A novel semantic feature enhancement network for extracting lake water from remote sensing images

The automatic lake water extraction method based on semantic segmentation is a research hotspot in the field of remote sensing image processing. In remote sensing images, the presence of complex noise information at the lake boundary hinders the normal expression of boundary information, which leads to methods cannot extract a coherent lake boundary. Moreover, partial small-scale lakes’ texture features are weak and easily masked by the background information. To address the above issues, an end-to-end semantic segmentation network is designed. The network uses a symmetric encoder-decoder architecture to extract lake water in remote sensing images. On the one hand, a directional noise reduction filtering algorithm is proposed to reduce the impact of noise information on the network segmentation process. The algorithm utilizes a preset directional guide map to guide the nonlinear propagation of boundary noise and suppress low-contrast halo artifacts in the image, thereby better preserving the boundary sharpness of the lake. On the other hand, for the problem of missing small-scale lakes, an attention gate compression module is embedded in the skip connection. This module can adaptively integrate the correlation features between different ground objects, and selectively assign more attention to small-scale lakes, thereby improving the network’s ability to recognize such lakes. In the experimental results, our method can produce more accurate lake water extraction results than the current mainstream methods. Besides it has an excellent performance in accurately identifying lake boundaries and small-scale lakes.