Development of an intelligent field investigation system for Liriomyza using SeResNet-Liriomyza for accurate identification

• Published in: Computers and electronics in agriculture 2023-11, Vol.214, p.108276, Article 108276
• Main Authors: Li, Hang, Liang, Yongxuan, Liu, Yongjian, Xian, Xiaoqing, Xue, Yantao, Huang, Hongkun, Yao, Qing, Liu, Wanxue
• Format: Article
• Language: English
• Subjects: Alien invasive species; Mobile application; Portable image-capturing device; SeResNet-Liriomyza model; SFPN
• ISSN: 0168-1699; 1872-7107
• DOI: 10.1016/j.compag.2023.108276

•Intelligent field investigation system to identify 5 Liriomyza species developed.•Magnifying lens/rotatable carrier assembled outside smartphone camera for images.•SeResNet-Liriomyza model with SFPN proposed for species detection.•Model showed an average recognition accuracy of 99.88%.•Proposed system a fast, convenient, and accurate tool for investigating Liriomyza. Liriomyza (Diptera: Agromyzidae) is genus of insects found worldwide. Among the species in this genus, L. sativae, L. huidobrensis, L. trifolii, and L. bryoniae are the four most globally invasive, with L. chinensis being native to China. Real-time and accurate monitoring is a prerequisite for reducing the spread and crop damage caused by this pest. Owing to the small size of Liriomyza, specialized technicians using a microscope are needed to identify them. We developed an intelligent field investigation system to conveniently and accurately identify these five species. The system includes a portable image acquisition device based on a smartphone, a deep learning-based image recognition model for Liriomyza, and a mobile platform-based investigation application. To capture magnified images of Liriomyza, we designed a magnifying lens and a rotatable carrier assembled outside the smartphone camera. To address the problem of local blurring in images captured by smartphones, all images were sharpened to enhance their features. To accurately distinguish these five species from similar species, we developed a SeResNet-Liriomyza model, which embeds spatial feature pyramid networks (SFPN) in the SeResNet model to prevent the false positives caused by different sizes of input images and generate new feature maps favorable for classification by combining global and local image features. Compared with the excellent classification models, DenseNet-201, ResNeXt-50, MobileNet-v3, EfficientNet-B4, and ResNeSt-101, SeResNet-Liriomyza showed excellent recognition performance, with an average recognition accuracy of 99.88%. We designed and developed a mobile application (app) for the intelligent investigation of Liriomyza. The app includes functions such as intelligent investigation, intelligent recognition, a biological dictionary, and investigation result retrieval. The proposed intelligent field investigation system provides investigators with a fast, convenient, and accurate tool for investigating Liriomyza, a pest that can cause serious damage to agriculture in the field.