Using explainable machine learning techniques to unpack farm-level management x climate interactions

Optimizing the management of maize production systems, including the milpa (intercropping of maize with beans and other species), is crucial for improving on-farm productivity and ultimately reducing food insecurity. This presentation showcases the results of a study aimed at identifying determinants of maize yield in Guatemala using agronomic and climate data. The study employs interpretability techniques in machine learning to explain the interactions between climatic factors and crop management in productivity. The study follows a three-step approach: (1) an Extract, Transform, Load (ETL) process of data, involving feature engineering and data standardization and cleaning; (2) identification of algorithms, metrics, and algorithmic tuning; and (3) delving into interpretability using techniques such as SHAP (SHapley Additive exPlanations), partial dependence plots (PDP), accumulated local effects (ALE) plots, and Friedman's H-statistic to evaluate interactions between features