CollembolAI, a macrophotography and computer vision workflow to digitize and characterize samples of soil invertebrate communities preserved in fluid

• Published in: Methods in Ecology and Evolution
• Main Authors: Sys, Stanislav, Weißbach, Stephan, Jakob, Lea, Gerber, Susanne, Schneider, Clément
• Format: Article in Journal/Newspaper
• Language: English
• Published: Wiley 2022
• Subjects: Mite; Springtail
• Online Access: http://dx.doi.org/10.1111/2041-210x.14001; https://onlinelibrary.wiley.com/doi/pdf/10.1111/2041-210X.14001; https://onlinelibrary.wiley.com/doi/full-xml/10.1111/2041-210X.14001; https://besjournals.onlinelibrary.wiley.com/doi/pdf/10.1111/2041-210X.14001

Abstract Soils are inhabited by communities of tiny invertebrates that participate in the essential functions of soils. Characterizing those communities in terms of species diversity and species abundance is part of investigating soil functions and response to perturbation. Dozens to hundreds of specimens can be extracted from a sample that need to be sorted, counted and identified. It involves an enormous amount of time, straining the workflow of soil zoologists. Deep learning‐based computer vision approaches have become increasingly popular to monitor biodiversity, as they can be applied to detect, count and classify organisms. In this work, we present CollembolAI, an open‐source prototype for a computer vision workflow. It includes a hardware system for acquiring high‐resolution pictures of soil species samples in fluid and a deep learning‐based application (Faster R‐CNN with Slicing Aided Hyper Inference) to train and evaluate models for the detection and classification of animals on those pictures. We evaluated the workflow using a mix of specimens belonging to 12 species of springtail and mite, picked from our taxonomic collection. Specimens were photographed multiple times under various angle of views. The model was trained using 5671 views on specimens on 30 images and tested on 442 views from new specimens on six images. CollembolAI is affordable, simple to build and allows the rapid digitization of mesofauna samples. Our deep learning model achieved a Precision of 0.940, Recall of 0.918 and a mAP@0.5 (Pascal VOC) of 0.868. The model showed a lower Recall for one species, but was performant on all others. Our prototype offers an operational workflow for the creation of soil fauna picture datasets needed to develop efficient deep learning‐based classifiers. The applications are numerous, for example, collection digitization, soil biodiversity analysis and monitoring, or automatic assessment of mesofauna‐based bioindicators. Computer vision is a rapidly emerging tool to handle efficiently the complexity ...