Snow depth on sea ice from passive microwave and altimetry
Snow on sea ice is the largest source of uncertainty in satellite thickness retrievals, yet it remains poorly measured. Passive-microwave spectral gradients and the freeboard residual between ICESat-2 and CryoSat-2 offer complementary, imperfect routes to a number.
Sensors
- ICESat-2 ATLAS (NASA, launched 2018): Six-beam photon-counting lidar measuring sea-ice freeboard at roughly 0.7 m along-track resolution, 91-day exact repeat. Measures the snow surface, not the ice-snow interface, so it captures total (ice + snow) freeboard.
- CryoSat-2 SIRAL (ESA, launched 2010): Ku-band radar altimeter with a 369-day sub-cycle and a 30-day drifting repeat near the poles. The radar pulse penetrates dry snow and reflects predominantly from the snow-ice interface, giving ice freeboard. Residual between ATLAS and SIRAL freeboards approximates snow depth, with caveats around penetration variability.
- AMSR2 on GCOM-W1 (JAXA, launched 2012): Passive microwave radiometer at 6.9 to 89 GHz. The spectral gradient ratio between 18.7 GHz and 36.5 GHz (vertical polarisation) is the standard proxy for snow depth over first-year ice, with a nominal spatial resolution of 10 km at 36.5 GHz. Revisit is roughly daily at polar latitudes.
- SSMIS on DMSP (USAF/NOAA, F-16 through F-19, 2003 onwards): Predecessor passive-microwave series providing a continuous archive back to SSM/I in 1987. Channels at 19, 22, 37 and 91 GHz underpin the long-term snow-depth climatology used for trend analysis. Spatial resolution 12.5 to 25 km depending on channel.
Why snow depth is the awkward middle variable
Converting sea-ice freeboard to thickness requires knowing snow depth and snow density. Both are poorly constrained from space. Snow density varies from roughly 200 kg/m³ for fresh wind-deposited snow to over 350 kg/m³ for aged, compacted layers, and no satellite sensor measures it directly. Snow depth itself ranges from near zero on young ice to more than 0.5 m on multi-year floes in the central Arctic. That range matters enormously: a 10 cm error in snow depth propagates to a roughly 10 cm error in ice thickness through the hydrostatic balance equation, which is comparable to the thickness signal of interest in thin-ice regimes.
The problem is not new. The Warren et al. (1999) climatology, derived from Soviet drifting-station data, was the community standard for decades. It works reasonably well in the central Arctic but is known to overestimate snow depth in the seasonal ice zone and in the rapidly changing Beaufort and Chukchi seas. Satellite retrievals are an attempt to do better, in near-real time, across the full ice pack.
What a spectral gradient ratio gives away, and what it hides
Passive-microwave snow-depth retrieval rests on a simple physical principle. Dry snow scatters microwave radiation, and the scattering efficiency increases with frequency. A deep snowpack therefore depresses brightness temperature more strongly at 36.5 GHz than at 18.7 GHz. The ratio of the difference to the sum of these two channels (the spectral gradient ratio, or GR) correlates with snow depth over first-year ice in published algorithms, including those distributed by the National Snow and Ice Data Center.
The limits are significant and worth stating plainly. The relationship breaks down over multi-year ice, whose rough surface and internal brine inclusions produce a GR signal that mimics deep snow even when snow is thin. Wet snow absorbs rather than scatters, so spring melt onset collapses the signal entirely. At 10 to 25 km resolution, the footprint averages over leads, ridges and level ice indiscriminately. Published retrieval uncertainties over first-year ice are typically quoted at ±5 to 8 cm root-mean-square against in-situ measurements, but that figure degrades substantially in deformed-ice regimes and in the marginal ice zone.
The freeboard residual method: elegant in theory, noisy in practice
ICESat-2 measures the elevation of the snow surface. CryoSat-2 Ku-band radar nominally measures the snow-ice interface. Subtract one from the other and you have snow depth. The method was formalised in a series of studies around 2020 and has become the most-cited approach for Arctic-wide snow-depth mapping at scales relevant to climate modelling.
The practical complications are considerable. CryoSat-2 radar penetration into snow is not complete: the effective scattering horizon sits somewhere between the snow surface and the ice surface, depending on snow grain size, temperature and salinity. Published estimates of the penetration bias range from a few centimetres to more than 10 cm. The two satellites also have different repeat cycles, 91 days for ICESat-2 and an effective 30-day sub-cycle for CryoSat-2 near the poles, so collocated retrievals require temporal averaging that smears out spatial detail. Over deformed ice and ridges, both altimeters struggle: ICESat-2 loses photons in shadow, and CryoSat-2 SIRAL waveforms become ambiguous. The method is genuinely useful but not a clean measurement.
