COMPARATIVE ALGORITHMS FOR OIL SPILL AUTOMATIC DETECTION USING MULTIMODE RADARSAT-1 SAR DATA Maged Marghany and Mazlan Hashim Institute of Geospatial Science and Technology (INSTeG), UniversitiTeknologi Malaysia 81310 UTM, Skudai, Johor Bahru, Malaysia Emails:
[email protected], :
[email protected]
ABSTRACT
2. DATA ACQUISITION
This study is utilized comparative algorithms for automatic detection of oil spill from different RADARSAT-1 SAR mode data (Standard beam S2, Wide beam W1 and fine beam F1). In doing so, three algorithms are implemented: Co-occurrence textures; post supervised classification, and neural net work (NN). The study shows that the standard deviation of the estimated error for neural net work of value 0.12 is lower than Entropy and the Mahalanobis algorithms. In conclusion, ANN performed accurately as automatic detection tool for oil spill in RADARSAT data.
The SAR data acquired in this study are from the RADARSAT-1 SAR that involves Standard beam mode (S2); W1; and (F1) beam mode data. SAR data are C-band and have a lower signal-to noise ratio due to their HH polarization with wavelength of 5.6 cm and a frequency of 5.3 GHz [1]. Further, RADARSAT-SAR data have 3.1 looks and cover an incidence angle of 23.7° and 31.0° [8]. In addition, RADARSAT-SAR data cover a swath width of 100 km. According to Marghany [1,3], Marghany and Hashim [9] oil spill occurred on 17 December 1999, along the coastal water of Malacca Straits.
Index Terms— RADARSAT-1 SAR,oil spill, Entropy, Mahalanobis neural net work (NN). 1. INTRODUCTION Oil spill or leakage into waterways and ocean spreads very rapidly due to the action of wind and currents. The study of the behavior and movement of these oil spills in sea had become imperative in describing a suitable management plan for mitigating the adverse impacts arising from such accidents. But the inherent difficulty of discriminating between oil spills and look-alikes is a main challenge with Synthetic Aperture Radar (SAR) satellite data and this is a drawback, which makes it difficult to develop a fully automated algorithm for detection of oil spill [1,3]. As such, an automatic algorithm with a reliable confidence estimator of oil spill would be highly desirable. The main objective of this work is to develop comparative automatic detection procedures for oil spill pixels in multimode (Standard beam S2, Wide beam W1 and fine beam F1) RADARSAT-1 SAR satellite data that were acquired in the Malacca Straits, using three algorithms namely, textures using co-occurrence matrix, post supervised classification, and neural network (NN) for oil spill detection with window size 7 x 7[1].
978-1-4577-1005-6/11/$26.00 ©2011 IEEE
2173
3. METHODS Three steps are required for automatically detect oil spill from SAR images: dark spot detection, dark spot feature extraction and dark spot classification. In doing so, three algorithms are implemented: Co-occurrence textures[1]; post supervised classification[2], and neural net work (NN)[3,56]. The co-occurrence textures involved Entropy algorithm [1]. Entropy is implemented to the different RADARSAT – 1 SAR mode data with window Kernel size of 7x7 pixels and lines. In addition, the post supervised classification is applied to RADARSAT-1 SAR data using Mahalanobis classifier[3]. Finally, Artificial Neural Net work technique (ANN) is also implemented to SAR data using back propagation algorithm[6]. 3.1. Entropy algorithm Co-occurrence is applied to categorize the image to “oil slick” and water. Entropy texture with 0° angular relationship and d=1 is employed. On other hand, RADARSAT-1 SAR gray tone can describe as texture (i.e., the microstructure). A co-occurrence matrix or cooccurrence distribution (less often co occurrence matrix or co occurrence distribution) is a matrix or distribution that is defined over an image to be the distribution of co-occurring values at a given offset. Mathematically, a co-occurrence
IGARSS 2011
matrix C is defined over an n x m image I, parameterized by an offset (ǻx,ǻy), as: n m 1, if I ( p, q) = i and I ( p +Δx, q +Δy) = j CΔx,Δy(i, j) = (1) ® p=1 q=1 ¯0, otherwise
¦¦
According to Marghany [1], the texture feature for oil spill and look-alike detections are computed by the following formula:
Ent = ¦¦ pij log pij
(2)
Where Ent is Entropy, i and j are the row and column, pxi and pyj are the marginal probability matrix obtained through the summation of pijin the direction of the row and column. In this study, the window size is 7x7 and 3x3 pixels and lines [1]. According to Marghany [1] and Marghany and Mazlan [8] the window size of 7x7 gives more details on an image.
dt2 = (v − M j )c −1 (v − M j ) 3.3. Artificial Neural Network (ANN)
Following Topouzelis et al.[3], the ANN’s and the pattern recognition (PR) technique, feed forward network with back-propagation algorithm are used in this study for both static and dynamic security assessment. For this application, a multi-layer feed forward network with error backpropagation has been employed [3],[5]. The major steps in the training algorithm are: Feed forward calculations, propagating error from output layer to input layer and weight updating in hidden and output layers [5]. Forward pass phase calculations are shown by the following equations between input (i) and hidden (j) [3], [6].
