**Advertisement**

ABSTRACT

Collocated active and passive remote sensing measurements collected at

(

1. Introduction

Detailed knowledge of cloud and precipitation microphysical properties is critical for determining many key aspects of the earth's climate system. This is especially true as numerical models of the climate approach resolutions that can begin to resolve cloud processes. While in situ measurements can be used to determine particle size distribution (PSD) characteristics, in situ data often lack the whole-cloud context that is needed to understand clouds at a process level. Remote sensors provide this context, but at the expense of having to infer geophysical properties via inversion algorithms. Several recent studies have documented success characterizing liquid and ice properties using a combination of Doppler radar and microwave radiometer brightness temperature (Tb) observations (LÖhnert et al. 2001, 2003; McFarlane et al. 2002; DelanoË andHogan 2008; Matrosov et al. 2008; Ebell et al. 2010; Wood 2011; Rambukkange et al. 2011). Several of these studies cast the retrieval problem in a probabilistic framework, in which uncertainties in observations and prior knowledge are each assigned a probability distribution with width proportional to the magnitude of the uncertainty (DelanoË and Hogan 2008; Ebell et al. 2010; Wood 2011). A solution is produced by combining all sources of information in a Bayesian context. The problem is made tractable by assuming all probability distributions are Gaussian, and solving the retrieval in a least squares optimal estimation (OE; Rodgers 2000) framework.

The strength of Bayesian inference in cloud and precipitation property retrieval algorithms is the ease with which multiple observational constraints can be applied to a problem. In liquid-phase clouds, Frisch et al. (1998) exploited the integral constraint that is provided by microwave radiometer-derived liquid water path and showed that liquid water content profiles can be inferred from profiles of radar reflectivity raised to an appropriate exponent. McFarlane et al. (2002) and LÖhnert et al. (2003) extended these ideas to nonprecipitating cumulus and stratocumulus, respectively, using in situ aircraftstatistics and cloud-resolving model output in an OE framework. Ebell et al. (2010) show that the uncertainties in these estimates depend fundamentally on the validity of a priori assumptions that must be derived from in situ aircraftdata. In thin ice clouds, an integral constraint equivalent to the liquid water path in warmclouds is provided by the downwelling infrared radiation (Matrosov et al. 1994;Mace et al. 1998; Zhang and Mace 2006). These ideas were extended to the use of radar Doppler moments by Deng and Mace (2006) and to constraints provided by both lidar and radar by Donovan (2003) and Zhao et al. (2011). DelanoË and Hogan (2008) combined active and passive measurements in an innovative OE algorithm applicable to cirrus. All of these ice cloud studies discussed limitations to accuracy that arose because of assumptions regarding the microphysical properties of ice crystals-specifically the mass- and area-dimensional relationships thatwere assumed. There have been few applications of Bayesian inference in mixed-phase clouds using active and passive sensors.

Bayesian OE has great utility in that it explicitly represents each piece of information and the associated uncertainty, and, if properly implemented, can produce a quantitative estimate of the retrieval error. Even so, the assumption of Gaussian uncertainty and a linear least squares framework is a limitation that can lead to misinterpretation of retrieval results (Posselt et al. 2008). If the relationship between retrieved quantities and observations is nonunique, the OE algorithm may not produce the optimal set of retrieved parameter values. In addition, Gaussian probability distributions are not suitable representations of uncertainty in cases for which parameters are constrained to be positive definite (Posselt et al. 2014).

Markov chain

In this paper, a Markov chain

The remainder of this paper is organized as follows. The case study and forward models for radar and microwave radiometer are described briefly in section 2 and in more detail in appendixes A and

2. Case study, forward models, and retrieval algorithms

a. StormVEx experiment

During the winter of 2010/11, the second Atmospheric Radiation Measurement Program (Mather and Voyles 2013) Mobile Facility (AMF2) conducted its maiden deployment to the

The SWACR collected data during;30-min sequences in zenith-pointing, RHI, and plan position indicator (PPI) scan modes. During the zenith periods, which occupied 18 min of every 30-min sequence, Doppler spectra were collected and archived, and Doppler moments were calculated from those spectra. The campaign benefitted from one of the snowiest winters on record in the northern

