Publication

Abstract : Automatic road extraction from high-resolution remote sensing images is essential for urban and rural development, as well as traffic management and environmental preservation, enabling informed decisions for sustainable planning. In remote sensing images, roads are densely packed and occupy only a small portion of the image, surrounded by complex backgrounds. This makes it challenging to preserve the integrity and connectivity of the extracted road network. To address this problem, we propose GADC-KANNet, to enhance the detection and extraction of road pixels from remote sensing images. The proposed gradient aware directional convolution layer enhances the model’s ability to capture directional features, improving road pixel detection by adapting to the varying orientations of roads; the dilated residual activation path extends the receptive field by introducing dilated convolutions into the residual connections, allowing the model to capture more complex spatial and semantic patterns; and the Kolmogorov–Arnold Networks-based Feature Selection Fusion module filters out redundant features and facilitates the seamless fusion of high-level and low-level information, optimizing the feature integration process for more accurate road extraction. Extensive experiments were conducted on the benchmark datasets MIT, DeepGlobe, and CHN6-CUG. These datasets feature varied road layouts, image resolutions, and environmental complexities. Experimental results on three public datasets show that our method achieves notable improvements compared with state-of-the-art methods, with the F1-score improved by 0.46% on the MIT dataset, 0.43% on the DeepGlobe dataset, and 1.23% on the CHN6-CUG dataset for road extraction.