Pulse
The city reveals its own drains
Every flood is an experiment already run: an ensemble Kalman filter turns traffic slowdowns and reports into a blockage map.
Every street. Three hours early.
VARUNA turns Doppler radar into street-by-street flood depth for the next three hours, learns the city's hidden drains from every flood, and routes emergency services around what is coming.
SIH 2026 · PS SIH26085 · Ministry of Earth Sciences
- 5-15 cm
- 15-30 cm
- 30-45 cm
- 45-60 cm
- > 60 cm
Forecasts stop at 12 km. Streets flood at 30 m.
A 12 km forecast cell gives all of Dadar one number. The water pools under one rail bridge.
VARUNA models the city at 30 m with its drains, and answers which street, how deep, and when.
Four ways a street floods
Each needs different physics. VARUNA models all four.
- Pluvial
- Rain falls faster than the drains can take it, and sits in the road.
- Hindmata junction, Dadar East, in almost every heavy spell.
- Fluvial
- A river or nullah overtops into the streets beside it.
- The Mithi at Kurla and Kalina, 26 July 2005 and since.
- Tidal lock
- High tide shuts the outfall, and drain water comes back up through the manholes.
- Mumbai's coastal outfalls at spring high tide.
- Invisible drainage
- A blocked pipe that no map shows floods a street the model calls dry.
- What VARUNA-Pulse learns from traffic anomalies and citizen reports.
Every five minutes, the whole city again
Each stage's measured time on the newest run, beside its budget.
Stage 1
Ingest
Radar frames decoded and checked, with gauges, tide, traffic and reports
Stage 2
Sky
Z–R, gauge merge, optical flow, 20-member STEPS ensemble
Stage 3
Twin
2D shallow water on 30 m terrain, coupled to the 1D drain graph
Stage 4
Flash
Reduced-order emulator, run beside the Twin for the street ensemble
Stage 5
Pulse
EnKF over pipe blockage, from traffic anomalies and citizen reports
Stage 6
Products
Segment depths, rasters, hotspots and surcharge
Stage 7
Command, Route, Public
Routes and reachability, alerts and pumps, then published to every screen
Reading the newest run's stage timings.
Six engines
Twin
Surface and sewer, solved together
Shallow water at 30 m, coupled to a 1D drain model through inlets and surcharging manholes.
Flash
Three hours of city in milliseconds
A reduced-order emulator calibrated to the Twin, with its measured error beside every answer.
Sky
Radar into a rain ensemble
pySTEPS on calibrated reflectivity: 20 members, three hours, five-minute steps.
Route
Prediction into an ambulance route
Time-dependent Dijkstra, costed at the depth the vehicle will meet when it arrives.
Command
One screen at three in the morning
Depth on streets, alerts, pumps and reachability, with every simplification labelled.
How we score ourselves
—
Probability of detection at 15 cm
Share of sourced pins where VARUNA forecast over 15 cm in the pin's time window.
—
Median warning time
How long before the civic log VARUNA first called that street impassable.
—
Ground-truth pins
Public records in the forecast window, each with a source URL.
Reconstructed replay of 2 July 2019, scored against sourced ground truth. See how we score ourselves, and where we fall short.
Where VARUNA sits
Others warn a ward about tomorrow. VARUNA says which street, how deep, and when, for the next three hours.
IFLOWS-Mumbai
- Resolution and horizon
- Ward, 6–72 h
- Drainage
- No
- Learns from events
- No
C-FLOWS (Chennai)
- Resolution and horizon
- Basin, hours to days
- Drainage
- Partial
- Learns from events
- No
IIT-B Mumbai Flood
- Resolution and horizon
- Ward to sub-ward
- Drainage
- Partial
- Learns from events
- No
IMD nowcasts
- Resolution and horizon
- Rainfall only, 0–3 h
- Drainage
- No
- Learns from events
- No
Google Flood Hub
- Resolution and horizon
- Riverine, global
- Drainage
- No
- Learns from events
- No
VARUNA
- Resolution and horizon
- Street, 30 m, 0–3 h
- Drainage
- Inferred and learned
- Learns from events
- Yes
What the data actually is
Anything synthetic says so on screen, beside the number it made.
- Copernicus GLO-30 DEMPublic30 m terrain, downloaded per city
- OpenStreetMapPublicRoads, buildings, waterways, hospitals, stations
- ESA WorldCoverPublic10 m land cover, for imperviousness and runoff
- BMC and news archivesPublic29 sourced ground-truth pins, each with a URL
- IMD Doppler radar volumesRequested from MoESThe prototype reconstructs the storm instead
- BMC drain GISRequested from MoESThe prototype infers the network and learns it
- Radar reflectivity framesSynthetic in the prototypeStorm designer calibrated to published gauge totals
- Traffic speeds and citizen reportsSynthetic in the prototypeLabelled everywhere they appear on screen
- Mobile pump inventorySynthetic in the prototypeTwelve pumps at plausible depots, labelled
V1, V10, V100
V1
This prototype
One Mumbai AOI: reconstructed radar, inferred drains, a learning blockage map, routes and alerts.
V10
A pilot city
Real IMD radar, the ward's drain GIS, LiDAR at chronic spots, and a monsoon of observations.
V100
Every city that wants one
City-in-a-box from open data, a public map in three languages, and a routing feed for apps.