Road Extraction from High-Resolution Remote Sensing Images Based on MRCBL-Net and Multi-attention Mechanism

Abstract

Road extraction from high-resolution remote sensing images is crucial for urban planning and geographic information systems (GIS). However, complex background interference, severe occlusions, and the inherent morphological complexity of roads often lead to discontinuities and insufficient accuracy in extraction results. To address these challenges, this paper proposes a Multi-scale Refined Convolutional Backbone Link Network (MRCBL-Net). The network adopts ConvNeXt as the backbone encoder to enhance feature extraction capability and embeds the proposed Multi-scale Dilated Convolution and Spatial-Channel Refinement (MULCSR)module into skip connections. The core of this module is the Channel and Spatial Relation (CSR)module, which captures multi-scale contextual information through parallel multi-branch dilated convolutions, and adaptively calibrates feature weights via channel attention and spatial attention mechanisms in sequence. This focuses on key road features, thereby effectively fusing shallow details and deep semantics. Experiments on the DeepGlobe dataset demonstrate that the proposed method achieves an accuracy of 98.61%,precision of 82.98%,recall of 82.46%,F1-score of 82.52%,and Intersection over Union(IoU)of 70.55%.Experiments on the CHN6-CUG dataset show that the method attains an accuracy of 96.13%,precision of 79.79%,recall of 79.66%,F1-score of 79.28%,and IoU of 66.16%.Compared with existing methods, the proposed method achieves the optimal performance in all five metrics. Qualitative results further confirm that the proposed method excels in maintaining road continuity and restoring fine boundaries, improving the robustness and accuracy of road extraction in complex scenarios.