Fully Automated Sar Based Oil Spill Detection Using Yolov4

The Eastern Mediterranean Sea is known as oil pollution hotspot because of high marine traffic and a growing number of oil and gas industrial activities inside, which makes efficient monitoring oil spills important in this area. Spaceborne Synthetic Aperture Radar (SAR) plays an important role for oil spill detection with its advantage of wide coverage and all-weather observations. However, discriminating whether the dark formations in the SAR imagery are from actual oil spills or look-alikes has been a challenging part. This study applied You Only Look Once version 4 (YOLOv4) object detection algorithm as an one-class (i.e. oil spill) object detector for learning oil spill features inside the Region of Interests (ROIs) and the background information from the rest of the image. The preliminary results pointed out that the pixel threshold for removing some tiny oil spills is suggested as they appeared regularly in the study area but are hardly visible. The average precisions (AP) of the trained model on validation and test sets are 67.80% and 65.37%, showing that the model is not overfitting on our training and validation sets. In addition, this study recommended some data augmentation strategies which might help improve the results.