Detecção e classificação de derramamento de óleo na superfície oceânica baseada em aprendizagem profunda via algoritmo YOLO

Abstract

Due to global growth, the importance of the world economy in relation to the use of petroleum-based raw materials has increased. As a result of this growth, the importance of issues related to socio-environmental and economic concerns has been highlighted in the scientific field. Consequently, the occurrence of oil spill incidents on the ocean surface requires the development of methodologies to mitigate the impact caused by the problem in the affected areas. With the availability of satellites equipped with Synthetic Aperture Radar, it is possible to monitor, detect and classify spills of oil and its derivatives at sea. This dissertation presents a proposal for a methodology based on deep learning, specifically using the YOLO family of algorithms. Therefore, according to the experiments carried out using the dataset obtained via radar and provided by the SENTINEL-1 mission, during the tests in the validation phase for YOLOv8 nano, small and medium, better performance was observed for medium, with accuracy metrics, mAP-50 and mAP50-90 equivalent to 0.891%, 0.85% and 0.716%, respectively. The result in the test phase reached a confidence level, according to the IoU (Intersection over Union) metric, of more than 70% of the objects classified as oil slicks.