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

Global agricultural markets are interconnected, with prices driven by aggregate supply. Accurately assessing the impact of weather shocks on production requires globally representative weather data. Recent advancements have yielded high-resolution daily and hourly datasets like GMFD and ERA5-Land. This study aims to evaluate the efficacy of these global datasets in explaining US agricultural yields, comparing them to the more localized PRISM dataset. The research question centers around whether these global datasets can effectively capture the known non-linear relationship between temperature extremes and crop yields, a relationship often masked by aggregated data. The study's importance lies in its implications for understanding the effects of climate change on global food security. Previous research using country-specific data has shown that temperature extremes, particularly heat, significantly affect agricultural yields and that incorporating the full temperature distribution (daily minimum and maximum) yields better predictions than using averages. However, the availability of globally consistent, high-resolution daily weather data has been limited until recently. This research addresses this gap by comparing the performance of readily available global datasets with highly accurate but regionally limited datasets, providing crucial insights for broader climate change impact assessments and informing global policy decisions.

Existing literature highlights the significant influence of temperature extremes, especially extreme heat, on agricultural yields. Studies using fine-scaled weather data for specific countries or regions have demonstrated this non-linear relationship. The use of daily data, rather than monthly or seasonal averages, is crucial for accurately capturing these non-linear effects, as averaging can mask the detrimental effects of extreme heat. However, most global datasets have historically only provided monthly data, hindering global analyses of the impacts of temperature extremes. The development of readily available daily datasets like GMFD and ERA5-Land provides an opportunity to re-evaluate this relationship on a global scale, which is crucial for understanding the dynamics of global food security and pricing, comparative advantage, and trade patterns. The use of these global datasets needs rigorous validation against more localized, higher-resolution datasets to ensure that the results are reliable and consistent.

The study employs three datasets: the fine-scaled US-specific PRISM dataset, and the globally representative GMFD and ERA5-Land datasets. The US analysis uses county-level yield data for corn and soybeans (1950-2019, with a shortened period for GMFD due to data availability). Data are aggregated to the county level, weighted by cropland area. Three functional forms are used to model the non-linear temperature-yield relationship: piecewise linear regression (with a data-driven cross-validation approach to determine the temperature threshold), an 8th-order Chebychev polynomial, and a semi-parametric specification using 3-degree temperature bins. The models control for precipitation, state-specific quadratic time trends (capturing technological change), and county fixed effects (accounting for time-invariant factors). Model performance is evaluated using a cross-validation procedure that calculates the root-mean-squared prediction error (RMSE), comparing the reduction in RMSE relative to a baseline model without weather variables. Climate change projections are generated using uniform warming scenarios (1-4°C) applied to the estimated response functions. For the Sub-Saharan Africa analysis, due to a lack of sub-national yield data, the study utilizes the Enhanced Vegetation Index (EVI) as a proxy for yield. The analysis compares ERA5-Land and GMFD with the CRU dataset, focusing on grid cells with and without nearby weather stations to investigate the impact of data scarcity. The same piecewise linear model is used, and out-of-sample performance is evaluated using RMSE.

The study finds that all three datasets (PRISM, ERA5-Land, and GMFD) produce quantitatively similar response functions for the relationship between temperature and US corn and soybean yields across all functional forms. The relationship is asymmetric, showing yield increases with moderate temperatures and sharp decreases at higher temperatures. However, PRISM models exhibit better out-of-sample predictive power than the global datasets, particularly for soybeans. Despite differences in out-of-sample prediction accuracy, projections of climate change impacts under uniform warming scenarios are remarkably similar across all three datasets when using daily temperature extremes. The effects of extreme heat (36°C) on yields are comparable across datasets, with reductions ranging from approximately 3.2% to 3.8% for corn and 2.7% to 3.5% for soybeans. In contrast, models using a quadratic function of average temperature produce substantially lower projected impacts, highlighting the importance of accounting for daily temperature extremes. The analysis in Sub-Saharan Africa shows that ERA5-Land and GMFD significantly outperform the CRU dataset in predicting crop yields, even in data-sparse regions. The difference in performance between locations with and without weather stations is relatively small, suggesting robustness even in areas with limited weather station data. Figure 1 visually compares the temperature-yield response across the three datasets, while Figure 2 illustrates the out-of-sample model predictions. Figure 3 displays simulated yield losses under various uniform warming scenarios.

The findings demonstrate that even though global gridded datasets have lower predictive skill than fine-scaled datasets in some regions, they can effectively capture the non-linear relationships between temperature extremes and agricultural yields. The consistent results across datasets, particularly those incorporating daily data, highlight the importance of correctly capturing temperature extremes for accurate climate impact assessments. The superior performance of ERA5-Land and GMFD in data-sparse regions suggests their wider applicability for global food security analyses. The discrepancy between models using daily extremes and average temperatures underscores the critical role of daily data for modeling climate change impacts. These results provide strong support for using the global daily datasets for large-scale analyses, even if higher resolution data might be preferable. These global datasets offer a standardized and globally consistent weather product, enabling unified global analyses of climate change impacts on agriculture and facilitating better policy decisions.

This study confirms that newly available global datasets (ERA5-Land and GMFD) accurately capture the non-linear effects of daily temperature extremes on US crop yields, yielding results comparable to those obtained using a high-resolution dataset. The consistent estimates of climate change impacts using daily extremes, across different datasets, demonstrate the value of these data for large-scale assessments. The superior performance of these datasets in data-sparse regions expands their applicability to global studies. Future research could explore the implications of different spatial resolutions and investigate other factors that influence the differences in predictive skill among datasets.

The study acknowledges that differences in predictive skill between the datasets may be influenced by measurement error related to spatial and temporal resolution, as well as interpolation methods. The reliance on EVI as a yield proxy in the Sub-Saharan Africa analysis is a limitation. Also, the uniform warming scenarios used for climate projections are a simplification of actual climate change patterns. Further research incorporating more complex climate scenarios would enhance the robustness of the projections. The study focuses on corn and soybeans in the US and corn in Sub-Saharan Africa, limiting its generalizability to other crops and regions.