Explainable Knowledge Distillation from ConvNeXt to MobileNetV3 for Efficient Plant Disease Classification

Categorizing plant diseases is crucial, as it influences the global food supply chain. Manual inspection-based methods are tedious and error-prone, as visual manifestations of these diseases often closely resemble one another. Regarding this, Deep Learning (DL) models may present an effective solution. In this work, we primarily selected the pre-trained ConvNeXtBase, a state-of-the-art (SOTA) Convolutional Neural Network (CNN), and fine-tuned it on the PlantVillage dataset, achieving 99.60% classification accuracy. The fine-tuned ConvNeXt was employed as the teacher model in a Knowledge Distillation (KD) setup to train the lightweight MobileNetV3Small model with ≈2.5M parameters. When combined with the K-Nearest Neighbor (KNN) classifier, fine-tuned MobileNetV3Small improves the classification accuracy to 99.71%. To interpret the model’s decision, we employed several modelspecific and model-agnostic Explainable AI (XAI) approaches. Therefore, through KD, we achieved better performance from the lightweight MobileNetV3Small, which may be deployed in edge devices.