The thesis addresses the problem of semantic segmentation of marine environments from Sentinel-2 satellite imagery, the focus being on detecting floating marine debris, which represents small, sparse structures that are difficult to detect. The initial analysis revealed limitations of the baseline UNet model in distinguishing rare classes and in detecting objects spanning only a few pixels. Based on a detailed visual and data-driven error analysis, several improvements were explored, including architectural expansion, the use of alternative loss functions, and more complex network architectures. The best performance was achieved by a modified UNet model with an increased number of hidden channels and a combined loss function (cross-entropy + Dice loss). The proposed improvement significantly enhances the segmentation of rare and challenging classes and outperforms the baseline model across all evaluated metrics.
|
