Oil Spill Detection using SAR Data: Feature-Based to Deep Learning Approaches

Abstract

Ocean surface monitoring is one of the important applications to analyze and ensure balance in the marine ecosystem. Several occasions of intentional and unintentional oil discharges have created discontinuity in marine life’s smooth existence and people dwelling in the coastal areas. It is impractical for the human force to timely monitor the disasters of oil spillage offshore. Due to the increasing distress concerning marine life protection for substantial development, a project under Indian Space Research Organization titled “Oceanic pollution and other ocean phenomenon monitoring us- ing feature extraction from multi-polarized SAR data” is developed with one of its dominant application as oil spill detection. This thesis aims to assist human forces in coastal relief zones to get timely alerts with on-time oil spill detection. The con- text of my thesis is to develop optimized feature extraction technique for extracting significant features from SAR images for oil spill detection. Optimized feature extrac- tion techniques are developed to classify the ocean surface features into five major classes, namely, oil, look-alike, ship, land, and clean sea. The ship provides ancillary information to confirm whether the detected dark region is oil or look-alike. For our experimentation we have considered SAR data from various sensors such as ALOS, RISAT-1, and SENTINAL-1. Both Level-1 and Level-2 SAR data are researched to boil down to the applicability of the type of data that suits the best for ocean surface monitoring, especially oil spill detection. Starting from the preprocessing of Level-2 data including speckle filtering and landmasking, the experiments were done using Otsu, Hysterisis 3D, Modified Otsu for dark spot detection. A modified version of Otsu was proposed were the seed values were automatically calculated. The classification was done using ANN. The work further progressed exploring the polarimertic SAR data. Various polarimetric features and decomposition were analysed using PolSARpro and later through our own implementation. These polarimeric feature images were further fed to VGG16 to experiment with the power of deep learning algorithms for final stage of automation. The issue of unavailability of large data for a specific phenomenon and a possibility of having systems with low computing power at grass-root level is mitigated by applying pre-trained networks and Generative Adversarial Networks. The final proposal uses Wasserstein GAN - Gradiant Penalty for performing the detection task. The solution can be plugged into software used by grass-root developers to avail the benefit.