Computers and Electronics in Agriculture 179 (2020) 105836 Available online 1 November 2020 0168-1699/© 2020 Elsevier B.V. All rights reserved. Detection and classifcation of soybean pests using deep learning with UAV images Everton Castel˜ ao Tetila a, e, * , Bruno Brandoli Machado b , Gilberto Astolf b, f , Nícolas Alessandro de Souza Belete c, e , Willian Paraguassu Amorim a , Antonia Railda Roel e , Hemerson Pistori b, e a Universidade Federal da Grande Dourados, Dourados, Mato Grosso do Sul, Brazil b Universidade Federal de Mato Grosso do Sul, Campo Grande, Mato Grosso do Sul, Brazil c Universidade Federal de Rondˆ onia, Cacoal, Rondˆ onia, Brazil e Universidade Cat´ olica Dom Bosco, Campo Grande, Mato Grosso do Sul, Brazil f Instituto Federal de Educaç˜ ao, Ciˆ encia e Tecnologia de Mato Grosso do Sul, Campo Grande, Mato Grosso do Sul, Brazil A R T I C L E INFO Keywords: UAV Remote sensing Soybean pests Precision agriculture Deep learning ABSTRACT This paper presents the results of the evaluation of fve deep learning architectures for the classifcation of soybean pest images. The performance of Inception-v3, Resnet-50, VGG-16, VGG-19 and Xception was evaluated for different fne-tuning and transfer learning strategies over a dataset of 5,000 images captured in real feld conditions. The experimental results showed that the deep learning architectures trained with a fne-tuning can lead to higher classifcation rates in comparison to other approaches, reaching accuracies of up to 93.82%. In addition, deep learning architectures outperformed traditional feature extraction methods, such as SIFT and SURF with Bag-of-Visual Words approach, the semi-supervised learning method OPFSEMImst, and supervised learning methods used to classify images, for example, SVM, k-NN and Random Forest. The results indicate that architectures evaluated can support specialists and farmers in the pest control management in soybean felds. 1. Introduction Vegetable soybean (Glycine max [L.] Merrill) is an oilseed with good nutritional profle and important to the world economic participation. The nutritional quality of vegetable soybean is determined by its content of protein, unsaturated fatty acid, minerals, vitamins, isofavone and other trace nutrients in the fresh seeds (Hou et al., 2011). From sowing to harvest, soybean cultivation is subject to the attack of defoliant pests such as insects and mollusks. Sampling methods, such as drop cloth, scanning net, visual plant examination, soil sampling and, more recently, pheromone-trapped have been employed to monitor the levels of pest control action in the soybean felds (Corrˆ aa-Ferreira et al., 2012). Early detection of pests allows a more effcient application of pesticides, since inputs can be applied to the right quantity and locations, thus reducing production costs and the environmental impact resulting from the application of pesticides in the total area, in addition to contributing to human health and food safety (Tetila et al., 2019b). As an alternative to manual sampling methods, technological in- novations have helped control pests and increase food production in the feld. UAVs equipped with high resolution cameras in data collection- missions are able to fy over a plant a few meters away and capture images rich in detail, which has helped to monitor the cultivation and harvesting of entire agricultural properties, with the aid of precision agriculture. Furthermore, the high cost of chemicals associated with low ecological impact actions lead to better precision agriculture practices. Thus, the use of UAVs in feld crops has been considered an important tool to detect problems in the feld, allowing experts and farmers to make better management decisions. In recent years, several neural network architectures have become popular due to impressive results in image classifcation and problem detection. Keyvan and Jafar (2013) proposed an artifcial neural network (ANN) with a 3 layers for identifcation of the insect Lepidop- tera Spodoptera exigua from other species of pests. Similarly, an ANN was trained by Leow et al. (2015) for the classifcation of Copepod species - a * Corresponding author. E-mail addresses: evertontetila@ufgd.edu.br (E.C. Tetila), brunobrandoli@gmail.com (B.B. Machado), gilberto.astolf@ifms.edu.br (G. Astolf), nicolas.belete@ unir.br (N.A.S. Belete), willianAmorim@ufgd.edu.br (W.P. Amorim), arroel@ucdb.br (A.R. Roel), pistori@ucdb.br (H. Pistori). Contents lists available at ScienceDirect Computers and Electronics in Agriculture journal homepage: www.elsevier.com/locate/compag https://doi.org/10.1016/j.compag.2020.105836 Received 17 July 2018; Received in revised form 29 May 2020; Accepted 13 October 2020