A Framework for Robust and Interpretable Yield-Risk Classification Using Country–Year–Crop Data (X-AMLF)

This study introduces the Explainable Agro-Intelligence Machine Learning Framework (X-AMLF), which is a structured evaluation framework for yield risk classification in agriculture using aggregated country-year-crop data. A harmonised dataset was created by combining data on crop yield, annual average temperature, rainfall, pesticide use, crop identity, and country indicators to create 28,242 complete observations on the yield of 101 countries (1990–2016) for 10 major crops. Multiple classification models were evaluated, such as linear baselines and tree-based ensembles, and hyperparameters were optimised with Optuna to investigate variations in prediction and error behaviour. Under random splitting, ensemble models were able to get very high scores (e.g. Random forest F1 ≈ 0.976), but under a stricter evaluation, performance dropped. With country-holdout validation, the best-performing models had F1 ≈ 0.843, attenuation of performance when tested on unseen regions.