Moving climate seasonal forecasts information from useful to usable for early within-season predictions of durum wheat yield

Abstract

Crop models fed by seasonal forecasts to deliver yield forecasts are becoming a valuable tool to tackle food production shocks and food price spikes triggered by climate change and extreme events. However, to what extent seasonal climate prediction can be coupled to crop models in an effective climate service to provide durum wheat yield forecasting at local, regional and national scales remains unknown. Within the H2020 MED-GOLD project, based on participatory action research and through a co-designing approach, a complete assessment of raw and bias-adjusted ECMWF-System5 seasonal forecasts data in feeding the Delphi model was performed and explicitly compared with a previous version of the Delphi model fed by a historical scenario, both benchmarked with yield observation data. The 5 and 3-months lead prediction accuracies of year-to-year variations in durum wheat yield in Italy for 31 crop years were explored and evaluated for which techniques to remove biases from input data to adopt and for which areas they performed better. The raw seasonal forecast (BC0) showed better yield forecasting correlation than the historical scenarios (0.42 vs 0.17) in southern Italy, but not better accuracy (53 % for both), while a first bias-adjusted seasonal forecast (BC1) had again a positive effect in terms of correlation (r = 0.43) especially in southern Italy and generally in terms of accuracy (59 %). Finally, by the implementation of a second bias-adjusted seasonal forecast (BC2) the best yield prediction system was reached, especially for the areas of central and southern Italy with a total accuracy of 62 % and r = 0.52.

Keywords

• Climate services
• Seasonal forecast
• Durum wheat
• Climate data store
• Delphi model
• Bias correction