Wide & Deep Learning Ranks Gulf Port Cyclone Impacts 85.5% Above Distance Baseline on Nicole Test
Two-stage model retrieves ports by proximity then ranks impact days via Wide & Deep Learning. Tests on Hurricanes Nicole and Ian show 55-85% gains over distance baselines with reduced sensitivity to track error. Framework supplies prioritized lists for inspection and recovery allocation.
The arXiv paper presents a two-stage system that first retrieves candidate ports by proximity to uncertain landfall then ranks them by predicted impact duration. It trains on pairwise interactions, tree combinations, and explicit crosses with embeddings. Retrieval recovered all observed affected ports in both test cyclones despite landfall variance.
On the ranking stage, Wide & Deep produced the highest agreement with observed orderings derived from 31 historical TCs. It showed lower sensitivity to landfall shifts than pure distance or tree-only models. The 37-port Gulf dataset supplies the only large-scale AIS impact labels available for this task.
Operational value lies in feeding ranked lists directly into port authority inspection schedules and resource allocation before landfall. This mirrors documented uses of learned embeddings in NOAA storm surge models and FEMA after-action reports that already ingest AIS tracks. The approach converts sparse port-level observations into actionable prioritization under the same uncertainty that defeats single-point forecasts.
Next steps require live integration with National Hurricane Center track ensembles and validation against 2025-2026 season data to test whether agreement remains above 80% when landfall errors exceed those in Nicole and Ian.
Sources (3)
- [1]Primary Source(https://arxiv.org/abs/2611.07296)
- [2]Supporting Source(https://www.nhc.noaa.gov/data/)
- [3]Supporting Source(https://arxiv.org/abs/1606.07792)