Abstract

A standard method of coral reef surveying is the collection of images along a transect typically ranging from 15-30 m which are then annotated to extract benthic composition. With the application of machine learning, larger scale phototransects are increasingly common collecting an order of magnitude more images per dive (hundreds) which are then processed through an automated image analysis pipeline. How does this difference in scale of reef data collection affect assessment of coral reef composition? Here, we investigate coral reef benthic data collected along the north coast of Timor-Leste, a small island country in Southeast Asia, using a range of methods, including in situ 15 x 2 m belt transects, 30 m phototransects, and 1.5-2 km phototransects. Comparison between the 30 m and kilometre-scale data shows that shorter transects generally overestimate hard coral reef cover. This could be from bias towards transect placement in high coral areas or that the scale of reef patchiness is on a similar scale. Transect placement also impacts time series data captured by in situ 15 x 2 m belt transects. The kilometre-scale benthic data was also modelled using a linear mixed-effect model with environmental drivers (wave exposure, distance to river), human population density, and reef structural complexity. Coral cover increased with an interaction between wave exposure and reef structure. A multi-variate model was also used to explore reef community relationships with the same predictors. While scale of data collection should be determined by the research objective, longer transects likely capture reef heterogeneity better for monitoring change through time.