Oil Spill Detection in Envisat Asar Images Using Radar Backscatter Thresholding and Logistic Regression Analysis

This paper presents a technique to detect oil spills in ENVISAT Advanced Synthetic Aperture Radar (ASAR) images using radar backscatter thresholding and logistic regression analysis. We developed and tested this technique using 4 Envisat ASAR images that were acquired many days after the M/T Solar I oil spill incident occurred on August 11, 2006 in Panay Gulf, southwest of Guimaras Island in Visayas, Philippines. A semi-automated approach by histogram analysis and radar backscatter thresholding was implemented to detect and segment dark formations in the Envisat ASAR images. Then, a logistic regression (LR)-based dark formation classifier was developed using 4 shape features, 11 contrast features, 2 homogeneity, and 2 slick surrounding features of the detected dark formations consisting of 154 verified oil slicks and 1,355 look-alikes. From this, a dataset consisting of 77 confirmed oil slicks and 77 look-alikes were randomly selected and used to train the classifier while the remaining dataset of 77 oil slicks and 1,272 look-alikes were used for validation. Features of the training dataset were fitted in a binary LR model and a backward stepwise-likelihood ratio approach was utilized to determine the sets of features that best discriminate an oil slick from its look-alike. Cross-validation of the LR classifier using the training dataset showed 84% accuracy for oil slick classification, 87% accuracy for look-alike classification, and an overall classification accuracy of 86%. An independent validation of the LR classifier revealed an above average performance, with 92% accuracy for oil slick classification, 76% accuracy for look-alike classification, and overall classification accuracy of 77%. The results of this study indicate that the combined radar backscatter thresholding and logistic regression analysis could be a promising approach in oil spill detection in Envisat ASAR images. The simplicity of the technique and its use of information readily available from the SAR images are advantageous in the rapid mapping of oil slicks right after an oil spill incident. Its improvement through consideration of prevailing wind conditions, the use of large training and validation datasets as well as inclusion of other relevant image features during classifier development could be a subject of future studies.