SUPERVISED PATTERN CLASSIFICATION TECHNIQUES FOR OIL SPILL
CLASSIFICATION IN SAR IMAGES: PRELIMINARY RESULTS
Andrea Montali(1), Giorgio Giacinto(1), Maurizio Migliaccio(2), and Attilio Gambardella(1)
(1)
(2)
Dip. Ingegneria Elettrica ed Elettronica, Università degli Studi di Cagliari, Piazza D’Armi 19, 09123 Cagliari, Italy.
Dipartimento per le Tecnologie, Università degli Studi di Napoli Parthenope, Via Acton 38, 80133 Napoli, Italy.
ABSTRACT
In this study, classical features extracted from Synthetic Aperture Radar (SAR) Images and used in oil spill
classification procedures have been examined/evaluated and ranked in function of their effectiveness. The best features
have been used to perform the classification of dark patches by means of two-class and one-class approaches. A number
of one-class and two-class classification techniques have been considered: Linear discriminants, 1-Nearest Neighbour,
Mixture of Gaussians, Parzen Windows and the Support Vector Machines. Better results have been achieved using the
one-class methods. The dataset has been extracted from a set of SAR images acquired by an airborne X band SAR
system mounted on board of a Laerjet 35A, during the Galitia Mission in January-February 2003 (TELAER
Consortium), after the sinking of the Prestige oil tanker.
1 INTRODUCTION
Oil spill detection by means of SAR images is possible because of the damping effect of the short wind waves caused
by the presence of oil on the sea surface [1-2]. As a consequence, an oil spill is physically a dark patch in SAR images
[1-2]. The sea radar image is a representation of the backscatter return, and the intensity of the pixel is proportional to
the surface roughness at the scale of radar wavelength (Bragg scattering) [3]. The radar backscatter coefficient is a
function of the viewing geometry of the SAR [3]. The backscatter coefficient decreases with the increase of the
incidence angle, and the scattering properties of a material depend on the polarization of the incoming radar signal [3].
The contrast between an oil spill and its surroundings depends on a number of parameters as the height of the waves,
the amount of oil that has been released, the speed of the wind, etc [1-2]. The shape of the oil spill depends on a number
of factors as whether the oil was released from a stationary object or from a moving ship, the amount of oil involved,
the speed of the wind, and the current history in the time interval between the oil release and the acquisition of the
image [4].
Unfortunately, several natural and atmospheric phenomena produce dark areas in SAR images similar to oil spills.
These dark areas are usually referred to as look-alikes, whose presence makes the detection of oil spills a challenging
task [1-2].
Oil spills may include all oil-related surface films caused by oil spills from oilrigs, leaking pipelines, passing vessels,
while look-alikes include natural films, grease ice, threshold wind speed areas (<3 m/s), rain cells, etc. Oil spills are
only man-made slicks due to crude petroleum and its products [1-2].
It is easy to see that oil spill detection over SAR images not only requires the detection of dark patches in the image, but
also requires post-processing techniques aimed at discriminating oil spills from look-alikes. Oil spill detection is thus
usually framed into three fundamental phases: dark patch detection, features extraction, and oil spill/look-alike
classification [1-10]. Features to discriminate among oil spills and look-alikes are typically based on geometrical
properties, as well as on radiometric measures and textures.
1.1 Segmentation techniques
As oil spills are characterised by low backscattering levels, the use of thresholding techniques has been investigated for
dark spot segmentation. In [5] an adaptive algorithm with the threshold set k dB below the mean value, estimated in an
image moving window, is applied. Image thresholding is then combined with a multiscale pyramid approach and a
clustering step to better separate the oil spill from its surroundings. In [6] the areas of low backscatter are located
automatically, without the intervention of a human operator. This procedure masks the land areas, and selects the dark
regions with a normalized radar cross section (NRCS) smaller than one half of the average NRCS computed for the sea
area in the image. It then rejects the selected areas that are too small or too large. The small regions are rejected because
these slicks are not significant from the coast guard point of view, and the large ones because these are most likely areas
with no wind. In [7] a dedicated oil spill processing and analysis tool based on edge detection is applied. Once defined
the region of interest by the user, the tool analyzes the overall backscattering of the region and produces a histogram.
