A deep learning approach to county-level soybean yield forecasting using large-scale environmental data

Agricultural systems continue to be disrupted by climate change, making crop yield forecasting a critical component towards informed agricultural planning. However, current yield forecasting approaches often lack in scale and are hard to interpret agronomically within real world scenarios. To fill the gap, this study developed an interpretable soybean yield forecasting framework using quantifiable county-level climate and soil attributes at scale for early, accurate yield predictions months ahead of harvest.

Here, 388 climate, soil, and elevation attributes were collected from a ~100-year time span over the continental US. After preprocessing, Shapley Additive Explanation (SHAP) values were computed for each factor to aid in the interpretability of forecasts within the framework. Using the combined climate factors, a deep learning neural network was tuned and constructed to produce yield forecasts in May, June, July, and August.

On a hold-out test set, these forecasts produced robust predictions, indicating the framework can reliably predict yields up to five months ahead of harvest. This study also indicated the importance of several environmental attributes towards yield forecasts; specifically, extreme temperatures during later summer months (DX90, EMXT) were found to have outsized negative contributions towards soybean yields, while timing of heat was also identified as a potential driver of yield (DYXT).

Overall, these findings indicate that our interpretable DNN model is capable of providing reliable, in-season soybean yield forecasts as early as May, offering a robust tool for agricultural stakeholders to adjust management strategies well before harvest.