Methods for Spatial Prediction of Crop Yield Potential

Opportunities for and constraints to crop production can be assessed with crop growth simulation models. Most crop simulation models require daily weather data as input but such data are generally not available at a high spatial resolution. Several approaches have been developed to estimate yield potential (Yp) at locations without daily weather data (weather stations) but these have not been compared. We used two crop simulation models (WOFOST and LINTUL) to compute Yp for two crops for the entire world. A global weather database was divided into 856 training and 12,808 testing sites. We predicted Yp at the testing sites by using five main methods (eight methods if one considers within-method variants): (i) weather interpolation followed by simulation; (ii) nearest neighbor interpolation; (iii) thin plate spline interpolation, either with or without covariates; (iv) Random Forest-based metamodels with either climatic or bioclimatic variables; and (v) weather generation from either climate data or interpolated climate data, followed by simulation. The metamodel with bioclimatic variables performed best [average root mean square error (RMSE) = 667 +/- 111 kg ha(-1)], followed by weather generation from climate data, weather interpolation, and spatial interpolation of yield with climatic covariables. The most commonly used method, nearest neighbor interpolation, performed worst (RMSE = 1763 +/- 472 kg ha(-1)). The optimal method for a particular study will depend on the simulation model, the region, weather station density, and other variables but these results suggest that for estimating Yp, alternatives to nearest neighbor interpolation should be considered.