AI-powered underwater imaging reveals multiscale occurrence and volume-based abundance patterns in coastal fishes

Long-term, high-frequency monitoring is increasingly required to detect ecological change in human-impacted coastal zones, yet fixed underwater imaging commonly produces relative counts that are difficult to standardize across time and sites. We analyzed a nine-year (2016–2024) time series from a cabled fixed camera deployed at the entrance of a Mediterranean harbor (Mallorca, western Mediterranean), imaging a small wreck at 8 m depth. An automated workflow combining image-quality filtering, Retinex correction, and deep-learning–based detection and classification generated continuous fish observations across 17 taxonomic categories. To improve comparability and ecological interpretability of camera-derived abundance signals, we developed two complementary indicator streams: (i) daily occurrence (presence-absence) from full-frame imagery to characterize seasonal and interannual structure, and (ii) daylight hourly fish density standardized by an estimated sampled water volume within a defined area of interest (individuals m −3 h −1 ). This volume-standardized density is intended as an index contingent on camera geometry and visibility conditions. Taxon-specific generalized mixed models and volume-based hurdle models related these indicators to environmental covariates (temperature, chlorophyll-a, irradiance, photoperiod, precipitation, and a sea-level–based flux proxy) and to a discrete post-2021 structural shift. Across taxa, results revealed pronounced season-dependent diel structure in volume-standardized density and a marked post-2021 shift in fish occurrence and/or density for several common coastal groups, with environmental effects varying by taxon and season. Overall, volume standardization provides a practical route to transform automated detections from fixed cameras into more comparable, indicator-oriented metrics for tracking abrupt change and environmentally driven variability in coastal fish assemblages. • Deep learning enables automated fish indicators from 9-year fixed-camera imagery. • Two indicators: daily occurrence and volume-standardized hourly fish density. • Volume standardization improves comparability across visibility and camera geometry. • Season-dependent diel structure emerges in volume-standardized fish density. • Several taxa show a coherent post-2021 step shift in occurrence and/or density.