Section 2A.1.4.4 Modelling snow structure

Modelling snow

Structure of the snowpack

The density and temperature of snow is not usually uniform throughout the snowpack.  Density at all snow levels is related to how much air is trapped, the ice or water content, and also linked to the temperature of the snow itself.  The upper snow layer, especially if of fresh snow, is largely uncompressed and has relatively low density.  Lower layers in the snowpack generally have greater density due to compaction by the snow above.  

Heat flux differs through each layer of snow according to its density and temperature.  Snow, especially new dry snow, is a good thermal insulator.  Percolation, freezing and melting of water also has an effect upon the transfer, release and absorption of heat in each layer and on the surface.  Nevertheless, the flux of heat through the snowpack, though relatively small, is important.  In particular, upward heat flux from the ground through the layers of snow influences surface snowmelt and sublimation, and of course, surface temperature.

In snow-free areas, there is a ready exchange of heat, moisture and momentum between the atmosphere and underlying surface.   Snow-covered regions have reduced heat conductivity, higher surface albedo, and reduced roughness compared to areas without snow.  Thus for snowy areas there is an effective thermal, hydrological, and mechanical decoupling between the overlying atmosphere and underlying soil.    

Skin temperature of snow and the 2m temperature

The skin temperature of the upper snow layer is governed by the balance between:

• sensible heat fluxes and latent heat fluxes
• downward long wave and shortwave radiation
• upward fluxes of heat through the layer(s) in the snowpack from the underlying ground.

Forecast air temperature at 2m is derived by interpolation between: 

• the model forecast temperature at the lowest model level (L137 at 10m) and 
• the model forecast temperature of the underlying surface (the skin (surface) temperature, T_skin). 

Generally, skin temperature is derived using HTESSEL.  This uses one or more “tiles” describing the land characteristics and evaluate heat fluxes into and from the underlying surfaces.

Calculation of skin temperature of snow is rather complex, however, and depends upon the characteristics of the snowpack throughout its depth.

To address this, a multi-layer snow model is used in two “tiles” within HTESSEL:

• exposed snow - for areas without vegetation above the snowpack, and therefore without interference to the weak incoming and outgoing heat and momentum fluxes.
• forest snow - for areas where there is tree cover sufficient to interfere with incoming and outgoing heat and momentum fluxes.   The heat flux at the snow/atmosphere interface is rather larger than over exposed snow. 

Widespread benefits are seen in the northern hemisphere winter troposphere, particularly the Rocky Mountains and Tibetan Plateau areas.  Results show in improvements in geopotential height, vector wind and humidity forecasts.

Forecasts of 2m air temperature use the skin temperature of the snow as if it were at the model ground surface rather than at the elevation of the snow surface.  This may lead to errors in derivation of 2m temperatures forecast in cases of deep snow.

The multi-layer snow model

The IFS multi-layer snow model uses up to five layers to represent the snowpack and the complex heat fluxes and interactions between them.   It represents the vertical structure and evolution of snow temperature, snow mass, density, and liquid water content in each layer.  The energy flux at the snowpack surface is the balance of the upward and downward energy fluxes, including the effect of any snow evaporation.   Albedo and surface fluxes vary according to the snow extent, depth and ground coverage (with account taken of trees in areas of forest), and age of the snow.  Heat flux from the underlying ground is also incorporated.  The fluxes are illustrated and explained in Fig2A.1.4.4-1, Fig2A.1.4.4-2, Fig2A.1.4.4-3.

The multi-layer snow model has a fairly realistic representation of the vertical density and temperature profiles of the snowpack.   This allows a good representation of its thermal properties.  

The model represents:

• thermal coupling across the snow-atmosphere interface and within the thin top snow layer.  This allows a realistic representation of sporadic surface snow melt and is especially important for warmer sites with wet snow.
• variations in snow density with depth in thick and cold snowpacks.
• percolation, freezing and melting of water which has an effect upon the transfer, release and absorption of heat in each layer and on the surface.
• heat fluxes across the snow-soil interface and within the thin bottom snow layer.  This allows a representation of changes in permafrost.   

