Abstract

Oil spills significantly threaten marine environment, damaging ecosystem, wildlife, and coastal communities. This thesis addresses these challenges by employing both advanced machine learning techniques and satellite imagery analysis technologies to enhance the accuracy and efficiency of oil spill detection and or source identification. By utilizing Synthetic Aperture Radar (SAR) images and examining semantic segmentation models, the research aims to accurately detect oil spills based on satellite images. Additionally, oil fingerprinting techniques, involving unsupervised classification are used to identify the sources of marine oil spills, providing a comprehensive framework for oil spill monitoring and management. The methodology involves the use of three distinct datasets: a multi-class dataset for detecting oil spills using satellite images, a binary dataset focusing on oil spill incidents in the Gulf of Suez from 2017 to 2021 as a case study for oil spill detection and a dataset for oil fingerprinting based on samples from the MV Manolis L shipwreck. For oil spill detection, semantic segmentation models were trained and evaluated using these datasets. Performance metrics such as Intersection over Union (IoU) were used to assess the modeling accuracy. Secondly for oil fingerprinting, PCA and HCA were applied to analyze the chemical composition data of the MV Manolis L. oil samples to identify their similarities and differences for oil source classification.
The results indicate that DeepLabv3+ and UNet++ models achieved the highest mean Intersection over Union (mIoU) scores for multi-class and binary segmentation tasks, respectively, demonstrating their robustness in detecting oil spills. Specifically, DeepLabv3+ achieved a mIoU of 68.3% in the multi-class dataset, excelling in complex categories like oil spills and look-alikes. UNet++ achieved a mIoU of 87.5% in the binary dataset, highlighting its effectiveness in distinguishing oil from non-oil regions. For oil fingerprinting, the Support Vector Classifier (SVC) model exhibited the highest accuracy, particularly in predicting the composition of n-alkanes, PAHs, and TPH, with F-scores of 1.0, 0.987, and 0.975, respectively. These findings underscore the effectiveness of coupling advanced machine learning models with established chemical analysis techniques, offering a reliable approach for oil spill detection and the subsequent effective cleanup.