Residual-Tail Boosting for Efficient Sentinel-2 Crop Classification: A Post-Training Approach

Accurate crop classification using Sentinel-2 imagery is essential for scalable and timely agricultural monitoring. While ensemble tree models such as XGBoost and LightGBM offer efficient and interpretable baselines, their performance often plateaus due to limited capacity for residual correction. In this work, we introduce Residual-Tail Boosting (RTB), a lightweight post-training enhancement that refines predictions by modeling the residual errors of trained classifiers using shallow learners. RTB operates modularly, requires no retraining, and introduces minimal computational overhead. Applied to the Sen4AgriNet benchmark, RTB improves macro F1-score by up to 2.2%, weighted F1-score by 1.6%, and AUC by 1.5 points, with under 5 s of additional CPU time. Compared to deep learning methods achieving \(\sim \)  0.95 F1 but requiring substantial resources, our RTB-enhanced pipeline provides a practical trade-off between performance, interpretability, and speed. RTB is particularly suited for real-time, resource-constrained, or explainability-critical crop monitoring systems.