Aiming at the problem that the water information in high-resolution remote sensing images is easily disturbed by non-water information such as vegetation, building shadow, and roads near the water, a water information extraction model for high-resolution remote sensing images is proposed in this paper. We introduced the Polarized Self-Attention (PSA) mechanism connected in parallel into the D-LinkNet to reduce the information loss caused by dimension reduction. In addition, we constructed a new water data set based on GF-2 satellite remote sensing images. The improved D-LinkNet model has achieved excellent performance in GF-2 satellite remote sensing images. Compared with other water extraction methods, the results show that the improved D-LinkNet model can achieve accurate and fast water extraction from remote sensing images.