Maize yield forecasting by linear regression and artificial neural networks in Jilin, China

Abstract

Forecasting the maize yield of China's Jilin province from 1962 to 2004, with climate conditions and fertilizer as predictors, was investigated using multiple linear regression (MLR) and non-linear artificial neural network (ANN) models. Yield was set to be a function of precipitation from July to August, precipitation in September and the amount of fertilizer used. Fertilizer emerged as the dominant predictor and was non-linearly related to yield in the ANN model. Given the difficulty of acquiring fertilizer data for maize, the current study was also tested using the previous year's yield in the place of fertilizer data. Forecast skill scores computed under both cross-validation and retroactive validation showed ANN models to significantly outperform MLR and persistence (i.e. forecast yield is identical to last year's observed yield). As the data were non-stationary, cross-validation was found to be less reliable than retroactive validation in assessing the forecast skill.