MSANet: A Multi-Scale Remote Sensing Road Extraction Network With Attention Mechanisms

Road extraction from high-resolution remote sensing imagery remains a challenging task due to complex background textures, significant intra-class scale variations, and frequent occlusions, all of which cause fragmented predictions and loss of road details. To mitigate these issues, this study leverages two publicly available benchmark datasets—DeepGlobe and Massachusetts Roads—and examines a set of modelling strategies involving multi-scale feature inputs, attention-enhanced feature representations, and pixel-wise road segmentation outputs.We propose MSANet, a Multi-Scale Remote Sensing Road Extraction Network with Attention Mechanisms, which integrates three core components for robust and accurate road extraction: (1) a multi-scale residual encoder that captures road structures of varying widths through hierarchical receptive-field expansion; (2) a dual-dimensional attention module that explicitly models long-range dependencies in both channel and spatial domains to suppress background interference and enhance road continuity; and (3) a lightweight decoder that generates refined high-resolution predictions. Quantitative experiments show that MSANet achieves 65.2% IoU / 79.5% F1 on DeepGlobe and 56.8% IoU / 72.6% F1 on the Massachusetts dataset, surpassing widely adopted baselines such as U-Net and LinkNet. These results indicate that the proposed architecture more effectively balances local detail preservation and global structural consistency. Compared with existing approaches, MSANet demonstrates superior robustness against occlusion-induced discontinuities and scale variations, underscoring its practical relevance and significance for road extraction in complex remote sensing scenarios.