Every mapped glacier and glacial lake in Nepal, joined into one dataset: 4,480 glaciers holding 4,215 km² of ice, 1,429 lakes covering 81 km², annual ice speed to 2025, and every recorded outburst flood. Open data, open pipeline, and explicit about what it cannot tell you.
Glacier outlines RGI 7.0 · Ice velocity ITS_LIVE · Lakes Kumar & Vijay 2026 and Hi-MAG · Floods HMAGLOFDB v4 · Built 1 September 2026
…
Lake fill shows area change 2016–17 → 2022–24: expanding stable shrinking not assessed
Ice and water occupy different elevations. Glacier area peaks between 5,000 and 6,200 m; lakes cluster about a thousand metres lower, on and beyond the debris-covered tongues. The gap between the two curves is where outburst floods start.
glacier area per 50 m band lake area per 100 m band (scaled ×20 for visibility)
Share of lakes that expanded between 2016–17 and 2022–24, grouped by distance to the nearest mapped glacier. Lakes at the ice margin expand around three times as often as detached ones, and the gradient is monotonic. Across all assessable lakes the net area gain is , close to the 5.5% the source study reports for High Mountain Asia as a whole.
Expansion means a gain above both 0.005 km² and the source study's own area uncertainty. Lakes under 0.01 km², or with fewer than three cloud-free images in either period, are not assessed.
RGI 7.0's Nepal outlines have a median source year of 1999. Adding ICIMOD's 2020 outlines for the same glaciers turns what was a caveat into a measurement: lakes touch the 1999 ice margin, but only touch the 2020 one. lakes have been left behind by a retreating terminus in between.
The earlier version of this page inferred retreat from 216 lakes sitting inside polygons RGI still called glacier. With 2020 outlines that inference is replaced by a direct count: of those lakes now lie outside the mapped ice entirely, a median of 234 m beyond it. The rest remain within the 2020 glacier and are genuine supraglacial or embayed lakes.
Judged against the 1999 margin, 32.9% of ice-contact lakes were expanding against 8.9% of detached ones. Judged against 2020, it is 46.4% against 4.0% at the far end, and the gradient with distance is now monotonic where before it flattened out.
That matters as evidence. A mapping artefact between two independently produced inventories would be expected to scatter lakes between the classes and wash the contrast out. Instead the contact group became markedly more distinct, which is what you would expect if the 1999 classification had been diluted by lakes whose glacier had already pulled away.
Ice contact means within 60 m of the mapped margin, about two Sentinel-2 pixels. Both fields ship: dist_to_rgi1999_m and dist_to_ice2020_m, with ice_contact_2020 as the flag to use.
It is tempting: RGI maps 4,215 km² of ice in Nepal around 1999 and ICIMOD maps km² in 2020, which looks like a 16.5% loss. Almost none of that gap is the differing minimum mapping area — only 1.1 km² of RGI's Nepal ice sits below ICIMOD's 0.02 km² threshold.
The problem is that the two disagree in both directions. 1,009 km² of RGI ice is absent from ICIMOD 2020, but 1,388 km² of ICIMOD ice is absent from RGI — more than the apparent loss itself. Different analysts, imagery, and treatment of debris-covered tongues and tributary splits all contribute, and none of it separates cleanly from real change. Only 76% of RGI's ice is mapped as ice by both.
A defensible change figure needs one methodology applied consistently. ICIMOD's publication contains exactly that — 1990, 2000, 2010 and 2020 layers from the same pipeline — but only the 2020 layer is included here. With the other three, this section becomes a real measurement.
Comparing glacier speed in 2000–2009 against 2016–2024 gives a median slowdown of about 46%, with 96% of glaciers apparently slowing. It is an artefact and should not be published.
Speed is the hypotenuse of two velocity components, so it cannot go below zero and random error pushes it upward. Early years rest on a handful of satellite image pairs; recent years rest on thousands. As the error collapsed, apparent speed fell with it. Median signal-to-noise across Nepal is in the 2000s against today — the older velocities sit at the noise floor.
Filtering for high early signal-to-noise makes it worse, not better: it selects the glaciers noise happened to inflate, so regression to the mean reproduces the same slowdown. velocity_change_interpretable is set to false for every glacier.
median image pairs per year (log scale) median velocity error, m/yr — shaded band is the epoch used for current speed
This is a statement about this dataset processed this way, not about Himalayan glaciers. Published work on High Mountain Asia velocity trends has not been verified here; check it independently.
Elsewhere, lake-terminating glaciers often flow faster than land-terminating ones. This data does not show that for Nepal. Raw medians suggest contact-lake glaciers are faster — 2.11 against 1.81 m/yr — but that reverses once you compare like with like, because glaciers with lakes are systematically larger.
| Glacier size, km² | Median m/yr with margin lake | Median m/yr without |
|---|
Read this as an absence of evidence, not evidence of absence: samples are small, Nepal's glaciers are slow and heavily debris-covered, and a glacier-wide mean speed is a blunt instrument for a terminus process.
| Province | Glaciers | Ice km² | Lakes | Lake km² | Floods |
|---|
Assigned by point-in-polygon against OpenStreetMap administrative boundaries — the representative point of each glacier, the centroid of each lake. Recorded floods are HMAGLOFDB v4 events matched to a lake within 3 km. Province totals cover the 4,480 glaciers centred in Nepal; the 190 whose centre falls outside the border are unassigned by design.
Nepal's ice is concentrated in a handful of the 77 districts. Manang, Taplejung, Solukhumbu and Dolpa alone hold about 40% of it.
| District | Glaciers | Ice km² | Lakes | Lake km² |
|---|
Fifteen districts with the most ice, of 24 that hold any.
RGI 7.0 source imagery for Nepal spans 1992–2010, median 1999. Sampling 2024 velocity inside a 1999 outline includes ground that has since deglaciated, biasing glacier-wide means low. The inventory is not a current map of ice extent.
4,679 glaciers intersect the border; 4,480 have their representative point inside it. Headline figures use the second. The first pulls in Tibetan glaciers such as Rongbuk that are under 2% inside Nepal. Both are in the files, with frac_within_nepal.
The Kalapani–Limpiyadhura area in the northwest is disputed between Nepal and India. This build follows the geoBoundaries gbOpen ADM0 line; a different boundary changes far-western counts.
There is no danger score, deliberately. Judging whether a lake is likely to burst needs moraine dam geometry, freeboard, ice content in the dam, slope stability above the lake, and downstream exposure. None of that is in these sources. The inventory supplies the objective ingredients and stops. A composite index built only from area, growth and ice contact would look authoritative and mean very little.
The lake source reports detection near 96% for lakes of 20,000–100,000 m² and 100% above that, but roughly 25% below 5,400 m². Small lakes have caused destructive floods, so do not read the count of tiny lakes as complete.
Every glacier in RGI region 15 carries term_type = 9, unassigned. That gap is why the lake inventory exists: matching lakes back to outlines supplies the flag RGI does not.
Everything is CC-BY-4.0 and rebuildable from source with the pipeline in this repository. The map above uses simplified geometry for speed; the files below are full resolution.
Full-resolution GeoPackages, the annual velocity and lake-area series, and the data dictionaries are in the data/ directory of the repository.