A Lagrangian Model for the Physical and Microwave Properties of Winter Snow on Arctic Sea Ice

The modelled evolution of snow depth on a selection of virtual ice floes for the cold season, 2016. Day 1 represents 1st September. One in every 25 ice parcels is shown. Parcels are spawned as open ocean freezes or existing ice diverges. Partials present on model initialisation (1st September) are given a snow depth from Warren et al. (1999), otherwise they are initialised with a 1cm ‘surface scattering layer’.

who cares about snow on sea ice?

Snow depth on sea ice is a critical climate variable but its distribution is relatively unknown on a Pan-Arctic scale. Thicker snow inhibits ice growth in winter by acting as a warming blanket, but in summer it forms a highly reflective shield, protecting the ice below from the sun’s rays. This reflective blanket also controls the amount of light that reaches primary produces in and under the ice.

As well as having competing physical effects, uncertainties surrounding snow on sea ice limit our ability to estimate sea ice thickness from space. Snow weighs down floating sea ice, ‘hiding’ its thickness below the waterline. It also affects the ability of radar altimeters like CryoSat-2 to ‘see’ the height of the ice surface. What are the spatial and temporal distribution of these effects? We don’t exactly know…

 

“Snow depth on sea ice is essentially unmeasured, limiting mass balance estimates and ice thickness retrievals” - IPCC SROCC 2019

 

The ratio of its scientific importance to our scientific knowledge is so high that the state of snow on sea ice was identified as one of the “Key Knowledge Gaps and Uncertainties” in a recent IPCC report on the oceans and the cryosphere.

how it’s been done in the past

Winter in the Arctic Ocean is cold, and snow on sea ice is pretty cold too. But the ice underneath is in direct thermal contact with the ocean. This means the ice is warmer than the snow.

Plot from Rostosky et al. (2018) of the gradient ratio against snow depth as measured by Operation Ice Bridge. A clear correlation, but not a tight fit.

This relative warmth means the snow emits much more thermal radiation that the snow. Being comparatively dull, the snow it acts like a ‘filter’ for the radiation of the ice. Snow depths on sea ice have historically been inferred by quantifying the amount of filtration occurring.

One technique to estimate the filtration is the ‘Gradient Ratio’: first measure a frequency of radiation that’s relatively unaffected by snow - 19 GHz. Compare this to a frequency that’s heavily filtered by snow - 37 GHz. By comparing the 37 GHz radiation to the 19 GHz baseline, we can estimate the filtering effect of the snow and thus its depth.

But the depth of snow isn’t the only thing affecting the Gradient Ratio (GR). The geometry and physical makeup of the underlying ice is also a controlling factor, in addition to the morphology of the snow itself.

modelling the snowpack stratigraphy

The evolution of snow grain shapes on an example virtual floe. Turquoise base layer is the ice surface, red represents melt forms, grey faceted grains, light green fresh snow, dark green wind slab, blue depth hoar and pink rounded grains. Snow height axis shifted by 400cm.

These processes have historically been modelled with a variety of one dimensional models. In this project we use a newly released sea-ice variant of the popular SNOWPACK model (Wever et al., 2020; GMD). But how can we upscale this 1D model to a Pan-Arctic scale?

 

we initialise thousands of 1d model instances on ‘virtual ice floes’, to develop a comprehensive coverage of the arctic ocean snow stratigraphy

 

These ‘virtual ice floes’ are advected around the Arctic using ice motion vectors from Tschudi et al. (2019). As these floes are pushed around the Arctic, they develop complex layered structure with depth hoar, wind slab and ice lenses. This technique is not new, it has been used by Merkouriadi et al. (2019) using the HIGH-TSI 1D model, and by Liston et al. (in revision) using SnowModel.

The novel aspect to this work is the subsequent modelling of each virtual floe’s microwave properties.

Modelling the snowpack microwave properties

Modelled evolution of typical grain diameter. A layer of thick grains (dark blue) develops after a warm period in this floe (trajectory plotted below), which may drive strong microwave scattering effects. Snow height axis shifted from the waterline by 400 cm.

Having modelled the snowpack stratigraphy in detail, we are in a position to model its microwave emissions. To do this we run the Snow Microwave Radiative Transfer emission model (SMRT; Picard et al., 2018). This model is capable of simulating microwave emissions and backscatter at a given frequency and incidence angle.

 

Each virtual floe therefore carries two models: one for its physical properties, and another for its microwave properties. The first feeds the second.

 

This allows us to model the ‘Gradient Ratio’, and compare it to the observed gradient ratio. This will allow us to explore where the method over- and under-estimates the true snow depth. Furthermore the method may allow us to correct for biases induced by, say, warm air intrusions that introduce icy layers.

The evolution of six model variables, but these results are very preliminary! Ice thickness is not (yet) responsive to ridging and rafting so is significantly underestimated. The 37 and 19 GHz brightness temperatures are those of the parcels themselves - not of a modelled satellite footprint which will also contain open water.

To obtain sufficiently continuous coverage to generate a gridded product of properties such as gradient ratio, snow surface temperature and snow-ice interface temperature, tens of thousands of floes must be modelled and analysed.

As such the model can be configured to run in two ways: the first involves the serial processing of SNOWPACK and SMRT - this is good for investigating individual floes as above. The second involves the simulation of all floes over a year with SNOWPACK, and the subsequent ability to process all the snow timeseries with SMRT in parallel on a high performance computer.

Even with parallel processing, running SMRT on thousands of floes in computationally intensive, which requires extensive use of high performance computing.

Making My Own Tracks

I recently began making my own Lagrangian ice-parcel tracks rather than using those of collaborators. This has allowed me to look much more closely at how well my parcel trajectories reflect reality by comparison with the International Arctic Buoy Program. So far the results of my algorithm are looking very positive!

A visualisation of one in every thirty tracks, with each track’s ‘tail’ thirty days long. New tracks are spawned as the ice edge expands and when ice diverges in the pack.

A visualisation of one in every thirty tracks, with each track’s ‘tail’ thirty days long. New tracks are spawned as the ice edge expands and when ice diverges in the pack.