Topology-Preserving Cross-Resolution Network for Semantic Segmentation of Complex Satellite Imagery

Abstract

Semantic segmentation of very-high-resolution‚ VHR‚ satellite imagery is important for land-cover mapping‚ urban planning‚ and disaster recovery․ However‚ existing deep networks have high pixel accuracy but fail on thin‚ elongated‚ connected structures‚ like roads‚ hedgerows‚ or building perimeters․ Most pixel-wise loss functions do not penalize them‚ causing these topological defects that make the maps less useful․ To overcome limitations of conventional pixel-wise losses‚ we propose a topology preserving‚ cross-resolution network (TP-CRNet) that combines a multi-branch encoder with two novel modules․ Cross Resolution Fusion Module (CRFM) passes information between resolution branches using bidirectional attention‚ thus preserving fine detail and global context․ Second‚ we use Boundary-Topology Head (BTH) to provide skeleton and persistent homology supervision through a composite loss that penalizes unnecessary disconnections and non-desired holes․ The model is evaluated on the ISPRS Potsdam dataset‚ the LoveDA dataset and the DeepGlobe dataset using mIoU‚ overall accuracy‚ boundary F1‚ clDice‚ and Betti error․ TP-CRNet produces large and consistent improvements of region accuracy and topological correctness compared to U-Net‚ DeepLabV3+‚ HRNet and UNetFormer‚ especially for narrow linear classes․ The ablation study shows that CRFM and BTH both provide complementary benefits and the cross-resolution experiments show that the model has graceful degradation on different ground sampling distances․