Abstract

Coastal marinas and high-traffic nearshore sites accumulate pollution from vessel movements, tourism, and shifting weather, yet routine monitoring rarely operates at the temporal resolution needed to catch emerging risks before they become acute. This study developed an AI-enabled decision support system and tested it across three European coastal sites over a 90-day window in summer 2025: an urban marina (Site A), a tourism marina (Site B), and a mixed-use port channel (Site C). A composite Water Quality Risk Index (WQRI), combining five normalised environmental and vessel-traffic stressor dimensions, fed a two-layer AI framework in which a gradient boosting model estimated short-term traffic-related stress and a random forest model classified next-day risk. Vessel traffic was heaviest at Site B, but water quality risk followed a different pattern: Site C returned the highest mean WQRI and logged the most hours under red alert despite intermediate traffic volumes, indicating that sustained moderate traffic mattered more than peak volume. Vessel intensity and WQRI were positively correlated at all three sites, most strongly at Site B, and the next-day random forest risk classifier, trained across all three sites, achieved strong discriminative performance (AUC 0.93). When the system indicated elevated risk, managers responded by deploying inspections, issuing traffic advisories and increasing monitoring activity. The pilot shows that connecting vessel tracking, environmental sensing, and ML-based classification into a single decision loop can move coastal pollution management from reactive to anticipatory.

Keywords