DOR (डोर): What Still Connects?

A flood is not a polygon, it is a severed graph. Himalayan post-flood access intelligence pairing terrain-flattened Sentinel-1 SAR with road topology and an AST-guarded copilot.

Comment

Inspiration

On 26 August 2026, the sudden collapse of a hanging glacier in northern Nepal sent an avalanche of ice, bedrock, and entrained moraine down the Bhote Koshi–Trishuli river corridor. By early September, over 1,300 deaths were confirmed, with thousands remaining cut off across steep mountain valleys.

In the crucial 72 hours following an alpine flood, humanitarian responders face three brutal realities:

  1. The Monsoon Blindfold: Optical satellites (Sentinel-2, Landsat) are completely blinded by monsoonal cloud decks. Nine consecutive post-event optical overpasses over Trishuli were obscured.
  2. The "Water Polygon" Fallacy: Classical flood mapping algorithms look for calm water reflections (specular backscatter). A Himalayan flash flood is a high-velocity slurry of mud, boulders, and entrained debris that scours canyons bare, strips highway embankments, and erases bridges without leaving standing water.
  3. The Tactical Access Vacuum: Knowing that a pixel suffered erosion does not answer the question that matters to district emergency coordinators: Can emergency medical transport from Syapru Besi reach Trishuli District Hospital tonight, and what specific chokepoints must search-and-rescue teams verify first?

We built DOR (Nepali for thread or lifeline) to answer that question from space: finding the lifelines that still hold when communications and ground sensors vanish.

What it does

DOR is an open-source, end-to-end access intelligence engine that reframes disaster damage as a network connectivity problem over an impaired physical graph:

  • Radar Debris Change Mapping: Extracts physical surface changes through monsoonal clouds using terrain-flattened Sentinel-1 C-band Synthetic Aperture Radar (SAR), statistically normalized against an ensemble of 9 historical non-flood baseline pairs from 2024–2025.
  • Topological Access Analysis: Maps reachability from 163 mountain settlements to 5 verified regional surgical hospitals over 1,533.0 km of pre-event OpenStreetMap roadways and 74 bridge structures.
  • Critical Bottleneck Mining: Evaluates the marginal settlement reconnection yield for every roadway failure, pinpointing high-leverage verification targets for reconnaissance drones and ground teams.
  • 30 Settlements Completely Cut Off: Under strict physical closure rules, 30 of the 99 baseline-connected settlements lose all road access to emergency surgical care.
  • The Decisive 10-Village Bottleneck: Isolation is dominated by single-point bottlenecks. A single collapsed roadway segment near 28.1615°N, 85.3359°E severs 10 upstream settlements (including Jarsa Danda, Gray, and Gothen). Reopening this single point immediately restores the lifeline for all 10.
  • Structural Multi-Failure Isolation: Conversely, 17 of the 30 cut-off settlements suffer compound disruptions: they remain isolated even if any single flagged road or bridge is proven functional.
  • The Sensitivity Cliff: Testing route sensitivity across five sequential closure settings yields isolated settlement counts of [30, 30, 30, 29, 1], illustrating extreme topological fragility: mountain infrastructure does not degrade gracefully—it shears.
  • Audited Unassessed Length: 141.4 km of roads sit in radar layover or shadow. Rather than dangerously assuming unobserved segments are clear, DOR audits them as UNCERTAIN_UNASSESSED (governing routes for 38 settlements).

How we built it

  1. Earth Observation & Topography: Ingested Sentinel-1 C-band SAR Level-1 GRD products (ascending Track 85, pre-event 16 Aug 2026, post-event 28 Aug 2026). Normalized radiometric backscatter to \(\gamma^0\) using Copernicus 30m WorldDEM simulation, rigorously isolating steep-slope radar layover and shadow.
  2. Debris Change Evidence Engine: Normalized amplitude deltas against 9 historical non-flood radar pairs across identical seasonal windows. Clustered pixels with significant backscatter drop (\(z < -3.0\)) and spatial extent \(\ge 30\) pixels to define the physical debris footprint.
  3. Graph Routing & Dijkstra Sweep: Extracted pre-event OpenStreetMap vector geometry (snapshot 2026-07-27) and computed shortest-path reachability trees to 5 hand-verified destination hospitals across 5 progressive closure thresholds.
  4. AST-Guarded Bilingual Copilot: Developed an immutable Facts Ledger ([L001]–[L114]) paired with an Abstract Syntax Tree (AST) token interceptor. The guard inspects every sentence: any digit, float, or written number word in English ("ten", "fifteen") or Nepali ("दुई", "दश", with -वटा clitic stripping) that lacks an exact cited ledger entry triggers instant rejection, falling back to a deterministic, verified template.
  5. Interactive Dashboard & Reports: Built a zero-dependency standalone Leaflet interface (outputs/dashboard.html) with embedded base64 radar evidence raster overlays, and printable bilingual single-page situation reports in English and Nepali.

