Multi-Modal Machine Learning Framework for Accurate Crop Yield Prediction Using Soil, Weather, and Agricultural Data

Abstract

Accurate predicting of crop yield is of greatest importance to the methods of precision agriculture and planning of food security. This study suggests a novel Multi-Modal Attention-Based Hybrid Deep Network (MAHDN) by combining heterogeneous data of agriculture, such as soil properties, weather parameters, and cropmanagement data, all to improve the accuracy of the prediction. A detailed multimodal database was created based on Soil Health Card Scheme, data on meteorological information (NASA POWER) and crop production data (Government of India). The proposed framework uses BiLSTM for the temporal modeling of weather sequences and another mechanism for cross modal attention fusion to dynamically weight the contribution of each modality. Experimental evaluation with conventional measures of regression and classifications shows that MAHDN is better than conventional machine learning and deep learning baselines regarding RMSE, MAE, accuracy and F1-score. Confusion matrix and ROC analyses further confirm that there is enhanced discriminative capability and less misclassification. The results showed the proposed approach as a robust and scalable decision support tool for intelligent and data-driven precision agriculture applications.