The case we examine in this paper was recorded on

b. Forward models

Ultimately the link between a set of measurements y that include profiles of Doppler moments and passive remote sensing constraints from collocated radiometers and inferences of the microphysical properties of the profile x is the set of radiative transfer models F(x). As noted above, the efficacy and error characteristics of any retrieval are strongly dependent on the details of these forward models. As the orographic clouds of interest have been shown to consist primarily of small liquid cloud droplets and a larger precipitating ice cloud mode, we assume in all cases that the PSD is bimodal and above cloud base include a liquid cloud PSD and a PSD of ice-phase hydrometeors (snow). Consistent with in situ measurements of both liquid and ice-phase clouds, we assume that the form of these PSDs can be accurately described by a modified gamma function:

... (1)

where the subscript i refers to either the small particle mode (subscript s-i.e., the liquid cloud mode) or the large particle mode (subscript l-i.e., the snow precipitation mode) so that n(D)5ns(D)1nl(D). In the remainder of this paper, the small mode will be assumed to be liquid and the large mode will be assumed to be ice, but it should be noted that there is flexibility in the forward model for the converse (ice small and liquid large), or for both small and large to be either liquid or ice. Below the precipitating cloud base, we assume a singlemode PSD that consists of snow. Our goal is to derive the PSD parameters of the gamma functions as well as the vertical air motion wair and the standard deviation of the air motion within the radar resolution volume sw so that x5[Nl , Dl , Ns, Ds, wair, sw] and ai are derived empirically, as described later. Integrating the PSD after multiplication by an appropriate empirical power-law assumption determines the microphysical properties of the PSD. The water content, for instance, can be calculated as follows:

... (2)

where the mass of a hydrometeor is mi(D)

To calculate radar measurables, we integrate across the bimodal PSDs to determine the radar backscattered power and attenuation through the cloudy column using scattering and extinction efficiencies as a function of particle size. To accomplish this, we neglect multiple scattering and derive the backscatter cross section as a function of size using Mie theory for liquid (Bohren and Huffman 1983), with refractive indices after Hale and Querry (1973) and T-matrix-derived cross sections for ice-phase hydrometeors provided by Matrosov et al. (2008) in size intervals ranging from 100mm to 2 cm in 100-mm increments. The T-matrix cross sections were calculated assuming oblate spheroids with an aspect ratio of 0.6 and a mass-diameter relationship of m50:003D2 in cgs units. The Maxwell Garnett formulas for mixtures of ice and air were used to compute the refractive indices. While T-matrix calculations are certainly not exactly representative of the scattering properties of the ice crystals present in the observed volume, we have no empirical information (e.g., on the habit distribution and the mass distribution within the habit) that might be used to specify a more exact representation of scattering (e.g., using a discrete dipole approximation). It is possible that the assumption of a single crystal shape and internal mass distribution will cause the retrieval variance to be smaller than it is in reality, but we have no specific information that might be used to confirm or deny this. It is equally possible that assumption of a mixture of crystal shapes and distribution of masses may in the end cause little increase in solution variability. A detailed study using various representations of scattering is certainly warranted, but is beyond the scope of the current study.

We have collected and adapted a set of published forward models that relate hydrometeor properties to measurements. With the exception of the radar forward model, we use published radiative transfer codes, including the microwave radiative transfer algorithm based on the Eddington approximation described by Kummerow et al. (1996) and modified by Lebsock et al. (2011). As the microwave radiometer of interest is vertically pointing (upward looking), the forward-modeled Tbs are nearly insensitive to the value of surface emissivity. We have used a value of 0.8 to strike a balance between the relatively low emissivity of fresh deep dry snow and high emissivity of aged wet snow (Hewison and English 1999).

c. Markov chain

Remote sensing retrievals by definition presume that a set of geophysical quantities can be estimated from a set of indirect measurements, and as such require a model (or set of models) that maps from geophysical parameter space to observation space. The forward problem describes the generation of simulated observations from a set of control parameters, while the inverse problem maps information from observation space into the control parameter space. Themost general solution to an inverse problem combines all pieces of information about the system of interest, taking into account their respective uncertainties. If uncertainties are represented as probability distributions, then the solution is defined as the joint probability distribution of the retrieved quantities conditioned on the observations, choice of model, and prior knowledge. Bayes's theorem provides a compact statement of the relationship between conditional probabilities (Tarantola 2005):

... (3)

where, as above, x is the set of retrieved microphysical properties and y are the observations. The term P(x) represents prior knowledge of the control parameters x, while P(y) represents the probability space containing all possible observations. The optimal estimate of the set of retrieved parameters is defined as the maximum likelihood point in the posterior conditional probability density function (PDF) P(x j y). Retrieval uncertainty can be quantified via calculation of the width of the posteriorPDF[posterior (co)variance, interquartile range, etc.], while relationships between retrieved parameters and observations can be determined via examination of the likelihood P(y j x) or by examining the forwardmodeled response function.

