Auto‐identification of two Sitophilus sibling species on stored wheat using deep convolutional neural network. Issue 5 (10th February 2022)

• Record Type: Journal Article
• Authors: Yang, Haiying; Zhao, Huiyi; Zhang, Dufeng; Cao, Yang; Teng, Shyh Wei; Pang, Shaoning; Zhou, Xiaoguang; Li, Yanyu
• Is Part Of: Pest management science. Volume 78:Issue 5(2022)
• Page Start: 1925
• Page End: 1937
• Publication Date: 2022-02-10
• Subjects: Sitophilus oryzae -- Sitophilus zeamais -- insect detection -- multilayer convolutional structure -- region proposal network -- wheat background; Pests -- Control -- Periodicals; Pesticides -- Periodicals; 632.9
• DOI: 10.1002/ps.6810
• Languages: English
• ISSNs: 1526-498X

Abstract

Abstract: BACKGROUND: Sitophilus oryzae and Sitophilus zeamais are the two main insect pests that infest stored grain worldwide. Accurate and rapid identification of the two pests is challenging because of their similar appearances. The S. zeamais adults are darker and shinier than S. oryzae in visible light. Convolutional neural network (CNN) can be applied for the effective differentiation due to its high effectiveness in object recognition. RESULTS: We propose a multilayer convolutional structure (MCS) feature extractor to extract insect characteristics within each layer of the CNN architecture. A region proposal network is adopted to determine the location of a potential pest in the wheat background. The precision of classification and the robustness of bounding box regression are increased by including deeper layer variables into the classification and bounding box regression subnets, as well as combining loss functions softmax and smooth L1. The proposed multilayer convolutional structure network (MCSNet) achieves the mean average precision of 87.89 ± 2.36% from the laboratory test, with an average detection speed of 0.182 ± 0.005 s per test. The model was further assessed with the field trials, and the obtained accuracy was 90.35 ± 3.12%. For all test conditions, the average precision for S. oryzae was higher than that for S. zeamais . CONCLUSION: The proposed MCSNet model has demonstrated that it is a fast and accurate method for detecting sibling species from visible light images in both laboratory and field trials. This will ultimately be applied for pest management together with an upgraded industrial camera system, which has been installed in over 100 000 grain depots of China. Abstract : Deep convolutional neural network is utilized to detect two Sitophilus sibling species on stored wheat automatically from visible light images, achieving accurate and rapid diagnosis. © 2022 Society of Chemical Industry.