AI-Powered Precision Agriculture: Integrating Smart Systems for Sustainable Yield Forecasting and Climate Resilience

Agricultural productivity faces ongoing challenges from climate variability, soil degradation, and constraints on resources. In this context, Artificial Intelligence (AI) is a new disruptive technology in precision agriculture enabling anticipated recommendations using data-based predictions for improved decision-making. This paper presented XTR-Net, a new hybrid agronomic decision-making framework which supports a temporal modeling using Transformer architectures and Reinforcement Learning (RL) for on-the-go agronomic decision-making, and involves SHAP (SHapley Additive Explanations) for model interpretation. The hybrid framework architecture is trained on a global, multi-year agricultural dataset made available from Kaggle, and includes key features including crop type, annual rainfall, temperature, pesticide use, and yield, all from multiple countries and climates. XTR-Net's performance was empirically evaluated, achieving a prediction accuracy of 96.3%, R2 score of 0.93, and RMSE of 0.019; XTR-Net outperformed benchmark models, Random Forest, LightGBM, and CNN-LSTM architectures. The Transformer model successfully described the temporal dependencies in year-to-year predictions, and the RL layer made it possible to optimally input data in real-time for the model to simulate the effect of adaptive interventions. The SHAP layer also provided clear, predicted feature explanations at the feature-level, so the model could address whether outcomes were only accurate, or both accurate and interpretable.