Regional heterogeneity modeling for spatiotemporal crop yield forecasting: a hybrid dynamic factor-neural network statistical model

Accurate crop yield forecasting is crucial for global food security and sustainable agricultural development, yet conventional models struggle to jointly address spatial heterogeneity, high-dimensional data, and nonlinear relationships. We propose a spatiotemporal hybrid framework, the Regional Heterogeneity Nonlinear Model (RHNM), integrating Dynamic Factor Analysis (DFA) and neural networks (NN) via decomposing into region-specific linear and cross-regional nonlinear components. The RHNM first employs DFA to construct a linear component with region-specific fixed effects for each geographical unit, explicitly capturing spatial heterogeneity. Subsequently, the residuals are collectively modeled by a cross-regionally shared NN to capture complex nonlinear patterns. Theoretically, we establish asymptotic statistical properties for this hybrid estimator to support valid inference. Empirically, using 35 years of grid-level data from China’s major rice-producing regions, encompassing yield, meteorological, and soil variables, our framework reduces prediction errors by 53.83–64.69% compared to conventional linear and standalone machine learning benchmarks, while exhibiting superior predictive stability across heterogeneous regions. Through theoretical analysis and empirical validation, this study offers a modeling framework that effectively balances interpretability with accuracy, thereby providing both a reliable data-driven tool for agricultural decision-making and a generalizable methodological framework for analyzing high-dimensional spatiotemporal data with pervasive regional heterogeneity.