Infrastructure

Ninety minutes of warning for a city built for drought.

Desert cities flood in hours, from rain that falls somewhere else. The drains were never sized for it and the warning, when there is one, names the city, not the street. We fuse radar, terrain, and drain sensors into a street level nowcast and evolve it against every flood the city has already had.

The problem

Three hours of rain, no warning, and a city underwater.

In arid cities the ground does not absorb, the wadis funnel, and the storm drains were built for a climate that no longer applies.

70 mmof rain in three hours in one coastal city, more than a year’s average, with no early warning system in place
113+deaths and more than ten thousand damaged properties from that single event
4 floodsin the same two cities within three years, the most recent one overwhelming drainage again

Flash flooding in a desert city is not a river rising slowly. It is rain on hard ground, often in the hills behind the city, arriving through dry valleys that are roads on an ordinary day. The water reaches a neighborhood an hour after the cloud passed over somewhere else. A city wide weather warning says rain is coming. It does not say which underpass will be impassable at four o’clock or which district’s drains will back up first.

The pieces of a street level forecast exist. Weather radar sees the rain cell. Elevation models know where the water will go. Drain and wadi sensors, where installed, report the level now. What is missing is the model that turns these into a map of water depth thirty, sixty, and ninety minutes ahead, and the loop that corrects it against the floods the city has already lived through.

In plain terms

Think of the city as a sink with the drain half blocked. The rain is the tap, but the tap is in the next room and you cannot see it. Radar is a window into the next room. Terrain is the shape of the sink. The drain sensors tell you how fast it is draining today. Put those together and you can say, an hour ahead, which part of the sink overflows first.

Why the standard approach fails

Hydrology models are built for rivers, and warnings are built for cities.

01

Models calibrated on rivers, not wadis

Standard flood models assume soil that absorbs and channels that behave. Desert catchments do neither, and the parameters that fit a temperate river are wrong here.

What we do. The model’s parameters are evolved against the city’s own recorded floods, so they describe this catchment and no other.

02

Warnings at the wrong scale

A warning for a city of millions tells nobody what to do. People need to know their street, their underpass, their school route.

What we do. The forecast is a map of water depth by street block at thirty, sixty, and ninety minutes, and the warning is issued per district with the places to avoid.

03

Too few events to train on

A city may have five serious floods on record. That is not enough for a standard learning approach.

What we do. We start from a physical model of how water moves and evolve its parameters and its corrections, which needs far fewer events than learning from scratch. Each new storm, even a minor one, adds to the record.

04

False alarms cost trust, misses cost lives

Warn too often and the city stops listening. Miss once and the result is the headline nobody wants.

What we do. Detection of every major past event is a hard floor. False alarm rate and lead time are the tradeoff the civil defense team chooses from.

Which streets, and when. A city wide warning says rain. A nowcast says the wadi corridor floods first, in half an hour, and the low districts beside it an hour later. Each block gets a depth and a time, so a road can be closed before the water arrives.
+30 minutes +60 minutes +90 minutes Dry Passable Close to cars Evacuate Illustrative
How we solve it

A physical model, evolved to the city.

We do not guess the water. We model it, then evolve the model until it matches what the city has already seen.

In plain terms

Water obeys physics, so we start with a physics model of how rain runs through this city’s hills, valleys, and drains. That model has dozens of settings nobody knows exactly: how fast water sinks into this ground, how much a blocked drain slows it, how the radar’s rain estimate should be corrected. We breed those settings. Dozens of versions of the model replay the city’s past floods; the ones that reproduce them best become parents of the next generation. The result is a model that knows this city, and it keeps learning from every storm.

Not one model. A population that is bred. Every generation is made from the last: the strongest candidates become parents, their settings are combined, small variations are introduced, and the children are scored on your own cases. Weak lines die out. Strong ones carry forward.
Generation 1Generation 6Generation 14 48 candidates, all differentWeak ones gone, strong ones combinedYour model, chosen by your team
This is an evolutionary algorithm, not a leaderboard. We do not train many models and keep the best. Each generation is bred from the last. The strongest candidates are selected as parents, their settings are combined to make children, small random changes are introduced, and the children are scored on your own scans. Settings that work together survive together.
1. Scoreon your own scans 2. Selectthe strongest as parents 3. Combineparents into children 4. Varysmall random changes Nextgeneration How a child is made Parent A Parent B Crossover: segments from each parent Child Mutation: one setting changed at random Child Each segment is one setting, such as window or threshold
  1. 01

    Assemble the city’s water map

    High resolution elevation, the drainage network as built, the wadis as they actually run, land cover, and every sensor the city has. Where sensors are few, the evolution also tells the city where the next ones would help most.

  2. 02

    Replay the floods on record

    Every past event with radar and damage records is reconstructed. The baseline shows how far a textbook model is from what actually happened on those streets.

  3. 03

    Evolve the model to the city

    A population of model parameter sets and radar corrections is scored on how well it reproduces the recorded floods, street by street, with lead time, and on false alarms in the storms that did not flood. Any candidate that misses a major past event is discarded. Civil defense chooses the operating point between early warning and false alarms.

  4. 04

    Nowcast in real time

    During a storm the model runs on live radar and sensor data every few minutes, producing water depth maps at thirty, sixty, and ninety minutes. District warnings, underpass closures, and school route alerts are generated from thresholds the city sets.

  5. 05

    Learn from every storm

    Every storm, including the ones that did not flood, becomes a new replay. The next generation is evolved on the larger record and proposed only if it beats the deployed model on every event.

What the evolution changes

The settings nobody knows for this catchment.

What we evolveWhy it matters in a desert city
Infiltration and runoff coefficientsDesert ground, paved city, and wadi bed each shed water differently. Textbook values are wrong; evolved values match the recorded floods.
Drain capacity and blockage factorThe drainage network on paper is not the one in the ground. Effective capacity is evolved per district from past events.
Radar rainfall correctionRadar estimates of rain are biased by distance and storm type. The correction is evolved against gauge records.
Routing through wadisHow fast water moves from the hills to the city, and which valley it takes, is evolved against arrival times in past floods.
Warning thresholds per districtWhat depth and lead time trigger a warning is a tradeoff between false alarms and safety, chosen by civil defense from the Pareto set.
Major event floorEvery major flood on record must be detected with lead time. Candidates that miss one are discarded.
What you get

A map of the next ninety minutes, street by street.

Street level nowcast

Water depth by block at thirty, sixty, and ninety minutes, refreshed every few minutes during a storm.

District warnings

Issued automatically from thresholds the city sets, with the underpasses and routes to avoid.

Sensor placement advice

Where the next drain or wadi sensor would most improve the forecast.

Replay of every past flood

Reconstructed hour by hour, so the city can see what the system would have said and when.

Drainage insight

Evolved capacity factors reveal which districts’ drains underperform their design.

Improves with every storm

Minor storms are training data too. The model gets sharper every season.

Where this fits

Built for cities where rain is rare and floods are not.

  • Arid and coastal cities with wadi catchments behind them
  • Municipalities and civil defense agencies with radar access and few sensors
  • Cities with a flood record but no street level warning
  • Infrastructure operators responsible for underpasses, tunnels, and low lying roads

A flood forecast for a desert city has every property that makes a problem hard: rare events, physics that defies the textbook, sparse sensors, and an asymmetric cost of error. It is also exactly the kind of problem where evolving a model to the place beats training one on the average. We would rather take on this one than ten easy ones.

Give us your city’s worst flood. We will replay it with radar and terrain and show you the warning the system would have issued, and when.
Partner With ALBA