Evaluating the Capability of VGG16 Trained on Kaggle Dataset for Detecting Tomato Diseases in Indonesian

Abstract

Tomato leaf diseases pose a significant threat to agricultural productivity in Indonesia, often leading to severe yield losses. This study evaluates the effectiveness of the VGG16 convolutional neural network in detecting tomato diseases, particularly when trained on the standardized PlantVillage dataset and applied to local agricultural conditions in Sragen, Central Java. The research involved data preprocessing using background removal and resizing techniques, model training via transfer learning, and deployment through a FastAPI backend and React Native frontend. The VGG16 model achieved high accuracy 82% on the PlantVillage test set but exhibited a sharp decline 25% accuracy when tested on locally sourced images, highlighting limited generalization capabilities. These findings emphasize the necessity of incorporating local datasets and domain adaptation strategies to develop AI-based plant disease detection tools that are effective in real-world settings. The study underscores the importance of contextualizing AI solutions for local agricultural environments to ensure their practical applicability and reliability.