uses the data sets SAR Independent of Speckle Noise
examined for the Andaman sea region. Fig.4 shows the Weather
internal waves in Andaman sea[16]. An ISW has short MODIS Consists different Can’t capture
wavelengths. The author uses SRAL to detect the ISW. To levels of images images when
identify the signatures of ISW they used SRAL and with the clouds are
cooperation of OLCI (Ocean Land Color Imager). Here present
Daubechies wavelet of order 3 is used. For the implementation
the author implements this method on Andaman Sea and
Amazon River. Fig 4. Show different methods to capture In SAR images the internal waves appear as dark stripes.
internal waves. Convolution neural network features are used in [19] for the
internal wave detection. The Fast R-CNN extracts the features
by using Convolutional networks.

Authorized licensed use limited to: IEEE Xplore. Downloaded on May 15,2021 at 06:37:00 UTC from IEEE Xplore. Restrictions apply.
 this apply sequence of filters such as Speckle, Sobel and
Different sizes of image travel to the Convolutional layer and majority
features are extracted by this layer. Using region of interest (ii) Edge detection using gradient filter
(ROI) they proposed region to the feature map. Fixed size of c. With different scales such as Internal waves detection using
image is traveled to fully connected layer. It replaces the fully local modulus maxima method wavelet transformation
connected layer with two output layers that are classification d. Internal wave packet enhancement and identification using
layer and regression layer. The results show that the use of edge linking
Convolutional networks gives accurate results. Energy e. Cluster based on the regions of neighboring features for
exchange is done by internal waves and it is very important. internal wave packet location
The active L-Band Microwave satellite sensor is used in [20] f. Create a binary image of the cluster
for the detection of internal solitons. This gives the accurate
results. Table 2 compares different sensors for the detection of
internal waves.

Table 2: Different Sensors for the detection of internal waves

Sensor Advantages Disadvantages

Acoustic sensor High sensitivity Not applicable
in real time
Microfiber- useful in harsh Cost is high
optic salinity environments
sensor
AUV It is very stable Can’t operate
and gives accurate everywhere
results

III. METHODOLOGY

This section provides complete functionality of proposed

method to the detection of internal waves automatically.
Optimization algorithm of Particle Swarm will be used with
data assimilation of HR satellite data. The proposed technique
is based on the multiscale edge representation of images,
which its basic idea is to represent an image with its edges at
multiple scales. Apriori knowledge is required in identification
algorithm to be able to automatically distinguish internal wave
features from internal wave look-alikes. Multivariate change
detection method such as k-nearest neighbour (k-NN)
estimation will be used. Fig.5 shows the process of the
proposed system.

1. Data Preprocessing:
a. Propose data assimilation forecast model with ocean state
variables such as wave height, tide height sea level etc for
ocean state reanalysis using Kalman filter. Fig 5: Process of Proposed Method
b. The main internal wave look-alikes observed in SAR
images are, for instance, ship and ship wakes that normally 3. Fourier analysis, to calculate internal wavelength
appear as quasicircular shapes followed by wake lines. Oil
slicks have a physical appearance given as black spots in the 4. Feature extraction such as temperature, salinity and
image whose contour is closed. Natural films are defined as a velocity variation, wave parameters from SAR imagery, the
single quasilinear dark lines. Eddies are defined as dark lines sea water density and oceanic internal wave pycnocline depth
in spiral form. are estimated from the Levites data, the surface currents
induced by the internal wave are calculated according to the
2. Wavelet Transformation for Internal wave detection and KDV equation as given in Eq(1) [24]
location from Satellite images
a. Reduce the image by averaging using Mean filter
b. Sea land classification by coastline detection using gradient + + + = 0,
filter and gray value thresholding technique
(i) Coastline detector using an appropriate selection of the
here t , x , g represents time , distance and elevation of the
threshold, used for land–sea segmentation, followed by
waves
contour tracing, and a refinement for contour precision. For

