Explore aligned LiDAR and UAVSAR observations, compare snow products, and inspect the data behind them.
Built by Anant Gahlaut at CryoGARS, Cryosphere, Geophysics and Remote Sensing research lab, Boise State University.
Research preview · 8 field sites · 1,664 display layers
Release v1.0.05 · enrichment 3.3.0 · MIT licensed.
Viewer · Get the archive · Status · Model results
Open the 3D Atlas → · Open the model results →
Choose a site and open its explorer. No sign-in, Python installation, or HDF5 download is required. Use a modern WebGL-capable browser; desktop viewing is recommended.
The Atlas home page with Grand Mesa selected: the site directory, a rotating elevation preview, and the link to that site's 3D explorer.
Banner Summit elevation in the 3D explorer, with a 2nd–98th percentile colour stretch. The data tree, colour legend, and comparison controls remain available beside the map.
- Explore the observations in 3D. View terrain, LiDAR snow depth and vegetation height, UAVSAR amplitude, coherence, phase, and available derived products.
- Find the right layer. Search the archive hierarchy, switch sites, and use the acquisition timeline to inspect survey dates and radar pairs.
- Inspect context and methods. Open product information and notes for units, grid spacing, data availability, source information, equations, and limitations.
- Control the display. Blend overlays, adjust terrain detail and lighting, choose palettes and colour stretches, and save custom palette presets by product type. Choose Smooth interpolation or Cells averaging at the selected Detail spacing, with one colour per displayed block.
- Compare two products. Select another same-product layer or temporarily import a numeric GeoTIFF. Switch between reference A, comparison B, and B − A on the existing 3D terrain.
- Export a 2D figure. Choose comparison colours and percentile limits, then export a north-up PNG with its legend, source labels, comparison direction, coordinate system, and analysis-grid resolution.
Imported GeoTIFFs and comparison results stay in the visitor's browser. They are not uploaded or added to the permanent archive.
Banner Summit LiDAR snow depth, 2021-03-15, with a 2nd–98th percentile colour stretch.
An actual browser comparison at Banner Summit: A = LiDAR snow depth on 2021-03-15; B = the snow-depth prediction from the 30 m random forest benchmark, imported as a GeoTIFF. The model never saw Banner Summit during training. Blue is where the model predicts more snow than the LiDAR measured, red is less. The screenshot uses a robust 2nd–98th percentile colour stretch with symmetric difference limits. This is an exploratory comparison on the explorer's 24 m display grid, not the formal evaluation. More details: model results →
- Prepare a common reference grid. The pipeline reprojects and resamples accepted rasters onto each site's 3 m DEM reference grid and stores the aligned layers, coordinates, dates, and metadata in HDF5. The same row and column represent the same mapped location across aligned raster layers.
- Create a browser-sized export. The exporter averages each array into display-sized blocks and quantizes them, then embeds them with the application in a self-contained HTML page. The browser does not read or download the full HDF5 archive.
- Render and compare locally. WebGL draws a DEM-based terrain mesh and colours it with the selected product. Automatic terrain detail adapts the mesh to the view; it does not increase the underlying data resolution. Comparisons align B to A's exported grid and use their shared valid cells.
The archive grid, exported colour grid, and displayed terrain mesh are different resolutions. A 3 m common grid does not create native 3 m radar information or guarantee common dates, vertical datums, or measurement accuracy. Current website colour values are 8-bit quantized, and some products have clipped tails. Comparison PNGs use the exported analysis grid, not the original 3 m archive or a perspective screenshot. Colour stretches change appearance, not the comparison's underlying samples or statistics.
Use the full archive for quantitative modeling. Read the comparison guide and limitations and scientific notes before interpreting differences.
The eight explorer pages total approximately 113.5 MiB; larger sites can take longer to load. Application code and display data are embedded. Optional Google Fonts have local fallbacks.
