The SMRT model (Picard et al., 2018) solves the radiative transfer equation for a stratified snowpack by integrating layer-specific electromagnetic properties with microstructure-dependent extinction. It simulates emission, absorption, and multiple scattering processes within the snow and ice layers for multiple frequencies. This forward modelling enables consistent simulation of how variations in snow parameters propagate into TB responses, providing an essential foundation for sensitivity and retrievability analyses.

In this study, we considered both single-layer and multi-layer snow over a representative first-year ice substrate. In addition, a representative multi-year ice substrate was included to assess the influence of sea ice type on snow depth retrievals. Snow is further classified into dry and wet conditions. SMRT provides multiple built-in options for microstructure representation and electromagnetic modelling, and the selected combination determines the parameterisation of snow. In this study, the sticky hard spheres (SHS) microstructure model is combined with the improved Born approximation (IBA) scattering solver (Mätzler, 1998), which is implemented as a standard configuration in SMRT and has been commonly applied in Arctic snow studies (Löwe and Picard, 2015; Picard et al., 2018; Vargel et al., 2020). Under the selected SHS+IBA configuration, each snow layer in SMRT is characterised by a set of snow parameters, including layer thickness, density, temperature, effective grain radius, and stickiness. Effective grain radius is determined by the size and shape of individual scatterers, whereas snow density is determined by the ice volume fraction. Together, these two parameters dictate the number of grains per unit volume. Stickiness further describes the degree of inter-granular bonding, clustering and sintering. The effective grain radius is linearly related to the optical grain radius which can be derived from in-situ measurements, with the corresponding scaling relationship described in detail in Sect. 3.3. Unless otherwise stated, all radius values reported in this study refer to the effective grain radius under microwave conditions. During snow metamorphism, grain growth typically occurs concurrently with densification (Colbeck, 1982). A notable example is the strong correlation between grain radius and density in wind slab, which often consists of small, fragmented grains that are densely packed (Domine et al., 2008). For wet snow simulations, liquid water content (LWC) is additionally prescribed for each layer. TBs are simulated using SMRT for the twelve AMSR2 channels, and all simulations are performed at a fixed incidence angle of 55° (see Sect. 3.2). For each channel, SMRT computes the upwelling TB at the snow–air interface.

Using the current SMRT model configuration, we simulated TBs based on in-situ snow and ice measurements collected during the MOSAiC expedition (see Sect. 3.2) and compared them with spatially and temporally collocated AMSR2 observations. The results show that this SMRT configuration can reproduce the general magnitude and temporal variability of the AMSR2-observed TBs, particularly at intermediate frequencies (18–36 GHz). Details of the simulation and comparison results are provided in Sect. S2 in the Supplement.

We apply EFAST (Saltelli et al., 1999) to decompose the variance of SMRT TBs and to rank the contributions from individual parameters as well as from their interactions. In the EFAST, each input parameter is assigned to a unique angular frequency, and the model response is evaluated along a one-dimensional sampling curve in the multidimensional parameter space. The resulting response is expanded into a Fourier series, such that variance contributions from each parameter are encoded in the amplitudes of the corresponding spectral components. Based on the Fourier coefficients, EFAST provides variance-based sensitivity indices including the single sensitivity index (SSI) and the total sensitivity index (TSI), which account for the isolated contribution of a parameter and its contribution including interactions, respectively. Accordingly, global sensitivity analysis is implemented in two steps. Firstly, for the forward model (y=fx,x=(x1,x2,⋯,xn)), the space of the input variables (xi(s)) is sampled along a curve defined as 1xis=12+1πarcsin[sinωis+φi], where i=1,2,⋯,n, and φi is a uniformly distributed random phase shift within the range of [0, 2π), s represents the sample order from 1 to the total number of samples N and ωi is the Fourier frequency, which can be calculated based on the sample size and interference factor: 2ωmax⁡=N-12M, where the N is the sample size for each input parameter, M is the given interference factor, which is usually set to 4.

