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

Area of Science:

• Agricultural Science
• Computer Science
• Artificial Intelligence

Background:

• Accurate crop yield prediction is vital for food security and resource management.
• Deep learning models offer improved accuracy but face challenges like gradient instability and limited generalizability.
• Existing models struggle with long physiological lags and vanishing/exploding gradients in agricultural data.

Purpose of the Study:

• To develop an innovative hybrid deep learning model for enhanced crop yield prediction.
• To address limitations of existing models, including gradient instability and poor generalizability.
• To improve the accuracy and interpretability of soybean yield predictions.

Main Methods:

• A novel MHCNN-LSTM-MHA model fusing Multi-head Convolutional Neural Networks, Long Short-Term Memory, and Multi-head Attention was developed.
• A comprehensive dataset of U.S. soybean crops including weather, soil, management, and historical yield data was utilized.
• The proposed model was benchmarked against single-component and existing advanced models.

Main Results:

• The MHCNN-LSTM-MHA model achieved an RMSE of 3.75 bushels/acre and an R² of 0.905, outperforming benchmarks by 9.86%.
• The Multi-head Attention mechanism improved generalizability and interpretability by prioritizing key features and time steps.
• SHAP analysis revealed weather (precipitation, solar radiation, vapor pressure) as highly impactful, followed by soil properties (pH, Clay content).

Conclusions:

• The MHCNN-LSTM-MHA model offers superior accuracy and interpretability for crop yield prediction.
• The model's ability to handle complex interactions and dynamic feature prioritization is key to its performance.
• This work advances explainable AI in agriculture, facilitating smart farming integration.