A Road Extraction Method of a High-Resolution Remote Sensing Image Based on Multi-Feature Fusion and the Attention Mechanism

Road extraction from high-resolution remote sensing images has a lot of practical value and significance and has been a research hotspot. Considering that methods based on deep learning and the attention mechanism have achieved good performance in road detection, this paper proposes a deep residual network and an attention mechanism based on the fusion of multiple road features. The encoder–decoder structure of the U-net network with strong multitasking generality is adopted as the basic network. It integrates the spatial multi-scale and multi-channel features of the road to enhance the robustness of feature extraction. Meanwhile, the decoder design based on the attention mechanism further improves the recognition accuracy and effectively curbs the increase in computing cost and time cost. A loss function based on the gradient coordination mechanism is introduced to address the imbalance of road sample data. Finally, experimental verification is carried out on two public road datasets and both qualitative and quantitative comparisons are conducted. Results show that the proposed method is satisfactory and outperforms other methods.