From local context-aware to non-local: A road extraction network via guidance of multi-spectral image

The use of artificial intelligence has led to an increase in road extraction projects from satellite images through deep learning. However, multi-spectral images (MSI) have been largely overlooked in road extraction algorithms due to their lower resolution compared to panchromatic or fused images. Additionally, deep learning faces the challenge of image content reasoning from distant contexts in data rule mining. To address these issues, we propose a new method for road extraction called the MSI-guided Segmentation Network, which utilizes all the data from the GF2 satellite to achieve optimal results. This study highlights the advantages of using MSI with low-resolution for obtaining deeper semantic information in a faster manner, while high-resolution fused images are better suited for extracting precise characteristics. The proposed method includes two sections: (1) a local symmetry feature fusion to enhance the network's local context-awareness for shallow details, and (2) a global asymmetric semantic fusion to improve the network's capability to comprehend the whole scene for deep semantic information. Moreover, to evaluate the robustness and generalization of this method, we have provided a GF2 Full-band China Road Dataset. The codes and datasets will be made public on https://github.com/ Dudujia160918/MSNet.