Explainable machine learning reveals diverse yield-determining factors among Thai rice farmer cohorts: Implications for targeted agricultural support

Abstract

Rice yield prediction and optimization remain crucial challenges in Thailand’s agricultural sector. This study presents an explainable machine learning framework for predicting farm-level rice yields and identifying key factors affecting productivity. We collected comprehensive data from 1,722 smallholder farmers in central Thailand, encompassing 58 agronomic and economic variables. Four automated machine learning (AutoML) frameworks – AutoGluon, auto-sklearn, h2o, and mljar – were evaluated using 5-fold cross-validation, with AutoGluon achieving the best performance (root mean square error: 0.532 tonnes/hectare, mean absolute error: 0.372 tonnes/hectare, R²: 0.538). Using global SHapley Additive exPlanations (SHAP) analysis, we identified farmers’ social networks, rental costs during harvest, and total harvesting expenses as the most influential predictors of rice yields. Notably, stronger social network connectivity was associated with higher yields, suggesting that information sharing and collective knowledge within farming communities play a key role in improving productivity. Clustering analysis based on individual SHAP values revealed six distinct farmer cohorts, each characterized by unique patterns of feature importance. These cohort-specific insights demonstrate the potential of combining AutoML with explainability techniques to move beyond uniform agricultural recommendations towards precision support tailored to the specific needs of different farmer cohorts.