MICROWAVE EMISSIVITY PARAMETERIZATION FROM 1.4 GHZ TO 90 GHZ FOR GLOBAL APPLICATIONS, USING MACHINE LEARNING

I. de Gelis^1,2, C. Prigent^2,1, C. Jimenez^1,2

¹Estellus, ²Observatoire De Paris

The importance of accurate microwave surface emissivity models across all surface types has long been recognized, both for retrieving surface parameters from satellites and for assimilating surface-sensitive observations in Numerical Weather Prediction (NWP) systems (Geer et al., 2022). To enable synergistic and robust use of multiple microwave observations across a wide frequency domain, emissivity models must be developed to span the full frequency range and encompass all Earth environments, from oceans and land to ice- and snow-covered surfaces.

Over ice-free ocean, physically based emissivity models exist, with acceptable accuracies (e.g., SURFEM-Ocean distributed with the RTTOV community model, Kilic et al., 2023). In contrast, despite decades of physical modeling efforts over land and sea ice, no existing model can fully reproduce the satellite observations in window channels, across all environments and microwave observing conditions. Surface emissivity exhibits large spatial and temporal variability, ranging from deserts to dense forests and snow-covered surfaces. The passive microwave signal emerging from the surface is influenced by numerous surface parameters (e.g., dielectric properties of surface and sub-surface layers, surface roughness, and topography), in addition to depend upon frequency, incidence angle, and polarization. The interactions of these factors with microwave radiation are highly complex and depend on many poorly constrained and variable parameters, making accurate modeling particularly challenging. For large-scale applications, physically based emissivity models over snow-free land and sea ice have shown notable limitations, even when tuned for specific frequency ranges (e.g., L-band over land, de Rosnay et al., 2020; C-band over sea ice, Burgard et al., 2020), as well as over snow (Hirahara et al., 2020).

Satellite-derived emissivity calculations (e.g., monthly-mean atlases such as TELSEM2, Wang et al., 2016) are publicly available. However, by definition, such climatologies do not capture day-to-day variability and cannot be directly linked to the actual surface conditions, limiting their use to providing average estimates of surface emission.

Here, a pragmatic parameterization of the emissivity of snow, sea ice and land surfaces is proposed, providing consistent emissivity parameterizations between 1.4 and 90 GHz, for both orthogonal polarizations. Satellite-derived microwave emissivities are calculated from the Soil Moisture Active Passive (SMAP), and Soil Moisture Ocean Salinity (SMOS), and Advanced Microwave Scanning Radiometer 2 (AMSR2) observations, subtracting the atmospheric contributions and the surface temperature modulation using ERA5 meteorological reanalysis. The resulting emissivities are analyzed, alongside geophysical parameters, to identify the relevant predictors for the emissivity parameterization. The geophysical variables are preferably extracted from ERA5 reanalysis, but can also be sourced from community model outputs or from auxiliary datasets when ERA5 reanalysis do not provide relevant information (for instance from neXtSIM for the sea ice). A training database of coincident satellite-derived emissivities and geophysical parameters has been constructed to develop a Neural Network (NN) parameterization of emissivity as a function of the geophysical predictors. This pragmatic approach establishes a direct link between calculated emissivities and physical surface properties, thereby eliminating the need for a priori assumptions.

Results have already been obtained over snow-covered surfaces and sea ice (de Gélis et al., 2025; Kilic et al., 2025). Errors are below ∼0.03 for most channels, and can reach ∼0.04 for the higher frequencies. Over snow, the most accurate results are achieved by combining ERA5 input parameters with satellite-derived emissivity climatologies. A comparison has been made with the SMRT physical model over snow, where the macro- and micro-structural properties of the snow could be estimated: reasonable agreement is obtained. The frequency dependence of the emissivity parameterization has been evaluated for possible interpolation for microwave sounders, close to the O2 band around 55 GHz.

The results will also be presented over snow-free land. Relevant predictors are expected to include the open water surfaces, the soil moisture, the surface temperature for different layers, the Leaf Area Index, and the topography.

References:

• Burgard et al., 2020, The Cryosphere, 14(7), 2387-2407.

• de Gélis et al., 2025, Remote Sensing of Environment, 328, 114821.

• de Rosnay et al., 2020, Remote Sensing of Environment, 237, 111424.

• Geer et al., 2022, ECMWF Technical Memoranda URL: https://www.ecmwf.int/node/20337, doi: 10.21957/zi7q6hau.

• Hirahara et al., 2020, Remote Sensing 12, 2946.

• Kilic et al., 2025, Earth and Space Science, in press.

• Kilic et al., 2023, Earth and Space Science, 10(11), e2022EA002785.

• Picard et al., 2018, Geoscientific Model Development 11, 2763–2788.

• Wang et al., 2017, Journal of Atmospheric and Oceanic Technology, 34(5), 1039-1059.