Confidence Thresholds: Towards an Automated Workflow for Unmarked Wildlife Population Modelling With Camera Trap Data
TL;DR
The ensemble approach provides a tool for ecologists who study underrepresented fauna that occur in less frequently surveyed locations and habitats, and provides a tool for ecologists who study underrepresented fauna that occur in less frequently surveyed locations and habitats.
Abstract
ABSTRACT While ecologists are exploring the capacity of artificial intelligence to automatically identify wildlife in camera trap and other remote sensing data, it has yet to be commonly adopted in workflows that produce density or abundance estimates and thus many image datasets remain under‐analysed. There is potential to enhance image classification with deep learning ensembles which remain uncommon in wildlife studies, and to synergistically connect the output with population modelling in automated workflows. Here we present a repeatable workflow that connects these elements in a case study of macropods (Bennett's wallaby Notamacropus rufogriseus and Tasmanian pademelon Thylogale billardierii) in Tasmania, Australia. We analysed a pre‐existing camera trap dataset with no contemporaneous field‐collected data. We used readily available computer vision algorithms to train a deep learning ensemble to classify the most common fauna in the dataset, achieving an overall classification accuracy of 94%. Once reviewed, we connected the output to a hierarchical abundance estimation process via a fully scripted workflow designed for repeat application, including the derivation of covariate data from spatial proxies. We then tested the use of classification confidence levels to subset raw (unverified) automated detections, by replacing the verified detection history with a series of automatically generated versions. We found that a detection history derived from a confidence level of 0.75 produced a highly comparable abundance estimate. Our analysis suggested that correctly identifying empty sampling occasions was more important to the estimation process than correctly identifying all occasions with detections. The potential to feed automated detection output that is subset with a classification confidence threshold into a modelling workflow, in place of costly manual verification, represents a high yield opportunity that is worthy of further investigation. In addition, our ensemble approach provides a tool for ecologists who study underrepresented fauna that occur in less frequently surveyed locations and habitats.