AI-Driven Maritime Traffic Risk Assessment for Smart Seaports: Multi-Vessel Conflict Reasoning and Runtime Safety Assurance

Smart seaports increasingly rely on AIS-based decision support, yet pairwise collision indicators and learned predictors do not by themselves determine when automated alerts remain trustworthy under multi-vessel coupling and degraded evidence. This study presents an interpretable scene-level extension of PortGuard-AIS that separates latent hazard estimation from decision-support authority. Pairwise encounters are assembled into a weighted vessel graph, aggregated into scene risk, filtered by persistence and hysteresis, and supervised by runtime evidence-quality gates that can cap alert authority, request human verification, or select fallback. The framework was evaluated in 16 controlled scenarios, a 30-seed geometry-jitter campaign, component and parameter-sensitivity analyses, an independent future-separation labeling test, expanded degraded-evidence stress tests, a literature-derived VCRO comparator, lead-time analysis, and computational scaling. Under the primary design labels, scene-temporal reasoning achieved F1 = 0.749 ± 0.015, precision = 0.689, recall = 0.821, and false-alert rate = 0.366; relative to pairwise temporal reasoning, its main effect was improved precision and false-alert control rather than a significant F1 gain. Under independent 0.30 NM future-separation labels, pairwise temporal and scene-temporal F1 values were 0.881 and 0.842, respectively, demonstrating that scene reasoning should not be interpreted as universally more accurate. Parameter sensitivity was limited except for persistence, which showed a clear stability-versus-lead-time trade-off. In 100 previously unseen mixed-degradation cases, none of the 78 materially degraded cases retained CRITICAL automated authority. At 40 vessels (780 pairwise encounters), P99 local processing latency was 23.85 ms. The results support controlled mechanism verification and authority-management robustness, not certified navigation safety or real-world VTS effectiveness.