Automatic Pest Identification System Based on Deep Learning and Machine Vision in the Greenhouse

Abstract

Identifying pest populations automatically in greenhouses is a challenging task. Image-based pest identification approaches provide a power tool with real-time pest information to facilitate agricultural management. Deep learning has shown the potential of image processing for automatic pest identification and counting. To build an accurate deep learning model for automatic pest identification, high-quality image datasets are indispensable. This study develops a trapped system with yellow sticky paper and LED light for automatic pest image collection and proposes an improved YOLOv5 model for pest recognition. The system is evaluated in a cherry tomato greenhouse and a strawberry greenhouse for 40 days of continuous monitoring. There are six diversities of pests in the greenhouse, including tobacco whiteflies, leaf miners, aphids, fruit flies, thrips, and houseflies. The proposed improved YOLOv5 model employed copy-pasting data augmentation obtains an average recognition accuracy of 96%, and demonstrates superiority in nearby pests identification over than the original YOLOv5 model. The pest number in the cherry tomato greenhouse is approximately 1.7 times that the strawberry greenhouse, and the trends of pest populations are also different in the two greenhouses. The pest monitoring system effectively provides dynamic changes in real-time, which provides insights for pest control and could be further applied in other greenhouses.

Keywords: Pest trapping system, Small pest detection, YOLOv5, Pest population dynamics, Greenhouse