Secondly, we input the samples into the SMRT framework and expand the resulting outputs using the Fourier series: 3y=fs=∑j=-∞∞Ajcos⁡js+Bjsin⁡js, where the Fourier coefficients Aj and Bj are defined as: 4Aj=12π∫-ππfscosjsds,5Bj=12π∫-ππfssinjsds, over the domain of integer frequencies j∈Z={-∞,⋯,-1,0,1,⋯,+∞}. The total variance of the model output is obtained as: 6D^=12π∫-ππf2sds-12π∫-ππfsds2≈2∑j=1+∞Aj2+Bj2. The partial variance D^i of the individual parameter xi is estimated based on the Fourier coefficients Aj and Bj with respect to its specific Fourier frequency ωi as 7D^i=2∑j=1+∞Ajωi2+Bjωi2. So, we can calculate SSI and TSI: 8SSIi=D^iD^,9TSIi=1-D^-iD^, where D^-i stands for the variance contribution from all Analysis of Variance terms that do not involve the ith parameter, defined as D^-i=∑Dk^,k≠i. The SSI represents the direct contribution of the ith parameter, whereas TSI measures its total contribution, including both its direct effect and all interaction effects with other parameters. If TSI is higher than SSI, it indicates that the parameter exhibits coupling effects with other parameters. The value obtained by subtracting SSI from TSI represents the magnitude of this coupling effect.

While EFAST ranks snow parameters using dimensionless sensitivity indices, it does not directly assess whether a parameter is retrievable under a given snowpack configuration. In a highly non-linear radiative transfer model, sensitivities are state dependent because coupling effects between parameters can cause the value of one variable to alter the response to others. Consequently, the practical retrievability of snow parameters from TBs can vary markedly across different snow structures. To address this issue, we introduce the concept of retrievable range to describe the conditions under which a given parameter can be effectively retrieved from passive microwave observations. We define the retrievable range as the subset of the multi-dimensional parameter space in which a specific parameter can be retrieved with adequate information content from TBs.

A parameter is considered retrievable when the observation provides sufficient independent information to reduce its uncertainty beyond what is implied by measurement noise and intrinsic variability. This notion of retrievability depends on the sensitivity of TBs to the parameter across channels, the observational noise of TBs, and the prior variability of the parameter itself. Following the uncertainty propagation and information content theory, we quantify this process using the averaging kernel matrix, a standard diagnostic within optimal estimation theory (Rodgers, 2000): 10A=KTSe-1K+Sa-1-1KTSe-1K, where K is the Jacobian matrix containing the partial derivatives of simulated TBs with respect to snow parameters, Sa is the prior covariance matrix, and Se is the observation error covariance. The diagonal elements of A, expressed as Aii, indicate the degree to which each parameter is constrained by observations rather than by the prescribed prior. In this sense, the Aii quantify the fraction of the retrieved variability attributable to observation information for parameter i, with values approaching 1 indicating strong observation-driven and values near 0 indicating a prior dominated. Thus, the averaging kernel diagonal provides a physically meaningful criterion for establishing the retrievable parameter space. Because the averaging kernel is rarely strictly binary, a practical threshold is required to classify whether a parameter is observational retrievable. Following the common practice in atmospheric and cryosphere studies (Hou et al., 2018; Stöffelmair et al., 2025), we adopt: 11Aii>0.5, as an operational criterion, indicating that the retrieval of parameter is predominantly driven by the observations rather than the prior, and therefore is retrievable with meaningful independent information content. Conversely, parameters with Aii below this threshold are regarded as weakly constrained by observations and are therefore not effectively retrievable from observations alone. This analysis not only identifies which parameters are retrievable but also delineates the specific regions in the parameter space, i.e. physical regimes, in which the observations become informative (see Sect. 4). By mapping Aii across the sampled multi-dimensional parameter space, this study defines the retrievable range for snow depth under varying snow and sea ice conditions.

In this study, we quantify the retrievable range for snow depth using single channel TBs and gradient ratios (GRs). For the prior uncertainty of snow depth, Sa is specified using an Arctic-wide estimate of natural variability. Following Rückert et al. (2023), we adopted σsnow depth=0.2 m for the prior uncertainty for snow depth retrievability analysis. For single channel TBs, the observation error covariance (Se) is specified using the nominal AMSR2 radiometric noise, assumed to be 1 K uncertainty for each channel. For the GR-based retrievals, Seis derived by propagating the TB measurement uncertainty through the GR formulation. For a GR defined from two TB channels, TB₁ and TB₂, 12GR=TB1-TB2TB1+TB2, the uncertainty of GR is approximated by first-order error propagation, 13σGR2≈∑i≤2∂GR∂TBi2σTBi2+2∑i<j≤2∂GR∂TBi∂GR∂TBjCovTBi,TBj, where σTBi denote the radiometric noise (set to 1 K here), and Cov(TBi,TBj) is the uncertainty covariance between the two channels. Assuming uncorrelated channel noise (Cov(TBi,TBj)=0) and equal variances (σTB1=σTB2=1 K), Eq. (13) reduces to 14σGR=2σTBiTBi+TBj2TBi2+TBj2(i<j≤2). This σGR will be used as the observation error for the GR-based snow depth retrieval in the retrievability analysis.

