Machine Learning Assisted Camera Trap Image Review
Michael Pearcy
Supervisor: Dr Emma Gray
Motion triggered ‘camera traps’ are widely used as a non-invasive and relatively low-cost tool for monitoring wildlife. A key limitation is the production of potentially millions of individual images. following deployment. The manual review and annotation of these images is laborious and tedious. Machine learning (ML) models have been widely developed to assist in the automation of camera trap image review and is a rapidly evolving field. This application of ML requires models trained specifically to recognise the fauna/environment of the project area. This often requires substantial technical computer experience – there is a lack of user-friendly capability to train custom ML models. The aim of this project was to investigate the feasibility of training and applying an open-source, user-friendly ML model to identify small mammals. Four ML workflows were initially investigated for suitability – Sherlock, Wildlife Insights, Mega Efficient Wildlife Classifier (MEWC), and MegaDetector (MD). Privacy/IP concerns (Wildlife Insights) and substantial technical requirements (Sherlock, MEWC) led to the selection of MD as the most suitable ML model. MD’s base model was trained to ID relevant animals using 34,098 manually reviewed camera trap images from an ongoing small mammal survey (mid-northern QLD) (11,719 empty images, 20,790 Rattus villosissimus, 476 Felis catus, 430 Sminthopsis douglasi, 320 Sminthopsis macroura, 262 Sus scrofa, 127 Planigale spp., 2 Leggadina forresti, 2 Mus musculus). EcoAssist (v4.3) (Addax Data Science); a user-friendly interface for utilising and training MD, was leveraged to manually annotate the training data and produce a custom MD model. The custom MD model developed is yet to be thoroughly applied and tested for ID accuracy. This MD model will likely have the capacity to semi-automate the review process by removing empty images and locating and identifying animals, greatly reducing workload. Furthermore, the accuracy of the model may be improved by compiling more training data, particularly of animals with fewer example images.