Markov chain

... (4)

where xi is the previously accepted parameter set and the acceptance ratio r(xi, x) is defined as

... (5)

Here, q(x, xi) is the proposal distribution and represents the probability of randomly transitioning from the current parameter set xi to the proposed parameter set x. Conversely, q(xi, x) represents the probability of randomly transitioning from the proposed parameter set x back to the current parameter set xi. If the proposal distribution is symmetric, q(xi, x)5q(x, xi), and Eq. (5) simplifies to the ratio of prior and likelihoods. Our implementation of the

Note that the form of likelihood and prior are completely general-any distribution shape may be assumed. The value of the acceptance ratio r(xi, x) determines whether the proposed set of parameters is stored as a sample of the posterior distribution. If the new set of parameters produces an improved fit to the observations, then this set is saved as the next sample in the distribution. If not, then a test value is drawn from a uniform distribution. If this value is less than the acceptance ratio, then the proposed parameter values are saved; if not, then the proposed set of parameters is rejected, the current set is stored as another sample, and new proposed parameter values are drawn. The accept/reject procedure is central to the operation of an

The sample of the posterior PDF generated by an

It should be noted that the implementation of

As in PV10 and PB12, the prior probability density function for all control variables is assumed to be uniform with minimum and maximum bounds set to physically realistic values of each of the PSD parameters (Table 1). The mass- and area-dimensional relationships are estimated by summarizing values reported in

3. Idealized retrieval experiments

Prior to conducting

a. Response functions

While it is ultimately the multivariate relationships that are of interest in the retrieval, it is useful to examine the univariate retrieval sensitivity. Such an analysis yields information as to the functional dependence of forward-model output on changes to control parameters, and can provide an early indication of nonlinearity and/or nonuniqueness in the retrieval. In the following analysis, radar and microwave radiometer forward models are run for a range of values of each input parameter (Table 1) one at a time, fixing the remaining parameter values at their default values. Forward-model sensitivities are presented in terms of the modal diameter and particle number, as these are the fundamental retrieved variables (Deng and Mace 2006). In the retrieval framework, the maximum likelihood estimate and associated PDF are converted to liquid and ice water content and total number via equations in appendix A.

Examination of the results (Fig. 3) reveals the following. W-band radar reflectivity (Fig. 3a) is insensitive to changes in liquid particle size up to modal diameter of approximately 10mm at which point liquid drops become large enough to produce changes to W-band reflectivity. However, the changes are less than a few decibels for a factor-of-3 change in model drop size, suggesting that realistic changes to drop size would be difficult to discern in reality. Doppler velocity (Fig. 3b) decreases as the liquid size increases. This counterintuitive result arises because the retrieval accounts for the effects of the snow as well as the liquid cloud mode. Recall that the Doppler velocity is a Z-weighted quantity, that is, Vd 5(ZlVd,l 1ZsVd,s)/(Zl 1Zs), where the subscripts s and l refer to the PSDmodes described above. With the snow (large mode l) properties held fixed and the modal diameter of the liquid (small s) mode increasing and contributing more and more significantly to the reflectivity, the bimodal Vd actually decreases because the small terminal velocities of the liquid mode figure more and more prominently in the weighted sum. Note that we have restricted particle sizes in the computation to modal values less than 20mmfor the sake of consistencywith the mixedphase orographic clouds observed during StormVEx. In our

Forward-modeled low-frequency microwave radiometer Tbs (Figs. 3c,d) display the well-known sensitivity to liquid water mass, while reflectivity changes little with liquid particle number (Fig. 3e), primarily because the liquid particles are assumed to be small (default modal diameter set equal to 5mm). As expected, liquid particle fall velocity, as viewed through the lens of mean Doppler velocity, is insensitive to changes in particle number (Fig. 3f). Liquid water content is linearly dependent on particle number (see appendix A for details), and this dependence is reflected in the microwave radiometer Tb (Figs. 3g,h). Reflectivity and Doppler velocity (Figs. 3i,j) are highly sensitive to changes in snow modal diameter, but, in contrast to the liquid particles, Doppler velocity sensitivity begins to saturate as particle size increases because of hydrodynamic drag. In comparison with the liquid diameter-Tb response function plots (Figs. 3c,d), low-frequency microwave radiometer Tbs are less sensitive to changes in ice diameter (Figs. 3k,l) for equivalent changes in ice particle size, although this sensitivity increases dramatically at large sizes because of scattering effects at these weakly absorbing wavelengths. Radar reflectivity is significantly sensitive to changes in the ice number (Fig. 3m), because the modal ice diameter is set by default to a relatively large 200 mm (Table 1). As with liquid particle number, Doppler velocity demonstrates a very weak dependency on the ice particle number density (Fig. 3n). Microwave radiometer Tbs are linearly related to changes in ice modal number, but the sensitivity is small because of the far smaller absorption of microwave energy at gigahertz frequencies in ice.

