Statistical spring wheat yield forecasting for the Canadian prairie provinces

A study to forecast regional spring wheat (Triticum aestivum L.) yields on the Canadian Prairies was conducted, based on simulated daily water use and soil water contents derived from the National Drought Model. Empirical linear regression models were calibrated from 1976 to 2006 spring wheat yield data for this purpose. Potential predictors assessed were mainly those indicators related to water stress conditions at different crop growth stages. Stepwise regression and cross-validation were employed for the selection of the predictors in multivariate linear regression models used for forecasting spring wheat yields from seeding to harvest. The cross-validated “forecasts” for 1976–2006, using data up to harvest, explained 77%, 64%, 63% and 70% of yield variances, respectively, for Alberta, Saskatchewan, Manitoba and the entire Prairie region. Root mean squared error of the “forecasts” ranged from 8% to 11% of the average yields. The prediction accuracy earlier in the season was often lower than later in the season. Usable prediction accuracy was found by the middle of the growing season (around heading or anthesis), but only marginally effective at seeding time, especially so for Saskatchewan.