Non-linear relationships between daily temperature extremes and US agricultural yields uncovered by global gridded meteorological datasets

Dylan Hogan and Wolfram Schlenker 

Department of Energy, Office of Science, Earth & Environmental Systems Modeling Program Acknowledged Support: yes

DOI:https://doi.org/10.1038/s41467-024-48388-w

Abstract

Global agricultural commodity markets are highly integrated among major producers. Prices are driven by aggregate supply rather than what happens in individual countries in isolation. Estimating the effects of weather-induced shocks on production, trade patterns and prices hence requires a globally representative weather data set. Recently, two data sets that provide daily or hourly records, GMFD and ERA5-Land, became available. Starting with the US, a data rich region, we formally test whether these global data sets are as good as more fine-scaled country-specific data in explaining yields and whether they estimate similar response functions. While GMFD and ERA5-Land have lower predictive skill for US corn and soybeans yields than the fine-scaled PRISM data, they still correctly uncover the underlying non-linear temperature relationship. All specifications using daily temperature extremes under any of the weather data sets outperform models that use a quadratic in average temperature. Correctly capturing the effect of daily extremes has a larger effect than the choice of weather data. In a second step, focusing on Sub Saharan Africa, a data sparse region, we confirm that GMFD and ERA5-Land have superior predictive power to CRU, a global weather data set previously employed for modeling climate effects in the region.

Caption:Figure compares out-of-sample prediction for piecewise linear regression models estimated using weather observations from PRISM (red), ERA5-Land (blue), and GMFD (yellow) data sets. Box shading indicates the functional form of temperature, including piecewise linear, 8th order polynomial, 3-degree bins, and growing season average temperature. The vertical axis shows the percent reduction in root-mean-squared error (RMS) relative to a baseline model that excludes all weather variables. A value of 0 zero implies the weather variable cannot explain any of the year-to-year variation in yields around the trend, while a value of 100 implies that the weather can explain the entirety. For each data set, response functions are estimated 1000 times, each time randomly sampling 85% of the years from the full panel. RMS is calculated based on each model’s prediction of the remaining 15% of years. Boxes, horizontal lines, and points represent the interquartile range, median, and mean of RMS reductions from the 1000 draws, respectively. Results are provided for corn (left panel) and soybeans (right panel) yields, and climate data sets are indicated by color, while the shading of a color refers to the functional form of the temperature variables. All specification include a quadratic in season-total precipitation.

Leave a Reply