How VayuNetra works, and how well

The whole system on one page, with every accuracy figure behind it: the satellite model on held-out test sites, the India field test, and the sensor early-warning model.

The system
From flag to verified reduction
Accuracy
Proof

The system

Sense

Sentinel-2 satellite
Free, every ~5 days. A U-Net AI with 33.9 million parameters reads 17 input channels and the MBMP methane signal.
Drone surveyplanned
Flies over a flagged cell on demand to confirm the plume and map it at metre scale.
Ground sensor networksimulated
Five solar nodes per landfill, a reading every 10 minutes, each compared with an upwind background node.

Decide

Physics gate
Spectral and wind tests sort every flag into T1 methane-confident, T2 probable or T3 surface change.
Confirmation tracker
Every T1 or T2 flag opens a task: EMIT match, hyperspectral tasking, OGI drone or ground survey.
Early-warning modelsimulated
A 257-weight network forecasts a methane rise 3 hours ahead from the pressure trend, the wind forecast and recent readings.

Act and prove

Dashboard and alerts
English and Hindi; emails to state officials; an action dossier per site.
Mitigation ledger
Who acts and by when, a capture calculator, metered tonnes.
Public scorecard
Monthly, per landfill, with CSV, JSON and GeoJSON downloads.
Integrity ledger
A hash chain of flags, results and readings, time-stamped on Bitcoin.

Built for MoHUA, CPCB and state pollution control boards, and municipal corporations.

From flag to verified reduction

1 Detect
Satellite AI flags T1, T2 or T3 on every new Sentinel-2 pass; ground sensors warn 3 hours ahead (simulated).
2 Confirm
Each T1 or T2 flag opens a task: EMIT match, hyperspectral tasking, OGI drone or ground survey.
3 Act
The only step that removes tonnes: each measure has a named agency, a start date and expected tonnes (modelled).
4 Verify
Only metered gas counts as reduced; the satellite tracks the site's upper bound.
5 Report
A public scorecard every month, and every record in the Bitcoin-anchored ledger.

If nothing changes after a measure, the loop goes back to Act with the next one.

Accuracy

Every figure below is measured. The sensor model's figures are measured on simulated data and say so.

Satellite model on held-out test sites

MethaneSET test split: sites the model never saw in training, compared with the classic MBMP method. Higher is better, except for false alarms.

Satellite model on held-out test sites
Measure	VayuNetra	Classic MBMP
ROC AUC (0.5 is chance)	0.759	0.513
Recall at 5% false alarms	0.24	0.10
Recall at 10% false alarms	0.34	0.15
Recall at 20% false alarms	0.52	0.24
Precision at the operating threshold	0.779	0.602
Recall at the operating threshold	0.303	0.144
False-alarm rate at the operating thresholdlower is better	8.6%	9.6%
Plume outline match (pixel IoU)	0.312	0.077
Recall for plumes of <1 t/h	0.09	0.27
Recall for plumes of 3–5 t/h	0.35	0.10
Recall for plumes of 5–10 t/h	0.34	0.10
Recall for plumes of >10 t/h	0.37	0.06

Emission rate: 58% of 839 real plumes within ±50% of the published MARS rate. Median ratio to the published rate: 1.02 (1.00 means no bias).

Accuracy matrices at the operating threshold

Rows are what was really there; columns are what the system said. Each share comes from the recall and false-alarm rate above, so these are rates, not counts.

VayuNetra

VayuNetra
	Flagged	Not flagged
Plume present	30.3%caught	69.7%missed
No plume	8.6%false alarm	91.4%correctly quiet

Classic MBMP

Classic MBMP
	Flagged	Not flagged
Plume present	14.4%caught	85.6%missed
No plume	9.6%false alarm	90.4%correctly quiet

Sensor early warning (simulated)

Sensor early warning (simulated)
	Warned	Not warned
Rise within 3 h	56.7%caught	43.3%missed
No rise	2.2%false alarm	97.8%correctly quiet

India field test, January 2024 to December 2025

Ghazipur, Bhalswa, Okhla, Deonar and Pirana, each with a control point 5 km away where no landfill is. A flag at a control point is a false alarm.

India field test, January 2024 to December 2025
	Clear scenes	Flags at model threshold 0.844	Flags at India threshold 0.793
Landfills	340	14	31
Control points	334	0	4
Landfill vs control (p)		< 0.001	≈ 1e-6

A flag is a scene scored above the threshold, before the physics gate sorts it.

Physics gate on the 14 landfill flags: T1 1, T2 2, T3 11 (2 burn-like at Ghazipur).
Per-pass detection over dense Indian landfills: about 25–35% for plumes of 10 t/h or more.
Known-truth test: the estimated rate was 0.98× the injected rate at 10 t/h.
Steady emission ruled out above (95%): Ghazipur 5, Bhalswa 5, Deonar 10, Pirana 10, Okhla 20 t/h.
Fine-tuning on Indian backgrounds did not improve held-out detection (AUC about 0.55): the limit is the Sentinel-2 signal over dense cities, which is why VayuNetra is a two-tier system.

Sensor early-warning model (simulated)

Trained and tested on simulated readings driven by real weather at the five landfills, tested on held-out Jul–Sep 2025. These numbers describe the simulator, not real air; the model is retrained once real nodes report.

Sensor early-warning model (simulated)
Measure	Model	Rate-of-change alarm
Rises warned in advance	89%	60%
Median warning ahead	150 min	60 min
False-alarm ratelower is better	2.2%	1.2%

Sensor early-warning model (simulated)
Measure	Model	Rate-of-change alarm
ROC AUC	0.962	—
Recall at 5% false alarms	0.77	—
Recall at 10% false alarms	0.90	—
Recall at 20% false alarms	0.97	—
Precision at the alert threshold	0.46	—

257 weights; about 3.2 false warnings per node per week. Comparison: logistic ROC AUC 0.953.

Try the sensor network in the simulator →

What these numbers do not show

Every satellite estimate is screening-grade: confirm with a hyperspectral satellite, an OGI drone or a ground survey before enforcement or carbon crediting.
On one pass at the operating threshold the satellite model catches about 3 in 10 plumes; repeated passes and the ground tier make up the rest.
T2 and T3 flags are not methane detections: T2 is probable and needs confirmation, T3 is surface change rejected as methane.
The sensor network is simulated until real nodes report; its figures will be re-measured on real readings.

Proof

Flags, confirmation results, mitigation commitments, metered readings and monthly scorecards go into an append-only hash chain, time-stamped on Bitcoin through OpenTimestamps.

Ledger entries: 4
Chain heads stamped: 2
Latest Bitcoin block: #970337

Check the chain yourself →Monthly scorecard →

Screening-grade satellite estimate: confirm with a hyperspectral satellite, an OGI drone or a ground survey before enforcement or carbon crediting.