Plant Disease Detection Using Convolutional Neural Networks

Plant diseases cause substantial crop losses, and timely diagnosis is difficult for growers without access to pathologists. This paper presents an image-based system that classifies 38 plant-disease and healthy-leaf categories across 14 crop species using transfer learning. A MobileNetV2 network pretrained on ImageNet is adapted in two phases: a new classification head is first trained on the frozen convolutional base, and the last 30 base layers are then fine-tuned at a lower learning rate with batch-normalization layers kept frozen. Training uses the PlantVillage dataset (54,305 images) with a stratified 70/15/15 split. On the held-out test set of 8,162 images the model achieves 97.35% accuracy and a macro-averaged F1-score of 96.27%, against 96.13% and 94.85% for a compact custom CNN baseline. Per-class analysis shows that errors concentrate in visually similar tomato diseases and in under-represented classes. The trained model is served through a Streamlit web application that returns the predicted class, a confidence score and the top three predictions. Because PlantVillage contains laboratory images and the split was made per image, the results are in-distribution estimates that require field validation. Key Words: plant disease detection, convolutional neural network, transfer learning, MobileNetV2, PlantVillage, precision agriculture.