Effects of Climate on Agriculture Unclarified by Data Downscaling

Under the specter of a warmer future, scientists must study the downstream effects of climate change on humans, including the impact on agriculture, the economy, and society. But the scale of global climate models and regional models of agriculture, hydrology, and other sectors may be orders of magnitude apart, forcing researchers to find novel methods of closing that gap.

In a new paper published in Proceedings of the National Academy of Science, a team of researchers from the Computation Institute’s Center for Robust Decision-making on Climate and Energy Policy (RDCEP) find that one common solution to this scaling problem in climate and agriculture fails to improve forecasts. 

“It is important for climate scientists to contribute to impacts studies because by now the Earth is committed to some significant level of climate change," said Elisabeth Moyer, assistant professor of geophysical sciences at the University of Chicago and senior author of the paper. "Not all effects of past CO2 emissions are yet realized and there is no prospect of immediate substantial reductions of future emissions. But our desire to help has to be balanced by checking to ensure that any model projections would lend value for the purposes they’d be put to.” 

Many global climate models project changes in temperature, precipitation, and other measures at a resolution of hundreds of kilometers -- imagine the Earth covered in a grid of 100-kilometer boxes. Popular agricultural models, on the other hand, operate at far smaller scales due to the sensitivity of plants to subtle, local changes in soil and weather. 

This tenfold or more difference of resolution creates challenges for researchers who want to use climate and agricultural models to estimate the impact of climate change on crop yields and food security. One solution for the mismatch of scales is termed “dynamical downscaling,” a method of using low-resolution global climate model outputs to drive higher-resolution regional climate models, the output from which can then be used to drive agricultural projections. 

The recently completed North American Regional Climate Change Assessment Program (NARCCAP) used dynamical downscaling in a multi-year project to generate these higher-resolution climate projections for North America, and a new ongoing international project plans to create dynamically downscaled projections for the rest of the world. But despite the high cost and computational demands of these efforts, how downscaling affects the results of agricultural models had only been evaluated in relatively limited cases and locations. 

“We don't necessarily know whether this is improving our ability to make future weather or climate projections, but we do know it’s costing quite a bit of money, and we're pushing forward on even bigger projects on the global scale,” said lead author Michael Glotter, a graduate student in geophysical sciences at the University of Chicago. “We thought, why don't we take a step back and see if this method is actually useful for agricultural impact assessments at national scales.” 

To test this question, the research team -- which also included Joshua Elliott, Neil Best, and Ian Foster of the Computation Institute, and David McInerney of the Department of the Geophysical Sciences at the University of Chicago (now at the University of Adelaide) -- compared the projections of a maize yield model in the United States when driven by climate data generated directly by a global model or by a regional model using downscaling. 

While the downscaled data created more accurate projections for some parts of the country, such as coastal regions and areas near mountain ranges, no significant differences between the projections were seen in yield projections for the Midwest, where the bulk of US corn agriculture occurs. Applying a statistical method to the climate data known as a “bias correction” improved the regional model’s performance in those Midwestern regions, but also removed the advantage of the regional model in other areas. Indeed, the projections driven by global and regional model data were “essentially indistinguishable” from each other after the correction, the authors wrote. 

The results suggest that low-resolution differences in temperature and precipitation from the GCM baselines – which are inherited by higher-resolution regional models – significantly impact the assessment of agricultural impacts. After correction for these differences, dynamical downscaling may not provide enough of an improvement in useful accuracy to justify the heavy computational expenses used to generate the higher-resolution climate projections. "As many new projects come online in the next few years to study the effects of climate at smaller and smaller scales, the temptation is high to push climate models to higher and higher resolutions through methods such as dynamical downscaling," said Joshua Elliott, research scientist and fellow at the Computation Institute and co-author of the paper. "This paper shows that researchers should take care before choosing which data products are necessary and sufficient for a given purpose." 

Based on these findings, the authors suggest that future research should concentrate on alternative methods for increasing the resolution of climate model projections for use in agricultural projections, or on improving the global climate models themselves. 

"With the challenge of impending climate change and growing impacts, an all-in approach is necessary to generate the tools required by scientists and stakeholders to plan for a changing world," said Glotter. "But given limited resources for research investment, we must continue to evaluate new and old methods, models, and datasets in a variety of applications, and use these assessments to help inform future research and investment decisions."

Article written by Rob Mitchum of the Computation Institute.