Estimation of Winter Wheat Yield From Sentinel-1A Time-Series Data Using a Dynamic Graph Convolutional Network-Transformer Hybrid Model
Abstract
Timely and accurate forecasting of winter wheat yield is crucial for ensuring national food security. Deep learning methods have been widely applied in crop yield forecasting. However, traditional deep learning models fall short in capturing spatial clustering and dependencies on key phenological stages, which hold significant potential for yield prediction. To address this issue, a novel hybrid deep learning architecture combining a dynamic graph convolutional network (DGCN) and Transformer architecture, named DGCN-Transformer, was proposed for winter wheat yield estimation using synthetic aperture radar (SAR) time-series data. The DGCN was used to capture dynamic clustering relationships among crops at different growth stages, and the multihead self-attention mechanism of the Transformer was used to capture the dependencies of winter wheat at each stage of its growth cycle, thereby estimating winter wheat yield. Experimental results showed that the DGCN-Transformer architecture achieved the highest accuracy in winter wheat yield estimation, achieving a coefficient of determination (R2) of 0.72, a root mean square error (RMSE) of 776.518 kg/ha, and a mean absolute error (MAE) of 615.800 kg/ha, thereby verifying its superiority in winter wheat yield prediction. Furthermore, through cumulative effects, the model accurately identified the key growth stages of winter wheat.