Abstract
Citrus Huanglongbing (HLB) poses a significant threat to the profitability of the citrus industry worldwide. In traditional agricultural practices, manually identifying citrus trees infected with HLB based on certain leaf characteristics is time-consuming, subjective, and inefficient. The initial automatic identification of citrus Huanglongbing (HLB) relies on traditional image processing and machine learning algorithms, exhibiting low accuracy and slow processing speed. In order to enhance both the detection accuracy and speed, researchers have introduced deep learning methods based on neural networks for the identification of citrus HLB. However, the neural network models currently used for citrus leaf HLB identification have large parameter sizes, high deployment costs, and require high computational power, making them unsuitable for deployment on edge devices for field detection. Therefore, in order to promptly detect and address diseased plants, improve farmers' agricultural operational efficiency, ensure the accessibility of deep learning in small-scale agriculture, and address the need for cost-effective measures, there is an urgent need for a low-cost deep learning framework. Therefore, we compared the performance of several commonly used deep convolutional neural networks in industry for citrus Huanglongbing (HLB) detection. We constructed image classification networks based on AlexNet, ResNet, MobileNet-V1, and MobileNet-V3, and evaluated the network models based on model size, parameter count, and classification performance. As a result, we proposed a deep learning-based method for detecting citrus HLB. This method has a small model parameter count, low computational cost, fast detection speed, and high detection accuracy. It can be deployed on edge devices or other embedded devices. This method has a small model parameter count, fast detection speed, and high accuracy. The classification task is achieved by training the overall feature extraction network and the classification network at the network's tail on the constructed training set. The actual detection results show that the detection accuracy for healthy citrus leaves reaches 99.02%, and for HLB-infected leaves, the detection accuracy reaches 99.07%. The overall accuracy is 99.04%. Both recall and precision rates are excellent, meeting the precision requirements for on-site detection.