A neural network method for the reconstruction of winter wheat yield series based on spatio-temporal heterogeneity

Abstract

Crop growth conditions and meteorological environments are observed and recorded by agro-meteorological stations, which, however, may fail to record crop yield data in some specific years. In this context, incomplete yield series data constrain their application and result in inconvenience in information mining. Accordingly, this study improves the existing spatio-temporal interpolation method and succeeds in interpolating wheat yield data observed and recorded by 56 agro-meteorological stations on the Huang-Huai-Hai Plain of China. In this study, pre-interpolation is first implemented to improve the completion rate of interpolation and eliminate the effect of absent neighboring values on the positions to be interpolated. Then, data reconstruction is performed in spatial and temporal dimensions based on spatio-temporal heterogeneity. Finally, the reconstructed results are combined with the back propagation (BP) neural network model for spatio-temporal integration. Moreover, this study analyzes the settings of key parameters and compares them with traditional interpolation methods. Corresponding results demonstrate the superiority of the proposed method in this study over traditional interpolation methods in terms of interpolation precision and the completion rate. Meanwhile, individual interpolation precision in each step of the proposed method is effectively enhanced.