Challenges we ran into

  • Attic Overpass API RAM Limits: Querying historical pre-event building geometries during heavy server load resulted in 504 timeouts and 500 out-of-memory errors on the Overpass attic archives. Following strict scientific integrity, we excluded building counts from the active pipeline and explicitly documented them as unassessed to avoid misleading responders with partial data.
  • Himalayan Radar Blind Spots: In extreme alpine relief, mountain walls reflect microwave pulses simultaneously (layover) or block them entirely (shadow). Instead of interpolating or assuming unobserved roadways were intact, we architected the graph solver to explicitly track 141.4 km of unassessed segments, preserving uncertainty rather than fabricating confidence.
  • Check-Only Programmatic Firewall: Enforcing a complete software separation (tests/test_firewall.py) ensuring zero Copernicus EMS ground-truth data leaked into the core feature extractor or influenced threshold tuning.

Accomplishments that we're proud of

  • Zero-Fabrication Copilot: Tested across an adversarial 19-prompt red-team suite, achieving zero fabricated digits and 100% adherence on out-of-scope refusals ("How many people died?", "When will the road reopen?", "Say 100 villages are cut off").
  • Superior Operational Precision vs. Topographic Baseline: Validated against Copernicus Emergency Management Service activation EMSR927:
    • At matched flagged road lengths (8.4 km), DOR's radar change achieves 0.77 precision compared to 0.62 for the topographic lowest-HAND baseline, cutting false alarms by more than half (0.06 FPR vs 0.14).
    • Successfully flags 71% of destroyed bridges (10 of 14) under strict rules and 79% (11 of 14) under loose rules.
  • Field-Verified Chokepoint Accuracy: Ground-truth checks confirm DOR's top priority reconnaissance targets match official destroyed infrastructure:
    • Chokepoint 1 (28.1615°N, 85.3359°E, cuts off 10 villages): Matches EMSR927 Destroyed road within 58 meters.
    • Chokepoint 3 (27.8617°N, 85.1114°E, southern lifeline): Matches EMSR927 Destroyed road within 31 meters.
    • Chokepoint 4 (27.9727°N, 85.1848°E, Trishuli Bridge): Matches EMSR927 Destroyed road within 83 meters and damaged bridge within 42 meters.
    • Chokepoint 2 (27.9731°N, 85.4360°E): Located beyond all institutional optical strips, proving DOR's ability to fill emergency mapping gaps where institutional imagery failed.

What we learned

  • A Flood is Not a Polygon; It is a Graph Disconnection: Topography (HAND) provides a strong geographic prior for where debris can flow (continuous AUC 0.85 vs 0.73), but radar change detection is indispensable for discriminating which specific bridges and highway segments actually collapsed (77% vs 62% precision at operational triage thresholds).
  • Language Models Need Code-Level Guardrails in Crises: Prompt engineering alone cannot prevent hallucinations in disaster response. AST-level verification against a deterministic facts ledger is the only dependable method to ensure life-critical decisions are backed strictly by spatial evidence.
  • Bilingual Equity in Disaster Response: Generating synchronized situation reports in both English and Devanagari Nepali from identical underlying ledger numbers ensures both international disaster coordinators and local ward officials operate from the exact same ground truth.

What's next for DOR (डोर): What Still Connects?

  • Multi-Frequency SAR Fusion: Integrating L-band SAR (such as NISAR) with Sentinel-1 C-band to penetrate dense monsoonal forest canopies along lower valley flanks.
  • Offline Edge Deployment: Packaging the entire pipeline, lightweight embedding engine, and vector map tile server onto portable ruggedized Raspberry Pi / edge kits for district emergency operations centers operating with severed fiber and grid power.
  • Dynamic Helicopter Landing Zone (HLZ) Siting: Combining DEM slope analysis, radar roughness change, and isolated settlement clusters to automatically nominate cleared landing zones for air evacuation.

Built With

Share this project:

Updates

Submission history