The method
SARtale uses plain statistics on radar backscatter: a darkness rule on HH (method D17), and alert levels from each pass against its own track's earlier passes and noise. No AI model runs in the pipeline or the app.
The data#
SARtale reads NASA's NISAR Level-2 GCOV product (geocoded, polarimetric covariance, about 10 m pixels) from the ASF archive. Passes up to 21 Oct 2026 are from the provisional release, CRID P05023, which NASA has not validated and will reprocess. Later passes are NASA's standard GCOV product, CRID P05036; we have no statement that it is validated. Each pass keeps its own CRID, and a lake whose status track mixes the two carries a limit, because a calibration change between them can shift areas. It uses two polarisations: HH (sent and received horizontally), which decides water, and HV (sent horizontally, received vertically), which helps screen out bad passes.
Six steps, per lake and per pass#
- Find passesSearch NASA's CMR catalogue for NISAR GCOV granules whose footprint covers the lake (
lakes-search). - Read a windowRead a 6 km × 6 km window around the lake straight from the archive with HTTP range requests, never the whole file (
lakes-fetch). Windows are cached indata/lakes/. - Find water (method D17)Smooth the linear HH power with a 7 × 7 Lee filter (4 looks), which calms the speckle but keeps a lake's edge, and convert to decibels. A pixel is water when its HH is at least 3 dB below the median HH of the window. Calm water reflects radar away from the satellite, so it looks dark. HV is not used for water: lake HV sits close to the radar's noise floor, as does the dry ground around it, and asking for both made small lakes read 40–90% low.
- Leave terrain out of the referenceA slope facing away from the satellite falls in radar shadow and looks dark; a steep slope facing it folds over (layover) and looks bright. Both fall on the valley walls, almost never on a flat lake, so nothing is masked inside the lake zone. They are left out of the window median: from the Copernicus GLO-30 terrain model, with the viewing direction fitted to GCOV's own number-of-looks layer (looking 255° from north on ascending passes and 105° on descending ones; NISAR looks left) at 40° incidence. The DEM is read once per lake and kept with its windows; a build without it says so on the lake.
- MeasureCount the water pixels inside the lake's zone and convert to km². The zone is the lake's outline buffered by 200 m where there is one (from OpenStreetMap), otherwise a circle around a published coordinate: 1 km, or smaller where the register sets
zone_radius_kmfor a small pond. - DecideCompare every valid pass, on any track, with the median of earlier valid passes on its own track, since a different viewing angle sees a lake differently, and give the lake a level: No change, Watch, Alert: growing or Decrease noted.
Noise and the alert levels#
Each valid pass is compared once its track has at least 3 earlier valid passes (not the pass itself, so a real change cannot widen its own threshold):
- Baseline: the median of up to 6 earlier valid passes on the same track.
- Noise: 1.4826 × the median of the absolute pass-to-pass differences of up to 8 earlier valid passes on the track, at least 0.005 km² and at least 2% of the baseline.
- z = (area − baseline) / noise.
- Alert: growing: two valid passes in a row, on any tracks, each with z ≥ +2. Neither may be a track's first pass on a processing version its baseline lacks, flagged frozen, or one of the first two passes per track after the lake opened.
- Watch: one valid pass with |z| ≥ 2, or a CUSUM over the z values (k = 0.5) reaching 4; the CUSUM starts again after it makes a Watch, and after a frozen spell.
- Decrease noted: two valid passes in a row with z ≤ −2. Information only, never an Alert: wet, smooth sediment flats are as dark as water, so the radar sees a drawdown at only about a tenth to a fifth of its size. It is the two-pass form of a Watch on a falling area, so it replaces that Watch.
