Design of an Improved Model Using Hyperspectral Imaging, Data Fusion, and Transformer-Based Deep Learning for Crop Disease Identification Operations

Accurate and timely diagnosis of crop diseases is crucial for food security and agricultural sustainability. Current methods lack accuracy, interpretability, data integration, and advanced imaging techniques. The current research aims to address such issues by proposing a broad-perspective approach that brings into play hyperspectral imaging in tandem with 3D reconstruction, multimodal data integration, advanced machine learning models, and finally, trickled-down with XAI techniques. Up to now, approaches in the detection of crop diseases have been basically dependent on RGB imaging information and single-information sources that are incapable of accounting for interactions amidst the variety of factors affecting crop health. These methods have poor spatial and spectral analysis; thus, leading to limited accuracy and also offering little elucidation on the implication of the models. Overcoming these limitations, we employed a hyperspectral imaging approach with the 3D reconstruction for achieving fine spectral and spatial quantification of the crops. Through this approach, accuracy in disease detection was improved from 85% to 92%, and this approach drives a holistic view of the crops. Through data fusion, hyperspectral images, soil health parameters, and weather data were integrated. Their integration significantly improved the predictive accuracy—all the way up to 89% from 78% overall accuracy. We also developed an explainable Transformer-based deep learning model with in-built attention mechanisms and Grad-CAM for explainability. Such an advanced model architecture allowed draft feature extraction and provided visual explanations on the decisions made by the model, increasing classification accuracy from 88% to 94% and interpretability by 70%. A combination of hyperspectral imaging, data fusion, and advanced machine learning with XAI improves the model's accuracy and its robustness and therefore improves the latter's transparency in the decision-making process.