Computer vision based taxonomic classification of Insecta using deep learning models

Abstract

Current taxonomic identification methods of Insecta rely heavily on human vision, meaning the conclusions experience error caused by human bias. Computer vision has been proposed as a bias-reducing alternative. For this study 16 Convolutional Neural Networks (CNNs) were trained using a data set of 292,197 images of the taxonomic class Insecta with order-level labels generated using DNA barcoding. Two different data splitting methods were used to create four unique training and testing data sets. Four CNN network designs were generated, and each design was trained once per data set. Models were evaluated using Matthew’s Correlation Coefficient (MCC), precision, recall and F1 scores. Results for the MCC score ranged from 0.2066 to 0.9180. Results for the precision, recall and F1 scores ranged from 0.2470 to 0.8933, 0.2253 to 0.8610 and 0.1436 to 0.8591, respectively. Among all the hyperparameters evaluated, class size was found to have the greatest influence on model performance.