Artificial intelligence annotations capture ecological patterns in sponge-dominated temperate mesophotic ecosystems

Manual image annotation is a major bottleneck for long-term benthic community monitoring, particularly as recent advances in marine imaging have generated volumes of images and video that far exceed the capacity for manual processing. Sponge-dominated temperate mesophotic ecosystems (TMEs) are widespread, but their ecological importance remains poorly understood, with their complex morphological characteristics creating challenges for automated image annotation. Here, we evaluated the performance of an existing AI-based annotation framework (CoralNet) for annotating images collected from a long-term monitoring program of sponge-dominated TMEs at Lough Hyne, Ireland (2018-2024). A classifier was trained on 453 manually annotated images and validated against an independent test set of 249 images. Classifier outputs were compared to expert annotations using point-level classification accuracy and taxonomic confusion patterns, percentage-cover estimates, community structure and temporal community trends. Point-level classification accuracy was 59 % (rising to 68 % at coarser taxonomic resolution), with higher performance for visually distinct and common taxa, and lower accuracy for rare or morphologically complex forms. Despite this, community-level compositional similarity between AI and manual annotations remained high (Bray-Curtis similarity ranging from 76 % at species level to 89 % at kingdom level) and showed no detectable drift over the six year monitoring period (mean similarity 87 % ± 7 % SD across all sites and sampling dates), indicating preservation of the ecological signals throughout the time series. Together, these results demonstrate that AI-assisted annotation can provide a reliable and scalable complement to manual image annotation for monitoring benthic community dynamics in sponge-dominated temperate mesophotic ecosystems, within the limitations identified here.