Which camera-trap images to review? Estimating occupancy from a small part of a survey

Michael Tabak 

Context. Camera-trap surveys produce more images than people can review. Computer-vision classifiers such as SpeciesNet label animals but miss some species, so people still review images.

Aims. To find how many and which images should be reviewed to reliably estimate species occupancy, and how a project can tell when it has reviewed enough.

Methods. On four fully labeled camera-trap surveys (in the USA, South Africa, and Malawi), I simulated partial reviews by choosing capture events in five ways: at random, with animals, spread across cameras, least certain first (as in active learning), and by confirming the suggestions an occupancy model counts. I then tested a two-step procedure for human-in-the-loop review: review two capture events per camera; then, in rounds, confirm the most confident suggestion of each unconfirmed species-site-week (a species at a site in a week), retraining a model of the project's species after each round.

Key results. SpeciesNet's suggestions alone gave occupancy within a mean absolute error (MAE) of 0.06 to 0.12 of a full review's where it knew the fauna, but missed Canada lynx and nyala; in Malawi, the MAE was 0.24. Reviewing 25% of images at random gave an MAE of 0.16 to 0.24 with only reviewed images counted as detections. Confirming suggestions came closest; reviewing the least certain first did little better than random. In the procedure, each capture event reviewed confirmed 0.6 to 1.0 new species-site-weeks until 92 to 98% were confirmed, then fewer than 0.2. Stopping there, after 15 to 27% of the images (1.1 to 1.8 capture events per species-site-week), occupancy, counting only reviewed images, had an MAE of 0.04 or less, lynx included.

Conclusions. Confirming suggestions approaches a full review's estimates because occupancy models need species-site-weeks, not images; reviewing the least certain images first spends effort where estimates gain least.

Implications. Projects can nearly match a full review's occupancy estimates by reviewing two capture events per camera, then confirming one suggestion per species-site-week until a round confirms fewer than one new species-site-week per five capture events. Researchers can expect to review about 1.5 capture events per species-site-week that the first step shows.