What can crop modelers learn from machine learning models about corn, sorghum and soybean?

Overview & Synopsis

An ingenious statistical analysis by Schlenker and Roberts (2009) of the county-level grain yield of cotton, corn and soybean in response to climate showed that these non-controlled experiments contain valuable and somewhat hidden information. Critically, these authors identified a temperature range over which grain yields increase (≈10 to 29°C for corn for corn) after which grain yields decrease sharply. Analogous analyses have been presented by Lobell et al. (2014) and Hoffman et al. (2017). Our goal was to apply machine learning (ML) tools to data panels like those used by Schlenker and Roberts (2009) to reveal the relationship between grain yield and climate variables for specific phases of the crop cycle. We analyzed the rainfed yield of corn, sorghum and soybean in response to climate in the US using Random Forest (RF), a non-parametric machine learning (ML) tool (Breiman, 2001). Unlike other ML tools, RF allows gleaning the functional form between predicted and predictor variables.