When clouds close in, the lakes stay in view.
Glacial lakes in Gilgit-Baltistan and upper Khyber Pakhtunkhwa, watched through cloud with NASA–ISRO NISAR radar.
NASA Space Apps Challenge 2026 · Dancing with the SARs · Islamabad local event
Team SARcastic Scientists: Saeed Ahmad, Shaharyar Sajid, Usama Hussain
Live site · More videos · Guide
- Problem. Pakistan's glacial lakes fill in summer, when cloud hides them from the optical satellites used to track them.
- What SARtale does. It reads every NASA–ISRO NISAR radar pass over each watched lake, turns dark radar pixels into a water area, compares it with that lake's own normal range on the same orbit track, and gives each lake one plain status with the evidence behind it.
- Try it. Lake Watch lists the 39 watched lakes; open any lake for its radar picture, change and history. Anyone can request another lake.
- Honest by design. Screening, not a warning: every status says how sure it is, and what it cannot see.
- Open. Apache-2.0 code, open NASA data, automatic weekly refresh on GitHub Actions.
Click a picture to play the video from the live site.
| 30 s demo | 60 s demo |
|---|---|
![]() |
![]() |
| Play · subtitles | Play · subtitles |
Both videos show the live site. Their narration voice and music are AI-generated (ElevenLabs), as each video says on screen and in its metadata. The 3½-minute how-it-works walkthrough explains the method; all three play with captions on the site.
Over 3,000 glacial lakes have developed in Gilgit-Baltistan and Khyber Pakhtunkhwa (UNDP GLOF-II), and SUPARCO tracks 131 of them as potentially dangerous, mostly with optical satellites. The lakes fill in summer, and summer brings cloud. SUPARCO's 2 July 2026 update says:
"Persistent cloud cover after 16 June 2026 has limited the availability of cloud-free satellite imagery, making it difficult to accurately assess lake area changes during the last three to four days."
Radar sees through cloud, day or night. NISAR has passed over the watched lakes more than 830 times since 18 June 2026, including the weeks when cloud stopped the optical checks; the site gives the current count.
The site opens on the story page, with Open Lake Watch as its main button. Lake Watch is a table of every watched lake, made for disaster-management teams, researchers and anyone following these lakes. It now watches 39 lakes in Gilgit-Baltistan and upper Khyber Pakhtunkhwa, with 503 more mapped lakes shown on the map as candidates.
-
One plain status per lake, with the reason on a second line.
Status Meaning ▲ Alert: growing Water area grew on two passes in a row, beyond twice the lake's own noise ◆ Watch One pass beyond twice the noise, or small changes adding up over recent passes ▼ Decrease noted Water area fell on two passes in a row (information only) ● No change Within the lake's normal range ⬡ Frozen Ice cover. The second line gives the evidence: "Seen in Sentinel-2 since …" or "By air temperature since …; no clear image" ⬡ Ice breaking up A frozen lake whose air has been above 0 °C for a week ◌ Not measured No usable reading; the second line says why (for example "Radar and Sentinel-2 disagree; area for information only" or "Waiting for a clear Sentinel-2 scene") ⊘ No lake now A site, such as a drained ice-dammed lake's basin, with no water in recent imagery; it becomes Lake may be forming when a clear Sentinel-2 view shows water A status holds until later passes clear it, and a newer pass that could not be read never hides it. Every Watch, Alert and Decrease noted carries a short message: what to do, what changed, how sure and what it is not. Each status has its own map marker (shape, text and colour, never colour alone), and What the statuses mean, a key that opens on click, explains them.
-
Puts the lakes that need attention first. Count tiles (Lakes watched, Needs attention, Open water, Frozen, Above mapped outline, No lake now, Not measured, plus Alert: growing, Ice breaking up and Changed since last build when there are any) filter the table; it can also be filtered by district or SUPARCO match, or searched by name or PDGL number. The lakes that need attention head the table under one line saying what to do for them: check recent optical imagery or the site, then contact NDMA (Gilgit-Baltistan) or PDMA (Khyber Pakhtunkhwa). Each lake shows its radar area, its share of the largest mapped outline (Hi-MAG, HKH-PK-2020, OpenStreetMap), its recent trend and its latest pass. Show map opens the lakes on a NASA MODIS picture. A phone layout has the same lists as cards, with filter chips and a List / Map bar.
