Geographically and temporally weighted neural network for winter wheat yield prediction

Accurate prediction of crop yield is essential for agricultural trading, market risk management and food security. Although various statistical models and machine learning models have been developed to enhance prediction accuracy, spatial and temporal non-stationarity, an intrinsic attribute of many geographical processes, is still rarely considered in crop yield modeling. From a statistical point of view, this study respectively provided evidence for the existence of spatial non-stationarity and temporal non-stationarity in winter wheat yield prediction based on geographically weighted regression (GWR) and temporally weighted regression (TWR). Then, a geographically and temporally weighted neural network (GTWNN) model was proposed by integrating artificial neural network (ANN) into geographically and temporally weighted regression (GTWR) using publicly available data sources, including satellite imagery and climate data. For a more credible evaluation, the leave-one-year-out strategy was adopted to make out-of-sample prediction resulting in a total of 12 test years from 2008 to 2019. The experiment results showed that the proposed GTWNN outperformed ANN, GTWR and support vector regression (SVR) achieving the average coefficient of determination (R2) values of 0.766, 0.759 and 0.720 at the three prediction times of end of July, end of June and end of May. Moreover, an extended Moran's I was adopted to assess the degree of spatiotemporal autocorrelation of the prediction errors. The error aggregation of GTWNN was lower than other models, indicating that GTWNN is applicable to addressing spatial non-stationarity in modeling the relationship between predictors and yield response. The methodology proposed in this paper can be extended to handle spatiotemporal non-stationarity in other crop yield predictions and even other environmental phenomena.