An Enhanced Deep Neural Network Framework for Accurate Tomato Disease Recognition in Real-time Environment

Agricultural production is a critical sector that directly impacts the financial and social well-being of a society. The identification of plant diseases in a real-time environment is a significant challenge for agriculture production. Conventional disease detection methods, which depend significantly on manual inspection, are time-consuming, labour-intensive, and susceptible to human error. Furthermore, many recently developed models struggle in real-time scenarios because their accuracy is compromised when trained on isolated leaf images but then used to analyse entire plants. To tackle these issues, this research offers an advanced, automated system from tomato leaf segmentation and disease detection to the automatic spray prescription in real-time environment. This research presents an integrated system to address these issues, focusing on tomato plants. In first part of the research after deeply analysing the YOLO (You Only Look Once) models we integrate two models, the YOLOv8 with SAM (Segment Anything Model), for leaf detection and masking in the tomato plants, and extraction of the individual leaves in a real-time environment for improving the performance of leaf disease detection. For leaf detection, the modified YOLOv8 is used, and for masking and extraction of the individual leaves, the SAM is used. The individual leaves are then, provided to the custom deep neural network for further disease detection. This ensures that subsequent disease detection is performed on isolated leaves, improving accuracy. We investigated deep learning models for precise and efficient tomato plant disease detection. All the models were trained and validated on individual and merged dataset of over 18000 and 25,000 images, encompassing 10 distinct classes (9 diseases and healthy plants). The performance of our custom CNNs (Convolutional Neural Networks) model was significant, achieving an accuracy of over 99%. Furthermore, it revealed higher efficiency, requiring less training time and computational resources than leading architectures such as VGG (Visual Geometry Group), ResNet (Residual Network), and DenseNet (Densely Connected Convolutional Network), making it a promising tool for real-world applications. In second part of this research work, we have designed, an automated system for tomato disease detection and spray prescription using an enhanced YOLOv9 model. By leveraging advance deep learning techniques, the proposed system accurately identifies and detect the nine tomato leaf disease in real-time by making efficient, precise and accurate decision including healthy leaves. This YOLOv9 model is modified for detecting tomato leaf diseases and optimized for getting higher accuracy and efficiency. Once disease is identified, the system automatically recommends a spray depending on the detected disease, which helps in reducing the pesticide use along with the environmental impact. This system helps in maximizing crop health and yield. After testing the system on the test dataset and real-time images demonstrates the system accuracy and efficiency, achieving detection accuracy of 97% and spray prescription accuracy of 94%. Integrating a YOLOv9 with spray prescription system provides a sustainable, cost-effective solution for managing tomato plant diseases. This thesis illustrates the efficacy of deep learning in the efficient and precise identification of tomato plant diseases, along with automated spray recommendations, thereby benefiting farmers and enhancing agricultural productivity. The effective performance observed in this thesis makes it promising for real-world agricultural applications.