Identifying the most influential climate predictors for crop yield using advanced statistical and machine learning models: A case study for mango crop in India

Accurately identifying the most influential predictors for crop yield is critical for the development of risk management tools. This study utilised advanced statistical techniques, namely stepwise selection in the d -vine copula regression and Mean Decrease Accuracy (MDA) in the random forest model, to determine the key climate variables influencing mango yield across 31 districts of Tamil Nadu state, India, between 2008 and 2019. Monthly climate and derived variables, corresponding to the mango growing season (January – June), were extracted from TerraClimate datasets with a spatial resolution of approximately 4 km using the district boundaries. The d -vine copula regression model employed a stepwise selection process to identify the most parsimonious set of predictors, while the RF model used MDA to rank variable importance. Although variations in predictor rankings were observed between the two models, consistent variables, such as Palmer Drought Severity Index ( PDSI ), minimum temperature ( tmin ), and actual evapotranspiration ( aet ), were identified across districts and phenological stages. Notably, PDSI emerged as a key predictor during all mango phenological stages, particularly during flowering and maturation. These findings provide valuable insights for improving climate resilience in mango production, offering a robust framework for the development of risk management tools such as yield forecasting and index-based insurance. The proposed methodology has broader applications in optimising agricultural practices under changing climate conditions.