__________________________________________________
Proceedings of SEASAR 2006, 23-26 January 2006, Frascati, Italy,
(ESA SP-613, April 2006)
Typically, histograms in presence of a dark region contain two peaks; the lower peak is located around the mean
backscattering value of the dark object, while the taller is located around the mean value of the background. The local
minimum value between the two peaks is used for the segmentation of the image. To this purpose, the darkest pixel in
the region is selected as the starting point for the region growing segmentation algorithm. The region grows until the
neighbouring pixels have a value grater than the threshold value given by the local minimum previously calculated.
This method allows defining the border around the dark region.
TABLE 1 – Features considered in automatic oil spill detection procedures. (adapted from [1])
Del Frate et
al., 2000
Fiscella et
al., 2000
Topouzelis
et al., 2002
Area
■
■
Big Areas
Perimeter
■
■
Features
Perimeter to Area Ratio
■
Normalized Perimeter to Area Ratio
■
Marghany,
2001
Espedal,
1999
Solberg.,
1999
■
■
■
Shape Index
■
Slick Width
■
Length to Width Ratio
Complexity
■
Form Factor
Spreading
■
Object Standard Deviation
■
■
Background Standard Deviation
■
■
Contrast
■
Gradient
■
Gradient Standard Deviation
■
First Invariant
Asymmetry
Planar Moment
■
■
Slick Radar Backscattering
■
Outside Slick Radar Backscattering
■
Intensity Ratio
■
Standard Deviation Ratio
■
Intensity Standard Deviation Ratio Inside
■
Intensity Standard Deviation Ratio Outside
■
ISRI to ISRO Ratio
■
■
Mean Difference to Neighbours
■
Power to Mean Ratio
■
Close to Big Areas
■
Close to Land
■
Close to Fractal Object
■
Layer Value Texture Based on Sub-objects
■
Shape Texture Based on Sub-objects
■
■
■
■
■
■
Texture
Distance to a Point Surce
■
Number of Detected Spot in the Scene
■
Number of Neighboring Spots
■
Homogeneity
■
Entropy
■
Energy
■
Correlation
■
■
1.2 Slick feature extraction
After extracting the thresholded dark spot area from the image, a number of features can be computed for each dark
spot. Table 1 summarizes the different features adopted by different authors. The individual features are typically
subdivided into the following classes:
Geometry and shape of segmented region: geometry and shape features are applied by all methods proposed in literature
(Table 1). If the oil slick is a fresh oil spill released from a moving ship (a tanker cleaning its tank), an important feature
is the elongatedness, which can be expressed as a ratio between the width and length of the slick. Among the shape
features, the perimeter and the area of the object are also considered.
Backscatter level of the spot and its surrounding: in [7] it has been shown that the features related to the gradient of the
backscattering value from background to slick provided the most valuable information when classification by means of
neural networks was used. Due to the heavy effect of wind speed, the background standard deviation was also shown to
be an important discriminant parameter. Similarly in [6] it has been shown that features related to the background area
surrounding the slick were also important.
Spot contextual features: spot contextual features depends on weather conditions at the moment of image acquisition, as
the speed of the wind, and also on the distance from possible oil spill source as ships or oilrigs. In [4] the distance to
bright spot (ships on SAR images) and to pollution source (i.e. oil platforms) is used. In [5] improved classification
results are attained by classifying dark spots in the context of their surrounding and weather information.
Texture: if the pixel intensity is compared to the backscattered signal, textures provide information about the spatial
correlation among neighbouring pixels. In [11] features based on grey level co-occurrence matrixes (GLCM) are used.
Power-to-mean ratio of the slick and surroundings is used in [5] as a measure of homogeneity.
1.3 Classification methods
As a number of phenomena can generate dark patches in SAR images, the purpose of a classifier is to distinguish oil
spills from other cases (look-alikes).
In [6] a Mahalanobis classifier in applied to estimate the probability p of a dark spot being an oil spill. The 82% of the
oil spill in the test set were correctly classified by this classifier. This result was compared with a compound probability
classifier, where the 91% of the oil spills in the test set were correctly classified. A training set of 80 oil spills and 43
look-alikes and a test set of 11 oil spills, 4 uncertain and 6 look-alike were used.
In [5] the probability was estimated by a multivariate Gaussian density function. The algorithm was combined with a
prior model for the number of look-alikes, a model for the presence of slick near a bright object and a rule-based
modification of the probability density function to take into account feature combinations indicating particular scene
conditions. The leave-one-out method with 84 scenes gave a correct classification of 94% for oil spills.
