Snowpack extent and snow water equivalent from orbit
Satellite data delivers two complementary snowpack measurements: snow-covered area at 10–30 m resolution from optical sensors, and snow water equivalent from passive microwave brightness temperatures, each with hard physical limits the other partially compensates for.
Sensors
- Sentinel-2 MSI: 10 m resolution in visible and near-infrared bands; 5-day revisit at the equator (2–3 days at mid-latitudes with both satellites). Band 3 (green) and Band 11 (SWIR at 1610 nm) support NDSI computation. Completely blind under cloud, which is the dominant data gap in mountain basins.
- Landsat 8/9 OLI: 30 m resolution; 16-day repeat per satellite, 8-day combined. Bands 3 (green) and 6 (SWIR at 1608 nm) used for NDSI. Archive extends to 1972 across the broader Landsat programme, enabling multi-decadal trend analysis of snow-covered area.
- MODIS Terra/Aqua: 500 m resolution snow products (MOD10A1, MYD10A1) with daily revisit and near-real-time delivery. Coarser than Sentinel-2 or Landsat but operationally useful for basin-wide daily compositing and cloud-gap filling via eight-day or monthly aggregates.
- AMSR2 (GCOM-W1): Passive microwave radiometer operating at 6.9 to 89 GHz. Snow water equivalent derived from brightness-temperature differences between 18 GHz and 36 GHz channels, exploiting volume scattering by snow grains. Native spatial resolution approximately 10 km at 36 GHz, roughly 25 km at 18 GHz. Daily global coverage.
What the normalised difference snow index actually measures
Snow is simultaneously bright in visible wavelengths and strongly absorbing in the shortwave infrared. Vegetation and most bare ground behave in roughly the opposite way. The normalised difference snow index, NDSI, exploits this contrast: it divides the difference between a green band and a SWIR band by their sum. Values above roughly 0.4 are conventionally mapped as snow-covered. At Sentinel-2's 10 m resolution, this means individual fields, forest clearings and valley-floor patches can be resolved, which matters enormously when a basin's snow line is retreating unevenly across aspect and elevation.
The method has a well-documented failure mode: wet snow in the melt phase absorbs more SWIR radiation, compressing the NDSI value and causing underestimation of snow extent at the precise moment hydrologists most need an accurate number. Forested pixels are also problematic because canopy interception masks the snowpack below. Published correction approaches exist, including sub-pixel unmixing and canopy adjustment factors, but they introduce their own uncertainties. Cloud cover is simply an outage: no optical retrieval is possible. In a basin that is persistently cloudy through winter, Sentinel-2 and Landsat may contribute only a handful of clean scenes per season.
Passive microwave SWE: what brightness temperature reveals, and where it fails
Snow water equivalent cannot be inferred from reflectance. You need to sense the volume of the snowpack, not its surface. Passive microwave radiometers do this indirectly: dry snow grains scatter upwelling microwave radiation from the ground below, and the degree of scattering increases with both grain size and snow depth. The brightness temperature difference between a higher-frequency channel (around 36 GHz on AMSR2) and a lower-frequency channel (around 18 GHz) correlates with SWE. The physical relationship has been formalised in algorithms such as the Chang algorithm and its successors, which convert that brightness-temperature gradient to a SWE estimate in millimetres.
Two physical limits constrain the method severely. First, the relationship saturates at roughly 150 mm SWE. Beyond that depth, the snowpack becomes opaque to the microwave signal and additional snow mass produces little further change in brightness temperature, so the retrieval plateaus and underestimates. Second, liquid water in the snowpack during melt events absorbs microwave radiation rather than scattering it, collapsing the brightness-temperature difference to near zero. A wet snowpack looks, to AMSR2, almost indistinguishable from bare ground. These are not instrument defects; they are consequences of microwave physics.
The spatial resolution problem is equally important for operational use. AMSR2's 36 GHz channel has a footprint of roughly 10 km. In a mountainous watershed where SWE can vary by hundreds of millimetres across a few kilometres of elevation and aspect, a single pixel averages across conditions that bear little relation to each other. Retrievals over patchy mountain terrain carry uncertainties that published studies routinely estimate at 30 to 50 per cent of measured SWE, and sometimes more.
The operational compromise: pairing extent with equivalent
No single sensor solves the snowpack problem. The standard operational approach pairs high-resolution optical snow-covered area with coarse-resolution passive microwave SWE, then uses the optical extent to constrain or disaggregate the microwave volume estimate. Several national hydrological services, including those operating under the WMO Global Cryosphere Watch framework, use variants of this combination for seasonal runoff forecasting.
In practice, the optical data defines where snow exists at fine spatial scales and tracks the retreat of the snow line through the melt season. The microwave data provides a basin-integrated SWE estimate that, despite its resolution and saturation limits, carries information about total water storage that optical reflectance cannot supply. Merging the two requires careful handling of their mismatched resolutions and different error characteristics. Downscaling microwave SWE using optical fraction maps and digital elevation models is an active research area, with published methods showing meaningful improvement over raw AMSR2 pixels in complex terrain, though the improvement is conditional on having sufficient cloud-free optical scenes.
