Bilayer U-Net with Depth-Wise Separable Graph-Convolution Network for Robust Road Extraction in Remote-Sensing Images

Based on remote-sensing imagery, road extraction has always been a common and challenging task. In this paper, we propose a Bilayer U-Net network based on a Depth-wise Separable Graph-convolution network (DSGNet). Compared to previous road segmentation methods, our model places more emphasis on the characteristics of small roads through the double U-Net design and balanced max pooling (BMP). We also introduce graph convolution to aggregate neighbor node features, enhancing the model's segmentation performance. Furthermore, adding depth-wise separable convolution before the graph convolution reduces complexity without affecting performance. The skip connections with the channel grouped cross fusion (CGCF) module enable better handling of road global and shape-related aspects. Owing to its simple and effective network and module design, DSGNet demonstrates strong performance, achieving 80.86% mIoU and 78.00% F1-Score on the Massachusetts road dataset.