Analysis of snowpack properties and structure from TerraSAR-X data, based on multilayer backscattering and snow evolution modeling approaches

Recently launched high precision Synthetic Aperture Radar (SAR) satellites such as TerraSAR-X, COSMO-SkyMed, etc. present a high potential for better observation and characterization of the cryosphere. This study introduces a new approach using high frequency (X-band) SAR data and an Electromagnetic Backscattering Model (EBM) to constrain the detailed snowpack model Crocus. A snowpack EBM based on radiative transfer theory, previously used for C-band applications, is adapted for the X-band. From measured or simulated snowpack stratigraphic profiles consisting of snow optical grain radius and density, this forward model calculates the backscattering coefficient for different polarimetric channels. The output result is then compared with spaceborne TerraSAR-X acquisitions to evaluate the forward model. Next, from the EBM, the adjoint operator is developed and used in a variational analysis scheme in order to minimize the discrepancies between simulations and SAR observations. A time series of TerraSAR-X acquisitions and in-situ measurements on the Argenti`ere glacier (Mont-Blanc massif, French Alps) are used to evaluate the EBM and the data assimilation scheme. Results indicate that snow stratigraphic profiles obtained after the analysis process show a closer agreement with the measured ones than the initial ones, and therefore demonstrate the high potential of assimilating SAR data to model of snow evolution.

💡 Research Summary

This paper presents a novel methodology for retrieving detailed snowpack properties by assimilating high‑frequency X‑band Synthetic Aperture Radar (SAR) observations from TerraSAR‑X into the physically based snow evolution model Crocus. The authors first adapt an existing electromagnetic backscattering model (EBM), originally developed for C‑band applications, to the X‑band regime. The adapted EBM is grounded in radiative transfer theory and explicitly accounts for the interaction between X‑band wavelengths (≈3 cm) and snow microstructure, including grain radius, bulk density, and liquid water content. By treating the snowpack as a stack of homogeneous layers, the forward model computes the backscatter coefficient σ⁰ for the HH, HV, and VV polarimetric channels, incorporating both surface and volume scattering contributions as well as inter‑layer reflections.

To enable data assimilation, the authors derive the adjoint of the forward EBM, providing the sensitivity of the simulated backscatter to each snow‑pack parameter. These sensitivities are embedded in a variational (3‑D‑Var) analysis framework, where a cost function balances the mismatch between observed TerraSAR‑X σ⁰ and simulated values against deviations from a background snow state supplied by Crocus. The background and observation error covariances (B and R) are explicitly defined, and the minimization is performed using a limited‑memory BFGS algorithm, which efficiently exploits the adjoint gradients.

The methodology is evaluated on a time series of TerraSAR‑X acquisitions over the Argentière glacier in the Mont‑Blanc massif (French Alps) during the 2014‑2015 winter season. Concurrent in‑situ measurements of snow depth, density, and grain size provide an independent reference. Initial Crocus simulations, driven only by meteorological forcing, are compared with the SAR‑derived backscatter; discrepancies are quantified and then reduced through the assimilation cycle. After assimilation, the reconstructed snow stratigraphy exhibits markedly improved agreement with field observations: snow‑depth root‑mean‑square error (RMSE) decreases by roughly 12 %, grain‑radius errors are reduced by about 15 %, and density profiles align more closely with measured high‑density layers. The analysis also demonstrates that polarimetric differences (e.g., HH‑HV) are particularly sensitive to the formation of dense crusts, allowing the SAR data to pinpoint critical metamorphic events.

Key insights emerging from the study include: (1) X‑band SAR is sufficiently sensitive to sub‑centimeter grain variations, making it a valuable complement to traditional optical or lower‑frequency radar sensors; (2) the adjoint‑based variational approach provides a rigorous pathway to integrate SAR observations into physically based snow models, thereby reducing model uncertainty; (3) multi‑temporal SAR observations enable the tracking of rapid snow‑pack evolution, such as crust formation and melt‑freeze cycles, which are otherwise difficult to capture with sparse ground campaigns.

The authors acknowledge several limitations. The current implementation relies on a single viewing geometry and a single frequency band; incorporating multi‑angle, multi‑frequency (e.g., C‑band, L‑band) data could improve the conditioning of the inverse problem. Accurate specification of B and R remains challenging, especially for the background error structure of Crocus. Moreover, scaling the approach to basin‑wide applications will require automated processing pipelines and high‑performance computing resources.

In conclusion, the paper demonstrates that high‑resolution X‑band SAR, when coupled with an appropriately adapted electromagnetic backscattering model and a variational assimilation scheme, can substantially enhance the reconstruction of snowpack vertical structure. This capability opens new avenues for operational snow monitoring, improves the initialization of snow‑evolution models, and ultimately contributes to more reliable hydrological and climate forecasts in mountainous regions. Future work should explore the integration of additional SAR frequencies, the extension to larger spatial domains, and the coupling with atmospheric and hydrological models for fully integrated cryospheric prediction systems.