Multi-scale boundary-aware network for remote sensing image semantic segmentation

Accurate semantic segmentation of high-resolution remote sensing images is crucial for urban planning, land cover mapping, and environmental monitoring. However, this task remains challenging due to the significant scale variations and blurred or unclear boundaries between objects in complex scenes. Conventional neural networks (CNNs) are effective in extracting local spatial details but have limited capability in modeling global context, while Transformer-based approaches capture long-range dependencies but often overlook fine structures and boundary cues and incur high computational costs. Therefore, we propose a network integrating CNN with Transformer, termed the Multi-Scale Boundary-Aware Network (MSBANet). The Multi-Scale Transformer Block (MSTB) extracts multi-scale semantic and boundary structural features through a Multi-Header Self-Attention (MHSA) mechanism and a Multi-Scale Convolutional Gated Linear Unit (MConvGLU). The Global-Local Fusion Module (GLFM) aligns deep semantic features with shallow spatial details during decoding, ensuring boundary structure integrity throughout multi-scale reconstruction. Furthermore, the Uncertainty Boundary Awareness Module (UBAM) employs an entropy-guided attention mechanism to enhance the model’s focus on uncertain regions, thereby optimising edge segmentation outcomes. Extensive ablation and comparative experiments demonstrate the effectiveness of each module and the overall architecture. The proposed MSBANet achieves 84.63% and 87.15% mIoU on the ISPRS Vaihingen and ISPRS Potsdam datasets, respectively, outperforming several state-of-the-art methods.