Classification of oil spill in the Krishna-Godavari offshore using ERS-1 SAR images with a fuzzy logic approach
R. Ramakrishnan1 and T. J. Majumdar2*
1Geo-Sciences Division, MPSG/UPSA, Space Applications Centre (ISRO), Ahmedabad –380 015, India
2Space Applications Centre (ISRO), Ahmedabad – 380 015, India
*[E-mail: tapan.j.majumdar@gmail.com]
Received 23 November 2011; revised 30 January 2012
In the present study new features are extracted from the dark spot and fuzzy based techniques were utilized for classifications of the dark spot into oil spill and look alike. Threshold is defined for detection of dark spot, which is obtained from histogram analysis, where the image histograms consist of two peaks. For scenes devoid of two peaks in the histogram, an empirical formula is developed from the scene statistics, which designate the threshold for segmentation of the dark spot. A centerline concept is introduced which bisects the slick along the major axis. This is further used to determine the width, shape and orientation of the suspected slick. Abrupt turn and curving of the slick is also monitored from the centerline. According to the characteristic of each feature, separate membership functions were assigned to obtain its fuzzy set. Few features were identified to have high discriminatory power, whose values were having marked contrast for oil spill and look alike. Five classes were defined: 1) oil spill, 2) tending to oil spill, 3) uncertain, 4) tending to look alike and 5) look alike. From 30 dark spots obtained from six ERS-1 SAR scenes, six were classified as oil spill and four dark spots as look alike. Six dark spots were classified as tending to oil spill and five as tending to look alike, whereas classification of nine dark spots was found to be uncertain.
Keywords: Synthetic Aperture Radar, Oil Spill Detection, Feature Extraction, Fuzzy Logic, Krishna-Godavari Offshore etc.
Introduction
In monitoring oil spills, space-borne Synthetic Aperture Radar (SAR) is found to be efficient than the other techniques used in remote sensing. Classifying oil spill from look alike in a Synthetic Aperture Radar (SAR) image still faces challenges owing to the discrepancy of a proper guideline to extract different feature from the suspected oil spill.
The brightness in a SAR image for ocean surface is due to the wind-generated short gravity-capillary waves. In SAR image, amount of backscattered radiation in the oil covered region is much less than the surrounding sea as the oil decreases the aerodynamic roughness of the ocean and consequently decreases the radar scattering. The degree of darkness depends on different sea state and to different dark forming characteristics. To decide whether the dark formation is due to oil spills or other phenomena called as look alike, still lacks a sufficient answer. Dark formation detection or segmentation is considered as the fundamental step in oil spill detection systems and constitutes the first step in oil spill detection approach.
The goal of segmentation step is to segment out all possible oil spill candidates. Dark spot segmentation is
usually done through threshold methods. Adaptive algorithm is applied where the threshold is set corresponding to mean value estimated in a moving window1. Threshold algorithms were developed from the mean and the standard deviation obtained from the DN value of the scene2,3. Threshold values were also obtained from histogram of the DN values4. Dark spot detection was done using Bayesian models based on curve evolution and geometric flow and from edge detection method by Sobel filter5.
Segmentation of dark spot is followed by feature extraction to classify the segmented dark regions to oil spill or look alike. Features can be grouped into three, which is generally described as shape: representing the geometry and orientation of the slick, contrast or homogeneity: referring to the physical behavior of the spill and the third one is the surroundings or contextual features of the segmented region6-9. Inclusions of contextual features were found to improve the classification of oil spill10.
Classifications are done to discriminate the dark region into oil spill and look alike based on the features extracted.
The most commonly used are statistical classifiers, where classification decisions are based on probability.
Probability assigned using Gaussian density functions11, Mahalanobis classifier7 and a statistical approach based on multi regression analysis3 were the kind of statistical classifiers used to classify oil spill and look alike. Neural network4 and fuzzy logic2 techniques were implemented to classify the detected dark region. Extraction of features and selection of features for classification lacks a proper guideline12. Features should possess good discriminatory power for improving the classification algorithms. New classification methods with evolutionary algorithms and automatized threshold selection techniques have already been attempted for various feature extraction13,14.