An approach based on multilayer perceptron (MLP) neural network with two hidden layer, was proposed in [7]. The net
was trained using the back-propagation algorithm to minimize the error function, and tested on ERS images. Using the
leave-one-out method with 139 oil spills, the 82% of oil spills were correctly classified.
The above automatic methods for oil spill detection, exhibit a percentage of correct classification of oil spills ranging
from 82% to 94%. These results are related to different data sets, different segmentation approaches to detect dark
patches, and different feature extraction processes. On the other hand, all the studies reported in the literature are based
on a classical two-class classification methodology, where examples of the two classes, i.e. oil-spills and look-alikes,
must be provided to train the classification model.
The main contribution of this study is twofold: i) classical features have been examined/evaluated and ranked in
function of their effectiveness; ii) the classification of dark patches has been performed using two-class approaches as
well as one-class approaches [12-13]. One-class approaches aim to model the class for which reliable examples can be
provided, e.g. the oil-spill class. As a consequence, a dark patch is classified as an oil spill if it fits the model, otherwise
it is classified as a look-alike.
2 FEATURE EXTRAXTION AND EVALUATION
The adopted segmentation process is a bottom up region-merging technique starting with one-pixel objects. In
subsequent steps, smaller image objects are merged into bigger ones. Throughout this pair-wise clustering process, the
underlying optimization procedure minimizes the heterogeneity of the resulting image objects. In each step, a pair of
adjacent image objects is merged, so that the heterogeneity level of the new object is below a fixed heterogeneity
threshold. The process stops if the heterogeneity level of the merged regions exceeds the heterogeneity threshold
defined by the human operator. The human operator can either accept the result suggested by the automatic procedure
or reject it and produce a new segmented image, by changing the threshold manually.
Once the border of the dark object is accepted, a number of morphological and physical features regarding the dark
object and its surrounding areas are computed (Table 1).
All these classical features have been examined/evaluated and ranked in function of their discrimination capability
between oil spills and look-alikes. Four features selection algorithms have been used, namely the Forward method, the
Backward method, the Branch & Bound method and the Individual method. For each feature selection method, four
different distance criterion have been employed, i.e. the Euclidean distance, the Mahalanobis distance, the Nearest
Neighbour distance and the Inter-Intra distance. Each combination of feature selection methods and distance criterion
produced different subsets of “good” features. In order to combine such different results, four “frequency plots” are
obtained by counting the times that each feature is ranked in the first five, ten, fifteen and twenty, respectively. These
frequencies allowed producing a cumulative ranking that has been used to finally assess the discrimination capability of
each feature (see Table 2).
TABLE 2 – Best features set after features discrimination capability evaluation.
(6) - Intensity Ratio µ obj µ sce
(1) - Mean Contrast ( µ - µ )
sce
obj
(2) - Outside Slick Standard Deviation ( σ sce )
(3) - (GLCM) Contrast
(4) - Inside Slick Radar Backscattering ( µ obj )
(5) - Outside Slick Radar Backscattering ( µ sce )
(
)
(7) - ISRO (µ sce σ sce )
(8) - (GLCM) Dissimilarity
(9) - (GLCM) Entropy
(10) - (GLCM) Correlation
(
(11) - Intensity St. Dev. Ratio σ obj σ sce
)
3 CLASSIFICATION TECHNIQUES
A classification problem can be formulated in terms of a one-class classification if one of the classes has to be
distinguished form all other classes. Typically, in a one-class classification problem two classes are considered: the
target class and the outlier class [12-13].
The target class is assumed to be well sampled, in the sense that either a large number of training examples are
available for this class, or this class can be more easily modelled. On the other hand, the Outlier class can be sampled
very sparsely, or examples can be totally absent. For example, examples of outliers may be very hard to collect, or its
features might be very expensive to measure, etc..
In our case, the oil spills are considered as the target class, as their characteristics have been thoroughly studied and
they can be easily reproduced in laboratory simulations. On the other hand, look-alikes (biogenic class) form the
“outlier class”, as they can be produced by a large variety of different natural phenomena. It is worth noting that
tipically different oil-spills will appear clustered in the features space because they are all generated from the same
man-made phenomena. On the other hand, look-alikes are considered the outlier class, as they can be produced by
several and very different natural phenomena, so that they are sparsely distributed in the feature space. This fact can be
motivated by observing that the values of features associated to different natural phenomena can be very different from
the values of the same feature associated to other natural phenomena.
