INVESTIGATIONS INTO THE METHODS FOR OIL SPILL DETECTION USING REMOTE SENSING DATA

Abstract

Among the different types of marine pollution, oil spill has been a major threat to the sea ecosystems. The source of the oil pollution can be located on the mainland or directly at sea. The sources of oil pollution at sea are discharges coming from ships or offshore platforms. Oil pollution from sea-based sources can be accidental or deliberate. Different tools to detect and monitor oil spills are vessels, aircraft, and satellites. Vessels equipped with specialised radars, can detect oil at sea but they can cover a very limited area. One of the established ways to monitor sea-based oil pollution is the use of satellites equipped with Synthetic Aperture Radar (SAR). The aim of the work presented in this dissertation is to implement methods at various stages for oil spill detection from Synthetic Aperture Radar (SAR) imagery. More than 60 images, mimicking ERS-2 SAR data have been generated and used as experimental data to assess the methodology of oil spill detection, which involves three stages: segmentation for dark spot detection, feature extraction and classification. Unfortunately oil spill is not only the phenomenon that can create a dark spot in SAR imagery. There are several others meteorological and oceanographic phenomena which may lead to a dark spot in SAR imagery. Therefore, these dark objects also appear similar to the dark spot due to oil spill and are called as look-alikes. These look-alikes thus cause difficulty in detecting oil spill spots. To get over this difficulty, feature extraction becomes important; a stage which may involve selection of appropriate feature extraction parameters. For segmentation, two methods; one based on edge detection segmentation and the other on artificial neural network (ANN) have been implemented. A total of 14 feature extraction parameters have been opted to form a feature vector for each spot. Ultimately, the final stage of methodology, the classification between oil spill and look-alike spots has been performed using ANN. The performance of several ANN architectures for segmentation as well as classification has been evaluated. An overall accuracy of 96.52 % in case of dark spot segmentation and almost 92 % accuracy in case of classification were obtained. The implemented methodology appears promising in detecting dark spots and discriminating oil spills from look-alikes.