-
Matches lakes to SUPARCO's list by map position and area, so an officer can line SARtale up with SUPARCO's GLOF updates. SUPARCO does not publish coordinates, so the match is ours, and it is marked probable where it is not certain.
-
Says when a lake is frozen. L-band radar cannot tell new ice from open water, so frozen lakes are found from a clear Sentinel-2 image or from NASA POWER air temperature at the lake's height, never from the radar area. A frozen lake gets no Watch or Alert, and passes taken while it was frozen are left out.
-
Shows the evidence on each lake's page. The message leads with What to do and its buttons: Open recent Sentinel-2 imagery, the official contact (NDMA for Gilgit-Baltistan, PDMA for Khyber Pakhtunkhwa) and Print lake brief. Where a change began with NASA's processing update, a note says it may be a processing effect. Below come the change against the lake's baseline (a gauge), the radar area against the largest mapped outline, the latest radar picture with the water read from it, the history chart and readings on the lake's status track with every pass left out and why, and the limits in plain words with the sources in a closed section. A frozen lake shows an Ice card in place of the gauge.
-
Prints a one-page brief per lake: the message first, then dates, looks, noise and limits.
-
Takes requests for new lakes through GitHub issues. In the story page's Add a lake section, a visitor fills in the form and presses Request a lake, which opens a pre-filled "Add a lake" issue (a GitHub sign-in is needed). A workflow checks the point, counts the NISAR passes over it and comments; a team member adds the
acceptedlabel to start the rebuild, or closes the issue to reject it. See Lake requests.
| A lake's page | The map |
|---|---|
![]() |
![]() |
The story page (The story) is the front page. Its opening gives the lakes watched, the lakes holding water now and the NISAR passes over them, with a link to the lakes that need attention. Seven numbered sections follow: the view (radar against optical, and mountain shadow), the 2022 Shisper lake outburst, the 2026 cloud gap, the measurement at Attabad, the levels of every site, the Add a lake form and the videos. How we know explains the method in four steps, with the technical detail in a closed section.
For every lake and every NISAR pass, SARtale reads a 6 km window straight from NASA's archive. Method D17 smooths the HH speckle with a Lee filter and marks pixels whose HH is at least 3 dB darker than the window median, leaving terrain in radar shadow and layover (from the Copernicus DEM) out of that median, and measures water inside the lake's zone. A lake that is frozen (seen in a clear Sentinel-2 image, or by NASA POWER air temperature at its height) gets no level. Otherwise each pass is compared only with the median of earlier passes on its own track: one pass beyond twice the lake's own noise, or small changes in one direction that add up over several passes, is a Watch; two passes in a row growing are an Alert.
For example, Attabad on 9 November 2026 held 4.43 km² on track 055A, a change of −0.01 km² against the median of its earlier passes on that track and a noise of 0.14 km² (z = −0.1), so it reads No change (from the lake data built 14 Nov 2026; the app shows the current numbers).
The full method, with every constant, is in the guide.
The app runs on the data in the repository; no NASA login is needed.
npm --prefix app install
npm --prefix app run dev # http://localhost:5173
npm --prefix app test -- --run # app tests, including the wording checkRebuilding the data needs conda and an Earthdata login in ~/.netrc:
conda env create -f environment.yml
cd processing
conda run -n sartale python -m sartale.run_pipeline lakes-search # find NISAR passes over each lake
conda run -n sartale python -m sartale.run_pipeline lakes-fetch # cache a 6 km window per pass
conda run -n sartale python -m sartale.run_pipeline lakes-build # write app/public/data/lakes.json
conda run -n sartale python -m sartale.run_pipeline lakes-check # compare radar water with Sentinel-2
conda run -n sartale python -m sartale.run_pipeline lakes-freeze # frozen and break-up states (re-run after every build)
conda run -n sartale pytest tests -qThe developer and user guide is a small website in docs/guide/. It is published with the site at https://ml-dev-hub.github.io/SARtale/guide/; you can also open docs/guide/index.html in a browser.
