Probabilistic yield forecasting of robusta coffee at the farm scale using agroclimatic and remote sensing derived indices

Timely and reliable coffee yield forecasts using agroclimatic information are pivotal to the success of agricultural climate risk management throughout the coffee value chain. The capability of statistical models to forecast coffee yields at different lead times during the growing season at the farm scale was assessed. Using data collected during a 10-year period (2008-2017) from 558 farmers across the four major coffee-producing provinces in Vietnam (Dak Lak, Dak Nong, Gia Lai, and Lam Dong), the models were built through a robust statistical modelling approach involving Bayesian and machine learning methods. Overall, coffee yields were estimated with reasonable accuracies across the four study provinces based on agroclimate variables, satellite-derived actual evapotranspiration, and crop and farm management information. Median values of prediction mean absolute percentage error (MAPE) ranged generally from 8% to 13%, and median root mean square errors (RMSE) between 295 kg ha(-1) and 429 kg ha(-1). For forecasts at four to one month before harvest, errors did not vary markedly when comparing the median MAPE and RMSE values. For farms in Dak Lak, Dak Nong, and Lam Dong, the median forecasting MAPE and RMSE varied between 13% and 16% and between 420 kg ha(-1) and 456 kg ha(-1), respectively. Using readily and freely available data, the modelling approach explored in this study appears flexible for an application to a larger number of coffee farms across the Vietnamese coffee-producing regions. Moreover, the study can serve as basis for developing a coffee yield predicting forecasting system that will offer substantial benefits to the entire coffee industry through better supply chain management in coffee-producing countries worldwide.

Automated summary

This study evaluated the accuracy of coffee yield forecasts using statistical models in major coffee-producing provinces in Vietnam. The models performed reasonably well, with stable errors for forecasts leading up to harvest. The flexibility and scalability of the forecasting system suggest potential for application to a larger number of coffee farms in the region.