A NEW OBJECT-ORIENTED METHODOLOGY TO DETECT OIL SPILLS USING ENVISAT IMAGES

A NEW OBJECT-ORIENTED METHODOLOGY TO DETECT OIL SPILLS USING ENVISAT IMAGES K. Topouzelis (1), V. Karathanassi (2), P. Pavlakis (3), D. Rokos (2) (1) DG Joint Research Centre (EC), Institute for the Protection and Security of the Citizen, Via Fermi 1, 21020, Ispra, VA, Italy, kostas.topouzelis@jrc.it (2) Laboratory of Remote Sensing, School of Rural and Surveying Engineering, National Technical University of Athens, Heroon Polytechniou 9, Zographos, 15780, Greece, karathan@survey.ntua.gr, rslab@survey.ntua.gr (3) Hellenic Centre for Marine Research, P.O.Box 712, 19013 Anavissos, Attika, Greece, ppavla@ath.hcmr.gr ABSTRACT Besides accidental pollution caused by ships in distress, different types of routine ship operations lead to deliberate discharges of oil in the sea. Illicit discharges are the greatest source of marine pollution from ships, and the one which poses a long-term threat to the marine and coastal environment. Satellite monitoring, in particular using Synthetic Aperture Radar (SAR) observations, may represent an effective tool for complementing traditional aerial surveillance. The capability of SAR in detecting oil slicks over the sea surface is well known and proven since a long time. A new automated methodology for oil spill detection was previously presented using ERS full resolution SAR data. In the present study a first attempt to examine the technique to ASAR ENVISAT medium resolution scenes was performed. The methodology relies on the object oriented approach and profits of image segmentation techniques in order for dark formations to be detected. A fuzzy classification method is used to classify dark formations to oils spill or look-alikes. 1. INTRODUCTION Oil spills seriously affect the marine ecosystem and cause political and scientific concern since they have serious effects on fragile marine and coastal ecosystems. The amount of pollutant discharges and associated effects on the marine environment are important parameters in evaluating sea water quality. While it is very well known that accidental pollution at sea can be reduced but never completely eliminated, illegal discharges from ships can indeed be eliminated by the strict enforcement of existing regulations and the control, monitoring and surveillance of maritime traffic. SAR systems are extensively used for the dark formation detection in the marine environment, as they are not affected by local weather conditions and cloudiness and occupy day to night. SAR systems detect dark formations on the sea surface indirectly, through the modification dark formations cause on the wind generated short gravity capillary waves [1]. Several manmade and natural ocean phenomena damp these waves which are the primary backscatter agents of the radar signals. For this reason, a dark formation appears dark on SAR imagery in contrast to the surrounding clean sea. Dark formations can be [1]: oil spills, low wind areas, organic film, wind front areas, areas sheltered by land, rain cells, current shear zones, grease ice, internal waves and upwelling zones. Several studies aiming at oil spill detection using SAR images have been implemented [2-15]. Most of these studies rely on the detection of dark areas which in a later step are classified as oil spills or look-alikes. Any formation on the image which is darker than the surrounding area has a high probability of being an oil spill and needs further examination. Although this process seems to be simple for a human operator, it contains three main difficulties if semi-automated or automated methods are used. First, fresh oil spills are brighter than older spills. They have a weak backscattering contrast relative to their surroundings and thus cannot be easily discriminated. Second, areas surrounding dark areas can have various contrast values, depending on local sea state, oil spill type and image resolution. Third, other phenomena may appear as dark areas. Further classification of the dark areas to oil spills and look-alikes is therefore required. The aim of this work is to detect mineral oil spills resulting from illegal ship discharges. In a previous study a new object methodology was developed using ERS PRI data. In the present study the method was tested to ENVISAT medium resolution data. Following the idea of thinking globally and acting locally, the methodology introduces four innovative points. The first is the segmentation of the image into large image segments with different statistical values, instead of the arbitrary cutting of the image in windows. Oil spills are thereupon not partially contained in image windows, and for each large segment a threshold adaptive to its local contrast can be estimated. Second, a very detailed scale segmentation is implemented and statistical values of each small segment are compared to the threshold of the big segment to which it belongs. Consequently, the methodology developed takes advantage of the different contrast and intensity values which are contained in a single ASAR image and Proc. Envisat Symposium 2007, Montreux, Switzerland 23 27 April 2007 (ESA SP-636, July 2007)

