Sophia Scipione, & Dr. Geoff Dilly
Monitoring ecosystems has always had the challenge of balancing data collection versus human analysis. The sheer volume of data required to represent ecosystem diversity and abundance accurately often exceeds the human capacity for manual analysis. One solution involves capturing field images and using machine-learning image recognition algorithms to make automatic annotations. Our approach involved utilizing a machine-learning tool called CoralNet, which was purposefully crafted to bolster marine research through collaborative image data annotation on benthic organisms. We developed a customized CoralNet application with a specialized label set and site-specific image training set. Our study focused on two sites along the Northeastern coast of Santa Rosa Island within the Channel Islands National Park: Bechers Bay and Skunk Point. From the years 2016 to 2023, we collected photos along six 20-meter transect lines at each site and grouped images by site, year, and zone to validate annotation accuracy. In each group, 20 images underwent manual annotation, while 60 were annotated automatically. We then compared the fidelity of automated analyses to manual annotations using regression analysis. While photo transects offer a data-rich environment snapshot, manual annotation is time intensive. When supported by human training and validation, automated annotation allows us to efficiently process large volumes of data with a high level of specificity.