Automated machine learning for identification of pest aphid species (Hemiptera: Aphididae)

Abstract

Effective crop protection requires accurate and rapid detection and identification of pest species. However, identification usually requires the involvement of experts. It would be useful to develop means to support pest identification. Although machine learning techniques have been applied to multiple fields including pest identification, expert knowledge for modeling was required to construct a high-accuracy model. In recent years, machine learning platforms that automatically construct a model are offered by IT firms. We tested whether automated machine learning using Google Cloud AutoML Vision was useful for identifying pest species. We trained machine learning models to identify aphids of three species—Aphis craccivora Koch, Acyrthosiphon pisum Harris, and Megoura crassicauda Mordivilko (Hemiptera: Aphididae)—sharing host plants and assessed accuracies. Models were constructed using 20, 50, 100, 200, and 400 images per species, with and without augmentation of training data volume by image inversion. The accuracy of identification increased with the number of training images and with the use of inverted images. Since the rates of correct identification were > 0.96 when the models were trained with 400 images per species with inversion, we consider automated machine learning to be useful for pest species identification.