Linking dynamic seasonal climate forecasts with crop simulation for maize yield prediction in semi-arid Kenya

By providing information about growing season characteristics in advance of the season, predictions of climate fluctuations at a seasonal time-scale offer opportunity to improve agricultural risk management, but only if forecasts are translated into probabilistic forecasts of production and economic outcomes of management alternatives. A mismatch between the spatial and temporal scale of dynamic climate models and process-level crop simulation models must be addressed if crop models are to contribute to the task. Methods proposed for linking crop models with dynamic seasonal climate forecasts include classification and selection of historic analogs, stochastic disaggregation, direct statistical prediction, probability-weighted historic analogs, and use of corrected daily climate model output. For a semi-arid location in Kenya, we demonstrate and evaluate methods to predict field-scale maize yields, simulated by CERES-maize with observed daily weather inputs, in response to downscaled seasonal rainfall hindcasts available prior to planting, derived from an atmospheric general circulation model, ECHAM. The methods we considered were statistical prediction by non-linear regression, probability-weighted historic analogs and stochastic disaggregation to predict field-scale maize yields simulated by CERES-maize with observed daily weather inputs. Downscaled ECHAM output predicted 36% of the variance of total precipitation and 54% of the variance of rainfall frequency in October–December at the site. non-linear regression showed the lowest, and stochastic disaggregation showed the highest overall error. All of the yield forecasting methods showed similar random error, predicting from 28 to 33% of the variance of yields simulated with observed weather. Incorporating the predictability of rainfall frequency into the stochastic disaggregation procedure did not improve yield predictions. Based on this study, stochastic disaggregation, direct statistical prediction and probability-weighted historic analogs all show potential for translating seasonal climate forecasts into predictions of crop response.