Efficient stabilization of crop yield prediction in the Canadian Prairies

This paper describes how spatial dependence can be incorporated into statistical models for crop yield along with the dangers of ignoring it. In particular, approaches that ignore this dependence suffer in their ability to capture (and predict) the underlying phenomena. By judiciously selecting biophysically based explanatory variables and using spatially-determined prior probability distributions, a Bayesian model for crop yield is created that not only allows for increased modelling flexibility but also for improved prediction over existing least-squares methods. The model is focused on providing efficient predictions which stabilize the effects of noisy data. Prior distributions are developed to accommodate the spatial non-stationarity arising from distinct between-region differences in agricultural policy and practice. In addition, a range of possible dimension–reduction schemes and basis expansions are examined in the pursuit of improved prediction. As a result, the model developed has improved prediction performance relative to existing models, and allows for straightforward interpretation of climatic effects on the model's output.

► We demonstrate improved performance of unified crop models over regional models. ► We examine the use of various sparsity-inducing bases for crop yield prediction. ► We test the inclusion of spatial dependence information in a crop yield model. ► We illustrate the improved prediction from a Bayesian spatial crop model. ► We develop a method for allowing for between-region (province) discrepancies.