DRD-SSALNet: A Dual-Representation Decoupled Self-Supervised Active Learning Network for Remote Sensing Semantic Segmentation

Abstract

The active learning reduces pixel-level annotation costs for remote sensing semantic segmentation, but classifier-dependent sampling is vulnerable to cold-start bias. This letter proposes a dual-representation decoupled self-supervised active learning network (DRD-SSALNet) that separates sampling from classification. The sampling path combines DINO-based semantic clustering for global diversity and MAE-based reconstruction responses for boundary-aware complexity, while DeepLabv3 + is used only for segmentation. Positive sample pairs weakly coordinate the two sampling branches. Experiments on LoveDA and ISPRS Potsdam show that DRD-SSALNet achieves 68.7% and 89.3% mean intersection over union (mIoU) and improves first-round mIoU by 5.6 percentage points, boundary $F1$ -score by 6.8%, and sparse-category recall by 12.1%, demonstrating its effectiveness for low-cost remote sensing segmentation.