Physics-guided fusion of coastal imagery and hydrodynamic forcing for hazard alert classification at a structurally evolving harbor breakwater

Le résumé fourni par la source

Wave overtopping at harbor breakwaters is an operationally critical hazard, yet visually energetic scenes (e.g., run-up, spray) do not map uniquely onto hydraulically hazardous states. To improve alert reliability, a physics-guided, decision-level fusion framework is proposed. A meta-classifier combines per-frame convolutional neural network (CNN) image inference with synchronized hydrodynamic descriptors, applying the physical state as a soft consistency constraint to re-weight, rather than override, the visual posterior. The framework was evaluated retrospectively on selected keyframes from 13 typhoon events recorded between 2023 and 2025 at an operational harbor breakwater, classifying the hazard state of the current hour across pre- and post-damage structural regimes (293 and 379 held-out event-hours, respectively). Relative to the two single-modality baselines (image-only and hydrodynamic-only), the random-forest (RF) late-fusion preserved severe-event recall at a fixed alert threshold in the image-aligned warning-period evaluation while delivering regime-dependent ranking gains. In that evaluation, the fusion raised average precision (PR-AUC) over the image-only baseline from 0.599 to 0.832 pre-damage and from 0.777 to 0.962 post-damage. In the full-window hourly evaluation, the fusion also improved PR-AUC over the hydrodynamic-only baseline in every regime-by-illumination cell, indicating a consistent contribution from the image score. The single-modality comparison yielding the larger improvement, however, varied with both structural regime and illumination; this is an observed single-site pattern rather than a validated deployment rule. These results provide a single-site field proof of concept for reliability-oriented monitoring under evolving structural conditions; multi-site transferability and uncertainty quantification remain to be established.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.