Spatial Transferability of Random Forest Models for Crop Type Classification Using Sentinel-1 and Sentinel-2

Large-scale crop type mapping often requires prediction beyond the environmental settings of training sites. Shifts in phenology, field characteristics, or ecological site conditions previously unseen area, may reduce classification performance machine learning classifiers that overfit to This study aims assess spatial transferability Random Forest models for across Germany. The effects different input datasets, i.e., only optical, Synthetic Aperture Radar (SAR), and optical-SAR data combination, impact feature selection were systematically tested identify optimal approach shows highest accuracy transfer region. selection, a combined with cross-validation, should remove features carry site-specific information data, which turn can model areas. Seven sites distributed over Germany analyzed using reference major 11 crops grown year 2018. Sentinel-1 Sentinel-2 from October 2017 2018 used as input. estimation was performed spatially independent sample sets. results combination outperformed those single sensors (maximum F1-score–0.85), likewise areas not covered by F1-score–0.79). forest based on SAR showed lowest losses when transferred regions (average F1loss–0.04). In contrast entire set, substantially reduces number while preserving good predictive Altogether, applying SAR-only is beneficial large-scale where evenly complete