A Hybrid Random Forest–LSTM Framework for Robust Crop Recommendation

Abstract

The integration of machine learning and deep learning in agriculture has significantly improved crop yield prediction, yet most existing models fail to jointly capture static soil conditions and temporal weather dynamics. This paper proposes a hybrid Random Forest–Long Short-Term Memory (RF–LSTM) framework that combines the interpretability and robustness of RF with the temporal modeling capabilities of LSTM. The model is trained on an augmented crop dataset incorporating both soil properties and synthetic weather sequences, and subsequently validated on an independent real-world Soil-Climate-data dataset to evaluate generalization. Experimental results show that the proposed model achieves 95.3% accuracy on synthetic data and 97.2% accuracy on real data, outperforming baselines (RF and LSTM) as well as comparable hybrid models. The minimal performance gap across domains demonstrates the model’s robustness and adaptability to natural environmental variability. By integrating static and temporal features in a unified architecture, the proposed RF–LSTM offers an effective and interpretable solution for crop recommendation and yield prediction under realistic agricultural conditions.