This means that only examples of the target class (oil spill) can be used for classification, while data from outlier
objects (look-alikes) could not provide useful information. As a consequence, one-class models estimates the boundary
of the area in the feature space containing the target class, i.e. the oil spill class. Therefore, one-class methods are
developed from mathematical models based only on examples from the target class, so that the large number of target
objects are accepted, while the chance of accepting outlier objects is minimised.
4
EXPERIMENTAL RESULTS
The training dataset has been extracted from a set of SAR images acquired by an airborne X band SAR system mounted
on board of a Laerjet 35A, during the Galitia Mission in January-February 2003 (TELAER Consortium), after the
sinking of the Prestige oil tanker.
In order to compare the performances of two-class classifiers and one-class classifiers, several experiments have been
carried out by varying the number of training objects and the number of features. A number of one-class and two-class
classification techniques have been used: Linear discriminants, 1-Nearest Neighbour, Mixture of Gaussians, Parzen
Windows and the Support Vector Machines [12-13].
Preliminary results show that the best classification performance is always achieved by one-class classification
techniques. In particular, once fixed at 2% the maximum acceptable oil-spill misclassification error (number of oil-spills
classified as look alike), the error rate on the look-alike class was evaluated (this error represents the percentage of
look-alikes classified as oil-spills). One-class classifiers allowed attaining an error of 1%, whereas two-class classifiers
attained an error of 3% (see Figs. 1-2).
ONE-CLASS - 7tr
TWO-CLASS - 7tr
0.4
0.35
0.35
0.3
FALSI POSITIVI (%)
FALSI POSITIVI (%)
0.3
0.25
0.2
0.15
0.25
0.2
0.15
0.1
0.1
0.05
0.05
0
1
2
3
4
5
6
7
8
9
10
0
11
1
2
3
4
5
Fig. 1 – Plot of the look-alike classification mean error, once
that a 2% oil-spill classification max error has been fixed, for
three different type of one-class methods.
4D
9
10
11
8D
0.08
Linear
1-NN
Mog
Parzen
SVDD
0.09
0.08
0.07
0.07
FALSI POSITIVI (%)
FALSI POSITIVI (%)
8
0.1
Linear
1-NN
Mog
Parzen
SVDD
0.09
0.06
0.05
0.04
0.06
0.05
0.04
0.03
0.03
0.02
0.02
0.01
0.01
7tr
9tr
10tr
0
12tr
7tr
NUM. OIL-SPILL NEL TR-SET
9tr
10tr
12tr
NUM. OIL-SPILL NEL TR-SET
Fig. 3 – Mean error plots of the considered one-class
classification methods for different training set dimensions (7tr,
9tr, 10tr, 12tr) and making use of 4 characteristics (4D).
Fig. 4 – Mean error plots of the considered one-class
classification methods for different training set dimensions (7tr,
9tr, 10tr, 12tr) and making use of 8 characteristics (8D).
8D
4D
0.1
0.1
Linear
1-NN
Mog
Parzen
SVDD
0.09
0.08
Linear
1-NN
Mog
Parzen
SVDD
0.09
0.08
0.07
0.07
FALSI POSITIVI (%)
FALSI POSITIVI (%)
7
Fig. 2 – Plot of the look-alike classification mean error, once that
a 2% oil-spill classification max error has been fixed, for three
different type of two-class methods.
0.1
0
6
DIMENSIONE DEL PATTERN
DIMENSIONE DEL PATTERN
0.06
0.05
0.04
0.06
0.05
0.04
0.03
0.03
0.02
0.02
0.01
0.01
0
Linear
1-NN
Parzen
0.4
Linear
1-NN
Parzen
7tr
9tr
10tr
12tr
NUM. OIL-SPILL NEL TR-SET
Fig. 5 – Mean error plots of the considered two-class
classification methods for different training set dimensions (7tr,
9tr, 10tr, 12tr) and making use of 4 characteristics (4D).
0
7tr
9tr
10tr
12tr
NUM. OIL-SPILL NEL TR-SET
Fig. 6 – Mean error plots of the considered two-class
classification methods for different training set dimensions (7tr,
9tr, 10tr, 12tr) and making use of 8 characteristics (8D).
As far as the feature selection process is concerned, all the features typically used in the literature have been ranked
according to their effectiveness. Results show that the highest accuracies can be attained by using either the first four or
the first eight features (see Figs. 3-6) obtained by the features selection phase (see Table 2). These features include the
mean contrast between the background and the dark patch, the outside dark patch standard deviation (as the dark patch
in the image depends also on the wind speed), the (GLCM) contrast, the inside slick dark patch radar backscattering,
the outside slick dark patch radar backscattering, the intensity ratio, the ISRO and the (GLCM) dissimilarity. The best
classification performance, for any training set and pattern dimensions, has been achieved with the Linear discriminants
one–class type method. The second best classification result has been achieved with the Support Vector one–class type
method. This latter technique has been the best two–class method.
