HVAF-XAI-Net: A multimodal explainable AI framework for accurate mulberry leaf disease detection

Mulberry leaf diseases, such as powdery mildew, leaf rust, and leaf spots, significantly impact silkworm health, silk yield, and quality. Traditional detection methods are labour-intensive, prone to errors, and ineffective for early disease identification. Existing automated approaches often fail to integrate multimodal data or provide transparent decision-making. To address these challenges, we propose HVAF-XAI-Net, a Hybrid Vision-Attention Fusion Network with Explainable AI (XAI) for accurate and interpretable detection of mulberry leaf diseases. The framework combines Vision Transformers (ViT) and Spatial Attention Modules to identify visual symptoms (e.g., lesions, discolouration) from UAV-captured images, Temporal Convolutional Networks (TCN) with Temporal Attention Layers to analyze environmental trends (e.g., humidity, temperature) from IoT sensors, and a Cross-Attention Mechanism to fuse spatial and temporal data for enhanced classification accuracy. XAI techniques, including SHAP, LIME, and Grad-CAM, provide interpretable insights, with visual case studies (e.g., Grad-CAM highlighting diseased leaf regions) aiding sericulture experts in validating predictions and identifying critical indicators. HVAF-XAI-Net achieves 97.0 % accuracy (95 % CI: [96.2 %, 97.8 %], p < 0.05), a 22 % reduction in false positives, and an inference time of 6.5 milliseconds, outperforming CNN-based (ResNet-50: 88.2 %), Hybrid CNN-LSTM (89.5 %), transformer-based (Swin Transformer: 90.1 %), and ensemble models (Random Forest + CNN: 89.8 %). Its scalable design, optimized for deployment on drones and low-resource devices (e.g., mobile apps via model quantization), enhances precision in sericulture, reduces manual inspection costs, and promotes sustainable silk production.