Abstract: In deep learning water extraction, there exists the problem that convolutional neural network has poor recognition effect on low-level semantic features, such as small lakes and small rivers. To solve this problem, a water extraction method based on Laplace edge enhancement is proposed. Synthetic Aperture Radar (SAR) data set is convolved with the pre-processed SAR data set, using the Laplacian operator to generate the Laplacian edge feature layer. Then the original image is fused with the generated edge feature layer to obtain the enhanced edge SAR data set, which makes the water edge clearer. On this basis, DeeplabV3+ and U-net semantic segmentation models are used for water extraction. The experiment shows that, compared with the unprocessed DeeplabV3+ and U-net models, the two models after Laplace operator processing have improved effect on water extraction in different regions. The U-net model after Laplace operator treatment has the best extraction effect on large water bodies, small lakes and small rivers.

Key words: water extraction, deep learning, SAR image, Laplacian edge enhancement, semantic feature