Multi-Teacher Divergence Perception Network for Semi-Supervised Semantic Segmentation of Remote Sensing Images

Semi-supervised semantic segmentation can exploit unlabeled remote sensing images to reduce the cost of pixel-level annotation, but its performance is strongly constrained by pseudo-label reliability. Multi-teacher predictions often exhibit structured disagreement around boundaries, small objects, and interclass-confusion regions, while a conventional softmax head may assign spuriously high confidence without sufficient supporting evidence. To address these limitations, we propose a Multi-Teacher Divergence Perception Network (MTDPNet). Its Multi-Teacher Adaptive Disagreement-Aware Module (MADM) explicitly measures pixel-level teacher disagreement using the Jensen–Shannon divergence, adaptively weights teacher predictions, and applies reliability-specific supervision to different image regions. In parallel, an Evidential Uncertainty Estimation Module (EUEM) fuses multiscale student features and constructs a Dirichlet evidence distribution to identify evidence-deficient predictions and prevent their over-reinforcement. On the ISPRS Vaihingen and Potsdam datasets with 5% and 25% labeled data, MTDPNet achieves mIoU scores of 63.44%/68.18% and 50.95%/64.96%, respectively, and corresponding mF1 scores of 76.51%/80.24% and 67.01%/78.63%, yielding the best overall performance in all four settings. These results demonstrate that jointly modeling teacher disagreement and student evidential uncertainty improves the reliable use of unlabeled remote sensing images, particularly for difficult regions and small objects.