Forward modelling of passive microwave emissivities over snow-covered areas at continental scale

To assimilate passive microwave data in numerical weather prediction, a comprehensive understanding of the components of the radiative transfer equation is essential. Given the significant variability of emissivity in snow-covered regions — affected by frequency, polarisation, and the macro- and microstructural properties of snow — attention must be paid to the design of a forward model. However, existing physical models are unsuitable for global-scale studies due to their reliance on numerous inputs, such as snow grain size across different layers, which are typically unavailable at larger scales. In this study, we propose a method that utilises geophysical properties accessible at the continental scale to derive accurate emissivity values for frequencies ranging from 1 GHz to 90 GHz, in both vertical and horizontal polarisations, with a focus on the incident angles of conical scanners (approximately 50°). Our approach employs neural networks to obtain a robust forward model using geophysical variables as input data. A training dataset was developed based on satellite-derived surface emissivity from the SMOS and AMSR2 instruments by subtracting atmospheric components and surface temperature modulation. The results, which accounts for the actual geophysical state of the surface and its temporal variability, outperform the emissivity climatologies. We achieved snow-covered surface emissivities at the continental scale with a correlation coefficient above 0.9 and a RMSE below 0.02 for frequencies up to 18.7 GHz, and around 0.03 for higher frequencies. Additionally, we demonstrate that, in a typical tundra snowpack where the macro- and microstructural properties of snow can be obtained, the emissivities retrieved by our neural network-based forward model are consistent with results from the physical model (SMRT). This proposed model will also support preparations for the CIMR mission.