Zooplankton play a key role in energy transfer and biogeochemical cycling in coral-associated coastal ecosystems, yet their short-term variability remains difficult to resolve using conventional net-based sampling. In this study, we conducted a 153-day high-frequency in situ imaging survey from 1 July to 4 December 2022 in the Dongshan coral habitat, near the northern distributional limit of reef-building corals along the western Pacific coast. An Imaging Plankton Probe was deployed at a fixed station and operated at 3 Hz, generating approximately 860,000 regions of interest. Images were classified using an EfficientNet-B3 convolutional neural network, followed by supervised correction and quantification. The resulting time series revealed pronounced taxon-specific variability in the mesozooplankton assemblage, with copepods dominating in July and August and amphipods becoming the dominant group from September to November. Several major taxa, including copepods and amphipods, exhibited higher occurrence and image-derived abundance at night than during the day. Tidal-cycle reconstruction and XGBoost–SHAP analysis further indicated that zooplankton detection probability was jointly regulated by diel light conditions, tidal phase, and tidal height. Detection probabilities were generally higher during nocturnal ebb tides than during nocturnal flood tides, whereas daytime light conditions were associated with reduced occurrence across multiple taxa. These findings suggest that mesozooplankton variability in this subtropical coral habitat is shaped by the combined effects of diel forcing and asymmetric tidal processes. Given the fixed-station design, the observed patterns likely represent an integrated signal of behavioral redistribution, benthic–pelagic exchange, and tidal transport, rather than the outcome of a single mechanism. This study highlights the value of high-frequency in situ imaging combined with supervised machine-learning analysis for resolving sub-daily zooplankton dynamics in complex coastal coral ecosystems.