Fusion and the path to a usable product
Neither approach alone is sufficient. The operational consensus is to use passive-microwave retrievals for their daily coverage and long archive, and altimetric residuals for their higher spatial specificity and physical grounding, then merge the two with a weighting that reflects their respective error structures. Published fusion schemes, including work from the Alfred Wegener Institute and from groups at University College London, typically produce gridded snow-depth fields at 25 km resolution with monthly compositing periods. Uncertainty estimates are included as separate layers and should be taken seriously.
For thickness-conversion workflows, the merged snow-depth product feeds directly into the hydrostatic balance calculation alongside a snow-density assumption. Many operational centres use a fixed climatological density (around 300 kg/m³) for lack of anything better. Some recent work uses ERA5 reanalysis snowfall fields to constrain density spatially, which improves results in regions with strong precipitation gradients but introduces a model dependency that is not always desirable.
Satellize applies these published fusion methods to AMSR2 and altimetry archives for clients who need snow-depth fields as an input to their own ice-thickness or shipping-risk workflows. The analytical approach is the same one underpinning the Tonga crop-estimation programme in a different domain: open-constellation data, published retrieval physics, honest uncertainty quantification.
Operational limits buyers should budget for
Snow depth over sea ice is not a mature operational product in the way that sea-surface temperature or ice extent is. The retrieval chain carries compounding uncertainties at each step. Passive-microwave algorithms are trained on in-situ data that is sparse, biased toward accessible regions and largely pre-2010. The altimetric residual method depends on two satellites whose missions were not designed to be flown together. CryoSat-2 is operating well past its design life of three years and has no confirmed successor in the same Ku-band configuration.
Buyers should expect gridded products at 25 km resolution as the standard output, with finer-scale detail available only where altimetric track density is sufficient. Temporal latency for merged products is typically several weeks to a month, not daily. In summer, passive-microwave retrieval is effectively unavailable due to surface wetness. Over deformed ice, uncertainty is large enough that the product should be treated as qualitative guidance rather than a measurement.
Typical figures
| Passive-microwave spatial resolution | 10 km (AMSR2 at 36.5 GHz) to 25 km (SSMIS at 37 GHz); standard gridded products at 25 km |
| Altimetric residual spatial resolution | Track spacing ~2.5 km (ICESat-2) and ~7.5 km (CryoSat-2) at 75°N; gridded products typically 25 km after compositing |
| Passive-microwave revisit | Daily to sub-daily at polar latitudes (AMSR2, SSMIS) |
| Altimetry revisit (effective for snow depth) | Monthly composites practical; ICESat-2 91-day exact repeat, CryoSat-2 ~30-day sub-cycle near poles |
| Key frequency channels | 18.7 GHz and 36.5 GHz V-pol (spectral gradient ratio); Ku-band 13.575 GHz (CryoSat-2 SIRAL); 532 nm photon-counting lidar (ICESat-2 ATLAS) |
| Typical retrieval uncertainty (first-year ice, dry snow) | ±5 to 8 cm RMS (passive microwave); ±5 to 15 cm (altimetric residual, depending on penetration correction) |
| Retrieval validity season | October to April (Northern Hemisphere); passive-microwave signal collapses at melt onset |
| Archive depth | SSM/I passive-microwave back to 1987; ICESat-2 from October 2018; CryoSat-2 from April 2010 |
| Standard product latency | Passive-microwave near-real-time: 1 to 2 days. Merged altimetric products: 2 to 6 weeks |
| Delivery formats | NetCDF-4 gridded fields with uncertainty layers; GeoTIFF on request; tabular time-series by region |
Analytics Satellize can run
| Monthly Arctic snow-depth grid with uncertainty | AMSR2 spectral gradient ratio (GR1937V algorithm) fused with ICESat-2/CryoSat-2 freeboard residual, following published AWI/UCL methodology | NetCDF-4 gridded field at 25 km, with per-pixel uncertainty layer, delivered monthly |
| Snow-depth anomaly map relative to AMSR2 climatology | Departure from 2003-to-present AMSR2 baseline; standardised anomaly in units of sigma | GeoTIFF anomaly layer and regional summary table, monthly |
| Ice-thickness input correction layer | Snow-depth and assumed-density field formatted for direct ingestion into hydrostatic balance thickness workflows (Warren or satellite-derived snow) | GIS layer compatible with client thickness-retrieval pipeline, with documented density assumptions |
| Seasonal snow-depth time series by defined region | Area-averaged passive-microwave retrievals over client-specified polygon (e.g. Northern Sea Route corridor, Beaufort Sea block) | CSV time series from 2003 to present, with confidence intervals, updated monthly |
| Melt-onset detection flag | Threshold change in GR signal indicating surface wetness, following published NSIDC melt-onset algorithm | Annual melt-onset date grid and regional earliest/latest bounds, delivered post-season |
| Multi-decadal snow-depth trend analysis | Linear trend and Mann-Kendall significance test on SSM/I and AMSR2 combined archive (intercalibrated), 1992 to present | PDF report with trend maps, regional breakdowns and caveats on sensor-transition artefacts |
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.