Considered as a whole, response functions indicate W-band radar reflectivity and Doppler velocity observations should contain enough information to constrain the ice particle size and ice particle number. Resolving the liquid drop size in the presence of snowfall would be possible only under certain conditions, such as low turbulence and large drop sizes, as recently shown by Luke and Kollias (2013). Addition of low-frequency microwave radiometer Tb observations provides constraint on the liquid number density and size, while penalizing large values of ice diameter. It cannot be overstated that the results presented above are only applicable to the specific combination of observations and forward models considered here (W-band radar and low-frequency passive microwave radiometer). The conclusions would almost surely differ for other radar wavelengths (and for multifrequency radar) and high-frequency microwave radiometer observations.

b. Idealized retrieval

We now examine the observational constraints on liquid and ice particle size distribution parameters using idealized PDF-based retrievals. We use the radar and microwave radiometer forward models to generate simulated observations of 23- and 31-GHz microwave radiometer Tb and radar reflectivity and Doppler velocity using specified (vertically homogeneous) profiles of liquid and ice content and number, as well as profiles of temperature and water vapor content consistent with conditions observed during StormVEx (Table 1). Gamma distribution width parameters and mass-dimensional relationships are fixed.We then retrieve themodal diameter and number assuming a homogeneous profile. As is the case in previous work, the width parameter and mass- and area-dimensional relationships are fixed. As in the response function analysis, PDFs of retrieved parameters are generated via a ''brute force'' or exhaustive perturbation method in which the forward models are run successively in increasing increments of each retrieved parameter over a specified range of values (Table 1). For each set of cloud parameter values the forward-model output is compared with the simulated observations via a Gaussian cost function

...

where S21 y is the observation error covariance matrix and F(x) is the vector of forward-modeled observations. As is commonly the case, we assume that the observation errors are uncorrelated, and as such, S21y is a diagonal matrix with the observation error variances (Table 2) on the diagonal. If we assume a uniform prior distribution P(x), then the characteristics of the forwardmodeled solution can be examined by plotting the unscaled likelihood function (the exponential of the cost function). If there is a unique maximum likelihood point, then the retrieval can be said to be well constrained by the observations.

The results reflect the sensitivity of the forward-model output to changes in the PSD parameters shown in the response function plots (Fig. 3), and by extension the information content of the simulated observations. Observations of W-band reflectivity alone are insufficient to constrain either liquid or ice particle size distribution, though large liquid and ice diameter are precluded (Figs. 4a,d). The addition of Doppler velocity renders a unique solution for the ice PSD parameters (Fig. 4e), and results in elimination of liquid modal diameters larger than about 12mm (Fig. 4b). Addition of low-frequency microwave radiometer Tb observations places additional constraint on the liquid PSD (Fig. 4c), but it is notable that the solution is still nonunique. There exist a near infinite number of combinations of modal diameter and number that will produce the identical set of forward observations. Examination of the response functions provides the explanation. Tb observations prevent large values of liquid diameter, but within a reasonable range of diameter and particle number, increases in diameter can be compensated by decreases in number. It is worth pointing out that this retrieval is perhaps the most idealized one can expect for the given set of observations, as the cloud fields were set to be vertically homogeneous, the true values were known, and the gamma width parameter and mass-dimensional parameters are specified. In the next section we examine the results of a retrieval in which real-world observations are used to constrain vertically varying profiles of liquid and ice content and number, and inwhich gamma distribution width parameter and mass-dimensional relationships are allowed to realistically vary.

