The response of maize, sorghum, and soybean yield to growing-phase climate revealed with machine learning

Abstract

Accurate representation of crop responses to climate is critically important to understand impacts of climate change and variability in food systems. We use Random Forest (RF), a diagnostic machine learning tool, to explore the dependence of yield on climate and technology for maize, sorghum and soybean in the US plains. We analyze the period from 1980 to 2016 and use a panel of county yields and climate variables for the crop-specific developmental phases: establishment, critical window (yield potential definition) and grain filling. The RF models accounted for between 71% to 86% of the yield variance. Technology, evaluated through the time variable, accounted for approximately 20% of the yield variance and indicates that yields have steadily increased. Responses to climate confirm prior findings revealing threshold-like responses to high temperature (yield decrease sharply when maximum temperature exceed 29 C and 30 C for maize and soybean), and reveal a higher temperature tolerance for sorghum, whose yield decreases gradually as maximum temperature exceeds 32.5 C. We found that sorghum and soybean responded positively to increases in cool minimum temperatures. Maize yield exhibited a unique and negative response to low atmospheric humidity during the critical phase that encompasses flowering, as well as a strong sensitivity to extreme temperature exposure. Using maize as a benchmark, we estimate that if warming continues unabated through the first half of the 21st century, the best climatic conditions for rainfed maize and soybean production may shift from Iowa and Illinois to Minnesota and the Dakotas with possible modulation by soil productivity.