Diagnosis of Plant Diseases using Image Processing and Machine Learning

Abstract

Citrus canker is a devastating bacterial disease affecting citrus trees and has significant economic implications for agriculture. The timely identification and precise diagnosis of citrus canker are critical components for the efficient management of the disease and the containment of its propagation. Machine learning algorithms and deep learning models have exhibited notable efficacy in automating the detection of citrus canker symptoms from visual data sources, notably images. This enables farmers, agricultural experts, and researchers to detect infected trees and implement necessary management measures promptly. In our present study, machine learning-based citrus canker identification and diagnosis were done. Developed our datasets utilizing two distinct environmental systems: natural field conditions, referred to as the 'attached leaf method,' and laboratory-controlled conditions, denoted as the 'detached leaf method. The dataset delineates the growth rate of citrus canker disease across various stages or levels of affliction (levels 1-6), contingent upon the appearance of lesions and the extent of the affected area. The defined stages within the dataset encompass 1) water soaking, 2) yellow chlorosis/initiation, 3) chlorosis, 4) blister formation, 5) canker development start (50% of the inoculated area), and 6) Canker infection (100% of the inoculated area). The entire processing includes pre-processing, segmentation, features extraction and perform classification using four different classifiers which include SVM, KNN, Naïve Bayes, and Neural Network. The efficiency of different classifiers was 96.77% by SVM, followed by KNN 90%, NB 84% and NN 82% for attached method, and 92% (SVM) followed by KNN (88%), NB (80.6%), and NN (79.2%) for detached method respectively. To enhance the efficiency of the detection systems, deep learning techniques such as MobileNet, DenseNet121, and a modified/proposed approach employing DenseNet121 were utilized. For MobileNet, the highest testing accuracy was found to be 89.4% for attached method. Similarly, in the detached leaf method, the highest achieved accuracy was 80.1%. The suggested approach entails converting images, reducing size, augmenting images, and employing DenseNet-121. The combined efficiency of all layers used was 98.97% and 98.1% for the attached and detached leaf image methods, respectively. This study introduces a novel approach for the identification and classification of growth stages in citrus canker, encompassing six distinct phases. Simultaneously, it quantifies temporal variations in the affected area of the disease within xx inoculated citrus leaves. Achieving an accuracy of 98.97%, the attached method demonstrates impressive performance metrics: macro precision of 97%, weighted precision of 99%, macro recall of 98%, weighted recall of 98%, macro F1 score of 97%, and weighted F1 score of 98%. The detached method similarly exhibits high accuracy at 98.1%, along with robust precision, recall, and F1 scores. Furthermore, the study proposes a mathematical model for predicting the disease's growth rate, offering valuable insights for disease management and prevention. The utilization of machine learning techniques in identifying and diagnosing various growth stages of citrus canker heralds a transformative era in citrus disease management. These techniques provide swift, accurate, and scalable solutions, contributing to enhanced disease control, agricultural sustainability, and global food security. The prospective trajectory of machine learning's progressive evolution and deployment within the domain of citrus canker detection and beyond significant promise for future scientific endeavors.