Amid escalating climate variability and the urgent pursuit of sustainable agricultural systems, forecasting crop yields has become a foundational element in strategic planning and food security efforts. Aiming to precisely predict crop yields by modeling rich spatial and temporal dependencies natural in agricultural data, this work offers a new hybrid deep learning architecture STAFNet-RAMC-BiLSTM (Spatio-Temporal Attention Fusion Network with Residual Attention Mechanism and Monte Carlo Dropout). The model integrates spatio-temporal attention fusion through convolutional layers, bidirectional LSTM units for temporal sequence learning, and residual attention modules to retain both raw and refined feature representations. Additionally, it incorporates Monte Carlo dropout to provide uncertainty-aware predictions, enhancing the reliability and interpretability of forecast outputs. The proposed model was evaluated on a real-world dataset comprising 21 years (1997–2017) of crop production data across 33 Indian states, covering 124 crop types and six seasonal classes. Compared to benchmark models such as STACNN-BiLSTM and CatBoost, the proposed model achieved superior performance, with an RMSE of 0.198, MAE of 0.135, and R2 of 0.99, reflecting both lower prediction error and higher variance explanation. The architecture also demonstrated strong generalization, stable convergence, and robust handling of diverse agricultural scenarios. Qualitative analysis further highlighted its scalability and domain interpretability, making it highly suitable for integration into real-time decision support systems. This research thus presents a comprehensive, scalable, and high-precision forecasting framework, offering a powerful tool for climate-resilient and data-driven agriculture.