Distilling 10-m Land Cover Maps From Multisource Consensus via AlphaEarth Embeddings and Noise-Aware Weak Supervision

Accurate 10-m land use and land cover (LULC) maps are essential for environmental monitoring, yet their scalable generation is often constrained by the high cost of manual annotation and the systematic label noise present in existing global products. To address this challenge, we propose a weakly supervised LULC mapping framework that exploits semantic consensus from multiple noisy products using foundation model embeddings. Instead of relying on raw optical imagery, the framework adopts AlphaEarth embeddings as input representations. A probabilistic consensus strategy is developed to transform conflicting discrete labels from seven global LULC products into soft pseudo-labels. To further suppress label noise, a symmetric dual-branch Co-teaching network is employed, guided by a dynamic curriculum learning scheme. The proposed method is validated using high-precision National Land Use Investigation data over Taiwan. Experimental results show that the framework achieves an Overall Accuracy of 80.61%, a Macro F1-score of 65.01%, and a Kappa coefficient of 0.7543, substantially outperforming traditional majority voting and individual global products. In particular, notable improvements are observed for minority and highly confused classes such as shrubland. These results indicate that combining foundation model embeddings with noise-aware Co-teaching provides an effective solution for regional-scale LULC mapping under weak supervision.