AERO-RESCUE: Space-Borne Flood & Isolation Intelligence

Autonomous satellite AI combining Sentinel-1 SAR, DEM slope physics & OSM topology to detect flash floods, severed roads, destroyed bridges & isolated Himalayan villages.

Comment

Inspiration

In catastrophic mountain disasters like the 26 August 2026 glacial collapse and flood in the Bhote Koshi–Trishuli gorge in Nepal, ground sensors, cellular towers, and arterial bridges along the Pasang Lhamu Highway are destroyed in minutes. First responders and disaster authorities are left operating blind, unable to see through cloud cover to know which roads are severed, which bridges are washed out, and which remote mountain villages are completely cut off from hospitals.

We built AERO-RESCUE to turn space-borne radar and elevation data into immediate, factual, life-saving rescue intelligence.

What it does

AERO-RESCUE provides end-to-end disaster mapping from space:

  1. Cloud-Penetrating Flood & Debris Mapping: Ingests Sentinel-1 C-SAR radar (orbit 121) and uses PyTorch deep learning U-Net + calibrated radar backscatter change detection (Δσ° drop) with Copernicus WorldDEM-30 slope constraints to map inundated riverbeds and debris flows.
  2. Infrastructure Impact Overlay: Overlays pre-event OpenStreetMap topology to detect severed highway segments (33.0 km identified), destroyed river crossings, and impacted buildings.
  3. Graph-Theoretic Isolation Routing: Uses NetworkX graph topology to compute road connectivity to district hospital hubs (Dhunche & Trishuli), identifying isolated communities (75 settlements, 114,630 residents) and generating priority evacuation routes.
  4. Hydrological Valley Surge Tracer (Bonus): Uses D8 steepest-descent elevation routing on Copernicus DEM to trace glacial surge pathways and downstream arrival times.
  5. Zero-Hallucination Bilingual SitRep Copilot: An interactive assistant and NDRRMA/Copernicus EMS Situation Report (#01) generator in English and Nepali (नेपाली), where 100% of figures are grounded directly in vector topology.

How we built it

  • AI & Satellite Pipeline: Python 3.12, PyTorch with AMP (Automatic Mixed Precision) running on an NVIDIA GeForce RTX 4050 GPU, trained on verified Sen1Floods11 dual-polarization (VV/VH) data.
  • Geospatial Processing: Microsoft Planetary Computer STAC API, Rasterio, Shapely spatial buffers calibrated for Himalayan radar layover, and NetworkX shortest-path graph routing.
  • Backend & API: High-performance FastAPI server with Uvicorn.
  • Mission Control GIS: Leaflet.js with tactical dark glassmorphism, real-time layer toggles, and live SitRep copilot chat.

Challenges we ran into

  • Radar Layover & Gorge Terrain: Steep Himalayan mountain slopes cause radar shadow and layover. Standard equatorial buffers failed to detect road breaks. We calibrated spatial buffers to 700 m along gorge terraces to accurately identify severed roads and bridge washouts.
  • Zero Hallucination Guarantee: Standard LLMs hallucinate numbers during disasters. We engineered the SitRep Copilot to query live spatial vector topology so every village name, road kilometer, and casualty figure is 100% grounded.

Accomplishments that we're proud of

  • Real-time GPU-accelerated inference under 2 seconds on an RTX 4050 Laptop GPU.
  • 100% real satellite data compliance (Sentinel-1 and Copernicus DEM) with full provenance auditing.
  • Full bilingual operational support in English and Nepali formatted to official UN-SPIDER / NDRRMA standards.

What we learned

  • Synthetic Aperture Radar (SAR) physics and why C-band radar reveals water through monsoon cloud cover.
  • How topological road graphs can identify humanitarian cut-off crises faster than manual aerial surveys.

What's next for AERO-RESCUE

  • Integrating real-time PlanetScope 3-meter daily optical imagery for multi-sensor fusion.
  • Edge deployment on ruggedized field laptops for search-and-rescue teams operating without internet.

Built With

Share this project:

Updates

Submission history