Cross-scale feature interaction and graph attention for semantic segmentation of high-resolution remote sensing imagery

Semantic segmentation of high-resolution remote sensing imagery faces persistent challenges arising from vast scale variation among ground objects, boundary blurring caused by repeated downsampling, and high inter-class spectral similarity. We propose an encoder–decoder network that integrates two complementary mechanisms to address these issues jointly. A Cross-Scale Feature Interaction Module (CSFIM) establishes bidirectional gated pathways between adjacent encoder stages, enabling fine-grained spatial detail and coarse semantic context to reinforce each other before fusion. A Graph Attention Segmentation Module (GASM) builds a per-image region-adjacency graph over the deepest feature map—its patch nodes and similarity edges are recomputed at every forward pass—and applies multi-head graph attention to capture non-local dependencies among semantically related regions. A joint loss function combining weighted cross-entropy and Dice terms provides complementary supervisory signals that mitigate class imbalance and promote region-level coherence. Experiments on the ISPRS Vaihingen and Potsdam benchmarks show that the proposed method achieves mIoU scores of 81.35% and 83.27%, respectively, outperforming several established methods including UNet, DeepLabV3+, PSPNet, DANet, ABCNet, and MANet. Ablation analysis confirms that CSFIM and GASM interact synergistically, with their combined gain exceeding the sum of individual contributions.