Spatially Aware Uncertainty Feedback for Hard Region Learning in Remote Sensing Semantic Segmentation

Remote sensing semantic segmentation often produces spatially concentrated errors around object boundaries, small targets, shadowed areas, and ambiguous land-cover transitions. This study investigates whether learned spatial uncertainty can guide additional supervision of such difficult pixels. We propose Spatially Aware Uncertainty Feedback for Hard Region Learning (SUF-HRL), which refines a maximum-softmax-probability prior through a residual uncertainty branch attached to a SegFormer backbone. Local alignment and boundary concentration objectives encourage a task-adaptive spatial difficulty map; after a warm-up stage, its highest-scoring pixels receive additional cross-entropy supervision. Experiments use the Potsdam and Vaihingen benchmarks of the International Society for Photogrammetry and Remote Sensing and the LoveDA land-cover dataset. Under the controlled SegFormer-B2 protocol, mean intersection over union (mIoU) increases by 0.69, 1.14, and 1.01 percentage points, respectively. Averaging boundary-region mIoU over 3-, 5-, and 7-pixel bands gives gains of approximately 1.44, 1.32, and 1.32 percentage points. Component ablations support contributions from both spatial objectives. Learned uncertainty improves the reported spatial uncertainty criteria on Vaihingen, although it does not achieve the best precision–recall or calibration score. On Potsdam, training time per iteration increases by approximately 68.5% while the reported parameter totals remain nearly unchanged. The additional inference latency is small in the tested configuration. These results support learned spatial uncertainty as a useful training signal for hard-region segmentation.