Enhancing crop yield prediction accuracy with a novel interpretable deep learning model: MHCNN-LSTM-MHA

Accurate and reliable crop yield prediction before harvest is essential for informed decision-making, food security, resource optimization, climate change mitigation, and smart agricultural systems. Deep learning has significantly improved accuracy in predicting crop yields as compared to conventional methods. However, some challenges such as unstable gradients, single-model constraint, limited generalizability, and lack of interpretability still persist. Gradient instability is particularly acute in agricultural settings, where early-season weather anomalies influence final yield through long physiological lags, yet back-propagated signals in deep hybrid architectures often vanish or explode before reaching the responsible input layers. These gaps necessitate an innovative and composite problem solving approach. This paper presents a novel MHCNN-LSTM-MHA model which fuses multiple deep learning paradigms of Multihead Convolutional Neural Networks, Long Short-Term Memory, and Multihead Attention Mechanism to improve the prediction accuracy. This study employed a comprehensive dataset on U.S. soybean crop with features like weather conditions, soil properties, management practices, and historical yields. We tested the proposed model against single component models (CNN, LSTM, CNN-LSTM) and benchmarked models (CNN-RNN, Interaction Regression, CNN-DNN). The proposed MHCNN-LSTM-MHA model achieved an RMSE of 3.75 bushels per acre and Formula: see text of 0.905, showing a 9.86% improvement over the previous best benchmark on the same dataset. This performance is attributed to the advanced hybrid architecture that effectively model complex feature interactions and dependencies. The four-headed MHA improved generalizability and enhanced interpretability by dynamically prioritizing important features and time steps, while mitigating bias, in alignment with crop growth dynamics. SHAP-based interpretability revealed agronomically consistent feature importance, showing that crop yield in more sensitive to variations in weather components (precipitation, solar radiation, vapor pressure) exhibiting higher impact on yield outcome and moderately sensitive to soil properties (pH, Clay content). The contribution of this work is the enhancement of predictive accuracy alongside SHAP-based interpretability, which is essential for advancing explainable AI in crop yield prediction and facilitating its integration into smart agriculture systems.