MTAU-Net: multi-scale feature fusion network for water body segmentation

Water segmentation has long been a key aspect of remote sensing image analysis. Water segmentation using optical remote sensing images has important practical significance in applications such as coastal mapping, marine resource management, and ship rescue. Optical remote sensing images contain rich information, but due to the complexity of spatial background features and the interference of noise, there are problems such as inaccurate tributary extraction and inaccurate segmentation when extracting water bodies. To achieve more refined water segmentation, we propose a U-shaped network for pixel-level water segmentation. The model contains two innovative components, namely, the multi-scale two-channel attention module and the multi-scale skip connection. Experimental results show that compared with other mainstream semantic segmentation networks, our method achieves higher segmentation accuracy and robustness. The IoU of the proposed MTAU-Net on the GF-2WS and LoveDA datasets reaches 92.57% and 77.78%, respectively.