θj =
3.2. Mahalanobis Classification Formally, the Mahalanobis distant of a multivariate vector is given as[9]: DM
=
(v )
where V
(v − μ )T S
−1
= (v1 , v2 , v3 ............, vn )t from group of values
with mean μ = ( μ 1 , μ 2 , μ 3 .......... .., μ n ) , and S, is covariance matrix. In order to apply Mahalanobis classification procedures to different remote sensing data, let v be the feature vector for the unknown input, and let M1, M2 be the two classes: oil spill, and look-alike [2]. Then the error in matching v against Mj is given by [v- Mj], the Euclidean distance. A minimum-error classifier computes [v- mj] for j= 1 to 2 and chooses the class for which this error is minimum. If the covariance matrix is the identity matrix, the Mahalanobis distance reduces to the Euclidean distance. If the covariance matrix is diagonal, then the resulting distance measure is called the normalized Euclidean distance [9]: t
→
d = (v) where
δi
¦
n i =1
1+ e
( ¦ j wijθi +θ j )
1 ( ¦ k w jkθ j +θ k )
.
(6) .
(7)
1+ e where θ j is the output of node j, θ i is the output of node i, is the output of node
between node i and j, and
w jk is the weight connected
θ j is the bias of node j, θ k is the
bias of node k. In backward pass phase, error propagated backward through the network from output layer to input layer as represented in equation (4). Following Topouzelis et al. [6]. The weights are modified to minimize mean square error (MSE).
1 n m 2 (8) ¦ i =1 ¦ j = a ( d ij − yij ) n th th where d ij is the j desired output for the i training MSE =
pattern, and yij is the corresponding actual output. Finally,
(4)
error standard deviation is used to determine the accuracy level of each algorithm has been used in this study. In addition, Error standard deviation is used to determine the accuracy of feature detections in RADARSAT-1 SAR data[4,7].
vi over the sample
4. RESULTS AND DISCUSSION
vi2 − v 2j
δ i2
is the standard deviation of the
1
θk =
θk
(3)
(v − μ )
(5)
set. The input parameters from SAR data can impact the Mahalanobis classifier running when the variability inner classes is smaller than whole classifier group variability. In this context, if the classes M are badly scaled and the decision boundaries between classes are curved, the classifier accuracy is reduced. Some of the limitations of simple minimum-Euclidean distance classifiers can be overcome by using the Mahalanobis distance
dt2 that in
covariance matrix C form is
2174
Cleary, Figure 1 shows the entropy algorithm isolates the oil spill from its surrounding pixels in three different RADASAT-1 SAR mode data. In fact, entropy used to separate between oil spill pixels, sea water and land. According to Marghany [1], entropy is measure of uniformity in SAR image. In general, the entropy is a measure of variability or randomness because the concentration of the backscatter changes in relatively few locations would be non-random essentially. On other words, entropy measures the absolute variability in backscatter
change over the selected window. This result confirms the study of Marghany and Hashim [8].
Fig.1. Entropy algorithm for (a) Wide mode (W1), (b) Standard mode (S2) and (c) Fine mode (F1) Data.
estimate probabilities and also consider the variability of brightness values in each class. It is the most powerful classification methods when accurate training data is provided and one of the major widely used algorithms [9]. In this context, this Classification uses the training data by estimating means and variances of the classes, which are used to estimate the probabilities and also consider the variability of brightness values in each class [2,9]. It is clear that neural network algorithm is able to isolate oil spill dark pixels from the surrounding environment. In other words, look-alikes, low wind zone, sea surface roughness, and land are marked by white colour while oil spill pixels are marked all black. Fig. (3b) does not show any class presence or existence of oil spill event. Further, Fig. 3 shows the results of the Artificial Neural Net work, where 99% of the oil spills in the test set were correctly classified that using multilayer perceptron (MLP) neural network with two hidden layers. The net is trained using the backpropagation algorithm to minimize the error function. 99% of oil spills are automatically. This study agrees with study of Topouzelis et al., [3,5].
