Abstract
Identifying benthic macroinvertebrates is a skill that requires time to master, resulting in fewer taxonomic experts and leading in part to the taxonomic impediment. The advancements of optical imaging and artificial intelligence (AI) technologies could prove useful in bridging this gap. Many researchers have turned to AI and machine learning (ML) techniques for identification of a wide range of living organisms, from algae to charismatic species, with great success. Recent work using ML or deep learning techniques, such as convolutional neural networks (CNNs) can achieve high accuracies in identifying various taxonomic groups of aquatic benthic macroinvertebrates. However, there are a few knowledge gaps in this application space. First, we questioned how image quality and quantity influence prediction accuracy. Second, we questioned how transferable are deep learning models across image domains, i.e. different courses of images? Here, we leveraged CNNs to accurately identify 23 families of the benthic macroinvertebrate orders Ephemeroptera, Plecoptera, and Trichoptera (EPT). Our CNN model achieved >99% accuracy on both EPT datasets with and without background. We found that minimal images are needed for accurate prediction when training and prediction occurs within the same image domain. However, we show that deep-learning CNN models for invertebrate taxonomic identification is only accurate when utilizing imagery collected within the same imaging domain as the model’s training imagery. We address the accessibility of these technologies in their current state and recommend future research be aimed at making these technologies more accessible to all researchers regardless of their taxonomic training.

This publication has 28 references indexed in Scilit: