MartiNet: An Efficient Approach For River Segmentation In SAR Images

The ubiquity of deep learning models in applications involving satellite imagery has seen steady progress particularly with Synthetic Aperture Radar (SAR) images. Known for its quality of image data regardless of the weather or time of the day, SAR images capturing water bodies has been discussed and extensively studied. With the focus shifting from inefficient image processing techniques to more accurate and bigger neural networks, the need for a computationally inexpensive model arises. In this paper we propose an end-to-end methodology to obtain higher segmentation accuracies, while also keeping the number of trainable parameters low. We recommend the method of gamma correcting and thresholding SAR images to obtain a cleaner dataset. We invert images to bring focus to thin streams and employ a triple learning strategy in the model architecture to improve outputs from the final layer. Finally, we compare our methodology with existing architectures and discuss the results. A quantitative analysis shows that our approach achieves an mIoU of 0.9903 and an F1 score of 0.9947.