Field sites and current viewer coverage
| Site | State | Snow-depth dates | Vegetation-height dates | Radar groups | Display layers |
|---|---|---|---|---|---|
| Grand Mesa | Colorado | 6 | 2 | 27 | 446 |
| Banner Summit | Idaho | 2 | 2 | 11 | 329 |
| Mores Creek | Idaho | 2 | 2 | 10 | 305 |
| Little Cottonwood Canyon | Utah | 1 | 1 | 6 | 177 |
| Fraser | Colorado | 2 | 2 | 5 | 159 |
| Cameron Pass | Colorado | 1 | 1 | 5 | 149 |
| Dry Creek | Idaho | 1 | 1 | 2 | 64 |
| Reynolds Creek | Idaho | 0 | 1 | 1 | 35 |
| Total | 3 states | 15 | 12 | 67 | 1,664 |
These counts describe the published viewer snapshot, not incoming data awaiting integration. Dates are counted within each site. Radar groups are site/date-pair/flight-line combinations; shared flights can appear at multiple sites. Display layers include radar channels and derived products, not just independent observations. Reynolds Creek currently supplies terrain, vegetation, and radar context, but no LiDAR snow-depth labels in this snapshot.
The full-resolution analysis archive is separate from the browser viewer and this source repository. It contains per-site HDF5 files with aligned LiDAR and UAVSAR products, a common 3 m grid, temporal match tables, and processing metadata.
This is how you download the archive. The 16 finished files total 289 GB. One script downloads the original NASA products and builds the archive in a folder you choose:
# 1. Set up the Python environment (once).
conda env create -f environment.yml && conda activate snowex-atlas
# 2. Run the tests.
python -m unittest discover -s tests -t . # pipeline tests
node --test tests/web/test_*.js # viewer tests (Node 22)
# 3. Print the total disk space the chosen sites need. Nothing is downloaded.
python get_dataset.py --dest /path/to/snowex --plan
# 4. Download and build. For each site the script:
# downloads the LiDAR and UAVSAR products from NASA,
# aligns them on the site's 3 m grid -> <site>.h5
# cleans them and adds the derived layers -> <site>.enriched.h5
# checks the result and writes SHA-256 checksums.
# If a run is interrupted, run the same command again and it resumes.
python get_dataset.py --dest /path/to/snowex # all eight sites
python get_dataset.py --dest /path/to/snowex --sites cameron_pass # one siteIt needs a free NASA Earthdata login in
~/.netrc (Windows: ~/_netrc). It downloads the exact products recorded in the
v1.0.05 archive (snowex_inventory.json). A test rebuild of Cameron Pass matched the published files dataset by dataset (same shapes, types and valid-cell counts, values within 1e-4 relative). Rebuilt files are not byte-identical, because HDF5 records creation times and software versions.
The script asks for about 314 GB of free space for all eight sites, which includes 25 GB of working space for downloads. Grand Mesa alone is 222 GB; --plan --sites grand_mesa prints the space for that site.
- LiDAR: terrain elevation, dated snow-depth and vegetation-height layers.
- UAVSAR: available polarizations, amplitudes, coherence, interferograms, unwrapped phase, and approximate geometry.
- Derived layers: slope, aspect, vegetation-height-based canopy fractions, and coherence masks, with product-specific qualifications.
- Context: CRS and affine transforms, source identifiers, survey/date-pair information, radar–LiDAR matches, and processing metadata.
Product availability varies by site. Base <site>.h5 files are already aligned
and processed; <site>.enriched.h5 files contain the enrichment stage. Neither
should be described as untouched sensor data.