Present study aims at improving the extraction of features with introduction of a centreline concept in feature extraction procedure and applying a fuzzy based classification technique for identification of oil spill near the Krishna-Godavari Offshore. In classification procedures relying on knowledge base, discrepancy occurs while using natural languages like “low”, “high”
etc and furthermore a sufficiently large set of training data are required15,16. To overcome these problems fuzzy possibility functions were assigned to each extracted features.
Methodology
Six ERS-1 IM (imaging mode) VV polarized image near Krishna-Godavari Delta in the east coast of India were analyzed to detect the oil spill (Figs. 1 and 2). Land masking was done and subsets with dark spots were taken.
Dark spot detection/Segmentation
Backscattering (sigma-0) values were generated from the amplitude. Two methods were introduced
and implemented for the segmentation of dark spot from the image. The histogram was analyzed and was observed to consist of two peaks as shown in Fig. 3a.
The smaller peak represents the pixels in the dark spot
Fig. 1Study area, FCC AWIFS image.
Fig. 2Location map of six ERS-1 images.
Fig. 3 a)Scenes having bimodal histogram b) scenes devoid of bimodal histogram.
having low sigma-0 value, and the larger peak represents the brighter pixels. The threshold τ is assigned to the sigma-0 value with minimum number of pixel lying between the smaller peak and the higher peak. A shift to the right side is given to the threshold value. Pixel with value below τ is identified as dark spot. In certain scene even though dark spots are visible, smaller peak in the histogram is not prominent (Fig. 3b), and histogram based segmentation is found to be inappropriate. The mean (µ) is calculated from the image and it was found that µ/4 approached the threshold value. The τ is shifted to the right with a value equal to µ/5, and the threshold is calculated from the experimental formula as
τ = µ/4 + µ/5 (1)
Figure 4 shows dark spot segmentation for images with bimodal histogram (a) using histogram analysis and for images devoid of two peaks in the histogram (b) which were segmented out using the statistical approach.
Feature extraction
Perimeter, area, complexity and shape index are commonly taken as shape features. A centerline concept is introduced here where centerline is a line that bisects the segmented region into almost two equal halves. Centerline is constructed by moving a small window from the upper side of the dark object,
which divides the perimeter in such a way that two points on the perimeters (positioned almost opposite to each other) are made to intersect on to the edges of the window. Bisect of the two points are marked. A series of such bisect points are obtained by moving the window. The bisect points obtained are connected to give the centerline. Figure 5 explains the concept for constructing the centerline, the window chosen is much smaller than that depicted. Figure 6 shows the segmentation and centerline extraction procedure.
Angle of turning of the slick is found to be an important parameter in describing a dark spot to be an oil spill or look alike17. Abrupt turns due to sudden
Fig. 4 a)Segmentation of scenes with bimodal histogram, b) segmentation of scenes devoid of bimodal histogram.
Fig. 5Illustration describing the centreline extraction.
Fig. 6 aSAR image b) dark spot segmentation c) separation of different segments and d) centreline extraction.
change in the wind direction or surface current is most likely to happen to oil slick whereas natural slick (look alike) tends to disappear in this condition due to its low viscosity. The angle of turning is calculated from the centerline. The centerline is further used to calculate the width of the segment. Width is measured from the lines perpendicular to the centerline intersecting with perimeter at both ends. Sobel operator is used to obtain the gradient of the intensity value along the width-line to determine the spread of the slick from the centerline. Mean Haralick texture and homogeneity is calculated from Grey Level Co-occurrence Matrix (GLCM). The segmented dark spots are separated out for individual analysis.
Winds for the corresponding date were obtained from the NCEP model results available at ftp://ftp.cdc.nova.gov/Datasets/encep.reanalysis/daily avgs/surface_gauss. From each dark region 18 features were extracted which are 1) Wind 2) Area 3) Thickness, given by L/W, L is the length of the centerline and W the width 4) Segment mean (µs) 5) Segment standard deviation (σs) 6) Segment power to mean ratio ((σs/µs) 7) Background mean (µB) 8) Back ground standard deviation (σB) 9) Background power to mean ratio ([(σB/µB) 10) Ratio of power to mean ratios:
(σs/µs)/(σB/µB)] 11) Mean Border Gradient 12) Standard Deviation Border Gradient 13) Slick Smoothness Contrast 14) Local Area Contrast 15) Homogeneity 16) Mean Haralick Texture 17) Centerline Turn Angle, obtained to monitor any abrupt changes in the flow of the slick 18) Backscatter gradient along the width-line, calculated for the spreading of the backscatter value within the slick. Classification based on fuzzy logic was applied to distinguish oil spill from the look alike. Membership function for each feature was defined according to the “lowness” or “highness” of the value of the features given in the knowledge base published4,7,8. The dark regions were distinguished to five classes namely 1) oil spill, 2) tending to oil spill 3) uncertain 4) tending to look alike and 5) look alike.
