<p>Accurate wheat yield prediction is essential for smart agriculture, food security, and sustainable resource management. However, modeling wheat yield remains challenging due to the complex interactions among spatial, temporal, and environmental factors contained within heterogeneous agricultural datasets. To address this issue, a novel dual-path hybrid deep learning framework for spatiotemporal wheat yield prediction is proposed. The framework integrates convolutional neural networks (CNN), long short-term memory (LSTM) networks, and an attention mechanism to simultaneously capture spatial patterns and temporal dependencies from multi-source agricultural data. A comprehensive preprocessing and feature engineering pipeline is implemented, including robust scaling, outlier clipping, temporal sequence construction, and Random Forest-based feature selection. The most informative features, including geospatial coordinates, cloud cover, humidity, pressure, dew point, and seasonal indicators, are selected to reduce dimensionality while preserving predictive information. The proposed architecture employs a gated feature fusion strategy to dynamically combine spatial and temporal representations and improve prediction robustness. Experimental evaluation demonstrates that the proposed model achieves superior predictive performance with an MAE of 0.41, RMSE of 0.74, R² of 0.994, and relative error of 1.7%. Comparative analysis further shows that the proposed framework outperforms existing machine learning and deep learning approaches for wheat yield prediction. The obtained results confirm the effectiveness of integrating hybrid deep learning, attention mechanisms, and domain-informed feature engineering for accurate and reliable agricultural forecasting.</p>