When and how to automate image analysis for wildlife monitoring? Guidelines and lessons from a worked example of seabirds in a dynamic coastal environment

Capsule:
Camera and automated image analysis technologies offer one solution to the challenges of seabird monitoring; however, their application is often not straightforward and there is currently unlikely to be a universally applicable method across sites and species.

Aims:
Evaluate a range of time-lapse camera and automated image analysis methods to monitor seabirds in various coastal settings. Provide recommendations and guidelines on the minimum equipment needed to obtain representative information from time-lapse images.

Methods:
To establish an efficient and effective method to generate count and/or presence/absence data from images of seabirds, we assess the application of two image analysis methods: (1) bespoke and site-specific object detection, and (2) open-source and generalised object detection. Using a worked example of time-lapse photography of three coastal seabird roosts, we compare data from each image analysis method, evaluating performance and overall success relative to manually collected data from the same images.

Results:
When detecting seabird presence/absence, we found no consistency in image analysis performance (3.2–94.7% correct detection of true presence; 9.8–98.2% correct detection of true absence). When generating counts, the performance of the different approaches varied (6.3–37.5% of automated counts were correct, relative to manual counts). Notably, whilst the bespoke method was the most successful when generating counts, both object detection methods consistently underestimated seabird abundance.

Conclusions:
Bespoke, site-specific analyses could be required where count data are needed, whereas open-source generalized detection methods may be suitable where presence/absence data are sufficient to meet study aims. Automating image analysis is unlikely to remove the need for manual analysis, whether to produce training data sets, validate output, and/or compliment observations. By offering lessons learnt from a worked example and resultant recommendations, this study informs future projects aiming to use inexpensive and accessible image technologies for ecological research.