Results and Discussion
In the present study, two methods for detecting dark region were applied successfully and centerline was extracted for each dark region. Fuzzy logic was applied to classify the dark regions into oil spill and look alike.
Dark region detection
Detection of dark region was done by histogram method and statistical method as explained. From the
six ERS-1 SAR images, subsets were taken which includes the dark region. Small clusters of dark regions obtained were combined together to form a single larger area. A total of 30 dark regions were detected. The statistical method was applied for the entire data sets (even with bimodal histogram) and found to be valid since for the entire data sets, dark region got separated from the background.
Feature extraction
Feature extraction was made more sensible using the centreline extracted for each dark area (Fig. 5). In previous studies11, width has been calculated as the minor axis of the ellipse representing the slick, where as the perpendicular along the centreline given an accurate estimation of the width. The backscatter variation is calculated along the width-line that gives more realistic spreading of the backscatter than calculated from the Sobel gradient filter. Haralick texture and homogeneity were calculated along the centreline using GLCM.
Classification
Fuzzy logic was found to be useful in classification for areas where there is a lack of training data sets available. Wind was taken as contextual feature and only dark areas above a threshold wind of 3 m/s is considered to be oil spill10. From the total 18 features extracted few features were found to possess good discriminatory power. Geometric features like area, thickness, turn angle, textural features like mean of gradient along the border, standard deviation of gradient along the border, local area contrast and backscatter gradient along the width are observed to be more useful to discriminate the oil spill from the look alike.
Two-sided Gaussian membership function was used to determine the fuzzy set for the feature representing the area of the dark region. Oil spill usually do not cover extremely large area. Areas that are too small have the possibility of being look alike and the darkness is considered as due to local disturbances and backscatter errors. For thickness, z-shaped membership function is defined since look alike have the possibility of being round shape. A sigmoid membership function is given to define the fuzzy set for angle of turning. Dark regions having abrupt turns are usually demarcated as oil spill. Due to change in wind direction or surface current, natural slick tends to disappear; but the oil slick, because of its higher viscosity, retains its shape17. For gradient along
the perimeter of the dark region, S-shaped membership function is assigned to calculate the fuzzy set. A sharp difference in backscatter value is expected for oil spill.
Each feature was treated independently and separate membership functions were assigned to derive the fuzzy set corresponding to each feature.
Certain features were also considered to be dependent on the other. Turn angle of a spill is counted as prominent for thinner spill. Homogeneity for a small dark area is assigned as slightly low for oil spill. For a dark area, fuzzy set values corresponding to each feature were added assigning additional weightages to prominent features. From the results obtained, six dark areas were classified as oil spill and four dark regions as look alike. Six dark areas were classified as tending to oil spill and five dark areas as tending to look alike. Classification of nine dark areas was found to be uncertain (Figs. 7, 8 and 9).
Conclusions
The dark spot detection from SAR images was done using two different methods. Segmentation of dark spot from histogram appears to be the best method for image subscene having bimodal histogram. An empirical formula was used to determine the threshold of the scenes devoid of bimodal histogram. Different features of the slick were extracted using simple methods and centerline concept was introduced in the feature extraction, which further improves the classification models. Oil spill and look alike were differentiated based on fuzzy logic applied to the extracted features.
Certain prominent features were identified that have high discriminative power to distinguish between oil spills and look alike. However nine dark regions were not able to be classified and were categorized as uncertain. For making the classification algorithm more efficient, inclusions of improved features were to be made like, difference in width along centerline of the slick. A large oil spill would probably have shape with larger width along the spill site, which gets thinner as it gets advected due to currents or wind. The variation of the backscatter calculated along the centerline can be used as a parameter to detect the flow of the slick.
