Surface water extraction from high-resolution remote sensing images based on an improved U-net network model

This paper proposes a surface water extraction method from high-resolution remote sensing images based on an improved U-Net network model. The GF-6 satellite is a significant achievement in optical remote sensing satellites in China, with a spatial resolution of 2 m. Using the high-resolution remote sensing images of a typical area in the southwest of Hefei City, Anhui Province, as experimental data, image cropping, label production, and data augmentation were carried out, and U-Net network was used to mine the deep and shallow features of the images. Firstly, add Dropout and BN layers to improve the training speed and robustness of the model while avoiding overfitting. Next, add the upper-level feature maps in the contraction path to the expansion path to form a dual feature channel fusion mechanism, preventing the loss of detailed information. Finally, rapid surface water recognition was achieved by continuously adjusting parameters to train the network model. The overall recognition accuracy of various objects was 97.99%, the recognition precision rate of target objects was 91.54%, the recall rate was 71.71%, the F1 value was 80.42%, and the Kappa coefficient was 0.90. In addition, the extraction results of this method were compared with other recognition methods such as U-Net, U-Net+, and NDWI, and the results showed that the accuracy, recall rate, and F1 value of the surface water extraction results were improved, with better accuracy.