Crop type classification in smallholder agriculture of central and South Asia using Sentinel-1/2 data fusion

Accurate crop monitoring in smallholder-dominated regions is challenging due to fragmented fields, high crop diversity, and often limited ground truth data. This study presents a systematic and transferable framework for crop-type classification in smallholder systems by jointly evaluating multi-sensor data fusion, temporal feature aggregation, feature selection, and model applicability. This study evaluates crop-type classification accuracy across smallholder agricultural landscapes in Central and South Asia (Kazakhstan, Tajikistan, Pakistan), leveraging Sentinel-1 radar and Sentinel-2 optical data separately and combined. Employing Random Forest models, we systematically compare temporal aggregation approaches (monthly, bi-monthly, quarterly) and evaluate the impact of feature selection on model performance. Across all study regions, combined Sentinel-1 and Sentinel-2 data achieved overall classification accuracies of approximately 80–96%, with substantial performance gains relative to single-sensor models, particularly in regions where individual sensors showed limited discrimination capability. Depending on region and sensor, accuracy improvements ranged from a few percentage points to more pronounced gains, reflecting strong benefits of data fusion in heterogeneous smallholder systems. Finer temporal aggregation schemes, including monthly aggregation, yielded additional accuracy gains of approximately 1–3 percentage points compared to coarser aggregations, while feature selection further improved model performance by roughly 2–5 percentage points. Sentinel-1 proved particularly effective for structurally distinct crops such as cotton, while Sentinel-2 substantially improved classification of more diverse crop classes. Application of the Area of Applicability concept enabled spatially explicit identification of well-supported and extrapolated predictions, providing a quantitative basis for uncertainty assessment and future sampling strategies. Together, these results demonstrate the value of an integrated and transferable methodological framework for robust crop-type classification in smallholder agricultural systems using freely available Sentinel data.