Simplifying crop models into statistical emulators of varying complexity

Process based crop simulation models can be used to anticipate how crop yields behave under different weather conditions which makes them useful for assessing the implications for yields of weather variability, uncertainty, and climate change. However, they can be cumbersome to use or embed in other models. Hence, developing statistical models that emulate the full models can provide a lighter-weight means of harnessing their power for rapid or high-volume assessments. We develop emulators for several major crops. Climate and fertilizer rates are matched to the yields simulated by a full crop model to provide a dataset covering the entire land surface of the globe. We use neural network specifications to be able to easily try out different levels of complexity. The final selection of the emulator models is a balance between complexity and performance. The large amount of data meant that the emulators could support complicated specifications. We ultimately settled on models with approximately 200 parameters. Neither rainfed nor irrigated conditions were consistently better performing. In general, the r-squared values of the estimated emulators ranged from 0.85 to 0.95. No crops appeared significantly easier or more difficult to model than the others. The emulators we developed provide a simple substitute for more complicated crop models when a high level of detail is not needed. The approach can be easily updated with new data or be applied to more specific circumstances if necessary by creating custom datasets that emphasize the particular conditions anticipated to be most important.