How sure is it?

What the satellite can and can’t see, how often our flags held up when we checked them, and how every lake is checked.

How sure is this?

Strong early warning. Then verify on the ground.

What we see: where water was, each year since 2019, from photos 10 metres sharp.
What we don’t: the legal lake boundary, small walls or sheds, or floods hidden between tall buildings.
Clouds: if a season is too cloudy, we say “not sure” instead of guessing.
Before you act: check the official lake map and ask a lawyer. We tell you what to ask.

How often we're right

We checked our own flags. Here's the score.

A satellite flag is a lead, not a verdict. So we went back to every one of the 46 flags on the lakes we've analysed and compared it with high-resolution historical photos: 11 by a person in Google Earth Pro, 35 by an AI-assisted review that a person can repeat.

14held up: the lake really did turn to land, fill, a road or a structure

21didn't: still water, usually at the edge where floating weeds or bright fill fooled a 10 m pixel

11couldn't tell from the photos available, often reed beds after desilting works

That's why every flag on a lake page says whether it held up. Use the ones that did; treat the rest as a reason to go and look.

How we check

Summer against summer. Only change that stays.

Every photoEvery Sentinel-2 satellite photo of the lake since 2019, free from the AWS Open Data registry.
No cloudsClouds and shadows are dropped; a summer counts only with at least 3 clear photos.
Summer vs summerJanuary–April of each year, so a dry summer is never mistaken for a lost lake.
Water or landOpen water and floating weeds count as lake; filled, bare, built or grassed-over ground doesn't.
Only what staysLand must stay land two summers running, in patches of at least 500 m² (about four 30×40 ft house plots).
Show and shareThen-and-now photos, a map of every change and an evidence pack with draft letters.

Runs on AWS: one Lambda container per lake (about two minutes), Step Functions across lakes, results in S3 and DynamoDB, a monthly EventBridge re-check and SNS email alerts.

The fine print

What the numbers mean, and what they don't.

What we compare: Sentinel-2 satellite photos, January–April each year from 2019 to 2026; post-monsoon photos only grade confidence.
Starting point: The 2019 and 2020 summers. Anything lost before 2019 isn't in this record.
Smallest change: 10 m pixels. A patch must be at least 500 m²; small sheds and walls are missed.
Lasting change: Land two summers running is “confirmed”; one summer is “new”. Water coming back resets it.
Weeds aren't land: Weeds floating on water reflect little shortwave infrared; grass on filled ground reflects much more. Weeds stay lake.
Cloudy years: Fewer than 3 clear photos in a summer means “not enough data”, never a guess.
Not a land survey: We measure against where the water used to be, not the revenue boundary. Legal works (desilting, walkways, sewage plants) show up too.
Check before you act: Every flag needs a look on the ground and in official records, whatever the photos say.

Satellites show that land changed, not who changed it or whether it was allowed. I can tell you where and when a lake changed; you and the records decide what it means.

Data and credits

Open data, credited.

Copernicus Sentinel-2Contains modified Copernicus Sentinel data (2019–2026). Level-2A COGs from the Registry of Open Data on AWS, found through Earth Search by Element 84.

Copernicus Sentinel-1Radar flood maps (Bengaluru 2022, Delhi 2023, Chennai 2015, Hyderabad 2020), from the Registry of Open Data on AWS.

ATREE-CSEI lake outlinesMap of Lakes in Bengaluru Urban, CC BY, via OpenCity.

Checking photosEsri World Imagery Wayback and Google Earth Pro historical imagery, used only to check flags, never republished.

Runs on AWS: Lambda, Step Functions, S3, DynamoDB, EventBridge, SNS, API Gateway, Amplify, Amazon Location, Amazon Translate (this site in Hindi, Kannada and Telugu), Amazon Polly (Jal's voice) and CodeBuild.