Pre-season prediction of regional rainfed wheat yield in Northern Greece with CERES-Wheat

The present study aims at forecasting hard wheat (Triticum turgidum L. var. durum) yield in northern Greece, a season prior to harvest. It is based on (a) crop simulated, with CERES-Wheat indicators at four planting dates and (b) reported crop yields at two regional levels (three NUTS2 [Nomenclature of Units for Territorial Statistics] and 16 NUTS3 regions), for the years 1979–2006. Principal component analysis (PCA) was applied to explore major patterns of joint variability in 20 crop simulated agroclimatic indicators of the growing season before harvest. Stepwise regression and hindcast were employed for the selection of the modes identified by PCA as predictors in multivariate linear regression models used for forecasting yield a season ahead of harvest. Forecasting skill varied to a large extent by spatial scale and planting date. When the simulation results aggregated to the larger spatial level (NUTS2), the yield forecasting skill, in terms of R2, was rated as high (ranging from 0.48 to 0.73) in three out of four planting dates for Central Macedonia and in one planting (R2 = 0.57) for Thrace. Harvest index, nitrogen leaching and related soil water crop simulated output of the previous season, were the most important predictors. No forecasting skill was found in the third NUTS2 region. The performance of the regression models substantially deteriorated at the higher resolution spatial level (NUTS3). In four regions only (including the one where CERES-Wheat was calibrated) yield forecasting skill was moderate (R2 > 0.25). The results demonstrate the potential of this approach for regional crop yield forecasting before the beginning of the cropping season. However, crop model calibration is required before its application.