Rethinking BiSeNet: A Lightweight Network for Urban Water Extraction

Urban water runs through and plays an important role in urban development, and understanding the spatial changes of urban water bodies is significant. However, the rapid growth of remote sensing data caused by rapid urbanization poses a challenge to the efficiency of water extraction methods. In this article, a high-precision method with a lightweight structure for urban water extraction was proposed in terms of the dilemma between efficiency and accuracy. We designed a squeeze-and-excitation attention refined module (SE-ARM), squeeze-and-excitation feature fusion module (SE-FFM), and depth-wise atrous spatial pyramid pooling (DW-ASPP) based on BiSeNet to improve it in the water semantic segmentation application, which resulted in SE-BiSeNet. To validate the proposed model’s effectiveness, we compared our SE-BiSeNet to other advanced water extraction methods on the Chengdu dataset in terms of accuracy and efficiency. The results showed that the proposed model performs well in both accuracy [0.96 intersection over union (IoU)] and efficiency (6.4 frame/s on a GTX 1060 card) with fewer parameters and calculations, which indicates an objective application potential.