Random Forests for Regional Crop Yield Predictions.

Accurate prediction of crop yield is critical for developing effective agricultural and food policies at regional or global scale. Multiple linear regression methods that are widely used for crop yield modeling have limitations especially when applied to predict crop yield over a large spatial scale involving complex interactive effects of climate and soil variables. We evaluated Random Forests (RF) regression as an alternative approach for predicting crop yield responses to climate and soil variables at regional scales in wheat, maize, and potato in comparison with multiple linear regressions (MLR). Crop yield data from two regions, mega-environment 6 for wheat and eastern seaboard region for maize and potato, were used for model training and testing. In all test cases, RF regression vastly outperformed MLR in predicting crop yield (wheat RF: pseudo R² = 0.93, RMSE = 0.218 ton/ha, wheat MLR: R² = 0.62, RMSE = 0.667 ton/ha, potato RF: pseudo R² = 0.78 ~ 0.94, RMSE = 1.689 ton/ha, potato MLR: R² = 0.33, RMSE = 4.636 ton/ha, maize RF: pseudo R² = 0.80 ~ 0.94, RMSE = 1.349 ton/ha, maize MLR: R² = 0.36, RMSE = 3.626 ton/ha). Our results suggest that RF regression is a powerful and versatile tool for crop yield predictions at regional scale for its high accuracy and precision, ease of use, and utility in data analysis. RF regression may result in a loss of accuracy when predicting the extreme ends or responses beyond the boundaries of the training data.