When fresh snow falls or melts away, it is added to or subtracted from the top of the snowpack.  Then the layers are reanalysed such that relatively shallow layers of snow are maintained at the top (5cm thick) and at the base (15cm thick) so that the atmosphere/snow and soil/snow heat fluxes can be best modelled.

The skin temperature (T_skin) over snow cannot rise above 0°C and any net positive heat flux at this temperature is used to warm or melt the snow layer.    The flux of heat might be:

• downwards from the atmosphere (e.g. by advection of warmer air,  by insolation, or beneath an area of warm-based cloud). 
• upwards from lower in the snowpack, though this could be offset by latent heat extraction where melting of snow occurs.

Vertical discretisation over flat terrains:

• for snow depth <12.5cm only one snow layer is modelled (Fig2A.1.4.4-1).  Note: Partial cover of the 'tile' is assumed if snow depth is less than 10cm.
• for snow on glaciers, land-ice or sea-ice only one snow layer is modelled. 
• for snow depth >12.5cm the number of snow layers varies up to a maximum of 5 layers according to the total snow depth (Fig2A.1.4.4-2).
• permanent snow is defined as snow water equivalent >= 10m with 5 snow layers (Fig2A.1.4.4-3).

Fig2A.1.4.4-1: Schematic representation of the multi-layer snow scheme. Shallow snow layer.  Snow depth <12.5cm.  (Note: Snow depth < 10cm implies only a partial cover of snow).  Snow on glaciers, land-ice and sea-ice is processed as a single layer.

Fig2A.1.4.4-2: Schematic representation of the multi-layer snow scheme. Deep snow. Snow depth >27.5cm. Any additional snow accumulation is added into the fourth snow layer in order to preserve the characteristics and thermal flux qualities of thinner layers at base and top of the snowpack.  For snow depths between 12.5cm and 27.5cm additional snow is added proportionately to the layers as they are introduced.

Fig2A.1.4.4-3: Schematic representation of the multi-layer snow scheme for permanent snow (e.g. Greenland, Antarctica).  Snow depth is defined as ≥10m.   Any additional snow accumulation is added into the fifth snow layer.  This is to preserve the characteristics and thermal flux qualities of thinner layers at the top of the snowpack.  

Table2.1.4.4-1: List of symbols for parameters shown in Fig2A.1.4.4-1, Fig2A.1.4.4-2, Fig2A.1.4.4-3.

Vertical discretisation over complex terrain areas

A different algorithm is applied to define the snow layers in regions of complex or mountainous terrain where snow depth >25 cm.  These layers are thicker than used for a snowpack with same depth over a flat region.  For example, in complex terrain an 85cm deep snowpack is discretised with layer depths: 16.00cm, 17.25cm, 17.25cm, 17.25cm, 17.25cm.

Complex terrain is defined as regions where the standard deviation of the sub-grid-scale orography is more than 50m.

Ground height data from internationally available datasets at 1km resolution are interpolated to a spectral representation at model resolution.  Also included are standard deviations of mean height, slopes, and direction of unresolved orography.  However,  in rugged mountainous areas smoothing misses important detail and mountain peaks may be under-represented and narrow valleys may not be represented at all.  

Permanent snow areas

In permanent snow areas (e.g. Greenland, Antarctica) a fixed snow layering it is used.  The top four layers of the five have a constant depths of 50 cm.  Any additional snow accumulation is added into the bottom layer (Fig2A.1.4.4-3).  Glaciers are handled differently

Sea-ice

Representation of snow on top of sea ice or ice on lakes is introduced in Cy50r1 in May 2026.  Coupling between the sea ice model to the atmosphere allows sea-ice and variable ice thickness to be represented in the atmospheric forecast model.  Snow cover on ice acts to increase its persistence by increasing the albedo and reducing the heat flux into the modelled ice.  Reduces the warm bias seen in winter over the ice surface, especially in cloud free situations.    Thin sea-ice or lake-ice covered by thin snow grows or melts much faster than does thick ice with deep snow.  