AMSR2 is a spaceborne passive microwave radiometer onboard the Global Change Observation Mission satellite, providing continuous global observations since 2012. AMSR2 measures upwelling microwave TBs at six frequency bands centred at 6.9, 10.7, 18.7, 23.8, 36, and 89 GHz (henceforth denoted as 6, 10, 18, 23, 36 and 89 GHz for simplicity), operating at a conical incidence angle of approximately 55°. The corresponding footprint sizes are approximately 62×35, 42×24, 22×14, 26×15, 12×7, and 5×3 km, respectively, from the lowest to the highest frequency. In this study, AMSR2 frequencies and polarisations are explicitly represented in the SMRT model simulations, forming the basis for the global sensitivity analysis in which channel specific responses of TB to snow depth variability and parameter interactions are quantified.

The snow parameter configurations in this study are based on MOSAiC in-situ measurements (Frickenhaus et al., 2022; Nicolaus et al., 2022). MOSAiC was a year-long drift expedition in the central Arctic Ocean from September 2019 to October 2020, with the icebreaker Polarstern serving as the central observational platform. The compiled MOSAiC snow pit database includes 576 snow pits and provides profiles of snow depth, density, temperature, snow water equivalent, salinity, and micro-computed tomography (micro-CT) samples, with measurements collected using a broad suite of instruments across the expedition (Macfarlane et al., 2021). In this study, we use in-situ measurements of snow depth, temperature profiles measured using a waterproof thermometer with a needle probe, snow density and specific surface area (SSA) profiles derived from micro-CT analyses, and snow salinity profiles from snow pit samples measured using a salinity, conductivity and temperature sensor.

Although this study primarily focuses on snow, the underlying sea ice also exerts a non-negligible influence on snow radiative transfer. We use the MOSAiC ice core datasets, providing co-located profiles of sea ice thickness, temperature, salinity, density and brine volume fraction (BVF) (Oggier et al., 2024, 2025). Ice cores were classified as first year ice (FYI) and multi-year ice (MYI) and sampled independently. During each coring event, ice temperature was measured using the thermometer inserted into pre-drilled holes. Ice salinity was measured from melted 5 cm core sections with a conductivity meter. Ice density was measured on a separate density core using the hydrostatic weighing method. Relative BVFs were derived from the measured salinity, temperature, and density following Cox and Weeks (1983) for cold ice.

We use the MOSAiC data together with existing Arctic snow literature (Table 1 in Fuhrhop et al., 1998) to define the snow stratification and to constrain physically plausible ranges for the key snow properties. We consider (1) single-layer snow and (2) multi-layer snow representations and perform the sensitivity analysis separately for each configuration. The multi-layer dry snow is subdivided into three characteristic layers: wind slab, depth hoar, and a saline interlayer (Fig. 1a). Wind slab typically forms through wind-driven redistribution and compaction, resulting in relatively high density and small grain radius. Depth hoar develops under strong vertical temperature gradients via metamorphism, producing relatively low density and coarse-grained structure. We include an idealised saline intermediate layer near the snow–ice interface. This layer represents a possible dense and saline transition between the snowpack and the underlying sea ice. For simplicity, we refer to this idealised intermediate layer as “saline interlayer” in the present study. The formation of saline interlayer may involve several processes. In Arctic sea ice, enhanced basal snow salinity may also arise from upward brine migration from the ice surface into the basal snow (Domine et al., 2004; Nandan et al., 2017; Mallett et al., 2024). Therefore, we treat this layer as an idealised sensitivity scenario to examine the potential radiometric impact of a dense saline layer at the snow-ice interface. Given the limitations of the SHS model in simulating high-density snow (Picard et al., 2022c), the saline interlayer is treated as a thin sea ice layer rather than a snow layer in the SMRT simulation. In this parameterisation, the saline interlayer is represented as an ice-rich medium with brine inclusions, rather than as a high-density snow layer composed of ice grains embedded in an air matrix. For wet snow conditions (Fig. 1b), the stratification was simplified into a wet snow layer and a basal superimposed ice layer, with the latter representing refrozen meltwater from snow that accumulates and freezes at the snow-ice surface. Figure 2 presents the vertical profiles of snow density, temperature, salinity and grain radius measured during MOSAiC, presented separately for the winter dry snow period (November 2019 to May 2020) and the summer wet snow period (June to September 2020). Coloured points represent individual snow pits profiles in each month, whereas the black curve provides the ensemble central tendency at each height, with horizontal bars indicating the corresponding standard deviations. Grain radius (rsnow, m) is derived from measured SSA (Domine et al., 2008): 15rsnow=3ρiceSSA, where ρice denotes the density of pure ice (910 kg m⁻³). It is worth noting that the grain radius used in the SHS model represents an effective microstructural parameter relevant to microwave scattering, which does not necessarily coincide with the optical grain radius derived from SSA. The relationship between these two radii depends on the stickiness. Löwe and Picard (2015) showed that, for a stickiness value of 0.13, the SHS grain radius is approximately equal to the optical grain radius. In this study, the stickiness was constrained to the range of 0.1–0.3, and the observed grain radius were used only to define the parameter range for the sensitivity analysis. Therefore, we directly adopted the grain radius estimated from Eq. (15), while also considering the applicable range of effective grain radius for the SHS model in the simulations.

