Automated quantification of ecological interactions from video

Ecological interactions, such as predation, are fundamental events that underlie the flow and distribution of energy through food webs. Yet, directly measuring interaction rates in nature and how they vary across space and time remains a core challenge in ecology.

To address this, we developed a machine learning pipeline that combines object detection, tracking, behavioural classification, and bias correction to automatically quantify interaction rates from video. We applied this pipeline to estimate feeding rates of the planktivorous reef fish Chromis multilineata in situ.

We show that the tool generates precise, unbiased estimates of planktivory at fine temporal and spatial scales, and use it to reveal how feeding rate changes in response to predator presence and proximity to refuge.

More broadly, this method provides a scalable, non-invasive framework for quantifying interactions in the wild, enabling new opportunities to test ecological theory and quantify energetic fluxes in nature.