- No change: none of these. A level holds until evidence clears it: Alert: growing until two valid passes in a row have z under +2; Watch until a valid pass is within ±2 with the CUSUM under 4; Decrease noted until two valid passes in a row have z above −2. A higher level replaces a lower one;
level_sincestays the date that set it, and the message gives the newest pass's z. A frozen spell suspends any level, and after it the first two passes per track rebuild the baseline. On one pass an Alert outranks everything, a Watch on a rising area outranks a Decrease noted, and a Decrease noted replaces a Watch on a falling area. - Precedence: No lake now, then Frozen and Break-up likely, then Not yet validated, Not measurable and Not enough NISAR looks yet (no level possible), then Alert: growing, Watch and Decrease noted, then Not observable and Not seen this pass, then No change. A newer pass that could not be read never hides a level: the status keeps the level and adds "The newest pass (date, track) could not be read: reason."
- Version rule: a track's first pass on a processing version (CRID) that its baseline lacks can raise a Watch but cannot be one of the two passes of an Alert. A second pass on that version, on the same track, enters the baseline. When every track's latest pass is its first on a new version, this means about two to three weeks without Alerts and the status says "Alerts paused"; otherwise it says "This pass is track <id>'s first on processing version <crid>; it can raise a Watch but cannot count toward an Alert. Other tracks can."
- Not measurable: no level when the lake's Sentinel-2 check (
lakes-check) overlaps radar water with an IoU under 0.4; the area is shown as information only and the status says why. Under half the zone observable on every pass is Not observable. - Not yet validated against Sentinel-2: no level without a current check, one made for this water method (
D17), this zone and this lake point (each check row stores this askey). A lake with no check, a stale one, or a check that found no usable pair or failed says why. The rebuild runslakes-check --stale-onlyafter the freeze check: it checks only these lakes (one whose last check found no usable pair only when a new open-water-season pass has come), reuses the current checks (kept in the publishedlakes.json), needs no login, and never fails the refresh for one lake. A newly accepted lake is checked in the same run against the summer archive (every pass since 17 Jun 2026 is fetched). It stays Not yet validated only when it has no valid NISAR pass between 15 Jun and 30 Sep, or no Sentinel-2 scene within ±3 days of one with at most 20% cloud over the window and at least 80% of the lake zone clear. - The message: every Watch, Alert and Decrease noted carries a short message, written by the pipeline from the decision values and stored in
lakes.json: what changed (radar water area, km² and % against the baseline), over which passes and dates, how sure (the z values; one pass or two), what it is not (not an outburst forecast; drawdown is hard to see; a change shows 5–12 days after it begins), and what to do (check recent optical imagery or the site; contact GBDMA or PDMA; official warnings come from PMD, NDMA, GBDMA and PDMA). Lake Watch shows it at the top of the lake, the printed brief starts with it, and a banner on the story page and Lake Watch names every lake at Alert. - No lake now: a site the register marks
lake_state: "no_lake"(recent imagery shows no lake, such as a drained ice-dammed lake's basin) gets this status whatever its dark area does. Its passes are still read and listed, but there is no level, no change, no noise and no water layer. Each refresh, all year, re-checks the site with Sentinel-2: the newest scene whose lake zone is at least 80% clear of cloud, shadow, snow and ice, with snow or ice on no more than 1% of it (a frozen or snow-covered pond reads as dry ground), and with pixels within 100 m of cloud or cloud shadow not counted as clear (their edges read as water). It records the scene, its date and the water and snow shares of the zone. When no scene qualifies, the last check stands and the status says that no clear, snow-free view has come since. When the two most recent qualifying scenes, at least 5 days apart, both show water above the site's threshold (4% of the zone by default; set per site where a river or pond shares the zone), the site becomes Lake may be forming: still no level, and its radar dark area is shown but not yet treated as lake area; the SARtale team reviews the site to add it as a measured lake. A single scene above the threshold is recorded as pending and confirmed or cleared by the next one; a forming site returns to No lake now only after two qualifying scenes below it.
