Taxonomic Classification of Ants (Formicidae) from Images using Deep Learning

Abstract

The well-documented, species-rich, and diverse group of ants (Formicidae) are important 1 ecological bioindicators for species richness, ecosystem health, and biodiversity, but ant 2 species identification is complex and requires specific knowledge. In the past few years, 3 insect identification from images has seen increasing interest and success, with processing 4 speed improving and costs lowering. Here we propose deep learning (in the form of a 5 convolutional neural network (CNN)) to classify ants at species level using AntWeb 6 images. We used an Inception-ResNet-V2-based CNN to classify ant images, and three 7 shot types with 10,204 images for 97 species, in addition to a multi-view approach, for 8 training and testing the CNN while also testing a worker-only set and an AntWeb 9 protocol-deviant test set. Top 1 accuracy reached 62% -81%, top 3 accuracy 80% -92%, 10 and genus accuracy 79% -95% on species classification for different shot type approaches. 11The head shot type outperformed other shot type approaches. Genus accuracy was broadly 12 similar to top 3 accuracy. Removing reproductives from the test data improved accuracy 13 only slightly. Accuracy on AntWeb protocol-deviant data was very low. In addition, we 14 make recommendations for future work concerning image threshold, distribution, and 15 quality, multi-view approaches, metadata, and on protocols; potentially leading to higher 16 accuracy with less computational effort. BOER & VOS

Pixel Intensity Resemblance Measurement and Deep Learning Based Computer Vision Model for Crack Detection and Analysis

“…The Xception model uses depth-wise separable convolutions and works as shown in Figure 8 a–c, redrawn from [ 27 ]. A general convolution step makes the spatial-wise and channel-wise computation in one single step.…”

“…In [ 24 ], to inspect the severity of the crack, the authors used crack magnifier. The deep learning models such as VGG-16, ResNet50 and Inception ResNet-V2 are discussed in [ 25 , 26 , 27 ]. Paramanandham et al [ 28 ] discussed about concrete crack detection using various deep learnbing models.…”

Deep learning as a tool for ecology and evolution

“…It is not surprising then that identification of individuals or species from image, video, and sound data is the most common use of deep learning in the field (Figure 3). These efforts already span many taxa, from bacteria (Kotwal et al, 2021), through protozoans (Hsiang et al, 2019), plants (Carranza‐Rojas et al, 2017; Schuettpelz et al, 2017; Unger et al, 2016; Younis et al, 2018) to insects (Boer & Vos, 2018; Hansen et al, 2020; Marques et al, 2018; Valan et al, 2019) and vertebrates (Norouzzadeh et al, 2018; Villon et al, 2018), both extant and fossil (de Lima et al, 2020; Liu & Song, 2020; Miele et al, 2020) and at scales ranging from local to global.…”

“…It is not surprising then that identification or classification of individuals or species from image, video, and sound data is the most common use of deep learning in the field (Figure 3). These efforts already span many taxa, from bacteria (Satoto et al, 2020), through protozoans (Hsiang et al, 2019), plants (Unger et al, 2016;Carranza-Rojas et al, 2017;Schuettpelz et al, 2017;Younis et al, 2018) to insects (Marques et al, 2018;Boer and Vos, 2018;Valan et al, 2019;Hansen et al, 2020) and vertebrates (Villon et al, 2018;Norouzzadeh et al, 2018), both extant and fossil (Liu and Song, 2020;Miele et al, 2020;de Lima et al, 2020) and at scales ranging from local to global. Intensifying efforts to digitize natural history collections provide troves of image data that can be used for this purpose (Smith and Blagoderov, 2012).…”