detects dark areas with various brightness values located in different sea-state environments. Third, once dark areas are identified, they are not isolated from the image but are still considered as parts of its whole. The classification stage is performed using features of the dark areas but also taking into consideration features of the surrounding areas. At this stage, dark areas are classified as oil spill or look-alikes, using fuzzy classification logic. Fourth, the method uses two distinct knowledge bases, each one adaptive to a different type of sea environment. In case of intense image fabric with local anomalies and/or instantaneous wind stop areas, the mean brightness of the large segment created by the methodology is reduced and the shape and boundaries of dark objects are modified. On the contrary, in a relatively bright and smooth sea environment, dark objects present high contrast, are more precisely defined and their detection is easier. 2. METHODOLOGY The methodology developed relies on an object oriented approach. More detailed presentation can be found at [13]. The methodology has been developed in the ecognition Software environment. The software introduces a new classification technology called Object Oriented Image Classification in which in first step extracts homogeneous image objects in any chosen resolution that are subsequently classified by means of fuzzy logic. The basic strategy is to build up a hierarchical network of image objects, which allows the representation of the image information content at different resolutions (scales) simultaneously. By operating on the relations between networked objects, it is possible to classify local context information. The developed methodology requires six main steps for its implementation (see figure 1). User imports the original ASAR image to a model for the 8-bit transformation and the filtering steps. The model s outputs (8-bit ASAR image and filtered images) are imported to ecognition and without any user contribution result is given through a specific protocol which is responsible for the next steps (segmentation, dark area detection, object union and object classification). Transformation from 16 to 8 bit was made using the simple rescale min-max algorithm with the max value to be equal with the mean value plus three times the standard deviation value (equivalent to 99.73% of the data distribution). Filtering was done using a specific combination of a 3x3 Lee filter to the original image, followed by a 5x5 Lee filter and a 7x7 Local Region filter. The combination has been previously used with success [16]. The three-layer image produced (original ASAR image, mean filter, combination of Lee and Local Region filters) was used for image segmentation. Segmentation occurred at two different scales: a very detailed and a very broad scale. The detailed scale factor was used in order to create very small objects capable of describing every formation in the SAR image. The broad scale factor was used in order to break off the ASAR image to large, as uniform as possible, areas. Small scale 8-bit transformation Knowledge base I Knowledge base II Filtering Segmentation Dark area detection Object Union Object-based classification Result Coarse scale Figure 1. Flow diagram of the used methodology Dark area detection occurred using a threshold algorithm which was implemented in two stages. In the first stage, a formula based on the statistical values of super objects was developed and small objects with low brightness and high contrast values relative to their surroundings were characterised as dark areas. However, fresh oil spills or broken parts of an oil spill were not satisfactorily identified. A second formula was applied, in which the statistical values of the super objects were based on pixels not characterised as dark areas in the first stage. Factors and formulas can be found in [13]. All small objects characterized as dark areas are then grouped in order for statistical features such as area, shape and texture to be estimated. Object union resulted in the classification of the image into two categories: dark objects and rough sea. Based on dark objects, several features can be derived, such as mean value, perimeter, complexity, texture, etc. The knowledge base which was created from them uses a fuzzy based logic in order for dark objects to be classified as oil spills or look-alikes. Thorough analysis of the sea environment has shown that discrimination between oil spills and look-alikes is unsatisfactory when a single base is used. Therefore two knowledge bases, depending on the brightness of the super object, i.e. the