Archive depth and what long records are actually good for
Landsat's archive from 1972 onwards is the longest continuous satellite record of snow-covered area available. Sentinel-2 data begins in 2015. MODIS snow products run from 2000. AMSR2 launched in 2012; its predecessor AMSR-E operated from 2002 to 2011, and SSM/I passive microwave records extend back to 1987, though inter-sensor calibration introduces non-trivial uncertainty across that span.
Long records matter for two reasons. Trend detection in snow-covered area or snow season length requires decades of data to separate climate signal from interannual variability. And basin-specific SWE retrieval algorithms can be calibrated against in-situ snow course measurements that have their own multi-decade histories, improving the microwave retrieval for a specific geography. Neither use case requires real-time data; both require careful radiometric consistency across the archive.
What a buyer should ask before commissioning snowpack analytics
The most important question is whether the basin of interest has a cloud climatology that makes optical data usable. A basin in the maritime Pacific Northwest or the Himalayas may have cloud fractions exceeding 80 per cent through winter, which renders Sentinel-2 and Landsat nearly useless for in-season monitoring without aggressive compositing. MODIS daily products help, but at 500 m resolution the mountain-terrain problem returns in a different form.
The second question is whether the snowpack regularly exceeds 150 mm SWE. In deep continental snowpacks, passive microwave retrievals are systematically biased low and the bias grows with depth. If the application is peak SWE estimation for reservoir inflow forecasting, that bias is consequential. If the application is tracking snow-line recession through spring, optical data alone may be sufficient and microwave SWE adds little.
Satellize runs snowpack analytics on open Sentinel-2, Landsat and MODIS archives, with AMSR2 SWE integration available for basin-scale water balance work. For clients with in-situ snow course networks, we can run site-specific algorithm calibration to reduce the microwave retrieval bias for their particular terrain. The Overhead column has covered seasonal snowpack anomalies in Central Asia and the western United States as case studies in how these sensor combinations perform under real conditions.
Typical figures
| Optical snow extent resolution | 10 m (Sentinel-2 MSI), 30 m (Landsat 8/9 OLI), 500 m (MODIS MOD10A1) |
| Optical revisit | 2–3 days (Sentinel-2 at mid-latitudes), 8 days combined (Landsat 8+9), daily (MODIS) |
| Passive microwave SWE resolution | ~10 km at 36 GHz, ~25 km at 18 GHz (AMSR2) |
| Passive microwave revisit | Daily global coverage (AMSR2 on GCOM-W1) |
| Key spectral bands for NDSI | Green (~560 nm) and SWIR (~1610 nm); NDSI threshold ~0.4 for snow classification |
| Passive microwave frequency channels for SWE | 18.7 GHz and 36.5 GHz brightness-temperature difference (AMSR2) |
| SWE retrieval saturation limit | ~150 mm SWE; deeper snowpacks systematically underestimated |
| SWE retrieval uncertainty in mountain terrain | Typically 30–50% of measured SWE; higher in complex topography or wet-snow conditions |
| Archive depth | Snow-covered area: Landsat from 1972; MODIS from 2000; Sentinel-2 from 2015. Passive microwave SWE: SSM/I from 1987, AMSR-E 2002–2011, AMSR2 from 2012 |
| Latency (operational products) | MODIS snow products: same-day to next-day. Sentinel-2 L2A: typically 1–3 hours after acquisition via Copernicus Data Space |
Analytics Satellize can run
| Daily snow-covered area map | NDSI computed from Sentinel-2 MSI or Landsat OLI bands 3 and 6/11; cloud masking via scene classification layer; gap-filling with MODIS eight-day composites | GeoTIFF raster layer, basin or watershed clipped, delivered daily or on acquisition |
| Snow-line elevation time series | NDSI binary snow mask intersected with a digital elevation model (SRTM or Copernicus DEM) to extract median snow-line elevation per acquisition | CSV time series and interactive chart; weekly update through melt season |
| Basin-integrated snow-covered area anomaly | Comparison of current-season snow extent against the MODIS or Landsat climatological mean for the same day-of-year; z-score or percentile ranking | Seasonal anomaly report with percentile ranking against the available archive |
| Passive microwave SWE estimate | AMSR2 brightness-temperature difference (18–36 GHz) converted to SWE via published retrieval algorithms (Chang-type or GlobSnow approach); flagged for wet-snow and saturation conditions | Gridded SWE raster (~10 km), daily, with quality flags for wet-snow and deep-pack saturation |
| Downscaled SWE field | Optical snow-fraction map used to spatially disaggregate AMSR2 SWE pixels; DEM-based elevation stratification applied; method follows published fractional snow cover downscaling literature | 500 m or 1 km SWE raster for mountain basins; uncertainty layer included |
| Melt-onset detection alert | Drop in AMSR2 brightness-temperature difference below wet-snow threshold, cross-checked against Sentinel-2 NDSI change; triggers when both sensors indicate melt onset | Email or API alert with date, basin name and estimated snow-line position at melt onset |
| Multi-decadal snow season trend | Annual snow-covered area and snow season length extracted from Landsat archive (1984 onwards at consistent quality) and MODIS; linear trend and Mann-Kendall significance test | PDF trend report with time-series plots and statistical summary; suitable for water-resource planning documents |
Who does the work
We can get this done for you. Satellize runs its own analyst desk and a strong science team. You do not buy a data feed and work out what it means; our people source the imagery, run the analysis described on this page, and hand you the answer with its confidence limits stated. Discuss this requirement.