AgriLite-XAI: An Explainable Ultra-Lightweight MobileNetV3–ECA Network for Multi-Species Plant Disease Classification with Grad-CAM++ Visualisation AgriLite-XAI: An Explainable Ultra-Lightweight MobileNetV3–ECA Network for Multi-Species Plant Disease Classification with Grad-CAM++ Visualisation

Plant diseases are estimated to destroy 20–40% of global crop production each year, with smallholder farmers in low-resource regions bearing a disproportionate share of this burden because accurate diagnosis traditionally requires expert visual inspection or computationally expensive deep learning models that exceed the capabilities of affordable mobile and embedded hardware. AgriLite-XAI shows that an ultra-lightweight network combining MobileNetV3-Small with a near-parameter-free channel-attention mechanism (ECA-Net) can reach 96.55% accuracy on a 15-class, multi-species leaf-disease benchmark using under one million parameters and 36 minutes of training on a budget consumer GPU. By pairing this efficiency with Grad-CAM++ explainability, the framework also addresses the trust gap that limits adoption of black-box diagnostic tools among agricultural extension workers and farmers, who can visually verify that predictions are grounded in genuine lesion symptoms. Together, these results indicate that accuracy, computational efficiency, and interpretability need not be competing design goals, offering a practical template for deployable, transparent AI diagnostic tools in precision agriculture and, more broadly, for other resource-constrained, trust-sensitive domains where black-box deep learning has previously been impractical. Plant diseases are estimated to destroy 20–40% of global crop production each year, with smallholder farmers in low-resource regions bearing a disproportionate share of this burden because accurate diagnosis traditionally requires expert visual inspection or computationally expensive deep learning models that exceed the capabilities of affordable mobile and embedded hardware. AgriLite-XAI shows that an ultra-lightweight network combining MobileNetV3-Small with a near-parameter-free channel-attention mechanism (ECA-Net) can reach 96.55% accuracy on a 15-class, multi-species leaf-disease benchmark using under one million parameters and 36 minutes of training on a budget consumer GPU. By pairing this efficiency with Grad-CAM++ explainability, the framework also addresses the trust gap that limits adoption of black-box diagnostic tools among agricultural extension workers and farmers, who can visually verify that predictions are grounded in genuine lesion symptoms. Together, these results indicate that accuracy, computational efficiency, and interpretability need not be competing design goals, offering a practical template for deployable, transparent AI diagnostic tools in precision agriculture and, more broadly, for other resource-constrained, trust-sensitive domains where black-box deep learning has previously been impractical.