Performance Evaluation of Deep Learning Architectures for Detecting Citrus Crop HLB Abnormalities

Abstract

Early detection of diseases like Huanglongbing (HLB), also known as citrus greening, is crucial for maintaining citrus crop health and productivity. HLB is a devastating disease that significantly impacts global citrus production, and effective, low-cost detection methods are needed for timely intervention. Convolutional neural networks (CNNs) have emerged as powerful tools for automated disease detection through computer vision, but distinguishing HLB from other citrus disorders remains a challenge. Traditional diagnostic methods, such as quantitative real-time polymerase chain reaction, are expensive and require large datasets for training CNN models. One of the promising solutions to the problem of a lack of labeled images is the utilization of transfer learning and pre-trained CNNs, which can still produce a good result with a smaller dataset. The study aims at comparing the performance of the series (AlexNet, VGG19) and the directed acyclic graph (DAG) CNN architectures, such as ResNet50, and Inception-V3, in three-class classification of Huanglongbing (HLB), healthy leaves, and 10 other citrus abnormalities. To do this, a dataset of 955 color images of Citrus leaves from the southern part of India was chosen, where 10-fold cross-validation was done. The results found that all CNN models could reach the sensitivity of HLB detection above 98%. Among them, VGG19 yielded the highest sensitivity in all experiments, even though it had fewer parameters than Inception-V3. These findings indicate that models with more parameters are capable of dealing with the problems of small datasets, thus allowing transfer learning to be effective for HLB detection. This work reveals the selection of CNN models for the detection of disease in citrus, which is a cost-effective and portable solution that can aid small-scale farmers in low-income areas to lessen crop losses.