| Page | For |
|---|---|
| Getting started | Install, run and test |
| Using the website | Lake Watch, a lake's page, statuses, the story page, Add a lake |
| The method | The six steps, noise, exclusions, constants |
| Pipeline and Data files | Commands, the Attabad gate, lakes_in.json and lakes.json |
| Web app | Modules, routing, copy rules, tests |
| Lake requests | The issue flow, labels, accepting and rejecting |
| Deploying (with the owner setup), Contributing, Questions and limits | Hosting, ground rules, FAQ |
The site is at https://ml-dev-hub.github.io/SARtale/, published by .github/workflows/deploy.yml on every push to main. A weekly workflow refreshes the data, and the lake-request workflows handle issues. The owner sets up three things once, in this order, before any lake request is opened (steps in the Deploying page):
- Settings → Pages → Source: GitHub Actions.
- Settings → Environments → new environment
earthdata, deployment branches limited tomain, with the secretsEARTHDATA_USERNAMEandEARTHDATA_PASSWORD; then delete the repository-level copies. - Actions → Lake requests → Run workflow once, to create the labels. The issue form drops labels that do not exist yet, so a request opened before this is never seen.
The default workflow permissions can stay at Read repository contents: every workflow declares the permissions it needs.
On a public repository GitHub pauses scheduled workflows after 60 days without activity.
- Indicative screening on provisional data. Passes up to 21 Oct 2026 are NASA's provisional GCOV release (CRID
P05023), which is not validated and will be reprocessed. Later passes are NASA's standard GCOV product (CRIDP05036); we have no statement that it is validated, and a lake whose track spans both carries a limit because a calibration change can shift areas. A status is not a measure of volume or hazard. - Optical agreement is rough. On clear days radar water is compared with Sentinel-2 water at each lake; How we know gives the overlap from the current data. Neither is a clean reference. A lake whose check overlaps under IoU 0.4 is shown as Not measured with that reason; its area is for information only.
- Zones are approximate. Attabad and Passu are measured in their OpenStreetMap outlines and Shazore in its 2026 Sentinel-2 outline; the others in a search circle around a published coordinate. Wet shore inside the zone reads as water.
- Drawdown is nearly invisible. Wet sediment flats are as dark as water, so the radar sees a lake drawing down onto them at only about a tenth to a fifth of the real decrease.
- Latency. An Alert comes 5–12 days after a change begins in Hunza and 8–12 days in Chitral. This is slow-growth monitoring, not outburst warning.
- Modelled rates. False-alarm and miss rates are modelled from one season's noise, not observed: no lake grew in the 2026 data. They will be estimated again after one more season. Replayed on 2026, the levels raised no growth Alert; Attabad's September drawdown shows as Watches, where the method study's replay (without the 2% noise floor) had a Decrease noted on 10 Sep.
- Small lakes. Lakes below about 0.05 km² are limited by speckle.
- Not an official warning service. Official warnings come from PMD, NDMA, GBDMA and PDMA.
| Path | Contents |
|---|---|
app/ |
The website: TypeScript, Vite, MapLibre, uPlot |
processing/ |
The Python pipeline and its tests |
media/ |
Videos, screenshots and the workflow figure |
docs/ |
The guide, data sources and licences, AI tool use |
SARcastic Scientists, NASA Space Apps Challenge 2026, Islamabad local event, challenge Dancing with the SARs.
| Member | Role | GitHub |
|---|---|---|
| Saeed Ahmad | Team lead | @saeedahmadicp |
| Shaharyar Sajid | Team member | @shaharyar-sajid |
| Usama Hussain | Team member | @UsamaHussain8 |
How AI tools were used is recorded in docs/SI_USE.md.
Code: Apache License 2.0 (LICENSE). Data keep their own licences, listed in docs/DATA_SOURCES.md. Video narration and music were made with ElevenLabs.







