Psyllid Detector: A Web-Based Application to Automate Insect Detection Utilizing Image Processing and Deep Learning

Abstract

Highlights An AI-enabled web application was developed to classify insects on sticky traps. The AI models were trained on images exposed to real environmental conditions. Automated insect identification can enhance integrated pest management. This technology could be used in other environmental applications. Abstract. The Huanglongbing (HLB) disease has had a devastating impact on several citrus production areas across the world. The disease is associated with three phloem-limited bacteria of the Candidatus Liberibacter group, which can be vectored by citrus-infesting psyllids. In Florida, it is vectored by Diaphorina citri, also known as Asian citrus psyllid (ACP), and in South Africa, by a native species, Trioza erytreae, or African citrus triozid (ACT). For better disease management, continued monitoring of these vectors is highly recommended. Furthermore, it is important to keep track of other related psyllids that reside in citrus environments (e.g., the South African native Diaphorina specimens DN, such as D. virgata, D. punctulata, and D. zebrana), as well as natural predators of psyllids (e.g., lady beetles, LB). Among the tools used to monitor pests in the field, sticky traps are one of the most common and widely used. Analyzing the traps and identifying the specimens of interest usually require specialized personnel, which can be costly and time-consuming. In an effort to automate this process, a web application called “Psyllid Detector” was developed. Three YOLOv7-based object detection algorithms (YOLOv7, YOLOv7-ResNet50, and YOLOv7-ResNeXt50) were selected to train six models using two datasets (South Africa and Florida datasets) composed of yellow sticky trap pictures taken from several citrus orchards. Three models were trained to detect ACP and LB using the Florida dataset, and three were trained to detect DN and ACT in the South Africa dataset. The models trained with the original YOLOv7 for both datasets showed the best overall results. For example, for the ACT detection in the South Africa dataset, precision, recall, and F1 scores of 90%, 70%, and 79% were achieved, respectively. By automating the process of insect detection, farmers can save time and resources and make informed decisions about pest control measures. Keywords: ACP, ACT, Huanglongbing, Machine learning, Object detection, YOLO.