4. Retrievals of cloud PSD from StormVEx observations

We now use W-band radar and low-frequency microwave radiometer observations obtained during the StormVEx field campaign to retrieve vertical profiles of liquid and ice using

a. Fixed gamma distribution width and mass-dimensional relationships

We first examine the constraint on the retrieval imposed by adding successively greater amounts of information to the system. In each of the experiments that follow, we begin with a reflectivity-based retrieval, then add information first from Doppler velocity then microwave radiometer Tb. Retrieval information content can be estimated via examination of the Bayesian posterior PDF produced by

Examination of the posterior PDF of the liquid number and content (Fig. 5) indicates, as expected, W-band reflectivity alone is not capable of producing a unique solution for the cloud of interest (Figs. 5a,d,g). While the liquid water content is restricted to relatively low values, the maximum likelihood region in the PDF is spread across a large range of values of number density. The addition of observations of Doppler velocity results in a strong constraint on the liquid PSD in the top layer (Fig. 5b) in accordance with our earlier discussion of Fig. 3b. Whern the ice and liquid modes both contribute significantly to the radar measurables, the reflectivityweighted velocity decreases as the liquid phase increases in mean size. Whereas the combined reflectivity changes in the same sense with increases or decreases in the ice or liquid PSD modes, the negative tendency in Doppler velocity for increases in the liquid mode provides a unique constraint on the microphysical properties. The solution is clearly bimodal in the lower and midlayers (Figs. 5e,h), especially for the liquid-phase clouds. A solution that consists of high number and low liquid water content is nearly as likely as lower droplet number but marginally higher water content. Note that the ice content in the top layer is low (Figs. 6a-c), and as such the amounts of liquid and ice are comparable.

Addition of low-frequency microwave radiometer Tb observations produces a preferred mode in the solution, but the secondary mode is still evident, particularly at low and midlayers (Figs. 5f,i). In retrieval algorithms that seek a solution via iterative convergence to a local maximum in the PDF (e.g., optimal estimation), it is possible that the algorithmmight select the incorrect (less likely) solution. In all cases it is clear that the posterior PDF departs significantly from a Gaussian shape. In select cases (e.g., observation of reflectivity alone, or the PDFs in the top layer (Figs. 5b,c), the PDF is unimodal and skewed, but in all other cases there are multiple possible solutions (multiple modes in the PDF). In each case, simply the physical requirement that PDFs be truncated at zero liquid content and number can present problems for methods that assume Gaussian PDFs.

In contrast to the liquid particle size distribution, use of reflectivity observations almost immediately produces a more well constrained solution for the ice water content and number density in all layers (Figs. 6a,d,g), though a bimodal PDF is still clearly evident. Addition of Doppler velocity observations serves to produce a unique solution in the upper layer and leads to a preferred mode in the PDF in the lower and midlayers (Figs. 6e,h). At first glance, addition of microwave radiometer Tb observations appears to introduce greater uncertainty into the retrieval of ice content and number in the upper and midlayers (a greater number of possible solution states is included in the posterior PDF). In the absence of Tb observations, the large liquid water content solution is coupled with the low ice content solution so that the forward-modeled reflectivity is consistent with the observations. Upon the addition of microwave radiometer observations, the relatively high LWC and low IWC solution emerges as preferred. As with the liquid PDFs, none of the posterior distributions for ice content and number can be said to be Gaussian in form, and most cannot even be claimed to be approximated by a Gaussian distribution, much less a distribution with a single mode.

In all of the results presented in this section, the gamma distribution width parameter and snow mass- dimensional relationships were assumed known. In reality, these may vary over a potentially large range of values introducing additional uncertainty into the retrieval. To be sure, in most real-world applications, the most appropriate values of these empirical parameters are usually not known, much less their variability. The effect of this additional uncertainty is explored in the next section.

b. Variable gamma width and mass-dimensional relationships

Prior to conductingMCMCexperiments in which the width and mass-dimensional parameters are allowed to vary, it is first useful to explore the response of the forward-modeled radar and microwave radiometer observations to changes in these variables. The results of similar response function experiments to those described in section 3 above are presented in Fig. 7. From Eqs. (A1) and (A3) it can be seen that the width parameter affects both the water content and total number via its presence in the gamma function, while the coefficient and exponent in the mass-dimensional relationship affect only the water content. Because the terminal velocity of particles depends on the ratio of particle mass to area, the mass-dimensional relationship parameters influence the forward-modeled Doppler velocity [see Eqs. (A5) and (B4)]. It is also clear that changes in the width parameter for liquid have negligible effect on W-band radar reflectivity and Doppler velocity (Figs. 7a,b) and a very small effect on 23- and 31-GHz microwave radiometer Tbs (Figs. 7c,d). In contrast, variability in the width parameter for ice has a relatively large effect on the radar forward observations (Figs. 7e,f), and on themicrowave radiometer Tb (Figs. 7g,h), especially at large gamma distribution widths (larger numbers of large particles).

