Design of Automated Model for Citrus Fruits and Leaves Disease Detection

Abstract

Diseases of citrus fruits and leaves are a serious challenge to farm production and economic viability. The need for effective, automated disease detection methods is on the rise due to the drawbacks of the conventional manual detection method. In this paper, a lightweight deep learning-based classifier using MobileNetV2 for real-time diagnosis of citrus diseases is proposed. The dataset, which was retrieved from Kaggle, consists of high-resolution images of citrus leaves and fruits and are classified into four categories: greening (HLB), black spot, citrus canker, and healthy. Rotations, scaling, shearing, zooming, and flipping were used to enhance model generalization by giving data augmentation methods. Categorical cross-entropy loss, early stopping, Adam optimizer, and learning rate reduction on plateau were all adopted during training so that convergence is strictly acquired. A web application has been developed in which an image is uploaded for the diagnosis of disease. The names of disease, symptoms, causes, preventive measures, and some external references can be viewed from the application. Comparative evaluation of InceptionV3 and an in-house built CNN model validated that MobileNetV2 gives better accuracy as well as processing efficiency, suggesting its applicability in smart farm applications. A number of recent research works have investigated deep learning methods for citrus disease detection, emphasizing the need for lightweight models for real-time agricultural use. Hybrid attention networks and hyper spectral imaging methods have also been found to yield good results in enhancing disease identification accuracy. The suggested method reduces human intervention while maintaining timely disease detection and intervention, thus enhancing yield and sustainable agriculture.