A lightweight and explainable CNN model for empowering plant disease diagnosis

Crop disease is a significant challenge in agriculture, requiring quick and precise detection to safeguard yields and reduce economic losses. Traditional diagnostic methods are slow, labor-intensive, and rely on expert knowledge, limiting scalability for large-scale operations. To overcome these challenges, a novel architecture called Mob-Res, combining residual learning with the MobileNetV2 feature extractor, is introduced in this work. Despite having only 3.51 million parameters, Mob-Res is lightweight and well-suited for mobile applications while delivering exceptional performance. The proposed model is assessed using two benchmark datasets: Plant Disease Expert, consisting of 199,644 images across 58 classes, and PlantVillage, with 54,305 images across 38 classes. Through a rigorous training strategy, Mob-Res demonstrates robust performance, achieving 97.73% average accuracy on the Plant Disease Expert dataset and 99.47% on the PlantVillage dataset. The cross-domain validation rate (CDVR) is computed to assess its cross-domain adaptability, with the model showing competitive results compared to other pre-trained models. Additionally, Mob-Res outperforms prominent pre-trained CNN architectures, surpassing ViT-L32 while maintaining a significantly lower parameter count and achieving faster inference times. The proposed model enhances interpretability by utilizing Gradient-weighted Class Activation Mapping (Grad-CAM), Grad-CAM++, and Local Interpretable Model-agnostic Explanations (LIME). These techniques provide visual insights into the neural regions influencing the predictions. The experimental results conducted in the current work highlight Mob-Res as a promising solution for automated plant disease detection, supporting large-scale agricultural operations and advancing global food security.