Worked example (from the lake data built 14 Nov 2026; the guide is rebuilt from lakes.json, and the app always shows the current numbers). Attabad on 9 November 2026, track 055A: 4.426 km² against a baseline of 4.436 km², the median of the 6 earlier 055A passes before it. The change is −0.010 km² and the noise 0.137 km², so z = −0.1; a level needs 2 × 0.137 = 0.274 km², and the lake reads No change.
Frozen lakes#
L-band radar does not see a mountain lake freeze: new ice under dry snow stays as dark as open water, while frozen ground around the lake darkens and reads as more water (at Shazore the area rose by 44% after freeze-up, while Sentinel-2 showed the lake under snow and ice). So lakes-freeze, run after lakes-build at each refresh, decides whether a lake is probably frozen from air temperature and, when one is clear, a Sentinel-2 image; never from the radar area. Only sites the register holds as lakes are checked; a site with no lake now gets no freeze state.
- Air temperature: NASA POWER daily mean (
T2M) and minimum (T2M_MIN) at the lake from 1 Sep, moved from the POWER grid cell's height to the lake's at −6.5 °C per km. The lake's height is the 10th percentile of the Copernicus DEM GLO-30 within 1 km of the lake point, measured once withlakes-elevation(the rebuild runs it before the build for any lake without one, so a newly accepted lake gets it in the same refresh) and stored inlakes_in.json. Without it the freeze is not assessed. - Frozen (likely): the daily mean below 0 °C on at least 7 days in a row and at least 30 freezing degree-days (the sum of −T on days below 0 °C) since 1 Sep. A missing day ends the run. The rule is late rather than early: where the air hovers around 0 °C it can be one to two weeks late without a clear Sentinel-2 image (in July 2026 it released Shazore about two weeks after Sentinel-2 saw it partly open).
- Frozen (seen by Sentinel-2): for a cold lake (frozen by the rule, or below 0 °C on the newest day, or with nights below 0 °C over the last week), the clearest scene of the last 7 days whose lake zone is at least 80% clear of cloud, cloud shadow and their 100 m edge and whose lake itself is at least half clear. The lake itself is its outline without the buffer, or, for a lake without an outline, its summer radar water core (the zone pixels dark in at least half of a track's open-water-season passes); both shares are measured over its clear pixels, so snow on the shore does not count. Snow or ice on at least 80% of it: frozen, from the freeze-up date when the air already had one, else from the scene date. Water on at least 50%: "Open water seen by Sentinel-2". Either reading holds from one refresh to the next until other evidence: a snow-covered lake stays frozen until a clear scene shows open water or the break-up rules release it; a lake seen open stays so until a new run of 7 days below 0 °C.
- Break-up likely: a frozen lake with at least 7 days in a row above 0 °C, or a radar pass whose lake core is at least 2 dB brighter against its window than in summer in both HH and HV, with a dark share at least 0.3 lower: a brightening that matches ice going out. The test uses core-minus-window contrast only, so a processing-version change cannot trigger it; with fewer than 3 summer passes on a track it is not run there. The lake is released after 14 warm days in a row and one radar pass back within its summer range (contrast within 1 dB, dark share no more than 0.15 below): "Radar back within its summer range on <date>"; or by a clear Sentinel-2 image of open water: "Open water seen by Sentinel-2 on <date>". A lake with no radar summer level is released by the 14 warm days alone and says "No freeze signal in air temperature since <date>; open water not confirmed by radar or Sentinel-2". Air temperature alone never shows open water. A release from an earlier season is not shown.
- Each pass gets a
frozenflag from the same rule on its own date. Frozen passes keep their area but are left out of the baseline, the noise and the summer level and core, and the app greys them and hides the radar water layer while the lake is frozen. A lake frozen today has the frozen status even when its last pass came before freeze-up. - The frozen statuses come before every status but No lake now: no level, and the area is shown greyed as not an open-water measurement. After a frozen spell the first two valid passes per track only rebuild the baseline.