5 CONCLUSIONS
In this study, one-class and two-class classification methods have been explored to classify dark patches in marine SAR
images, i.e. the capability to distinguish among oil-spill and look-alike has been explored. In particular, the study has
been focused on two aspects:
i) analysis of the classical features used in the oil-spill classification procedures, and their ranking in function of their
effectiveness;
ii) classification of dark patches using a two-class approach as well as a one-class approach.
The features which have shown the best discrimination capability between the oil-spill and the look-alike classes are
principally related to the variations of the contrast in correspondence of the transition from free sea surrounding area
and dark patch, i.e. the Mean Contrast feature and the Contrast (GLMC) feature. Another feature which has shown a
good discrimination capability is the Outside Slick Standard Deviation since its strong dependence the wind. It is
important to keep in mind that the features have a good discrimination capability in the features space only if they are
jointly considered
Once that the best features have been selected, a set of one-class and two-class classification techniques have been
tested for different dimensions of the training set: Linear discriminants, 1-Nearest Neighbour, Mixture of Gaussians,
Parzen Windows and the Support Vector Machines. The best classification performance has been achieved with the
Linear discriminants one –class type method.
ACKNOWLEDGEMENT
The authors acknowledge the TELAER Consortium and Telespazio S.p.A. for providing the airborne SAR data and Dr.
D.M.J. Tax, Delft University of Technology for providing the DD_tools for Matlab, ver. 1.4.1.
REFERENCES
1. M. Migliaccio M. Tranfaglia, Oil Spill Observation by SAR: A Review, Proc. US-Baltic International Symposium,
Klaipeda (Latva), June 2004.
2. C. Brekke and A. H. S. Solberg, Oil Spill Detection by Satellite Remote Sensing, Remote Sens. Environ., no.95, pp.113, 2005.
3. C. Elachi, Introduction to the Physics and Techniques of Remote Sensing, New York (NY) USA, John Wiley & Sons,
1987.
4. H. Espedal and T. Wahl, Satellite SAR Oil Spill Detection Using Wind History Information, International Journal of
Remote Sensing, 20, pp. 49-65, 1999.
5. A. H. S. Solberg, G. Storvik, R Solberg and E. Volden, Automatic Detection of Oil Spills in ERS SAR Images, IEEE
Trans. Geosci. and Remote Sensing, vol. 37, pp. 1916-1924, 1999.
6. B. Fiscella, A. Giancaspro, F. Nirchio, P. Pavese, and P. Trivero, Oil spill detection using marine SAR images, Int. J.
Remote Sensing, vol. 21, no. 18, pp.3561-3566, 2000.
7. F. Del Frate, A. Petrocchi, J. Lichtenegger, and G. Calabresi, Neural Networks for Oil Spill Detection Using ERSSAR Data, IEEE Trans. Geosci. and Remote Sensing, vol. 38, pp. 2282-2287, 2000.
8. A. H. S. Solberg, S. T. Dokken and R. Solberg, Automatic Detection of Oil Spills in ENVISAT, RADARSAT and ERS
SAR Images, Proc. IEEE Symp. Geosci. Remote Sensing (IGARSS’03), July 2003.
9. K. Topouzelis, V. Karathanassi, P. Pavlakis and D. Rokos, Oil Spill Detection: SAR Multiscale Segmentation and
Object Features Evalutation, Remote Sensing of the Ocean and Sea Ice, pp. 77-78, 2002.
10. M. Marghany, RADARSAT Automatic Algorithms for Detecting Coastal Oil Spill Pollution, JAG, vol. 3, pp. 191196, 2001.
11. H. Assilzadeh, S. B. Mansor, Early Warning System for Oil Spill Using SAR Images, Proc. ACRS 2001 – 22nd
Asian Conference on Remote Sensing, Singapore, vol 1, pp. 460-465, 2001.
12. R. Duda, P. Hart and D. Stork, Pattern Classification, second ed., New York, NY (USA), John Wiley & Sons, 2001.
13. D. M. J. Tax, One Class Classification, Phd Thesis, Delf University of Technology, http, June 2001.