Multi-scales feature extraction model for water segmentation in satellite image

This paper proposed multi-scales feature extraction architecture and loss function focusing on water segmentation. In this paper, Res2Net is used feature extraction on the encoder, which incorporates receptive field of different sizes to facilitate the extraction of water areas and surrounding information such as land, cloud. The skip connection is used as the input of the attention gate, which can increase the weight of the water to highlight the water area. In this study, IoU of water areas are 0.76, which is a significant improvement compared to other architectures.