COMPARATIVE EVALUATION OF MOBILENETV2 FOR SEVEN-CLASS PLANT DISEASE CLASSIFICATION: A LIGHTWEIGHT TRANSFER LEARNING APPROACH
Keywords:
plant disease classification, transfer learning, MobileNetV2, convolutional neural networks, precision agriculture, deep learning, ROC-AUC, confusion matrix.Abstract
Plant diseases are responsible for an estimated 20–40% of global crop losses each year, making rapid and accessible diagnosis essential for food security. This paper presents a lightweight transfer-learning system for plant disease classification built on MobileNetV2, pre-trained on ImageNet. Unlike prior 38–41-class PlantVillage studies, this work deliberately narrows the problem to a focused 7-class, 3-crop subset (Tomato, Potato, and Corn/Maize, covering healthy foliage and their most prevalent diseases) drawn from the public PlantVillage dataset on Kaggle, comprising 5,602 training, 1,201 validation, and 1,203 held-out test images. With the MobileNetV2 backbone frozen and a compact classification head (Global Average Pooling → Dense-256 → Dropout → Dense-7 Softmax) trained for 10 effective epochs under early stopping (out of a 30-epoch budget), the model reaches a peak validation accuracy of 97.09%, a macro-average ROC-AUC of 0.9986, and a mean F1-score of 0.966 across all seven classes. Confusion is confined almost entirely to the visually similar Tomato Early Blight and Tomato Late Blight pair, while Corn and Potato classes are classified with near-perfect precision and recall. These results indicate that restricting the class space to agronomically related, visually distinguishable categories allows a lightweight, edge-deployable CNN to substantially outperform the accuracy typically reported for full 38–41-class PlantVillage benchmarks, while retaining MobileNetV2’s suitability for smartphone and embedded deployment.