Variability in the mass-dimensional relationship is only relevant to the ice phase, as liquid particles in this particular case are small and can be assumed to be spherical. Note that in the current version of the radar forward algorithm, reflectivity is rendered insensitive to changes in the mass-dimensional relationships by design. We are well aware that this is a limitation of the current forward model, and an improved version of the model, in which the radar backscatter cross sections are functionally related to the mass-dimensional parameters, is currently in development (

We interpret the sensitivity in the microwave Tbs to the assumed ice parameters to be primarily a scattering phenomenon. We note that the effects are larger at 31GHz than at 23GHz even though the complex index of refraction taken from Sadiku (1985) is smaller at 31GHz (1.0 3 1023) than at 23GHz (1.4 3 1023). The scattering cross section varies approximately as the inverse wavelength raised to the 24 power and directly as the particle size raised to the exponent 6 in the Rayleigh regime. Therefore, the stronger response at 31GHz may be ascribed to the inverse wavelength dependence. As the shape parameter increases (Figs. 7g,h) the ice size distribution broadens and ice mass is increasingly moved to large particle sizes where the D6 relationship results in increased scattering cross sections. For the mass- dimensional relationship parameters am and bm, we find a positive change with increasing am and a negative change with increasing bm. This can be understood by considering that mass changes as Dbm for changes in am while the mass changes in a manner inversely proportional to D for changes in bm. Larger am results in increased mass at larger sizes, while changes in bm result in less mass at the larger sizes. In any case, these microwave Tb sensitivities should serve as a caution against simple interpretations of microwave radiometer measurements in mixed-phase clouds since ice scattering can have a considerable influence on the observations.

Comparison of the retrieved joint PDFs of liquid water content and number density (Fig. 5) with those retrieved in the presence of fixed a, am, and bm (Fig. 8) reveals the following. In all but a few cases, the fundamental shape of the probability density function does not change significantly when additional PSD parameters are allowed to vary. This reflects the fact that the underlying functional relationship between changes to parameters and forward-modeled observations is robust to changes in the shape of the PSD and particles themselves. However, variability in the liquid and ice gamma width parameters (Fig. 8, left) causes a fundamental change in the center of mass and variance of the solution PDF for the liquid cloud retrieval. The optimal number concentration decreases near the top of the cloud (Fig. 8a), increasing in lower and midlayers (Figs. 8d,g). The mode in the PDF of liquid water shifts toward larger values in the upper layer; while the set of solutions at low number and relatively high liquid water content in the lower layer disappears. The reason for the elimination of this mode (and a corresponding mode in the ice PSD PDF) is explored below, following a discussion of the ice PSD PDFs. In contrast to variability in the width parameters, changes in the ice mass-dimensional parameters (Figs. 8b,e,h) lead to an increase in the liquid water content solution variability, but little change in the optimal solution. As was mentioned above, the exception in both cases is the removal of the mode in the PDF at small number concentration and relatively large liquid water content. When both width and mass parameters are allowed to vary (Figs. 8c,f,i), the result is a solution space that exhibits greater variability than when parameters are fixed, but for which there remains an optimal (unique) solution.

As with the liquid cloud properties, the shape of the retrieved ice PDF (Fig. 9) is consistent across all experiments; IWC and ice number are positively correlated and the magnitude of the correlation is similar. However, in contrast to the liquid experiments, the location of the mode in the PDF does not change as the gamma distribution width and ice mass parameters are allowed to vary. Variability in a (Figs. 9a,d,g), as in the liquid cloud case, changes the PDF structure, leading to removal of the mode in the PDF associated with very low ice water content and large ice number concentration. When the mass parameters are allowed to vary (Figs. 9b,e,h), the structure of the solution space changes little, but the width of the posterior PDF increases significantly, especially in the lowest layer. Solutions are now relatively likely (e.g., large ice water content and number density) that were prohibited when am and bm were fixed. When liquid and ice gamma distribution width and ice mass parameters are allowed to vary (Figs. 9c,f,i), the location of the primary solution remains the same, but the variability increases further. A wide range of ice water content is now feasible in the lower and midlayers and there has been an expansion in the range of possible solutions near the top of the cloud as well.

