TabPFN for crop classification using multi-temporal ALOS-4/PALSAR-3 SAR data under imbalanced conditions

Abstract

Accurate crop classification using synthetic aperture radar (SAR) data is essential for operational agricultural monitoring but remains challenging due to complex scattering mechanisms, correlated features, and class imbalance in reference datasets. This study evaluates the applicability of Tabular Prior-Data Fitted Networks (TabPFN) for SAR-based crop classification and compares its performance with Random Forest, Support Vector Machine, and Extreme Gradient Boosting. Field-level classification was conducted using multi-temporal SAR-derived features, and model performance was assessed through cross-validation with fold numbers ranging from 2 to 10. Overall Accuracy, Cohen's Kappa, class-wise User's Accuracy, Producer's Accuracy, and F1-score were examined under four sampling strategies: no oversampling, and three Synthetic Minority Over-sampling Technique (SMOTE)-based approaches (Borderline-SMOTE, DeepSMOTE, and KMeans-SMOTE). Permutation feature importance was used to ensure consistent interpretation across models. Results show that TabPFN consistently achieved the highest classification performance across all validation settings, reaching an Overall Accuracy of up to 91.9% and a Cohen's Kappa of up to 0.896, outperforming the best-performing conventional machine learning model. Results also demonstrated strong robustness to class imbalance, with the best performance obtained using the original imbalanced dataset without any SMOTE-based oversampling. These findings suggest that model selection may be more influential than synthetic oversampling in SAR-based crop mapping. Furthermore, the robustness of TabPFN may reduce the reliance on additional oversampling procedures, thereby simplifying the preprocessing workflow for operational applications. Taken together, the results highlight the potential of TabPFN as a practical and effective approach for agricultural monitoring and crop mapping.