Multimodal water body extraction network based on remote sensing satellite images and digital surface model

Multispectral remote sensing satellite images exhibit characteristics such as small objects, complex scenes, significant changes in object scale, and difficulty in distinguishing regions with similar spectral features. As the spectral reflection characteristics of water bodies vary with factors like season and geographical location, different background information affects the accuracy of the water body extraction. Therefore, extraction of broken and discontinuous water bodies is still challenging. Recent studies have shown that using multimodal information can represent the features of targets from different perspectives, thereby improving the robustness of semantic segmentation. To address these issues, this paper utilizes the spectral index’s ability to recognize water to drive the extraction accuracy of neural networks for water recognition. A multimodal remote sensing semantic segmentation network (MRSSNET) is proposed, which integrates water index method to fuse images with Digital Surface Model (DSM) images. We use deep learning-based segmentation models to perform water segmentation, such as Fully Convolutional Networks (FCN), U-Net, Segformer, SegNext and Deeplabv3+, representatively. Experimental results demonstrate that MRSSNET outperforms the other four algorithms in identifying water bodies within complex and discontinuous geographical environments.