NU-TSC: Weakly Supervised Forest/Nonforest Mapping Using Public Land Cover Products

Abstract

Accurate forest-cover mapping is essential for large-scale resource monitoring; however, acquiring dense pixel-level labels remains costly and limits practical scalability. This study investigates the use of publicly available land-use/land-cover products (PLCPs) as supervisory signals to alleviate the scarcity of high-quality labeled data and support reliable forest/nonforest mapping under limited supervision. Specifically, pixel-level training annotations are derived from PLCPs through a multistage label refinement procedure that reduces structural noise with minimal human intervention. To mitigate the influence of residual labeling errors, weakly supervised learning is implemented within a teacher–student architecture, incorporating a noise-aware masking strategy to identify potential annotation errors and suppress gradient contributions from unreliable labels. Pixel-level uncertainty is further estimated using stochastic teacher predictions to improve data utilization and reduce error accumulation. We also develop PAU-NeXt, a segmentation backbone that incorporates spatial coordinate attention and explicitly embeds normalized difference vegetation index priors into skip connections, thereby enhancing the discriminative capability of forest representations in complex backgrounds. Experiments on Sentinel-2 datasets demonstrate that the proposed approach achieves competitive performance compared with existing label-efficient methods, improving mean intersection over union from 89.37% to 92.39% and increasing nonforest F1-score from 88.75% to 92.23%. Further evaluations in two independent regions show that the method maintains stable performance across geographically diverse areas, highlighting its potential for practical large-area forest monitoring with limited manual annotation.