European Journal of Sustainable Agroecosystems

Background: Timely and accurate prediction of wheat (Triticum aestivum L.) grain yield well before harvest is essential for precision management, market planning, and food security decision-making, yet conventional prediction approaches relying on single-date imagery or purely empirical vegetation-index regressions often fail to capture the nonlinear, temporally dynamic relationship between canopy development and final yield. 
Objective: This study developed and evaluated a hybrid deep learning framework integrating convolutional, recurrent, and attention-based components with multi-temporal satellite imagery to predict wheat grain yield at progressively earlier crop growth stages. 
Methods: Multi-temporal, multispectral satellite imagery spanning emergence through maturity was acquired over two growing seasons across replicated wheat cultivar trials at a semi-arid research site. Convolutional neural network (CNN) layers extracted spatial canopy features from each image date, long short-term memory (LSTM) layers modeled temporal growth trajectories across the acquisition sequence, and a transformer-based attention mechanism weighted the relative contribution of different growth stages to final yield. Model predictions were benchmarked against conventional vegetation-index regression and standalone CNN and LSTM models using root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and coefficient of determination (R²). 
Results: The hybrid CNN-LSTM-Transformer model achieved the highest prediction accuracy among all evaluated approaches, with progressively improving accuracy from tillering through grain filling, and meaningful predictive skill as early as the stem elongation stage. Feature importance analysis identified the normalized difference red-edge index (NDRE) and enhanced vegetation index (EVI) during heading and grain filling as the strongest predictors, while the attention mechanism assigned greatest weight to mid-to-late reproductive stages. 
Significance: Integrating spatial, temporal, and attention-based learning substantially improved early-season yield prediction relative to conventional single-stage approaches. 
Conclusion: Hybrid deep learning architectures coupled with multi-temporal satellite imagery offer a robust, scalable pathway toward operational early-season wheat yield forecasting to support precision agriculture and climate-resilient decision-making.