HDF5 layout of an enriched file and a small Python example
identification/ site, grid, and processing metadata
matches/ LiDAR–radar matching records
science/
LIDAR/
DEM/grids/elevation
SD/<survey_date>/snow_depth
VH/<survey_date>/veg_height
DERIVED/
UAVSAR/
<date1>_<date2>/<flight_line>/
<polarization>/
GEOMETRY/
Read a small window without loading an entire site:
import h5py
with h5py.File("path/to/archive/banner_summit.enriched.h5", "r") as archive:
info = archive["identification"].attrs
print("CRS:", info["common_crs_epsg"])
print("Grid:", info["common_grid_shape"])
surveys = archive["science/LIDAR/SD"]
survey_date = sorted(surveys.keys())[0]
depth = surveys[survey_date]["snow_depth"]
window = depth[:256, :256]
print(survey_date, window.shape, dict(depth.attrs))Use the recorded transform and CRS to locate pixels. Preserve missing-value masks, and keep readers closed while an archive writer is active. See the workflow guide for acquisition, generation, verification, and building a viewer from an existing archive.
v1.0.05, released 2026-10-08: all eight sites re-enriched to version 3.3.0 and the explorers exported from them. See the release notes.
| Area | Limitation |
|---|---|
| Labels | Snow depth exists for 15 site-surveys at seven sites, from 11 site-winters; five are Grand Mesa in February 2017. Reynolds Creek has none. Label uncertainty is not established. |
| Radar geometry | Incidence is approximate: straight-track, constant-altitude platform; navigation and height compatibility unverified; no terrain occlusion. |
| Vertical reference | Six sites NAVD88/GEOID12b, Grand Mesa WGS84 ellipsoidal, Reynolds Creek unresolved; no conversion is applied. |
| Grand Mesa SWE/density | The builder lists two measured Grand Mesa rasters at 1 m, snow water equivalent and snow density. Neither is in this release. |
| Grand Mesa 2020 snow depth | Covers 2.3% of the site grid. |
| Grand Mesa canopy | Vegetation height and canopy fraction exist only for 2020. The five 2017 snow-depth surveys have no same-year canopy layer. |
| Unwrapped phase | 19 unwrapped-phase layers are empty. The provider stored zeros where unwrapping failed, and enrichment masks them. |
| Rebuilding | get_dataset.py depends on ASF's original UAVSAR URLs, which still redirected on 2026-10-08, after ASF's 2026 reorganisation; fresh catalogue discovery needs a builder update. |
The full list, with the reasoning behind each, is in the technical reference.
The archive is the training and test data for snow-depth models. The first baselines, random forest and gradient boosting, predict LiDAR snow depth from terrain, vegetation and UAVSAR features. Each is tested on a site it never saw during training (leave-one-site-out), at block sizes from 3 m to 99 m.
On unseen sites, mean R² falls from about −0.1 at 3 m to between −0.4 and −0.5 at 99 m, while a random pixel split scores 0.54 to 0.78. The models beat predicting the training sites' average depth by only about 12% in RMSE: they capture part of the spatial pattern but miss each site's overall depth, and terrain explains far more than radar. The results site has the predictions, error maps, feature importance and every training run; the code is in the results repository.
Cameron Pass in the model results viewer: prediction error (model − LiDAR snow depth) for 2021-03-19 from the 30 m gradient boosting model, trained on the other six sites. Blue is where the model predicts more snow than the LiDAR measured, red is less; the colour stretch is symmetric from the 2nd–98th percentiles. Over the 23,528 mapped blocks the RMSE is 0.39 m, the bias −0.07 m and the correlation 0.54.
No model in this release is a validated snow-depth or SWE retrieval.
Documentation index · Design document · Technical reference · Comparison guide · Scientific notes · Data sources · Workflow guide · Release notes
Observations are provided by the NASA SnowEx community, NASA/JPL UAVSAR, and the NSIDC, ASF, and ORNL archives. Cite the original datasets and identify the site, dates, subset, and processing version used in your analysis.
Code license: MIT. Cite this software with CITATION.cff. Source datasets retain their own attribution and use requirements. Developed in the CryoGARS lab with HP Marshall. The source repository, viewer distribution, hosted Atlas and model results are public.