Mobile application for Solanaceous crop health diagnosis and treatment using a lightweight YOLO Model

• Main Author: Khan, Asar
• Format: Thesis
• Language: English; English
• Published: 2025
• Subjects: Q Science; QA Mathematics
• Online Access: http://eprints.utem.edu.my/id/eprint/29319/

Agriculture, an extremely important industry for developing countries, but it is facing difficulties by plant disease, water insecurity, and rising temperatures. Traditional disease detection procedures is time consuming, require professional expertise, and can end in misdiagnosis, leading to poor treatment. In addition, the crop diseases looks similar although the diseases class is different. Therefore, this leads to the usage of Artificial Intelligence (AI) approach. With advances in AI especially deep learning, automated solutions appear to be a potential strategy to address these issues. There are remains a gap in accessible, accurate, and lightweight diagnostic tools that function efficiently on resource-constrained mobile devices. The goal of this research is to analyze and identify the best-performing optimized lightweight deep learning model among YOLOv5n, YOLOv7t, and YOLOv8n for identifying Solanaceous crop diseases specifically in these four types of plants used, peppers, potatoes, eggplants, and tomatoes. A total of 23 disease and healthy classes were considered, including Chili Anthracnose, Eggplant Cercospora leaf spot, Potato Common scab, Tomato Leaf mold, and others, which can result in significant financial losses. The models' performance is evaluated using important measures such as mean average precision (mAP), precision, recall, F1-score, inference time, and model size. Optimization was performed by fine-tuning hyperparameters such as (mini batch size, learning rate, loss functions, and weight decays), applying these techniques on data, and balancing performance with computational efficiency to enable on resource-constrained mobile devices. Using the PlantVillage dataset, which was processed through Roboflow for annotation and augmentation, the models were trained and tested on Google Colab, with YOLOv8n achieving the highest mean average precision (99.1% mAP), followed by YOLOv7t (98.6%) and YOLOv5n (98.5%). Based on these findings, a mobile application was developed to integrate the best-performing model, enabling real-time disease diagnosis from static images and providing treatment recommendations. This research contributes to precision agriculture by offering a cost-effective and efficient tool that empowers farmers with accurate disease detection and treatment guidance, ultimately improving crop health, increasing productivity, and supporting food security.