ASSPSE-Net: A Novel Deep Network with Multifeature Fusion for Extracting a Building Contour from High-Resolution Remote Sensing Images

Abstract As urbanization accelerates, automatic building contour extraction from high-resolution images is crucial for urban planning and disaster assessment. Traditional methods struggle with complex backgrounds and occlusions. We propose a multifeature fusion deep network ASSPSE-Net. First, it improves based on the U-Net structure, using the residual network ResNet50 as an encoder to enhance the deep feature detail extraction ability of the model. Second, the squeeze-and-excitation networks (SE-Net) in the channel attention mechanism are added to make the model pay more attention to the detailed features of the building, and Atrous Spatial Pyramid Pooling (ASPP) is added to the last layer of the encoder to improve the ability to recognize multiscale buildings. Finally, the loss function combining focal loss and Dice loss is used to balance the ratio of the target to the background to enhance the extraction of difficult sample information. Experiments on the Wuhan University (WHU), Massachusetts, and self-made Changsha Building Dataset show that the accuracy of ASSPSE-Net on the three datasets is 98.17%, 93.98%, and 90.84%, and the mean intersection over union (MIoU) is 93.36%, 79.30%, and 71.02%, respectively. Compared with other semantic segmentation methods, ASSPSE-Net can accurately extract building contours and local detail features in complex backgrounds and multiscale buildings.