Using Machine Learning and Earth Observation to Forecast Crop Yields in Kenya

Low crop productivity and weather variability contribute to food insecurity in Eastern African countries. Knowing what end-of-season crop yields will be early in the season is critical for decision-makers to mitigate impacts. Recent advances, unprecedented access to Earth Observation (EO) satellites, and machine learning technology have provided new opportunities to develop better methods of forecasting crop yields. In this study, machine learning models were applied to forecast end-of-season maize crop yields in Kenya using agrometeorological EO data from 2002 to 2016. Input data included: precipitation totals, temperature, evaporative stress index, soil moisture, and Normalized Difference Vegetative Index. Ridge Regression, Random Forest, and Mixed Effect Random Forest (MERF) were evaluated. Input data was processed to favor prediction on lower test yields, i.e., the years when crop failure occurred. After the optimization of feature calculation and selection from the various EO products, results showed RMSE to be as low as 0.2 and 0.3 for the Random Forest and MERF model, respectively. A consistent range of R2 values between 0.75 - 0.95 suggests that these models hold promise for providing early warning predictions. Analysis of feature contribution showed that total seasonal precipitation had the highest feature importance across the MERF and Random Forest models. Mid-range percentiles of NDVI and daily minimum temperature also had strong contributions. Future work will include building the model at lower administrative levels and incorporating data on field management practices as features in the model. The predictions produced by these models can be used to inform decision-makers when crop yields are impacted negatively by adverse weather conditions.