Integrating deep learning with active contour models in remote sensing image segmentation

Semantic image segmentation using deep learning is a crucial step in remote sensing and image processing. It has been exploited in oil spill identification in this work. Remote sensing Synthetic Aperture Radar (SAR) images have been used to identify oil spills due to their capability to cover wide scenery irrespective of the weather and illumination conditions. Oil spills can be seen by radar sensors as black spots. Nonetheless, the discrimination between the oil spills and looks-alike is challenging in the case of semantic segmentation at pixel level. To overcome this problem, the active contour without edges models take into account the length of boundaries, the areas inside and outside the region of interest to be integrated in the deep learning image segmentation model. For this purpose, a loss function, which includes the area and the length of object, is back propagated into the semantic segmentation architecture to optimize the deep learning process. The method is evaluated on a publicly available oil spill dataset. The experiments show that the proposed approach outperforms other state-of-the-art methods in terms of Intersection over Union (IoU).