Castopini

During the October 2024 Valencia DANA floods, more than 220 people died. Many did not evacuate in time — not because the alert failed to reach authorities, but because it failed to reach people who could act on it. Emergency broadcasts get distorted through word-of-mouth. Some residents don't trust official warnings. Roads believed to be passable were flooded. And when the first message fails, there is no automatic recovery.

We built an AI-driven flood evacuation coordinator — Synoptis — that addresses this last-mile communication problem. It does three things that don't exist in traditional emergency management systems:

• Global, Real-Time Mapping: The system is completely grid-agnostic. Operators can input any bounding box coordinate and timestamp globally, and ECHO-SWARM will pull the local road network and live Copernicus Sentinel-1 SAR imagery. It detects flooded roads at ~10m resolution and injects that information into a geospatial knowledge graph (Neo4j). Flood-blocked roads are immediately removed from routing, ensuring every evacuation path is physically passable.

• Resilient Evacuation Orders: The Hermes engine uses Claude (Anthropic) to produce structured evacuation messages following the CERC (Crisis and Emergency Risk Communication) protocol — specifying who should move, where, via which exact route, and citing the satellite timestamp as a verifiable source. A second LLM pass scores message clarity (1–10); messages below 7 are automatically revised.

• Autonomous Learning Loop: A multi-agent simulation (MiroFish) models how residents — Compliant, Skeptical, Panic-prone, and Immobile — propagate and distort the message over multiple hops. The Critic engine diagnoses failures and writes updated Standard Operating Procedures (SOPs). As seen in the Mission Evolution graph, across 10 simulation runs, the AI's auto-corrections improved the survival rate from 5.0% to 39.0% (a 34.0% gain) by specifically adapting to the needs of Skeptical agents who require verifiable satellite citations.

The system runs end-to-end: satellite -> graph -> message -> simulation -> learning -> better message next time. It is not a dashboard. It is a closed feedback loop.

🛰️ EU Space Technologies

Primary data source: Copernicus Sentinel-1 (SAR)

ECHO-SWARM connects directly to the Copernicus Data Space Ecosystem (CDSE) via OAuth2 and the Sentinel Hub Process API. It requests Sentinel-1 GRD VV-polarization imagery for the user-defined bounding box and applies adaptive backscatter thresholding to classify flooded pixels (configurable for turbulent conditions). The resulting flood mask is vectorized into WGS-84 Shapely polygons and injected directly into the road network graph.

Why SAR specifically matters for disaster response:

• Sentinel-1 is all-weather and day/night capable. Flood events are almost always accompanied by cloud cover that renders optical satellites blind; SAR penetrates clouds.

• The C-band VV signal has a well-established backscatter signature for open water detection.

• The CDSE provides open access with a ±3-day search window — the system automatically finds the closest scene if no exact-date image exists.

Value added by EU space data:

• Flood polygons from Sentinel-1 are the only input that dynamically mark roads as impassable. Without them, routing is static and dangerous.

• The satellite timestamp is embedded in the evacuation message as a verifiable credibility signal.

• The system's information decay metric directly measures whether satellite-sourced information survives word-of-mouth propagation.

🌊 EU Space for Water — Resilience to Water-Related Disasters

ECHO-SWARM targets the gap between satellite flood observation and human evacuation outcomes. Copernicus provides excellent maps, but the translation from "here is where the water is" to "here is what each resident should do" is still largely manual and degraded by bureaucratic relay.

The system contributes to resilience in three concrete ways:

• Real-time Passable Routing: Flood polygons are spatially joined to the road network graph using an adjacency buffer. Every shortest-path query to the shelter automatically excludes flooded roads, reflecting actual ground conditions at the exact time the message is sent.

• Communication Optimized for Compliance: The CERC-structured messages include a source_justification field citing Sentinel-1 data. This specifically addresses Skeptical residents who require independent confirmations before acting.

• Autonomous Improvement Across Events: The Critic engine makes diagnostic reasoning explicit. It identifies which message dimensions caused failure, writes new SOP rules for Hermes, and saves them. The system gets better because it observes what worked in the population model.

https://github.com/fanceapa1/CASSINIHackathon