Author(s): Shylaja Chityala

Abstract: Accurate crop yield prediction is fundamental to sustainable agriculture, informed policymaking, and global food security. Traditional statistical models are inadequate in harnessing the increasing volume of complex, high-dimensional agricultural data, including satellite imagery, granular weather records, and detailed soil profiles. To address this, we propose MCYP-Net, a novel multi-modal AI framework that integrates these heterogeneous data sources using a hybrid deep learning architecture. The model combines convolutional neural networks for spatial feature extraction, recurrent neural networks for temporal modeling, and a cross-attention-based fusion mechanism to learn inter-modal dependencies. Comprehensive experiments were conducted across diverse agro-climatic regions in the USA (Iowa, Nebraska) and India (Punjab, Maharashtra) on three staple crops—maize, wheat, and soybean—using multi-year datasets comprising Sentinel-2 imagery, NOAA weather data, and ISRIC soil profiles. MCYP-Net consistently outperformed traditional machine learning (Linear Regression, Random Forest) and unimodal deep learning baselines (CNN-only, LSTM-only), achieving an R² of 0.91, RMSE of 0.42, and MAE of 0.35. Ablation studies confirmed that removing any modality reduced performance significantly, validating the synergistic effect of multi-modal integration. Cross-attention fusion proved more effective than simpler alternatives, boosting R² by 6%. Region-wise feature importance analysis revealed that weather features dominated in temperate zones, while soil and vegetation indices were more critical in semi-arid regions, highlighting the model’s context-aware adaptability. Visualizations demonstrated strong alignment between predicted and actual yields, underscoring the model’s robustness. Overall, MCYP-Net advances state-of-the-art in crop yield prediction with high accuracy, interpretability, and scalability for real-world precision agriculture applications.

Pages: 67-74 | Views: 154 | Downloads: 89

Download Full Article: Click Here