GreenDefender: Predictive plant health and care system

Tan, Lok Wei (2025) GreenDefender: Predictive plant health and care system. Final Year Project, UTAR.

PDF Download (6Mb)

Abstract

This project explores the intersection of artificial intelligence (AI) and mobile technology to provide home gardeners with a personalized, user-friendly solution for maintaining decorative plant health. Existing applications such as Plantix, Agrobase, PlantVillage and Blossom largely focus on agricultural crops, and offer limited disease coverage, minimal personalization, and inadequate real-time support for ornamental plants. GreenDefender addresses these gaps by integrating a multi-stage AI pipeline, an interactive recommendation system, and a cloud-enabled user experience into a single mobile platform. The proposed application was trained using two publicly benchmarked datasets: the Leaf Detection Dataset from Kaggle, containing 1130 annotated images for verifying the presence of leaves, and a combined disease dataset comprising the Rose Lead Disease Dataset, Flower Leaf Diseases Dataset New and Update, Tomato Leaf Disease Dataset, and Pumpkin Leaf Diseases Dataset from Mendeley Data, providing 9683 images across six disease classes: Fresh Leaf, Mosaic, Powdery Mildew, Downy Mildew, Black Spot, and Rust. Both datasets were split using an 80:10:10 ratio for training, validation, and testing. The application employs a YOLOv8-based leaf detection model as a gatekeeper to verify the presence of leaves before disease classification, achieving results with an mAP50 of 0.961 (training), 0.748 (validation), and 0.681 (test).As for disease classification, 3 deep learning architectures – Attention U-Net, Standard U-Net, and ResNet50 – were trained and compared. Attention U-Net achieved the highest accuracy, with 97.94% test accuracy (random splitting) and 96.42% (stratified splitting). Beyond detection and classification, GreenDefender provides interactive care plan generation powered by Gemini Artificial Intelligence (AI). Users answer context-specific questions to receive highly tailored recommendations. The application also supports diagnosis history tracking, profile management, and a community discussion module to encourage knowledge sharing among users.

Actions (login required)