Data fusion using climatology and seasonal climate forecasts improves estimates of Australian national wheat yields

• Published in: Agricultural and forest meteorology Vol. 320; p. 108932
• Main Authors: Mitchell, Patrick J., Waldner, François, Horan, Heidi, Brown, Jaclyn N., Hochman, Zvi
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.06.2022
• Subjects: Data fusion; Forecast verification; Seasonal climate forecasts; Wheat yield; Yield forecasting; Forecast verification; Data fusion; Yield forecasting; Wheat yield; Seasonal climate forecasts
• ISSN: 0168-1923; 1873-2240
• DOI: 10.1016/j.agrformet.2022.108932

•Yield forecasts based on climatology, model and data fusion methods were assessed.•Climatology and climate model forecasts were biased in opposite directions.•Forecast bias was reconciled using a data fusion approach applied to these forecasts. National and regional yield forecasts can provide important insights into agribusiness beyond the farm gate. The incorporation of dynamical climate models into these forecasting systems strengthen their predictive performance in many cases but may contribute inherent biases to the final yield estimates. Downscaling the native climate model output so that is suitable for crop simulation modelling can also present challenges in representing realistic conditions for plant growth from a climate model. This study evaluated the performance of an operational national wheat yield forecast system for the Australian wheatbelt using climatology and seasonal climate model-based input data, and introduces an alternative approach using a data fusion method. The crop forecasting system uses the APSIM wheat model to estimate water-limited potential yield. The climatology-based forecast tended to over predict national yield (high yield bias; 1.5 to 7% across forecast months), while the model-based method (using ACCESS-S1 dynamical model) tended to under predict yield (low yield bias; -5.9 to -0.5% across forecast months) and had a lower spread than climatology (10 to 50% lower across forecast months). The model-based forecast had skill in terms of accuracy and reliability during the second half of the season. The newly developed data fusion method used a weighting method calibrated for separate forecast locations (stations) to remove bias in the mean forecast yield and reduce ensemble spread. This resulted in improvements in the Australia-wide yield forecasts across all forecast start dates. This study provides a demonstration of how a data-driven approach can be applied to a crop forecast to improve accuracy and resolution of crop yield forecasts without the need for more computationally intensive downscaling approaches.