Climate-Routed Low-Rank Adapters for Label- and Compute-Efficient Cross-Country Cropland Mapping in Africa

Cropland mapping in African countries faces problems due to scarcity of local labels, agro-climate diversity, and limited GPU re- sources in research organizations. We present CAPA, a climate-guided parameter-efficient adaptation framework for a frozen Presto remote- sensing encoder, with a sparse mixture of four low-rank residual adapters guided by the ERA5 climate, Sentinel-2 phenology, and SRTM elevation information, along with soft-mixture training and hard top-1 inference routing in the frozen state. We test our approach with the AgroShift- Africa Leave-One-Country-Out (LOCO) setup, for seven African coun- tries in CropHarvest [9] using a geographically disjoint support/validation/test split with a one-cell geohash buffer between them. With 50 labels for the target countries, CAPA attains a mean Macro-F1 score of 77.9 ± 1.0% and a lowest-country (Mali) F1 of 68.2%, showing better performance compared to baseline LoRA by 4.0 pp mean and 6.1 pp lowest-country, as well as full fine-tuning by 2.7 pp mean. The parameter-matched single- adapter baseline (4,354 parameters) scores 73.7%, thereby establishing that the performance boost is not caused by increased capacity alone. CAPA updates ≈0.15% of the model’s parameters and needs 1.4 GB GPU memory peak and 1.8 minutes adaptation on an RTX 3090.