Water Segmentation from SAR Images with the Presence of Speckle Noise

Abstract

Paper presents water segmentation algorithm in satellite SAR images. One of the purposes of water segmentation is flood monitoring and assessment its scale. Flood monitoring is complicated by the presence of severe weather conditions such as raining and cloudness. For this reason, C-band images were chosen which ignore atmospheric conditions and time of day. For the investigation, a dataset of 27 satellite images form Sentinel-1 with a spatial size of approximately 200 by 300 kilometers with a resolution of 10 meters per pixel was collected. U-ResNet-34, SegFormer_b5 and SegNeXt_l neural networks are used as segmentation algorithms. Training using balanced batch and augmentation invariance was used to improve the quality of the algorithms and the highest Dice value was equal 0.90. The paper also considers speckle noise filtering using a median filter and BM3D which allowed increasing the F1 value by 0.01.