Wheat yield forecasts with seasonal climate models and long short-term memory networks

Abstract

The potential of seasonal climate forecasts (SCFs) within machine learning models to forecast crop yields remains unexplored. We propose a workflow for integrating SCF data into a long short-term memory (LSTM) network to forecast wheat yield at the county level across the Great Plains in the United States. Each month, past predictors were filled with observations and future weather predictors were forecasted using the seasonal climate model of the European Centre for Medium-Range Weather Forecasts (SCF approach). This approach was benchmarked with the truncate approach that only used observed predictors. Using all observed predictors at harvest, the model achieved an R² of 0.46, an NRMSE of 0.24, and an MSE of 0.46 t/ha on the test set. The SCF approach and truncate approach performed poorly from January to March. The SCF approach outperformed the truncate approach in April and May. At the beginning of May, three months before harvest, the SCF approach achieved an MSE of 0.6 t/ha, improving the truncate approach by 10 %. In June, the SCF approach further improved but did not outperform the truncate approach. Predictor importance analysis revealed the critical role of SCF data at the beginning of May for the latter half of May. This study suggests that weather forecasts issued at the right time, when both crop development and forecast skill align, could be as short as 16 days and still significantly improve the accuracy of sub-national wheat yield forecasts over other approaches.