Methodology

How the scores are built

This is an open-data demonstrator. Every number on the site is derived from free, public sources on a daily schedule and baked into a static build — there is no live backend and no proprietary feed.

Sources

• GDELT DOC 2.0 — global news coverage volume and headlines per chokepoint. Drives the "news intensity" and 14-day trend. Free, no key.
• NASA GIBS — a daily wide-area true-color satellite snapshot per zone (VIIRS / MODIS). Context imagery only. Free, no key.
• NASA FIRMS — near-real-time thermal anomalies (fires, flares, hot spots) inside each box, as an activity proxy. Free key, optional.

Planned extensions: Copernicus Sentinel-1 SAR (all-weather dark-ship detection) and Global Fishing Watch AIS-derived events (gaps, loitering, encounters).

The risk model

Each zone's risk score (0–100) is deterministic and transparent. It blends:

• a slow-moving structural prior for the chokepoint (its baseline geopolitical exposure);
• news intensity — how much of global coverage currently concerns the zone;
• momentum — whether that coverage is rising or falling vs its 14-day average;
• thermal anomalies in the box over the last 24h, when FIRMS is enabled.

The score is weighted toward the structural prior so a quiet news day still reflects real geography. Bands: critical ≥78 high ≥60 elevated ≥42 calm.

The vision-language model

When a VLM is configured (via Ollama Cloud), it narrates — it turns the computed signals, headlines and the daily snapshot into readable analyst prose. It never sets the risk score, and it is explicitly instructed not to count or identify vessels from the coarse daily imagery. If the model is unavailable, a deterministic rule-based narrator is used instead. The current build's narrator is rule-based.

This is deliberate: generalist VLMs are excellent narrators of a scene-plus-numbers, but unreliable as sensors on low-resolution satellite imagery. Hard detection (dark-ship counts, wake analysis) belongs to purpose-built detectors on SAR/EO — a planned upgrade, not the VLM.

Limitations

• News-coverage volume reflects attention, not ground truth; it can spike on anniversaries or unrelated stories that match the query.
• Daily optical snapshots are frequently cloud-covered and far too coarse to resolve ships.
• Scores are heuristic and for demonstration — not navigational, operational, or investment advice.

Built by MarineAware — AI-native maritime intelligence.