DEEP FEATURE EXTRACTION FOR PLANT DISEASE CLASSIFICATION: A RESNET50-BASED TRANSFER LEARNING FRAMEWORK WITH WEB INTERFACE

Abstract

This study presents a lightweight and modular pipeline for automated plant disease classification, capitalizing on deep feature extraction and ensemble machine learning. We leverage the PlantVillage dataset, featuring over 54,000 images categorized into 38 disease classes, to develop a scalable workflow. Images undergo preprocessing and augmentation, then are passed through a ResNet50 pre-trained network to obtain 2048-dimensional feature vectors, compressing the storage requirements significantly. These features serve as input to a Random Forest classifier, which achieves over 95% accuracy on the test data. The solution’s efficiency enables real-time web deployment using Streamlit, offering single or batch predictions with confidence scores, bringing advanced diagnostics to end users with minimal computational demand.