Quantifying the impacts of pre-occurred ENSO signals on wheat yield variation using machine learning in Australia

Australia is one of the top wheat exporting countries in the world and the reliable prediction of wheat production plays a key role in ensuring regional and global food security. However, wheat yield in Australia is highly exposed to the impacts of climate variability, especially seasonal rainfall, as wheat is mostly grown in the drylands. Previous studies showed that El Niño Southern Oscillation (ENSO) has a strong influence on Australia's climate and found the ENSO-related phenomena have prognostic features for future climatic conditions. Therefore, we examined the predictability of state-scale variation in Australian wheat yields based on ENSO-related large-scale climate precursors using machine learning techniques. Here, we firstly established a set of random forest (RF, a machine learning method) models based on pre-occurred climate indices to forecast spring rainfall for the four major wheat producing states of Australia, the forecasted rainfall was then combined with selected precedent climate drivers to predict yield variations using another set of RF models for each state. We explored the most influential variables in determining spring rainfall and yield variation. We found that the first set of RF models accounted for 43-59% of the change in spring rainfall across the four states. By incorporating forecasted spring rainfall with selected ENSO climate indices, the RF model accounted for 33-66% of the variation in yield which was greater than the 22-50% of yield variations explained by ENSO-related indices alone. The results suggest that wheat yield variation at a state level could be reliably forecasted at lead-times of three months prior to the commencement of harvest. We also found that forecasted spring rainfall and precedent Southern Oscillation Index (SOI) in July were the most important factors in estimation of crop yield in the winter dominant rainfall states. ENSO climate indices are easy to obtain and can be rapidly used to drive the forecasting model. Therefore, we believe the proposed models for predicting wheat yield variations at three-month lead time would be helpful for state governments and policy makers to develop effective planning to reduce monetary loss and ensure food security