A deep learning and Grad-Cam-based approach for accurate identification of the fall armyworm (Spodoptera frugiperda) in maize fields

• Published in: Computers and electronics in agriculture 2022-11, Vol.202, p.107440, Article 107440
• Main Authors: Zhang, Haowen, Zhao, Shengyuan, Song, Yifei, Ge, Shishuai, Liu, Dazhong, Yang, Xianming, Wu, Kongming
• Format: Article
• Language: English
• Subjects: EfficientNet; Grad-CAM; Identification system; knowledge Distillation; Spodoptera frugiperda
• ISSN: 0168-1699; 1872-7107
• DOI: 10.1016/j.compag.2022.107440

•1. A large-scale image dataset of 36 maize pest species was constructed.•2. A two-stage classification model based on EfficientNet was proposed.•3. The proposed model provided the superior and low-cost identification ability.•4. Real-time solutions aiming at fall armyworm identification were developed. Population monitoring and early warning to the world's important invasive pest, the fall armyworm (FAW), Spodoptera frugiperda (J.E. Smith), is a basis for its management in the field. Deep learning is a useful intelligence approach for identifying Spodoptera frugiperda; however, it has a few limitations, such as too many model parameters, and the mismatch between training in simple backgrounds and complex real scenarios. Herein, we established a two-stage classification approach called MaizePestNet, which is based on EfficientNet, to address these concerns. It combines the Gradient-weighted Class Activation Mapping (Grad-CAM) technique for removing background interference with the knowledge distillation strategy for reducing model parameters and size. A dataset of adults and larvae of 36 common maize field pests, including Spodoptera frugiperda, was collected. The accuracy, precision, recall, and F1-score of the new model were acquired as 93.85%, 93.56%, 96.23%, and 93.85% through training and testing on the dataset. The new model’s accuracy was 5.65% and 9.04% higher than that of the ResNet101 and DenseNet161 models, respectively; the model size was 90% lower and 85% lower than the ResNet101 and DenseNet161 models, respectively. A WeChat applet and online real-time identification system (http://migrationinsect.cn/maizepestidentification) based on our classification model were developed for practical applications to accurately identify Spodoptera frugiperda adults and larvae. This study gave public an accurate, efficient, and straightforward method for detecting Spodoptera frugiperda in maize fields, and the findings can guide and assist plant protection agencies and farmers in monitoring, early warning, and control activities of the pest.