HDRoad: An encoder-decoder architecture with hybrid attention and directional prior for efficient road extraction from remote sensing images

Road extraction from very high resolution (VHR) remote sensing images (RSIs) presents significant challenges due to the varied morphology and high semantic complexity of road structures. Many existing methods struggle to consistently perform well across diverse and complex scenarios. Additionally, balancing efficiency and performance remains an unresolved issue in prior research, particularly those employing transformers. To address these challenges, we propose HDRoad, a novel encoder-decoder architecture that improves both model performance and computational efficiency, enabling training and inference on high-resolution inputs with a single GPU. The encoder, Hybrid Attention Network (HA-Net), combines dense and sparse spatial attention to effectively distill semantic information. The decoder, Directional Augmented Road Morphology Extraction Network (DARMEN), uses morphological priors to accurately refine and reconstruct road features. The model is validated on the DeepGlobe dataset and SouthernChina12k, a newly developed road segmentation dataset comprising 11,791 images from various remote sensing sources. Experimental results demonstrate that HDRoad achieves an IoU of 75.09 % on the DeepGlobe dataset and is the only model exceeding 60 % IoU on SouthernChina12k, setting new benchmarks for state-of-the-art performance in the field.