Crop Disease Detection And Classification

Authors: Saniya G S, Maaz Ahmed, Murali Karthik Nayaka P J

Abstract: This paper presents a web-based system for classifying tomato leaf images as bacterial spot, early blight, late blight, or healthy. It uses MobileNetV2, initialized with ImageNet weights, as a frozen feature extractor, with a small trainable classification head. The model was trained on 6,644 images from a PlantVillage-based dataset. Images were resized to 128 × 128 pixels, and training images were augmented with random flips, rotations, and zoom. On a fixed, random 20% image-level validation split, the saved evaluation report shows 94.43% accuracy and a weighted F1-score of 94.32%. However, performance varied by class: early blight had the lowest recall at 79.34%. The accompanying Streamlit app allows users to upload a leaf image and view its predicted class, confidence, and available disease information. Since the validation images come from the same controlled dataset and were split randomly by image, these results do not indicate how well the system would perform on field photographs or images from independent farms..

DOI: