Case study: From replica seabirds to koalas in forests

Updated: 31 October 2026 · Research synthesis & operational design

Sir Airffic Go’s checkpoint
Use controlled tests to evaluate measurement and field references to evaluate observation. Detection and abundance estimation remain separate stages.

2018: A benchmark with known totals

Hodgson and colleagues compared ground counts and drone imagery for ten replica seabird colonies with known totals. Image-based counts were more accurate and precise in that experiment. Stationary replicas are an important boundary: the study did not establish equivalent performance for moving animals or all forms of habitat occlusion.

2019: Validate detections against field locations

Corcoran and colleagues checked automated thermal detections against same-day ground observations of an already tracked koala population. This provided a reference for identifying genuine detections. False candidates remained part of the observation process. The implementation lesson is to preserve the distinction between locating candidate heat signatures and establishing an abundance estimate.

2020: Test the estimator after detection

A later comparison evaluated N-mixture and modified Horvitz–Thompson approaches, including detection, false detections and duplicates. The modified approach’s 95% intervals contained the reference counts in four test surveys. That result is validation on those data, not a guarantee of interval coverage for another species or site.

Airffic proposal: Use three acceptance stages

First test capture and counting against a known reference. Then combine independent field verification with surveys in the intended habitat to evaluate availability, misses and movement. Finally assess whether the design can detect the management-relevant change at an acceptable effort and cost. Define acceptance criteria before examining the results and revise the design when criteria are not met.

Show a decision record instead of one accuracy score

A useful report brings together species, place, season, independent test units, verified detections, omissions, false positives, deduplication, uncertainty and review effort. Many adjacent frames do not equal many independent field tests. Traceability to imagery, model versions, reviewers and estimation procedures makes later comparisons possible.

Evidence & further reading

Check the original study conditions, species and location before using a result in a plan. Do not transfer findings universally across species and settings.

← Back to the guide