Contextual features like ocean currents (which influences the advection of the slick), waves (creates turbulence and hence the dispersion of the oil), suspended sediment concentration (induces sedimentation and thus settling of the oil) are to be incorporated for the improvement of the classification.
Fig. 7Classification of the dark spot into Oil spill, Tending to oil spill and Uncertain
Fig. 8Classification of dark spot to Tending to look alike.
Fig. 9Classification of dark spot to look alike
Acknowledgements
The authors wish to thank Shri A.S. Kiran Kumar, Director, SAC for his keen interest in this study. RR is thankful to Dr. J.S. Parihar, DD, EPSA, Dr. Ajai, GD, MPSG and Dr. A.S. Rajawat, Head, GSD for providing opportunity to carry out this work. TJM is thankful to CSIR, New Delhi for awarding Emeritus Scientist Fellowship since January 2011.
References
1 Solberg A, Brekke C & Husoy P O, Oil spill detection in RADARSAT and ENVISAT SAR images, IEEE Trans Geosci and Rem Sens, 45 (3) (2007) 746–755.
2 Karathanassi V, Topouzelis K, Pavlakis P & Rokos D, An object-oriented methodology to detect oil spills, Int J Rem Sens, 27 (23) (2006) 5235-5251.
3 Nirchio F, Sorgente M, Giancaspo A, Biamino W, Parisato E, Ravera R & Trivero P, Automatic detection of oil spills from SAR images, Int J Rem Sens, 26 (6) (2005) 1157–1174.
4 Frate F D, Petrocchi A, Lichtenegger J & Calabresi G, Neural networks for oil spill detection using ERS-SAR data, IEEE Trans Geosci and Rem Sens, 38 (5) (2000) 2282–2287.
5 Karantzalos K & Argialas D, Automatic detection and tracking of oil spills in SAR imagery with level set segmentation, Int J Rem Sens, 29 (21) (2008) 6281–6296.
6 Ardhuin F G, Mercier G & Garello R, Oil slick detection by SAR imagery: potential and limitation, Ocean 200.
Proceedings, V. 1, pp. 164-169, September 2003.
7 Fiscella B, Giancaspro A, Nirchio F, Pavese P & Trivero P, Oil spill detection using marine SAR images, Int J Rem Sens, 21 (18) (2000) 3561–3566.
8 Topouzelis K, Stathakis D & Karathanassi V, Investigation of genetic algorithms contribution to feature selection for oil spill detection, Int J Rem Sens, 30 (3), (2009) 611–625.
9 Topouzelis K, Karathanassi V, Pavlakis P & Rokos D, Detection and discrimination between oil spills and look- alike phenomena through neural networks, ISPRS J Photogramm & Rem Sens, 62 (2007) 264-270.
10 Brekke C & Solberg A, Oil spill detection by satellite remote sensing, Rem Sens Environ, 95 (1) (2005) 1–13.
11 Solberg A H S, Storvik G, Solberg R & Volden E, Automatic detection of oil spills in ERS SAR images, IEEE Trans Geosci and Rem Sens, 37 (4) (1999) 1916–1924.
12 Kubat M, Holte R C & Matwin S, Machine learning for the detection of oil spills in satellite radar images, Machine Learn, 30 (2) (1998) 195–215.
13 Bandyopadhyay S & Maulik U, Genetic clustering for automatic evolution of clusters and application to image classification, Pat Recog, 35 (2) (2002) 1197-1208.
14 Majumdar T J & Bhattacharya B B, Extraction of shoreline and drainage patterns from aerial MSS thermal IR data over Cambay basin, India - an attempt to automatize the threshold selection, Pat Recog, 24 (2) (1991) 157-164.
15 Klir G J, St. Clair U & Yuan B, Fuzzy Set Theory, Foundation and Applications, [Prentice Hall, New Jersey], 1997.
16 Blonda P N, Pasquariello G, Losito S, Mori A, Posa F &
Ragno D, An experiment for the interpretation of multitemporal remotely sensed images based on a fuzzy logic approach, Int J Rem Sens, 12 (3) (1991) 463–476.
17 Topouzelis K, Oil spill detection by SAR images: dark formation detection, feature extraction and classification algorithms, Sensors, 8 (2008) 6642-6659.
