LAUSR

Automatic extraction of water bodies from various satellite images containing complex targets is a very important and challenging task in remote sensing and image interpretation. In recent years, convolutional neural networks (CNNs) have become an important choice in the field of semantic segmentation of remote sensing images. However, generic CNN models present many problems when performing water body segmentation, such as: 1) blurred water body boundaries; 2) difficulty in accommodating different scales of rivers, often losing information about many small-scale rivers; and 3) a large number of trainable parameters. This article proposes an end-to-end CNN structure based on multiscale residuals and squeeze-and-excitation (SE)-attention for water segmentation, called MRSE-Net. MRSE-Net consists of an encoder–decoder and a skip connection, which captures contextual information at different scales using the encoder, and then passes the encoder feature mapping through the improved skip connection, while localization is achieved by the decoder is implemented. With the multiscale residual module, the number of parameters in our model can be significantly reduced and water pixels can be extracted accurately. The SE-attention module is used to enhance the prediction results, mitigate the blurring effect, and make the segmented water boundaries more continuous. Landsat-8 images are used to train our model and validate our proposed method’s performance and effectiveness. In addition, we evaluate our method on Landsat-7 and Sentinel-2 images and obtain the best water segmentation results. Preliminary results on Sentinel-2 images show that the cross-sensor generalization capability of our model is beyond the range of the Landsat sensor family.