Generalization and Feature Attribution in Machine Learning Models for Crop Yield and Anomaly Prediction in Germany

A recent study has analyzed the generalization performance and interpretability of machine learning models for predicting crop yield and anomalies in Germany's NUTS-3 regions. The research compares ensemble tree-based models like XGBoost and Random Forest with deep learning approaches such as LSTM and TCN, revealing significant performance degradation on temporally independent validation years despite strong accuracy on conventional test sets.
This development is crucial as it highlights the limitations of current machine learning models in agricultural predictions, particularly their inability to generalize effectively over time. The findings suggest that reliance on seemingly credible SHAP feature importance values may be misleading, raising concerns about the reliability of model interpretability in practical applications.
The study underscores a broader challenge in machine learning regarding generalization and interpretability, echoing issues found in other domains such as active learning and time series forecasting. The persistent vulnerabilities in model performance and evaluation methods point to a need for improved calibration and validation strategies across various applications, including agriculture and beyond.

Generalization and Feature Attribution in Machine Learning Models for Crop Yield and Anomaly Prediction in Germany