Glaciers

Representation of glaciers is introduced in Cy50r1 in May 2026 by use of a new land-ice tile within HTESSEL.  This allows representation of the proportion of land-ice or glaciers that partly (or wholly) cover the grid box.  The parametrisation scheme uses four layers of equal thickness on top of the soil column.  There is an exchange of energy and water fluxes between the ice and the atmosphere and also at the bottom where the soil (ground) temperature is used as boundary condition and to compute the ice basal heat flux.

The snow layer on top of the glacier is modelled as only a single layer.  The snow is allowed to melt if skin temperatures rise above zero.  Once the snow is depleted, the underlying ice can also melt as it becomes exposed to weather elements.  Incoming energy can melt the snowpack.  If ice is exposed (e.g. the snow completely melts) the ice itself can melt if the ice surface temperature rises above zero.  The melted water can re-freeze, but otherwise ice-melt runoff is added to the surface runoff over the grid-box.  Meltwater can re-infiltrate the ice column (e.g. moulins in glaciers).  But the additional melting of snow and ice over glacier points generally leads to an increase of river discharge.   These processes can significantly contribute to the total streamflow, particularly during late spring and summer.

These modifications to the snow scheme allow a more realistic representation of the snowpack properties over glaciers and ice-sheets.  Basically the land-ice tile is similar to that used for sea-ice.

The new glacier scheme replaces the binary glacier mask (either wholly covering a grid square or not) used in Cy49 and earlier.

Fig2A.1.4.4-4: Schematic representation of the multi-layer ice scheme for glacier (e.g. Greenland, Antarctica), and-ice and sea-ice representation in the sea-ice scheme SI³.  There are four layers of ice in the model.  Any accumulation of snow is added to the surface and the snow is modelled as a single layer.

Fig2A.1.4.4-5: Glacier and permanent ice cover is predominantly in the coloured areas.  These are predominantly in Greenland and other Arctic mountainous island areas.  Permanent snow is located in the Himalayan region. Isolated locations are in the Alpine region, Caucasus mountains and southern Norway.  Elsewhere the grid squares can have partial ice or glacier cover and this is taken care of by a glacier/land-ice tile in HTESSEL.  Many coastal regions of Greenland, Ellesmere Island and Baffin Island have less permanent snow and ice.  Snowmelt and warming of exposed rock allows better modelling of temperature and precipitation changes. 

Snow Indication

Snow depth

The model snow depth changes when fresh snow falls, or when snow on the ground melts, evaporates or is compressed.  The response in dry periods at different altitudes is shown in Fig2A.1.4.4-9. 

Fresh snowfall is added to the top layer, with a new snow density depending on air temperature and wind speed.  Both convective and dynamic snowfall is considered to be homogenous over the grid box.  Melted snow is removed from the top layer as meltwater runoff.  The snow mass is then redistributed across the different layers.  However, relatively shallow layers of snow are maintained at the top and at the base of the snowpack.  This is so that the atmosphere/snow and soil/snow heat fluxes can be best modelled. 

Liquid water from rainfall onto snow or some snow-melt percolates downwards and can refreeze on a different level, releasing latent heat.  If snow already exists on the ground then freezing rain and ice pellets are accounted for as rainfall that has frozen.

Snow depth water equivalent is the sum of frozen and liquid water within the snowpack.  Snow density considers meltwater refreezing, so the density will vary but the snow water equivalent should not change.

The snow depth of each layer is calculated by snow depth water equivalent divided by snow density for each layer.  

Snow depth is computed using:

• the liquid water equivalent of snow lying on the ground.
• the average density of the snow layer (typically lower for fresh uncompacted snow, higher for compacted old snow).

The total snow depth is the sum of the snow depth of each layer.

