The ability of simple statistical models to accurately capture changes under climate change: a case study in crop yields

Crop yield responses under climate change are often projected using empirical models, which extrapolate historical yield-weather relationships to future conditions. However, it is unclear whether such historically trained empirical models accurately capture crop biophysical processes and predict long-run climate change effects.

In this interdisciplinary project, we test this accuracy for commonly-used empirical crop models using simulated-data experiments. By leveraging yields simulated by process-based crop models, we train an empirical model on historical yield simulations and then evaluate whether the model can accurately emulate the simulated target yield responses under climate change scenarios.

We find that empirical models, despite a high goodness-of-fit within the historical climate, are overly sensitive to severe heat increases and consequently overproject target yield losses under high-warming scenarios by a factor of two. An empirical model trained not only on historical yields, but also on yields across all warming scenarios, accurately emulates the target yield response, indicating that projection bias in the empirical model is primarily due to the interannual yield-temperature relationship in the historical climate not matching the yield-temperature relationship across climate shifts.

This work cautions against projecting future yields based on historical interannual yield-weather relationships and demonstrates the promise of increased collaboration between process-based and empirical modeling groups to improve the robustness of future yield projections and adaptation planning.

This project has been done in collaboration with Haynes Stephens, María D. Hernández Limón, Harshil Sahai, James A. Franke, Christoph Müller, Jonas Jägermeyr Jonathan Proctor, Alex C. Ruane, and Elisabeth J. Moyer. Please look for our paper coming out soon!