Figs. (2a) and (2c) shows that the slick has a large contrast to the gray-values surroundings. In addition, Figs. 2a and 2c shows the ability of Mahalanobis classification in determining the level of oil spill spreading. Mahalanobis classification can identify oil spill pixels from the surrounding environment (Fig. 2).
Fig. 3. Neural Network for Automatic Detection of Oil Spill from (a) W1, (b) S2, and (c) F1 Mode Data.
Fig. 2. Mahalanobis Classifier (a) Wide mode (W1), (b) Standard mode (S2) and (c) Fine mode (F1) Data.
This Classification uses the training data by estimating means and variances of the classes, which are used to
2175
Fig. 4, however, shows the standard deviation of the estimated error for neural net work of value 0.12 is lower than Entropy and the Mahalanobis. This suggests that ANN performed accurately as automatic detection tool for oil spill in RADARSAT data. The good performance of the neural algorithm encouraged a second phase where an optimization of the net from the point of view of the number of its adaptive parameters (units and connections) has been carried out by using a pruning procedure. Accordingly, a network is examined to assess the relative importance of its weights, and the least important ones are deleted. Typically, this is
followed by some further training of the pruned network, and the procedure of pruning and training may be repeated for several cycles. Clearly, there are various choices to be made concerning how much training is applied at each stage, what fraction of the weights is pruned, and so on. In the present work, every time a weight was removed, the new net until is trained, as in the case of the initial training. The overall error value approaches a value of convergence, and, since the study started with a net committing no errors, the pruning procedure was continued until it was realized that new removals involve errors in the classification task. The most important consideration, however, is how to decide which weights should be removed [3]. To do this, some measure of the relative importance was needed, or saliency of different weights. This result agrees with Topouzelis et al., [5],[6].
Fig.4. Error standard deviation of different algorithms.
5. CONCLUSIONS This study has demonstrated a comparative algorithms for oil spill automatic detection from different RADARSAT-1 SAR different mode data. Three algorithms are involved: Entropy, Mahalanobis, and Artificial Neural Network (ANN) algorithms. The study shows that ANN provide automatically oil spill detection with error of standard deviation of 0.12 which is lower than Entropy and the Mahalanobis algorithms. 6. REFERENCES [1] M. Marghany, “RADARSAT Automatic Algorithms for Detecting Coastal Oil Spill Pollution”. Int. J. of App. Earth Obs. and Geo. 3, 191-196, 2001. [2] B. Fiscella , A. Giancaspro, F. Nirchio, P. Pavese, and P. Trivero, P, Oil Spill Detection Using Marine SAR Images”. Int. J. of Remote Sens. 21,3561-3566, 2000. [3] K.Topouzelis, V. Karathanassi,P. Pavlakis, and D. Rokos, “Potentiality of Feed-Forward Neural Networks for Classifying Dark Formations to Oil Spills and Look-alikes”. Geo. Int. 24, 17919, 2009. [4] M. Marghany, A.P. Cracknell, and M. Hashim, “ Modification of Fractal Algorithm for Oil Spill Detection from RADARSAT-1 SAR Data”. Int. J. of App. Earth Obs. and Geo. 11,96-102, 2009. [5] K.Topouzelis, V. Karathanassi, P. Pavlakis, and D. Rokos, “Detection and Discrimination between Oil Spills and Look-alike Phenomena through Neural Networks. ISPRS J. Photo. Remote. Sens. 62, 264-270, 2007.
2176
[6] K.Topouzelis, “Oil Spill Detection by SAR Images: Dark Formation detection, Feature Extraction and Classification Algorithms”. Sens. 8, 6642-6659,2008. [7] M. Marghany, A.P.Cracknell, and M. Hashim, “Comparison between Radarsat-1 SAR Different Data Modes for Oil Spill Detection by a Fractal Box Counting Algorithm”. Int. J. of Dig. Earth, 2, 237-256, 2009. [8] M. Marghany, and M. Hashim, “Texture entropy algorithm for automatic detection of oil spill from RADARSAT-1 SAR data”. Int. J. of the Phys. Sci. 5(9), pp. 1475-1480, 2010. [9] M. Marghany, and M. Hashim, “Comparison between Mahalanobis classification and neural network for oil spill detection using RADARSAT-1 SAR data”. Int. J. of the Phys. Sci. Vol. 6(3), pp. 566-576, 2011.
