Improving regional crop yield prediction by incorporating intra-regional phenological variations

Ground crop yield records are typically published only at regional scales to protect farmers’ privacy. To predict regional crop yields, remote sensing and meteorological variables are commonly aggregated (e.g., averaged) to match the spatial scale of ground yield records. However, spatial aggregation of remote sensing data—such as the simple averaging of pixel-wise vegetation index (VI) time series—ignores intra-regional phenological variations. Several vegetation phenology studies unveiled a phenological bias caused by the simple average, which distort the representation of crop-specific phenological characteristics and subsequently compromising the accuracy of crop yield modeling. To address this critical limitation, we therefore proposed a novel yield prediction framework that explicitly incorporates intra-regional phenological heterogeneity. Instead of relying on aggregated VI time series, our method first applies adaptive clustering to automatically group pixels according to their growth dynamics. The resulting subregions and their respective area proportions are then integrated into a customized loss function within a convolutional neural network (CNN)-based modeling framework, termed Group-CNN. We evaluate the performance of Group-CNN in predicting soybean yields across 13 states in the U.S. Corn Belt, comparing it against multiple benchmark models. Results demonstrate that Group-CNN consistently outperforms the benchmarks, achieving a lower root mean square error (RMSE) (5.16 vs. 5.94–6.76 bu/ac), a higher coefficient of determination (R²) (0.83 vs. 0.72–0.76), and a lower mean absolute percentage error (MAPE) (9.78% vs. 11.72–12.81%). Ablation experiments confirm a notable decline in model performance when intra-regional phenological variations are excluded. This study highlights the importance of accurately representing intra-regional crop phenological characteristics—an often-overlooked factor in current yield prediction models.