When snow depth is:

• < 12.5cm, fresh snow produced by IFS is added to the single snow layer.
• 12.5cm < 27.5cm fresh snow produced by IFS is added proportionately to the layers as they are introduced.
• >27.5cm, only the layer 4 is used as the snow accumulation layer.

For permanent snow areas (e.g.Greenland, Antarctica):

• the snow depth is defined as 10m of water equivalent.
• there is a fixed snow density for all 5 layers.
• the depth of each of the upper four layers is 0.50m; fresh snowfall is added the bottom layer (i.e. layer 5 is used as an accumulation layer). See Fig2A.1.4.4-3.

It is common for snow depth to be extremely high at grid points within these areas of permanent snow.  

Fresh snow seems to be too dense and compacted in the model.  As a rough rule of thumb:

• generally 10mm of precipitation approximates to 10cm of snow depth. 
• for less dense newly fallen snow, 10mm of precipitation approximates to 10-15cm of snow depth.
• with wind effects the density of snow increases quite quickly, 10mm of precipitation approximates to 7cm of snow depth.  
• very low negative temperatures are required to considerably decrease the density and increase the equivalent snow depth in cm.

The current snow scheme tends to melt snow too slowly.

Snow cover

Snow cover is diagnosed from the water equivalent of the modelled snow: 

• Total snow cover is assumed where snow depth is diagnosed as >10cm.  Only exposed snow or forest snow "tiles" are used by HTESSEL.
• Partial snow cover is assumed where snow depth is diagnosed as <10cm.  A snow water equivalent of 6cm is considered to be associated with 60% snow cover (Fig2A.1.4.4-7).   In addition to the exposed snow or forest snow "tiles" the remainder of the grid-box is described by other HTESSEL "tiles".

Snow is not intercepted by a tree canopy and will accumulate on the ground.  Snow accumulation on glaciers, land-ice, sea-ice and lake-ice was introduced in Cy50r1 in summer 2026. 

The albedo of snow in forested areas is given by a look-up table depending on (high) vegetation type (Table 2.1.4.4-2).  The albedo of exposed snow decays with time between 0.85 for fresh white snow to 0.5 for older snow.  It is reset to 0.85 after large snowfall events.

Table2.1.4.4-2: Mean values of Northern Hemisphere five-year (2000-2004) broad band surface albedo (in the presence of snow) aggregated by high vegetation type.

Data assimilation of snow on the ground

Snow cover, snow depth and snow compaction affect all IFS atmospheric forecast models.  It is important the IFS monitors actual values and updates the background fields accordingly.  Any discrepancy will cause errors in the forecast as several physical properties of snow influence:

• the energy and water exchanges between snow surface and atmosphere.
• the upward heat flux from the ground into the atmosphere, which in turn influences surface snowmelt and sublimation.
• the albedo.

Model variables of snow need to be reanalysed at each analysis cycle.  These are:

• snow temperature,
• water equivalent of snow,
• liquid water content.

Snowfields are initialized every day at 00UTC from continuous offline data.

Snow data assimilation at ECMWF relies on:

• an Optimal Interpolation method which adjusts the model-analysed snow water equivalent and snow density prognostic variables.
• conventional measurements of snow depth (from SYNOP and other national networks) with additional national snow depth observations, particularly in Europe and North America.  These are generally an important and reliable source of information.  However, snow depth observations from many other regions of the world remain unavailable to IFS.  Thick hoar frost (which can look like a dusting of snow) can be incorrectly reported as very shallow snow.  This can be assimilated by the model despite no supporting evidence from other sources.  
• snow extent data from the NOAA/NESDIS Interactive Multi-sensor Snow and Ice Mapping System (IMS).  This combines satellite visible and microwave data with weather station reports.  It gives snow cover information and sea ice extent over the northern hemisphere at 4km resolution.  There is some manual intervention and quality control.  The IMS product only shows where at least 50% of the grid cell is covered by snow.  This is converted to snow depths using relationships shown in  Fig2A.1.4.4-8 and Fig2A.1.4.4-9.  IMS data is not currently used by the IFS at altitudes above 1500m.
• SNOTEL ~900 automated observations of snow depth in USA.  These have to satisfy stringent QC before assimilation but have proved useful.