The absence of the mode in the midlayer liquid and ice PSD PDFs can be explained via examination of the PDF of the modal diameter (D0) and number (N0) of liquid and ice (Fig. 10). Recall that these are the parameters that are, in actuality, varied in the retrieval process. The mode in the PDF ofLWCandN0liq (Fig. 5f) at largeLWC and small N0liq is associated with the mode in Fig. 10a at relatively large modal diameter (approximately 30mm) and small number (by inspectionN0.1200 cm24).When the gamma distribution width is specified to be zero, it is possible to fit the available radar and microwave radiometer Tb observations with a small number of relatively large cloud droplets. The fit is only reasonable in a very small range; however, because of the sensitivity of the W-band radar reflectivity to small changes in particle diameter once the particles are large. Inspection of the joint PDF of modal liquid diameter and ice number (Fig. 10b) reveals there is a large range of ice numbers compatible with the large liquid modal diameter values, but these are necessarily associated with a small ice diameter (Fig. 10c). This explains the secondary mode in the IWC/N0ice PDF (Fig. 6f) centered at small IWC values. The liquid particles are providing the major portion of the radar signal and necessitate the retrieval of small ice particles. Since the particles are small, a relatively large range of sizes will produce a reasonable fit to the observations. The mode in the joint PDFs of liquid and ice exists only because the liquid PSD width is specified and set equal to zero. Once the width parameter is allowed to vary, there is no preference for a specific set of values of liquid modal diameter and number. A large number of combinations of gamma distribution width, modal number, and diameter will fit the observations and as such there will be no localized mode in the PDF.

Note that the localized mode is also eliminated when the mass-dimensional parameters are varied (Fig. 9e). Changing am and bm changes the assumed ice-particle shape. This has absolutely no effect on either the microwave radiometer Tb or radar reflectivity because of our experimental design, but significantly affects the Doppler velocity (Fig. 7). What this reveals is that it is not only the PSD but also the particle shape that is the cause of the localized mode in LWC/Nliq. If ice crystal shape is allowed to vary, the fall velocities respond accordingly. With fixed gamma distribution width and crystal shape and large liquid droplets, Doppler velocity can only be satisfied over a small range of liquid particle modal diameters. Once the crystal shape is allowed to vary, an additional degree of freedom is introduced into the system and the localized mode in the probability distribution disappears.

The fact that a localized mode (and hence a nonunique solution) is produced when the system is placed under tighter constraints is a result that might not have been expected a priori, and it is one of the reasons a detailed examination of the retrieval probability structure is useful.

5. Summary and conclusions

This paper provides an analysis of the probability structure associated with a combined surface-based radar and passive microwave radiometer retrieval of a bimodal cloud particle size distribution for a case of snowfall from a mixed-phase orographic cloud. While optimal estimation type retrievals have been used to retrieve the PSD parameters of interest for combined active-passive observing systems, the true nature of the solution space has remained unknown. The nature and degree of constraint of snowfall retrievals by radar and/ or microwave radiometer observations has also been unclear. Characterization of the full solution space allows an evaluation of the robustness (or lack thereof) of optimal estimation retrievals, and evaluation of the sources of uncertainty and their effect on the resulting retrieval. Markov chain

This paper focused specifically on an

The major conclusions of this study are the following:

1) W-band radar reflectivity alone is insufficient to constrain the ice PSD; at minimum Doppler velocity is needed if a unique solution is to be obtained.

2) If Doppler velocity is included as an observable, radar-only ice retrievals may return a robust solution, provided the mass-dimensional relationships are well constrained. Retrieval of liquid cloud properties requires observations of microwave radiometer Tb, but in the case of low-frequency microwave radiometer observations there remains a nonunique relationship between number and liquid water content.

3) When the constraint on the assumed gamma distribution width parameter is relaxed, the retrieval may still be able to produce a unique solution, but the solution mode will differ from that obtained with a specified value.

4) When assumed mass-dimensional parameters are allowed to vary, the retrieved liquid and ice water path and number become far less certain, and retrieval of a single unique value is rendered more difficult.

5) In select cases, specification of the mass-dimensional parameters and the PSD width produces a nonunique solution that does not exist when these parameters are allowed to vary. Specifically, when the gamma distribution width is set equal to zero and the mass- dimensional parameters are fixed, a combination of a small number of large liquid cloud droplets with generally small ice particles provide a fit to the combined radar and microwave radiometer observations over a very narrow range. The large sensitivity of the observations to small changes in the liquid particle size distribution in this region of the solution space leads to the production of a highly localized solution mode that disappears when gamma distribution width and ice crystal shape are allowed to vary.