sea state, were developed. They include the same features but use different rules for the characterization of oil spills. Red colour presents the oil spills detected, light blue look-alikes, and blue the sea. A set of 10 features was introduced in the knowledge bases. A special study was performed in order to choose which features will be used [17]. The proposed combination was found very efficient in distinguishing oil spills from look-alikes. Classification was performed using fuzzy logic. Fuzzy rules are easily understood and can be applied on every feature selected. Each of the 10 features was considered a separate class. Each class consists of a set of fuzzy expressions, allowing the estimation of the specific values and their logical operation. The method can be used in operational scale for detecting oil spills from medium resolution ASAR data. The user has only to contribute in terms of importing the original ASAR image and products derived from them. The computational time of the proposed methodology on a medium resolution ASAR image is 2-5 minutes. 3. DATASET The methodology developed has been applied on a dataset of four ENVISAT ASAR medium resolution images. The dataset contains several sea states and all images contain a certain number of dark objects. Two of the images contain at least one oil spill, while the other two contain only look-alike features. For illustration reasons the implementation of the methodology is presented on two ASAR images: the first one (Figure 2) captured on 28/7/2006 (orbit 23050, frame 2242) and the second one (Figure 3) captured on 21/7/2006 (orbit 22950, frame 1464). The two images present different sea states and contain oil spills and look-alikes. Moreover, each image presents a spatial sea-state variation. The images have been thoroughly analyzed by three different experts using visual photo-interpretation techniques. The first image presents a rough sea surface, efficient to produce a strong contrast signal in the presence of oil spills. It contains four old oil spills in the left part and a fresh oil spill in the right part. It also contains a look-alike on the left part, very close to land. Sea state is ideal for oil spill detection, as wind speed is estimated 3.4 m/sec. The second image contains very complicated features. Wind speed varies from 0 m/sec to 4 m/sec, making oil spill detection a rather difficult task. The image contains several look-alikes towards the area next to the land. 4. RESULTS The result of the analysis for the above mentioned ENVISAT ASAR images are presented in figures 2-5. Figure 2. ENVISAT ASAR medium resolution image containing oil spills and the classification result From figure 2, describing the results on the ASAR image with the oils spills, it can be seen that the four old oils spills in the left part were successfully detected. Figure 4 focuses on the bigger oil spill in the left part of the image. On the contrary the fresh oil spill at the right part of the image in figure 2 was not detected. The failure occurred at the dark detection phase. The method successfully separated the image in four big segments but the difference of the small segments, describing the fresh spill, from their bigger segment was not enough in order to detect the dark object. Therefore there was no object describing the fresh spill in the stage of classification. The backscattering values of the fresh spill are not into the limits derived from the fresh spills which were previously examined on ERS PRI scenes. Figure 3 illustrates the result of the analysis of the second image which contains only look-alikes. Here the