Incorrect analyses and forecasts of snow are possible:

• in data sparse areas and the representativeness of observations.
• after a prolonged period without observations.
• at altitudes above 1500m.
• snow on glaciers is modelled as a single layer which may not be sufficiently thick.  In Cy49 and earlier, glaciers were considered as very deep snow rather than ice.  Either could cause nearby correct observations to be rejected.
• at some high-latitude grid points in the model it is common for snow depth to be extremely high and may not be assimilated.

At high levels (altitudes >1500m) IMS data is not used and observations of snow depth are sparse or non-existent.  In these cases snow depth prediction depends only upon the short range IFS evolution.  Thus there can be little or no decrease in snow water equivalent (if it remains cold enough), though there might be an increase after further forecast or actual snowfalls.   Snow depths may also reduce because the density of the snow tends to increase with time through compaction in the model (and also in reality).  Snow depths in such regions rise in response to forecast snowfall but may not decrease sufficiently at other times (See example in Fig2A.1.4.4-6 and Fig2A.1.4.4-10).

Interpreting snow forecast information

Some considerations should be made before issuing forecasts when interpreting model output of snow.

Fig2A.1.4.4-6:. Weather station at Røldalsfjellet (Norway).  The temperature sensor is mounted at 5m above the ground (left picture).  This allows sufficient clearance beneath the sensor with high snow accumulation (right picture). Photos:MET Norway.

Fig2A.1.4.4-7: Snow depth (cm) and sea-ice cover (%) at the resolution of Ensemble Control Forecast.  DT 12UTC 07 Feb 2023 T+00.  Note frozen lakes (e.g. NW Russia, north Caspian Sea, Uzbekistan) are also plotted as "sea ice".  FLake represents or generates ice on coastal or inland  water.  

Fig2A.1.4.4-8: Conversion of background and forecast snow water equivalents to snow cover.  Forecast snow water equivalents of 10cm or greater are considered as associated with full cover of snow on the ground.  Snow water equivalent of 5cm is considered to be associated with half cover.

Fig2A.1.4.4-9: Conversion of IMS information into an estimate of snow water equivalent for data assimilation.   IMS delivers binary information on the presence of snow for each grid cell but does not give information on snow depth.

• If the background snow water equivalent is 0cm and IMS shows snow cover then the updated snow water equivalent is set to 5cm.
• If IMS shows no snow cover then the updated snow water equivalent is set to 0cm.

IMS strongly impacts upon any updates to the background snow depth field.  Only if both IMS and background fields indicate snow is the IMS information not used.

Fig2A.1.4.4-10: Forecast snow water equivalent at high level stations (blue) and low level stations (red) during the winter of 2019/20.

• At low levels background fields are updated using IMS data and numerous observations of snow depth.  Forecasts show a gradual decrease in snow water equivalent during a dry period.
• At high levels observations are more sparse and IMS data is not used (>1500m).  Background fields rely on earlier snow depth forecasts.  Forecasts show constant snow water equivalent during a dry period.

Additional sources of information

(Note: In older material there may be references to issues that have subsequently been addressed)

A description of the structure and evaluation of multi-layer snow models and associated consequences can be found at:

• IFS DOCUMENTATION – Cy48r3 Operational implementation PART IV: PHYSICAL PROCESSES.
• Improving ERA5-Land soil temperature in permafrost regions using an optimized multi-layer snow scheme
• Hydrological Impact of the New ECMWF Multi-Layer Snow Scheme
• Impact of a Multi-Layer Snow Scheme on Near-Surface Weather Forecasts 
• Read more on the model snow depth and sea ice with example chart and model sea-surface temperature with example chart.
• Read more on snow data assimilation or land surface monitoring by satellite (LDAS).
• Reads more on modelling of glaciers.

(FUG Associated with Cy50r1).