Based on the MOSAiC winter dry snow measurements, the layered snowpack is clearly expressed in the vertical increase of grain radius with depth (Fig. 2d). The upper snow is characterised by small grain radius, consistent with a dense, fine-grained wind slab formed under wind compaction. With increasing snow depth, the grain radius progressively increases, ultimately transitioning to a coarse-grained layer that is characteristic of depth hoar. Snow salinity is often assumed to be negligible for Arctic snow, and the MOSAiC observations broadly support this assumption for snow layers above approximately 0.1 m, where salinity remains close to zero (Fig. 2f). However, MOSAiC also reveals a distinct regime in the near to basal snow (< 10 cm), where salinity can increase sharply, reaching values as high as 10 (Fig. 2f). We conclude from this pronounced salinity enrichment that the snowpack cannot always be treated as a purely fresh medium near the snow–ice interface. In this study, we refer to this idealised intermediate layer as the saline interlayer. Given that the SMRT model derives the BVF from salinity and temperature and subsequently uses it to simulate scattering, we directly parameterise the saline interlayer using the BVF. The stickiness in the SHS is an effective parameter describing short-range spatial correlation among snow grains. In SMRT, it is used together with grain radius to control the simulated microwave scattering and emission. Since stickiness is not a strictly physical parameter itself, it is not discussed in further detail in this study. The parameter ranges (See Tables 1 and 2) were primarily constrained by the MOSAiC observations, while also accounting for the applicable range of the SMRT configuration and the ranges adopted in Fuhrhop et al. (1998). For example, the snow density upper limit was set to 500 kg m⁻³, rather than including MOSAiC measurements reaching 700–900 kg m⁻³ (Fig. 2b and c), because the performance of the IBA+SHS model deteriorates at high snow densities (Picard et al., 2022c). This is also broadly consistent with the snow density upper range of 546 kg m⁻³ used by Fuhrhop et al. (1998).

For wet snow conditions, the presence of liquid water markedly alters both the microstructure and the thermodynamic properties of the snowpack. Once melt begins, liquid water increases inter-grain bonding and promotes rapid grain coarsening, while percolation and refreezing processes redistribute mass vertically and tend to homogenise the snowpack (Colbeck, 1974; Wever et al., 2014; Vuyovich et al., 2017). We represent wet snow in our simulations using a single snow layer, characterised by depth, density, LWC, grain radius and stickiness. Based on the MOSAiC measurements, wet snow temperatures remain close to the melting point (273.15 K) with only a small standard deviation throughout the melt season (Fig. 2i). Therefore, in the sensitivity analysis, the wet snow temperature was treated as a fixed parameter and set to 273.15 K. To ensure the validity of the Rayleigh scattering regime at 89 GHz, the grain radius was constrained to be less than 0.5 mm, following the established practices (Mätzler, 1998; Picard et al., 2018). This choice improves numerical stability and maintains physical consistency of the IBA-based scattering calculations in our simulation. In addition to wet snow melting, there often exists a superimposed ice layer at the snow–ice interface. Superimposed ice develops when meltwater generated within the snowpack percolates downward and refreezes upon reaching a colder underlying ice surface (Nicolaus et al., 2003; Granskog et al., 2006). Repeated melt–freeze cycles and sustained percolation can thicken this layer, which effectively modifies the lower boundary condition of the snowpack by increasing reflectivity and reducing permeability (Arndt et al., 2021; Wever et al., 2016). Table 3 presents the physically plausible ranges of the corresponding wet snow parameters.