Authorized licensed use limited to: IEEE Xplore. Downloaded on May 15,2021 at 06:37:00 UTC from IEEE Xplore. Restrictions apply.
a. Compute multiscale gradients as a Wave Tail (WT) of the Data Requirements
SAR image containing possible internal wave features. Remote and in-situ measurements from sensing images,
b. Detect local gradient maxima and obtain the edge of the moored buoy stations for internal waves , Bathymetry,
image. Latitude φ and longitude λ of the measurement site, wind
c. Compute the wavelet maxima to obtain only the significant speed Wa and wind direction ϕa, air temperature Ta, and
edge points relative humidity r above the sea surface. The Seabird
d. Compute the automatic thresholds from the histogram near‐surface (z = 5 m) , Winds stress, Surface heat fluxes, and
analysis of the wavelet maxima potential density
e. Apply an edge-linking method to enhance possible internal S: Salinity
wave features and remove short segments in the image. T: Temperature
f. Identify the internal wave features using geometric shape of p: pressure
the internal wave signatures. min_w: lower vertical wavelength limit for integration
g. Apply a k-NN clustering algorithm on the detected internal max_w: upper vertical wavelength limit for Integration
wave features in the SAR image. nfft: number of points in fft calculation (default=length of
data)
5. PSO Algorithm for modelling axis: axis to perform calculations default=0
a. Initialization of each particle with lower and upper bounds window: window function to reduce variance loss in FFT
b.Initialize pbest to its initial position (default='hanning')
c. Initialize to the minimal value of the swarm: detrend: detrend method for fft (default='constant')
d. Initialize the Velocity with_strain: option to calculate eta from strain instead of
e. Repeat until a termination criteria is met from density surfaces. (default=True)
For each particle do
Pick random numbers
Update particle’s velocity. IV. CONCLUSIONS AND FUTURE WORK
Update particle’s position.
Update the best known position of particle i Mariners observe internal waves while sailing in the oceans in
Update the swarm’s best known position the form of bands. Research on internal waves in the coastal
f. Output that holds the best found solution. ocean is one of the most important tasks both in physical
oceanography and ocean monitoring network. Remote sensing
6. Error Estimation for internal waves detection in SAR images, multispectral
The evaluation is based on the computation of pixel-based satellite images and space-shuttle photographs is extensively
confusion matrices per tile, and an accumulated confusion studied and a novel system using Particle swarm optimization
matrix. From those matrices, quantization indexes such as is proposed in this paper to automatically detect internal
False Alarm Rate (FAR), per class completeness (recall), waves. Such system can differentiate with surface wavelets
correctness (precision) and F1 score and overall accuracy. and slicks. This paper described extraction of internal wave
The True Positive, tp, pixels are derived from the main parameters from SAR images. Initially ocean internal wave
diagonal elements, while the False Positive, fp, is computed detection is modeled based on optimization algorithms.
from the sum per column, excluding the main diagonal In the proposed method, two models such as feature
element. Likewise, the False Negative, fn, is the sum along the representation model and prediction model will be build.
row, excluding the main diagonal element. From those tp, fp, Future work is to implement and test the models on open
fn per object the usual measures are derived: source datasets that are available for free and also from
datasets such as ENVISAT data.
(i) Confusion matrix: It is a table to visualize classification
algorithm performance
(ii) Classification Accuracy Rate (CAR): This is used to REFERENCES
measure basing on confusion matrix as given in Eq(4). [1] Qinghong Dong, Shengke Wang, MuweiJian,Junyu Dong, An Internal
accuracy  (tp  tn ) / (tp  tn  fp  fn ) (4)
Waves Detection Method Based on PCANet for Images Captured from
UAV, China: ICST Institute for Computer Sciences, 2018
[23] [2] J. Rodenas and R. Garello, “A new automatic internal wave detection
(iii) Precision: It is used to measure relevancy of the result and characterization method for SAR images,” IEEE Oceanic
generated as defined in Eq(5). Engineering Society. OCEANS98. Conference Proceedings (Cat.
No.98CH36259),1998.
Pr ecision  tp / (tp  fp ) (5) [23] [3] Jie Zhang, Xudong Zhang, Observation of internal waves with OLCI
(iv) Recall: It is used to measure relevancy of the result and SRAL on board Sentinel-3, China: Chinese Society for
generated as given in Eq(6). Oceanography and Springer-Verlag GmbH Germany, 2018
Re call  tp / (tp  fn ) (6) [23]
[4] J. Kusche1, L. Drees, MULTI-MODAL DEEP LEARNING WITH
SENTINEL-3 OBSERVATIONS FOR THE DETECTION OF
(v) F-Measure as given in Eq(7). OCEANIC INTERNALWAVES, Germany: ISPRS Annals of the
F1=2*(precision*recall)/(precision +recall ) (7) [23] Photogrammetry, Remote Sensing and Spatial Information Sciences,
(vi) FAR is FAR  fp / ( fp  tn) Volume V-2-2020, 2020 XXIV ISPRS Congress, 2020
[5] Wang, Jing , Study on Inversion Amplitude of Internal Solitary Waves
Applied to Shallow Sea in the Laboratory, Germany: IEEE
GEOSCIENCE AND REMOTE SENSING LETTERS, 2020
[6] A. Kurekin, Andrey , InternalWaves at the UK Continental Shelf:

Authorized licensed use limited to: IEEE Xplore. Downloaded on May 15,2021 at 06:37:00 UTC from IEEE Xplore. Restrictions apply.
 Automatic Mapping Using the ENVISAT ASAR Sensor, UK: MDPI, Portugal: MDPI, 2018
2020 [17] M. Santos-Ferreira, Adriana, SAR Mode Altimetry Observations of
[7] WANG, TONGCHEN, Detection and Parameter Estimation of Solitary Internal Solitary Waves in the Tropical Ocean Part 1: Case Studies,
Internal Waves Using Distributed Acoustic Sensors, China: IEEE Portugal: MDPI, 2018
Access, 2020 [18] Zheng, Quanan , Theoretical expression for an ocean internal soliton
[8] Liao, Yipeng , In-situ Detection for Ocean Internal Solitary Waves Based synthetic aperture radar image and determination of the soliton
on a Miniaturized Microfiber-optic Salinity Sensor, China: IEEE characteristic half width, USA: JOURNAL OF GEOPHYSICAL
SENSORS JOURNAL, 2020 RESEARCH, 2001
[9] Y Gu, K Fan, An improved technique for SAR internal wave signal [19]SUN Lina, BAO Sude, Detection of ocean internal waves based on Faster
detection, China: Indian Journal of Geo Marine Sciences,2018 R-CNN in SAR images, China: Journal of Oceanology and Limnology,
[10] Petillo, Stephanie, Exploiting Adaptive and Collaborative AUV 2019
Autonomy for Detection and Characterization of Internal Waves, [20] Nora G. Glembocki, Juan P. Pisoni, Internal solitary waves from L-band
China: IEEE JOURNAL OF OCEANIC ENGINEERING, 2014 SAR over the Argentine inner Patagonian shelf, China: Taylor &
[11] Pan, Xiaoyi , A deep-learning model for the amplitude inversion of Francis, 2020
internal wave based on optical remote-sensing images, FLORIDA: [21] Internal Waves off Northern Trinidad,
INTERNATIONAL JOURNAL OF REMOTE SENSING, 2017 https://earthobservatory.nasa.gov/images/80337/internal-waves-off-
[12] R´odenas, and Ren´eGarello, Wavelet Analysis in SAR Ocean Image northern-trinidad Last Accessed on 20-1-2020
Profiles for Internal Wave Detection and Wavelength [22] Jackson, C. R., “An Atlas of Internal Solitary-Like Waves and Their
Estimation,USA:IEEE TRANSACTIONS ON GEOSCIENCE AND Properties, 2nd ed.,” Global Ocean Associates, Alexandria, VA, 2004.
REMOTE SENSING,1997 Available at http://www.internalwaveatlas.com/Atlas2_index.html., Last
[13] Christopher, Jackson, Internal wave detection using the Moderate Accessed on 20-1-2020
Resolution Imaging Spectroradiometer (MODIS), China: [23] S. Vasavi, Reshma Shaik, Sahithi Yarlagadda, "chapter 12 Moving
JOURNAL OF GEOPHYSICAL RESEARCH, 2007 Object Classification in a Video Sequence Using Invariant Feature
Extraction", IGI Global, 2018, pp:1-25
[14] LEE-LUENG FU, BENJAMIN HOLT, Internal Waves in the Gulf of
[24] Zihua Liu *,Roger GrimshawOrcID andEdward Johnson, The Effect of a
California: Observations From a Spaceborne Radar, California:
Variable Background Density Stratification and Current on Oceanic
JOURNAL OF GEOPHYSICAL RESEARCH, 1984
Internal Solitary Waves, MDPI Fluids 2018, 3(4), 96
[15] Surampudi, Samvedya , Internal Wave Detection and Characterization
with SAR data, Tamil Nadu: IEEE, 2019
[16] M. Santos-Ferreira, Adriana, SAR Mode Altimetry Observations of
Internal Solitary Waves in the Tropical Ocean Part 1: Case Studies,

Authorized licensed use limited to: IEEE Xplore. Downloaded on May 15,2021 at 06:37:00 UTC from IEEE Xplore. Restrictions apply.