- Freeze-up is called only between 1 Sep and 31 May, but a lake that is frozen or breaking up stays so after 31 May until the break-up and open-again rules release it: a high lake's ice can go out in late June or July (Shazore, 2026). Freezing degree-days start again each 1 Sep; a lake still frozen then stays frozen and says so. When the newest POWER day is more than 10 days old, or POWER cannot be reached, the last check stands (its state, its pass flags and its reason) and says since when it has not been refreshed; a lake never checked is not assessed and keeps the usual status.
- Limits: winter radar is untested (the data so far end just after freeze-up), so the rule is re-tested with late-January data, including whether a radar pass back within its summer range in late winter or early spring really means the ice has gone (new ice under dry snow reads as dark as open water); the air temperature is a reanalysis estimate for a cell about 50 km across with one lapse rate; and nothing here says how thick ice is.
Passes that are left out#
A pass is excluded, with its reason shown in the app, when:
- its window-median HH is more than 3 dB from its track's median (too bright or too dark overall);
- on a track with at least 4 passes, its dark share or window-median HV is an outlier: more than 3.5 robust z-scores (median and 1.4826 × MAD) from the track, with floors of 1 percentage point and 1 dB;
- it is on the manual list in
processing/sartale/config.py. In this build that is2026-09-05_156A, anomalous over the Hunza lakes.
A pass where less than half of the lake zone can be seen is not used for the status either. The screens use the 3 × 3 boxcar and the share of pixels dark in both HH and HV, as before method D17; only the water rule changed.
Constants#
All of these live in processing/sartale/config.py and are copied into lakes.json under constants, so the app's text always matches the build.
| Name | Value | Used for |
|---|---|---|
METHOD_ID | D17 | The water method, recorded in lakes.json under method with its parameters |
DROP_DB | 3.0 dB | How far below the window-median HH a pixel's HH must be to count as water (the screens' dark share: HH and HV both) |
LEE_BOX, LEE_ENL | 7 pixels, 4 | The Lee filter on linear HH before the water rule |
DEM_LOOK_AZIMUTH_DEG, DEM_INCIDENCE_DEG | 255° (A), 105° (D); 40° | The viewing geometry of the layover and shadow mask on the window median |
SPECKLE_BOX | 3 pixels | Box average on linear power for the screens and the pictures |
CHANGE_FACTOR | 2.0 | |z| that makes a Watch (one pass) or, two in a row, an Alert or a Decrease noted |
CUSUM_K, CUSUM_H | 0.5, 4 | The CUSUM over the z values that makes a Watch |
BASELINE_PASSES, NOISE_PASSES | 6, 8 | Earlier passes on the track in the baseline and in the noise |
NOISE_FLOOR_KM2, NOISE_REL_FLOOR | 0.005 km², 2% | Smallest noise a pass is measured against: the larger of the two (2% of the baseline) |
AFTER_OPEN_PASSES | 2 | Passes per track after a frozen spell that only rebuild the baseline |
MIN_VALIDATION_IOU | 0.4 | Sentinel-2 check IoU under which a lake is Not measurable |
MIN_LOOKS | 3 | Earlier valid passes on its track a pass needs before it is compared |
MIN_OBSERVABLE | 0.5 | Share of the lake zone that must be visible |
BRIGHT_PASS_DB | 3.0 dB | Window-median HH distance from the track median that excludes a pass |
OUTLIER_Z | 3.5 | Robust z-score that excludes a pass |
CLOUD_WINDOW | 17 Jun – 4 Aug 2026 | The period counted in the count of valid looks on How we know and in the guided story |
OPEN_WATER_SEASON | 15 Jun – 30 Sep | The only dates the Sentinel-2 cross-check pairs are taken from |
Checking ourselves#
lakes-check compares radar water with Sentinel-2 water (MNDWI) on days when both saw a lake under a clear sky, at every lake that holds a lake now. It uses only pairs in the open-water season (OPEN_WATER_SEASON in config.py, 15 June to 30 September), the clearest one per lake: in October and November a lake can be frozen or under snow, which optical water never counts, so such a pair is no test of radar water. A lake with no clear in-season pair gets a note instead of a number. Neither source is a clean reference: radar counts wet shore inside the zone and is limited by speckle at small lakes, and optical water misses shaded or very dark lakes. The check's IoU also decides whether a lake can be given a level (under 0.4: Not measurable), so lakes-check sets each checked lake's status again.