Notably, the sampling strategy used in the global sensitivity analysis assumes that all parameters are uniformly distributed within their prescribed ranges, which may differ to some extent from the actual distributions of snow parameters. We therefore also considered an alternative sampling strategy based entirely on Gaussian distributions derived from the MOSAiC observations. A comparison of the sensitivity analysis results obtained using the two sampling strategies shows that they are largely consistent. The results of this comparison are provided in the Sect. S3.

In the snow sensitivity analysis, the basal sea ice layer needs to be specified with special care because it provides the lower boundary condition for radiative transfer through the snowpack. Arctic sea ice is commonly categorised into FYI and MYI, which differ primarily in their salinity and BVF. FYI typically retains higher salinity and larger BVF, whereas MYI is generally desalinated through brine drainage and melt processes, resulting in lower BVF (Cox and Weeks, 1974; Notz and Worster, 2009). For winter dry snow analysis, the underlying basal sea ice was assumed to be FYI, whereas for summer wet snow analysis it was assumed to be MYI.

The porous, air-rich upper layer of MYI can produce volume scattering similar to that of snow, thereby introducing a bias in snow-depth retrievals based on GR algorithms (Markus and Cavalieri, 1998; Rostosky et al., 2018). We therefore used the MYI fraction (MYIF), defined as the areal proportion of MYI over sea ice within the sensor footprint, to quantify and evaluate the potential impact of ice type variability on the snow depth retrieval. Guided by MOSAiC observations and accounting for the microwave penetration depth (< 0.5 m in sea ice at the lowest frequency), we represent the basal sea ice medium using a two-layer structure. The upper ice layer extends from the ice surface to 0.5 m depth, while the lower layer represents the ice below 0.5 m. In SMRT, the primary distinction between FYI and MYI is represented through the dominant scattering inclusions. For FYI, scattering is mainly associated with brine inclusions, whereas for MYI it is primarily governed by air bubbles, and MYI density is computed from the porosity in SMRT. Based on the contrasting physical characteristics of these scatterers, we prescribe scattering properties in the SMRT simulation (Shokr and Sinha, 1994; Tonboe, 2010; Picard et al., 2018). Each layer is parameterised by thickness, temperature, and either density/BVF (FYI) or porosity (MYI), as summarised in Table 4. The ice temperature is assumed to be constant in both FYI and MYI at 265 K.

Global sensitivity analysis offers a framework for identifying the snow parameters that most strongly influence TB and characterising how their interactions shape that influence. However, a critical remaining question is whether parameters exhibiting high sensitivity in specific channels can be retrieved robustly. Equally important is to quantify how parameter coupling impacts retrieval performance. This section builds on uncertainty propagation theory and introduces a new metric, the retrievable range, to define the parameter space where a given quantity can be reliably retrieved under varying background conditions.

In recent years, a wide range of snow depth retrieval algorithms have been developed using different passive microwave channels and modelling assumptions (Markus and Cavalieri, 1998; Rostosky et al., 2018; He et al., 2024). However, retrieval ambiguity remains because other factors like snow grain radius and the radiometric contribution from the underlying sea ice also exert strong impacts on TBs. In this section, we focus on the three snow physical parameters that exhibit the strongest coupling effects in the sensitivity analysis (Fig. 3): snow depth, density, and grain radius. We evaluate the retrievable range of snow depth using V-pol TBs and V-pol GRs, because V-pol TBs are widely used in existing retrieval algorithms and exhibit greater sensitivity than H-pol TBs (Markus and Cavalieri, 1998; Rostosky et al., 2018; He et al., 2024). The evaluation is conducted across a range of snow microstructural states and underlying sea ice conditions. This section is based on additional simulations, rather than on the EFAST sample sets used for the global sensitivity analysis in Sect. 4.1. Specifically, the analyses in Sect. 4.2.1 and 4.2.2 use an idealised dry single-layer snowpack over typical Arctic FYI. Here, snow temperature and stickiness are fixed at 260 K and 0.1, respectively, and only snow depth, density, and grain radius are varied. The snow depth range is intentionally extended to 0–1.0 m in Sect. 4.2 in order to diagnose the saturation behaviour of single-channel TBs and GRs and to define the retrievable range.

