Automatic pest identification system in the greenhouse based on deep learning and machine vision

Monitoring and understanding pest population dynamics is essential to greenhouse management for effectively preventing infestations and crop diseases. Image-based pest recognition approaches demonstrate the potential for real-time pest monitoring. However, the pest detection models are challenged by the tiny pest scale and complex image background. Therefore, high-quality image datasets and reliable pest detection models are required. In this study, we developed a trapping system with yellow sticky paper and LED light for automatic pest image collection, and proposed an improved YOLOv5 model with copy-pasting data augmentation for pest recognition. We evaluated the system in cherry tomato and strawberry greenhouses during 40 days of continuous monitoring. Six diverse pests, including tobacco whiteflies, leaf miners, aphids, fruit flies, thrips, and houseflies, are observed in the experiment. The results indicated that the proposed improved YOLOv5 model obtained an average recognition accuracy of 96% and demonstrated superiority in identification of nearby pests over the original YOLOv5 model. Furthermore, the two greenhouses show different pest numbers and populations dynamics, where the number of pests in the cherry tomato greenhouse was approximately 1.7 times that in the strawberry greenhouse. The developed time-series pest-monitoring system could provide insights for pest control and further applied to other greenhouses.