Abstract
Purpose Accurate and timely prediction of wheat yield is crucial to apply precision agriculture management and ensure food security. Since large-scale regional studies integrating multi-source data with deep learning remain limited, this study aims to propose a deep learning framework combining Sentinel-2 with environmental data for large-scale, field-level wheat yield prediction across central Spain. Methods Based on wheat yield data from central Spain over two seasons (2022–2023), we evaluated the potential of the Transformer model and compared its performance against partial least squares regression (PLSR), K-Nearest Neighbors Regression (KNN), and deep neural networks (DNN). Additionally, we assessed the relative importance of environmental variables versus spectral features and examined how single and combined phenological stages influence prediction accuracy. Results The Transformer achieved the highest accuracy when integrating all phenological stages and multi-source inputs (R² = 0.93, RMSE = 200 kg/ha), compared with DNN (R² = 0.91, RMSE = 225 kg/ha), KNN (R² = 0.88, RMSE = 257 kg/ha), and PLSR (R² = 0.77, RMSE = 358 kg/ha). Models using environmental variables performed better than those based on spectral data. Moreover, combining the tillering and heading stages provided a balance of high accuracy and timeliness, with the Transformer achieving R² = 0.91 and RMSE = 230 kg/ha. Conclusion This study demonstrates the improved performance of the Transformer model and underscores the critical importance of integrating multi-source environmental data and phenological information for robust crop yield forecasting. Impact This study proposes a data-driven precision agriculture framework that combines Sentinel-2 and environmental data for accurate, early-season wheat yield forecasting with relevant applications in precision agriculture.