Brahmaputra Flood Predictor

Prototype for the DPR, Ministry of Jal Shakti pilot

Brahmaputra Valley Flood Predictor

Enter sensor readings for a location and get the probability of flooding, the alert decision and what is driving it.

A 5-year, Rs 49.2 crore pilot to predict floods across 24 districts of the Brahmaputra valley, using AI.

Rs 1,600 crand 120 lives
lost per year in the valley
28 lakh
people affected in a typical year
43% – 149%
return over five years, cautious to favourable

01 · The problem

Floods are expected. The damage is not inevitable.

Warnings today arrive “without a place, without an hour and without belief.” People do not move, and the same losses repeat every monsoon.

Ten years of loss, 2015 to 2024

Brahmaputra valley compared with the national total.

MeasureValley
Lives lost1,200
Economic lossRs 14,800 cr
Share of national total loss16.1%
National total lossRs 99,080 cr

Totals include a notional Rs 1 crore per life lost.

Where the money goes

Economic loss in the valley, Rs crore per year.

  • Crops520
  • Roads, bridges, embankments430
  • Houses330
  • Livestock70
  • Utilities70
  • Business60

Rs 1,480 crore of property, plus 120 lives at a notional Rs 1 crore each: Rs 1,600 crore a year.

02 · How it works

From a rain gauge to a siren, in five layers.

Data flows top to bottom. Hover over any box, or move to it with the Tab key, to see what it does.

03 · Live predictor

Set the conditions, see the risk.

Nine readings go in, one flood probability comes out. Every change updates the prediction, the drivers and the map.

Presets:

Sensor readings

Prediction: all valley

0% 100%
0%
Low
Low
below 30%
Moderate: alert
30% to 70%
High: alert
70% or more

The alert threshold is 30%.

What is driving this prediction

Each bar shows how far a reading pushes the risk up (red) or down (teal) compared with a typical Brahmaputra observation. Logistic regression, so every alert can be explained.

← lowers riskraises risk →

04 · Valley map

Twelve sites, one set of conditions.

Each site applies the current readings through its own profile. Select a site to see its prediction and drivers.

Low: below 30%
Moderate, alert: 30% to 70%
High, alert: 70% or more

Alert threshold is 30%. Circle size grows with probability. The river line is an approximate course. Site profiles (distance to river, drainage, blockage, rain and river-level factors) are illustrative, not surveyed.

Flood probability by site, from upstream to downstream
SiteProbabilityBand

05 · Batch file

Score a whole file at once.

Upload a CSV with any of these columns: . Missing columns use typical values.

06 · Cost and return

Rs 49.19 crore over five years. It pays back in every scenario.

What it costs

Rs crore over five years. Unit prices are planning assumptions.

ItemBasisRs cr
Capital
Sensors700 × Rs 1.5 lakh10.50
Gateways50 × Rs 5 lakh2.50
Control rooms25 × Rs 8 lakh2.00
Platform and model build7.00
Capex22.00
Operating, 5 years
Cloud6.00
Connectivity2.25
Data subscriptions2.00
Core team10 people11.00
Sensor maintenance2.60
Outreach and drills1.00
Opex24.85
Contingency 5%2.34
Total49.19

What it returns

Avoidable loss per year = Rs 1,600 cr × controllable share × accuracy. Benefit starts in year 3 at half weight, giving 2.5 full-benefit years.

5-year benefit, Rs cr
Net, Rs cr
Return on investment

Cumulative net by year, Rs crYour inputsOther scenarios

Break-even, in three facts

1.23%
of baseline loss needs to be avoided to recover the full cost.
43%
cost overrun the cautious case can absorb and still break even.
~56%
accuracy is enough to break even, against the 80% the plan assumes.

Break-even falls in year 4 for the moderate and favourable cases, and in year 5 for the cautious case.

07 · Roadmap

60 months, four gates. No gate, no next tranche.

  1. G1
    End of M6Data agreements signed
  2. G2
    End of M1480% accuracy in shadow mode
  3. G3
    End of M24Build accepted
  4. G4
    End of M36All districts live

Shadow mode: the model runs on live data and its predictions are checked against what actually happens, but no public alerts are issued from it.

08 · Who is accountable

A machine raises the alert. People decide.

  1. Steering committeeSets direction, approves each gate, releases funds.
  2. Nodal agencyOwns the pilot, the data and the contracts.
    Program managerRuns delivery day to day; reports to the committee.
  3. Technical advisory groupIndependent experts who check the model and the evidence.
  4. Delivery teamCore team of 10: builds and runs sensors, platform and model.
  5. District operationsControl rooms in 24 districts; act on every alert.

The three-line rule

  1. 1
    The model raises the alert.Every score is logged with the data behind it.
  2. 2
    The duty hydrologist confirms or downgrades.A named person, with the reason recorded.
  3. 3
    The District Collector issues the warning.Legal authority stays where it is today.

09 · Risks and answers

What could go wrong, and what we do about it.

Four questions the Department of Expenditure will ask

What if the model is less accurate?
It still breaks even at about 56% accuracy. Gate G2 stops the rollout if shadow mode does not reach 80%.
What if it costs more?
Even the cautious case absorbs a 43% overrun. A 5% contingency is already included.
How will we know it worked?
Months 37 to 60 measure lead time, floods caught, false alarms and losses against the 2015–2024 baseline.
Who is answerable if a warning is missed?
The three-line rule. Every alert, review and decision is logged against a named officer.

10 · About the model

What is under the hood.

Cross-validated AUC
Floods caught at the 30% threshold, all basins, cross-validated
Floods caught in the Brahmaputra basin alone (28 floods, 85 rows)
9Predictors (data dictionary selection)
1,000Training observations, 9 basins

What moves the odds most

    Caution. Synthetic one-year sample; real accuracy will be proved in shadow mode. Planning assumes 80%.

    Models compared

    Cross-validated AUC. Axis starts at 0.85.

    The simplest model scored best, and its reasoning can be explained to a District Collector.

    Trained on the synthetic sample dataset supplied with the case (Flood_Prediction_v2.1). Outcome fields (water depth, NDWI, risk score, severity) were excluded to prevent leakage. The threshold is set at 30% because a missed flood costs lives and a false alarm costs trust. At 30%, cross-validated recall is about 95% with about 84% precision; at 40% it would be 92% and 89%. Brahmaputra alone has only 85 rows, so its recall is noisy. Field accuracy will be lower than on synthetic data; this is a prototype, not an operational warning system.