Proximal detecting invertebrate pests on crops using a deep residual convolutional neural network trained by virtual images

Abstract

Detecting invertebrate pests on crops at early stages is essential for pest management. Traditionally, traps were used to sample pests and then human experts undertook classification and counting to estimate the levels of infestation, which is subjective, error-prone and labour intensive. Recently, semi-automatic pest detection is possible by using computer vision technologies to classify and count pest samples in laboratories or insect traps, however, the decision made by the laboratory-based or trap-based approaches are still too late for more optimised pest management decisions. Today, precision agriculture needs detection of pests on crops so that real-time actions can be taken or optimised decision can be made based on accurate information of time and location pest occurs. In this study, we used computer vision and machine learning technologies to detect invertebrates on crops in the field. We first evaluated the performances of the state-of-art convolutional neural networks (CNNs) and proposed a standard training pipeline. Facing the challenge of rapidly developing comprehensive training data, we used a novel method to generate a virtual database which was successfully used to train a deep residual CNN with an accuracy of 97.8% in detecting four species of pests in farming environments. The proposed method can be applied to a robotic system for proximal detection of invertebrate pests on crops in real-time.

Keywords

• Insect detection
• Deep convolutional neural network
• Precision agriculture