The principal conclusions of this study were obtained using the SHS microstructure model. SMRT provides several alternative representations of snow microstructure, and the selected configuration may influence both the magnitude of the simulated TBs and their sensitivity to snow properties. In several recent studies, snow microstructure has instead been represented using the exponential model (Wivell et al., 2023; Meloche et al., 2024, 2026; Murfitt et al., 2024). In contrast to the SHS model, which idealises the snow microstructure as spherical ice particles characterised by an effective radius and a stickiness parameter, the exponential model describes the spatial organisation of the ice and air phases using an exponentially decaying autocorrelation function characterised by a correlation length. It therefore does not explicitly assume spherical grains and does not include a parameter equivalent to SHS stickiness. Some studies indicate that the suitability of a microstructure model may depend on snow type, microwave frequency, and the electromagnetic theory employed. The SHS model can provide a physically meaningful description of snow as an assembly of clustered ice particles and has shown performance comparable to, or slightly better than, the exponential model for some weakly metamorphosed or rounded-grain snow (Löwe and Picard, 2015). However, in a direct evaluation using 119 Arctic and sub-Arctic snow pits, Vargel et al. (2020) found that the exponential model produced lower errors than the SHS model, reducing the RMSE by up to 24 % for Arctic snow.

To assess the dependence of our results on this modelling choice, we repeated the sensitivity analysis using the exponential microstructure model for single-layer dry snow. Snow depth, density, temperature, and grain radius were retained as the parameters, with the exponential correlation length derived from the sampled SHS radius through the conversion given by Debye et al. (1957) and Mätzler (2002). Stickiness was fixed at 0.2 in the SHS model because it has no equivalent in the exponential model. Details of the experimental design and complete results are provided in the Supplement. The principal frequency-dependent sensitivity characteristics are generally preserved between the two configurations. Snow density dominates the TBs at 6–10 GHz, whereas the grain radius becomes increasingly important at the higher frequencies, particularly at 36 and 89 GHz. Differences occur mainly within the 18–23 GHz channels, where the dominant parameter and the magnitude of the sensitivity indices vary between the SHS and exponential model. Importantly, the strong coupling effects associated with snow radius, depth, and density, which are a central focus of this study, are evident under both microstructure models and exhibit highly consistent frequency- and polarisation-dependent patterns.

Although this study combines SMRT with EFAST to conduct sensitivity and snow depth retrievability analyses, several limitations should be considered when interpreting the results. First, the MOSAiC measurements represent a spatially and temporally restricted sample of Arctic snow and sea ice conditions rather than the full variability of the Arctic Ocean. Consequently, our results may not capture the full range of conditions across the Arctic, particularly in marginal ice zones, where the snow-covered sea ice is often thinner, more fragmented, and more spatially heterogeneous. However, for studies of microwave radiative transfer from snow-covered sea ice, the MOSAiC measurements remain the most suitable choice because it provides invaluable integrated observations of snow depth, snow microstructure, and sea ice conditions. Second, the simulations were performed specifically at the AMSR2 frequencies and a fixed incidence angle of 55°, and the resulting sensitivity patterns may not be directly transferable to other sensor configurations. Finally, the simulations assume horizontally homogeneous snow and ice conditions. Although variations among sea ice types within a satellite footprint are considered, sub-footprint heterogeneity in snow depth, density, temperature, and microstructure is not explicitly represented. Because the microwave response to these properties may be nonlinear, additional footprint-scale simulations and observational validation may be required before directly applying these results to operational retrievals.

An additional limitation of this study is that all snow and ice interfaces were treated as smooth boundaries. This simplification does not fully represent natural snow-covered sea ice, where roughness may occur between snow layers, and at the snow-ice boundary. SMRT can explicitly represent rough interfaces, and this capability has been applied in several previous studies (Picard et al., 2018; Meloche et al., 2024, 2026; Murfitt et al., 2024). These studies demonstrate that roughness can modify microwave reflection, transmission, and scattering, thereby affecting the magnitude and polarisation characteristics of the simulated microwave signal. Surface and interfacial roughness may consequently influence the retrieval of snow and sea ice properties. (Stroeve et al., 2006; Lee et al., 2018). However, explicitly accounting for roughness would require additional parameters and substantially expand the multi-dimensional parameter space, which is beyond the scope of the present study, which focuses primarily on the effects of snow microstructure. Future work should incorporate surface and interfacial roughness into the SMRT framework to assess whether the sensitivity rankings and snow depth retrievability identified here remain robust under more realistic interface conditions.