A Geo-AI Framework for Label-Efficient Agricultural Land-Cover Mapping with Climate-Aware Active Learning

Multi-source agricultural land-cover mapping is important for food-system monitoring, sustainable land management, and climate-adaptation planning, yet obtaining reliable labels remains difficult in heterogeneous landscapes. This paper presents a climate-aware, label-efficient active learning framework for reducing annotation demands in agricultural mapping. The framework uses multi-source predictors derived from Sentinel-1, Sentinel-2, ERA5 climatological variables, and topographic context to prioritize informative samples under constrained labeling budgets. In the main benchmark, uncertainty-guided and diversity-aware sampling achieved a macro F1-score of 0.827 while using fewer than 40% of the training labels used by the same-pool supervised baseline. Additional feature-group and cross-regional experiments showed that agro-environmental context improves robustness, particularly under regional domain shift. Overall, the results indicate that uncertainty-guided, diversity-aware active learning is a practical Geo-AI strategy for reducing annotation requirements in climate-relevant agricultural monitoring within the scope of binary crop/non-crop mapping over ML-ready feature time series.