Improving maize yield prediction at the county level from 2002 to 2015 in China using a novel deep learning approach

Abstract

Accurate and timely maize yield monitoring from satellite imagery is in great demand in developing countries. The spatial heterogeneity deprived of the large territory of China makes it a challenge. In this article, we developed a novel deep learning model for maize yield prediction at the county level based on multiple satellite data. The two-stage feature learning structure integrated data from disparate sources, enhancing feature representation. The discriminative features were optimized from two levels: the dictionary matrix learned the spatial diversity of the provinces, and the improved optimizing formulation fit the distribution of the unbalanced records. The cross-validation results showed that our approach could explain 82 % of the variation in maize yield, achieving state-of-the-art. The model was robust when predicting the future, with the average root-mean-square error of 1006 kg/ha and the mean-absolute-percentage error of 17.1 %. The ability of early maize yield prediction clarifies the tremendous application value, showing the data from the first two months can already explain 75.6 % of yield variation. It was the first effort to improve county-level maize yield prediction in China, providing a potential framework for advancing the use of multi-source datasets for maize yield estimating.