Method D17 was chosen in a study of 59 variants against 38 radar and optical pairs from summer 2026 at Attabad, Passu, Karambar and Shazore. It raised the median IoU from 0.45 / 0.10 / 0.40 / 0.62 to 0.62 / 0.58 / 0.57 / 0.69 and brought the median area error from −38% / −88% / −49% / −11% to +1% / −16% / −6% / +20%. Its pass-to-pass noise is lower at Passu, Karambar and Shazore and the same at Attabad. Rebuilt from SARtale's own code on the current register (Shazore now has its 2026 Sentinel-2 outline, not the older OpenStreetMap one), the IoU is 0.62 / 0.58 / 0.57 / 0.59. The app words the agreement from the IoU values of the current build, in How we know and in each lake's panel; the full numbers are in lakes.json under checks.
Earth Search marks Sentinel-2 scenes whose pixel values already have the bottom-of-atmosphere offset applied (earthsearch:boa_offset_applied). For those scenes lakes-check uses an offset of 0; applying the −0.1 offset a second time had pushed turbid lakes below the water threshold and made the earlier checks look far worse than they were.
Known limits#
- Drawdown is nearly invisible. Wet, smooth sediment flats are as dark as water, so a lake drawing down onto them shows at only about a tenth to a fifth of the real decrease (Attabad lost about 1 km² of Sentinel-2 water between August and October 2026, and Passu about half its area; their radar areas hardly moved). Growth onto rough ground, ice or vegetation is seen much better.
- Latency. An Alert comes 5–12 days after a change begins in Hunza (three tracks, about five passes in 24 days) and 8–12 days in Chitral (two tracks).
- Slow-growth monitoring, not outburst warning. An outburst takes hours.
- Modelled, not observed. The false-alarm and miss rates (about 0.1 false Alerts per lake per 20 passes; 20–25% growth of Attabad, Passu or Karambar caught within four passes in 90% of trials, 45% at Shazore) come from a noise model fitted to one season, not from events: no lake grew in the 2026 data. With 7 to 13 pass-to-pass differences per lake the noise itself is uncertain by about ±40%. They will be estimated again after one more season.
- The 2026 replay. Replayed on the 2026 series, the levels raised no growth Alert. Attabad's September drawdown shows as Watches around 10 Sep, not as a Decrease noted: with the noise floored at 2% of its 4.3–4.7 km² baseline, its 063D pass of 10 Sep is z = −1.4. The method study's own replay, which had no 2% floor, called a Decrease noted there; the floor binds only at Attabad and does not change how soon its growth would be seen (25%, about 1.1 km²).
- Small lakes. Below about 0.05 km² any per-pixel method here is limited by speckle; at Karambar (0.08 km²) the 0.005 km² noise floor is 6% of the lake.
- Wet shore and reeds inside the 200 m buffer read as water (Shazore's radar area is about 20% above Sentinel-2's in the study's zone).
- Frozen lakes, lakes under 0.05 km² and wind-roughened water are not validated. The frozen status comes from air temperature and Sentinel-2, not from radar; winter radar is untested.
- Only 69% of Attabad's outline falls inside the 6 km window, so its area is the part inside the window. (Lake data built 14 Nov 2026.)
- A 1 km circle around a published coordinate can include land, or miss part of a large lake. Lakes with an outline are measured inside it.
- A site with no lake now (status No lake now) gets no level; its dark radar area is ice, slope or river, not a lake.
SARtale Guide