CM-UNet++: A Multi-Level Information Optimized Network for Urban Water Body Extraction from High-Resolution Remote Sensing Imagery

Urban water bodies are crucial in urban planning and flood detection, and they are susceptible to changes due to climate change and rapid urbanization. With the development of high-resolution remote sensing technology and the success of semantic segmentation using deep learning in computer vision, it is possible to extract urban water bodies from high-resolution remote sensing images. However, many urban water bodies are small, oddly shaped, silted, or spectrally similar to other objects, making their extraction extremely challenging. In this paper, we propose a neural network named CM-UNet++, a combination of the dense-skip module based on UNet++ and the CSMamba module to encode different levels’ information with interactions and then extract global and local information at each level. We use a size-weighted auxiliary loss function to balance feature maps of different levels. Additionally, features beyond RGB are incorporated into the input of the neural network to enhance the distinction between water bodies and other objects. We produced a labeled urban water extraction dataset, and experiments on this dataset show that CM-UNet++ attains 0.8781 on the IOU (intersection over union) metric, which indicates that this method outperforms other recent semantic segmentation methods and achieves better completeness, connectivity, and boundary accuracy. The proposed dense-skip module and CSMamba module significantly improve the extraction of small and spectrally indistinct water bodies. Furthermore, experiments on a public dataset confirm the method’s robustness.