classification method worked successfully as all the dark features were correctly classified as look-alikes. In figure 5 a more detailed example is given for a lookalike which is located in the upper left part of figure 3. are not the same in high and medium resolution scenes, therefore a specific study is required to adapt them in each resolution. The proposed methodology yields two main advantages. First, its results are not affected by sea state conditions and second, the method is independent of the original ASAR data quality (sensor calibration status, speckle, atmospheric conditions etc). Moreover, dark objects are not isolated from the image and therefore neighbouring rules can be used for classification phase. The importance of the last point can be seen from figures 4 and 5. Even an expert can not distinguish the oil spill from look-alike when only these parts of the images are examined. On the contrary, analysis is much easier when consideration is taken for the neighbouring features. Further research on this issue includes validation of the method on more images with various sea states and types of oil spills. Moreover, a comparison between the proposed method and a statistical based one has to be done, using as inputs the same ASAR data which should contain verified oil spills and look-alikes. Figure 3. ENVISAT ASAR medium resolution image containing look-alikes and the classification result 5. CONCLUSIONS - DISCUSSION In the present study a new automated methodology for oil spill detection using ENVISAT ASAR medium resolution data is presented. The methodology was previously developed for ERS PRI data. Several factors had to be changed, i.e. scales segmentation and fuzzy limits in order to work satisfactorily in several oil spill detection cases for ENVISAT scenes. Oil spills morphological characteristics and backscattering values Figure 4. An oil spill example and its classification

Figure 5. A look-alike example and its classification 6. REFERENCES 1. Alpers, W., Wismann, V., Theis, R., Huhnerfuss, H., Bartsch, N., Moreira, J. & Lyden, J. (1991). The damping of ocean surface waves by monomolecular sea slicks measured by airborne multi-frequency radars during the saxon-fpn experiment. Proc. IGARSS'91 2. Martinez, A. & Moreno, V. (1996). An oil spill monitoring system based on SAR images. Spill Sci. & Tech. Bul. 1-2, 65-71. 3. Kubat, M., Holte, R. C. & Matwin, S. (1998). Machine Learning for the Detection of Oil Spills in Satellite Radar Images. Mach. Learn. 2-3, 195-215. 4. Benelli, G. & Garzelli, A., (1999). Oil-Spills Detection in SAR Images by Fractal Dimension Estimation. Proc IGARSS'99, 218-220. 5. Solberg, A., Storvik, G., Solberg, R. & Volden, E., (1999). Automatic Detection of Oil Spills in ERS SAR Images. IEEE Trans. on Geos. and Rem. Sens. 37(4), 1916-1924. 6. Del Frate, F., Petrocchi, A., Lichtenegger, J. & Calabresi, G. (2000). Neural networks for oil spill detection using ERS-SAR data. IEEE Trans. on Geos. and Rem. Sens. 38(5), 2282-2287. 7. Gade, M., Scholz, J. & Viebahn, C. (2000). On the delectability of marine oil pollution in European marginal waters by means of ERS SAR imagery. International Geoscience and Remote Sensing Symposium (IGARSS 2000), VI, 2510-2512. 8. Fiscella, B., Giancaspro, A., Nirchio, F. & Trivero P. (2000). Oil spill detection using marine SAR images. Int. J. of Rem. Sen. 21(8), 3561-3566. 9. Topouzelis, K., Karathanassi, V., Pavlakis, P. & Rokos, D. (2002) Oil Spill Detection: SAR Multi-scale Segmentation & Object Features Evaluation. 9th International Symposium on Remote Sensing (SPIE), 77-87. 10. Wu, S. Y. & Liu, A. K. (2003). Towards an automated ocean feature detection, extraction and classification scheme for SAR imagery. Int. J. of Rem. Sen. 24(5), 935 951. 11. Brekke, C. & Solberg, A. (2005). Oil spill detection by satellite remote sensing, Rem. Sens. of Env. 95, 1-13. 12. Nirchio, F.,. Sorgente, M, Giancaspro, A., Biamino, W., Parisato, E.,. Ravera, R & Trivero, P. (2005) Automatic detection of oil spills from SAR images, Int. J. of Rem. Sen. 26(6), 1157-1174. 13. Karathanassi, V., Topouzelis, K., Pavlakis, P. & Rokos, D. (2006). An object-oriented methodology to detect oil spills. Int. J. of Rem. Sen. 27(23), 5235 5251. 14. Keramitsoglou, I., Cartalis, C. & Kiranoudis, C. (2006). Automatic identification of oil spills on satellite images, Env. Mod. Soft. 21(5), 640 652. 15. Solberg, A &.Brekke, C. (2007). Oil spill detection in Radarsat and Envisat SAR images, IEEE Trans. on Geos. and Rem. Sens. 45(3), 7467-55. 16. Rio, J. N. & Lozano-García, F. D. (2000). Spatial Filtering of Radar Data (RADARSAT) for Wetlands (Brackish Marshes) Classification. Rem. Sen. of Env. 73(2), 143-151. 17. Stathakis, D. Topouzelis, K. & Karathanassi, V. (2006). Large-scale feature selection using evolved neural networks. 13th SPIE Int. Symposium on Remote Sensing, 11-14 September 2006..