Inspiration

During Bengaluru monsoon cloudbursts, flood-control rooms only have minutes to decide where to send mobile pumps, which underpasses to lock, and how to reroute public buses. Standard GPS ignores flood depth and routes vehicles straight into submerged streets, while traditional 2D hydraulic simulators like EPA SWMM take far too long to run during a live emergency. We built UrbanFLOW to bridge that gap—turning real-time flood predictions into actionable dispatch in milliseconds.

What it does

  • Real-Time Flood Risk Scoring: Evaluates 12,000+ Bengaluru road junctions in ~300 ms for any rainfall rate and storm duration.
  • Actionable Dispatch Orders: Automatically clusters nearby flooded streets and generates WHERE, WHAT, and WHEN directives (mobile pumps, flood barriers, traffic wardens), assigning the nearest reachable unit and drawing its flood-safe route.
  • Flood-Safe Bus & Point-to-Point Corridors: Compares standard bus routes (e.g., BMTC 500D / 500A) against flood-safe green detours that avoid submerged road links—and supports custom two-point routing anywhere on the map.
  • Defensive Red-Alert Mode: Lowers action thresholds defensively during extreme \(\ge 100\text{ mm/hr}\) tail cloudbursts.

How we built it

Under the hood, we trained HydroGINE-v5, a 4-member, 6-layer Graph Isomorphism Network (GINE) with a Directional Gravity Gate and Topological Mixture-of-Experts on 12,000+ junctions and 29,000+ directed edges using 32 scale-invariant node features (zero raw coordinates).

  • Decoupled FiLM Dual-Head: A Margin-Focal classifier (\(\gamma = 2.0\)) generates affine \((\gamma_v, \beta_v)\) conditioning for a Huber depth regressor (\(\delta = 0.10\)), solving 80.9% zero-inflation collapse.
  • Physical Continuity Bounding: A 2.1 ms post-processing step enforcing hydraulic mass-conservation and water-surface-elevation (WSE) backwater envelopes across the 4-checkpoint ensemble.
  • Command & Dispatch Stack: FastAPI backend + interactive 3D map frontend for live rainfall simulation, fleet management, and corridor rerouting.

Challenges we ran into

  • Zero-Inflation Collapse: Over 80% of road junctions remain dry in typical storm events, causing naive regression models to predict zero depth everywhere. Our Decoupled FiLM Dual-Head solved this cleanly.
  • 10 m Public DEM Smoothing: Public 10-meter digital elevation models smooth out micro-topography like curbs and underpass ramps, shifting localized ponding depths.
  • Extreme Tail Events: \(100+\text{ mm/hr}\) cloudbursts are rare in historical training data, which motivated adding Red-Alert Mode to lower dispatch thresholds defensively.

Accomplishments that we're proud of

Across 16 global catchments (76k nodes), HydroGINE-v5 achieves:

  • 0.913 catchment Nash-Sutcliffe Efficiency (NSE)
  • 2.44 cm Mean Absolute Error (MAE)
  • 90.8% F1 flood classification score
  • Up to 8,690× faster inference than EPA SWMM 5.2, enabling interactive <300 ms command-center dispatch.

What we learned

  • Combining physics-informed inductive biases (directional gravity gates, continuity bounding) with scale-invariant graph features lets a single GNN generalize across 16 diverse cities without overfitting to local coordinates.
  • Speed only matters in disaster response if it translates into clear, actionable decisions—pairing raw flood depths with fleet routing and corridor detours made the AI immediately useful.

What's next for UrbanFlow

  • Sub-Meter LiDAR & Drain Surveys: Ingest 0.5 m – 1 m DEMs and municipal drain-invert geometry to capture curbs and underpass ramps accurately.
  • Live Telemetry & Data Assimilation: Fuse live IoT rain gauges and BBMP/IMD Doppler radar nowcasts for dynamic boundary conditions.
  • Tail-Event Fine-Tuning: Augment training with synthetic 100–130 mm/hr EPA SWMM 5.2 cloudburst and culvert-blockage scenarios.

Built With

Share this project:

Updates

Submission history