Do models trained on sampled landscapes still work in the areas nobody sampled?
Ph.D. research at Kyushu University maps coffee production systems in Dak Lak, Vietnam, from Sentinel-2, Landsat, Sentinel-1 and terrain data, and is developing forecasts of Robusta coffee production. The models are interpreted with SHAP, and spatial block cross-validation tests how far their accuracy carries beyond the training areas.
Mapping tree-crop landscapes
Coffee production-system classes that were well separated within sampled landscapes were poorly separated in areas held out from training.
Multi-sensor satellite data are combined into yearly maps of coffee production systems, and the same data support production forecasts.
2026 · Remote Sensing Applications: Society and Environment
Coffee-system classes transfer poorly beyond sampled landscapes
A 2024 map of coffee production systems in Dak Lak reached 95.9% overall accuracy and a coffee-subclass macro F1 of 0.91 under cross-validation within sampled landscapes; when whole 10–20 km blocks were held out, the coffee-subclass macro F1 fell to 0.14–0.37.