Pomegranate disease diagnosis with severity estimation and treatment remedies using deep learning and RAG-based LLM

Pomegranate cultivation faces significant challenges due to fruit diseases that significantly impact crop yield and farmer income. Traditional methods for disease detection are often slow and prone to errors, delaying timely intervention. This paper proposes a deep learning-based system for automatic, multi-class disease classification in pomegranates using transfer learning. A dataset comprising 5099 annotated images was used to train and evaluate several CNN models, including DenseNet121, EfficientNetB0V2, MobileNetV2, ResNet50, VGG16, and InceptionV3. DenseNet121 emerged as the top performer, achieving an accuracy of 99.35%. To enhance practical value, a novel Healthy-Based Deviation Scoring (HBDS) method was developed to estimate disease severity using Grad-CAM ++ for lesion localization and Mahalanobis distance-based scoring, followed by Gaussian Mixture Model clustering. The severity predictions of the system were verified against manually labeled images, and the system has shown superior accuracy compared to pixel-based methods. Also, a recommendation module was integrated using a retrieval-augmented language model, which provides disease-specific treatment suggestions based on the predicted severity. The complete pipeline is implemented as a user-friendly web application that delivers real-time diagnosis, severity estimation, and actionable treatment plans, which offer a practical and scalable solution for modern precision agriculture.