Investigation of the spatio-temporal transferability of large-scale random forest crop classification models and potential influencing factors

Training samples are crucial for remote-sensing-based crop classification; however, field surveys for data collection over large areas are costly and challenging. Model transfer offers a viable solution. This study evaluates the spatio-temporal transferability of machine learning models for large-scale crop classification and examines factors influencing classification performance across different training regions and years. A Random Forest model trained on Landsat/Sentinel-2 time series and CDL crop labels from three U.S. states was applied to six European countries for crop mapping. Results showed that models trained in Kansas achieved high overall accuracy (88.1–96.8%), with minimal accuracy variation (0.6–2.1%) across training years, demonstrating strong adaptability. However, model performance varied significantly across training regions (up to 21.6% accuracy difference), with Kansas outperforming Minnesota and South Dakota. Interestingly, models trained on merged regions or years performed worse than the best individual models, indicating that broader spatial or temporal training data cannot guarantee better transferability. Transferred models achieved accuracy comparable to locally trained models (≤2.0% difference). Temperature similarity between regions is a key factor, with greater annual temperature alignment leading to better transferability. This study offer a practical solution to large-area crop classification without local sample and guide the selection of optimal training areas.