This study illustrates the importance of a rigorous exploration of the retrieval solution space, and indicates that care is needed when interpreting the results of radar-only and combined radar-passive cloud property retrievals. This is especially true when the empirical parameters describing the ice crystal habit are specified, if a priori information is not available to actually constrain those parameters. The results also illustrate the important role of prior knowledge in optimal estimationtype retrievals. In those cases for whichMCMC returned a multimode (nonunique) solution, prior knowledge could serve to restrict the solution space to a unique set of PSD parameters. However, global specification of prior information is a difficult exercise at best, and at worstmay lead to the retrieval of the wrong set of PSD parameters when the solution space is by nature nonunique.

In closing, it is appropriate to point out a number of limitations to the current study. First, the temperature profile was obtained from a nearby sounding and was not allowed to vary as part of the retrieval. Undoubtedly, variability in temperature would have had an effect on the microwave radiometer Tbs as well as a (small) effect on the retrieved Doppler velocity, which exhibits a weak dependence on the air density. Averaging of the observations into three broad layers had an unknown effect on the retrieval error characteristics.While it may be reasonable to expect that averaging (as a smoothing operator) leads to smaller noise in the solution, this does not necessarily mean that the uncertainty is reduced.Additional retrieval error is also present because radar attenuation was not represented in the radar forward model, and changes in the ice particle shape were not functionally related to changes in the radar reflectivity. While many in situ analyses of hydrometeor populations indicate a gamma function provides a close match to observed particle size distributions, the assumption that PSDs adhere to any particular functional form introduces additional (and unknown) uncertainties into the retrieval. It should also be emphasized that the results reported here are preliminary, as the analysis has only been applied to a single vertical profile on a single case day. In the near future, we plan to explore the characteristics of the solution space for a time series of profiles, as well as for other case days observed during the StormVEx field campaign.

Last, and perhaps most significant, it should be emphasized that the results and conclusions presented in this study are only applicable to the particular observations, forward models, and physical environment we have chosen. Much of the nonuniqueness observed in the posterior PDFs may not exist for other cloud system types; and in particular for bimodal liquid PSDs. In addition, it is expected that high-frequency microwave radiometer observations (far more sensitive to scattering from ice crystals) may provide enough unique information to constrain aspects of the snow PSD. The use of multifrequency radar observations and use of observations of Doppler spectral width would also likely render a unique solution for many of the parameters studied here. Of course, the converse is also true; in other types of ice cloud, W-band reflectivity and Doppler velocity may not be sufficient to uniquely constrain the retrieval of ice content and number.While we view themethodology and results to be robust for our chosen case, care must be taken not to assume these results apply globally to all cloud systems. We are currently using the framework described above to explore retrieval sensitivity to forwardmodel assumptions, potential increase in information provided by other sensors, and the error characteristics of retrieved PSDs for other cloud system types.

Acknowledgments. This research was supported primarily by the

REFERENCES

Bohren, C. F., and

Borys, R. D., and

DelanoË, J., and

Deng, M., and G. Mace, 2006: Cirrus microphysical properties and air motion statistics using cloud radar Doppler moments. Part I: Algorithm description.

Donovan, D. P., 2003: Ice-cloud effective particle size parameterization based on combined lidar, radar reflectivity, and mean Doppler velocity measurements. J. Geophys. Res., 108, 4573, doi:10.1029/2003JD003469.

Dorsi, S. W., 2013: Airborne and surface-level in situ observations of wintertime clouds in the

Ebell, K., U. LÖhnert, S. Crewell, and

Frisch, A. S., G. Feingold,

Gossard, E. E., 1994: Measurement of cloud droplet size spectra by Doppler radar. J.

Haario, H., E. Saksman, and

Hale, G. M., and

Hallar, A. G., G. Chirokova,

_____,

Hansen, J. E., and

Hastings, W. K., 1970:

Hewison, T. J., and

Heymsfield, A. J.,

Kummerow, C.,

Lebsock, M.D., T. S. L'Ecuyer, andG.

Locatelli, J. D., and

LÖhnert, U., S. Crewell,

_____, G. Feingold, T. Uttal,

Luke, E. P., and

Mace, G. G.,

Mace, J., and Coauthors, 2010:STORMVEX: The Storm Peak Lab Cloud Property Validation Experiment science and operations plan.

Mather, J. H., and

Matrosov, S. Y.,

_____,

McFarlane, S. A.,

Metropolis, N.,

_____, and

Posselt, D. J., 2013: Markov chain

_____, and T. Vukicevic, 2010: Robust characterization of model physics uncertainty for simulations of deep moist convection. Mon. Wea. Rev., 138, 1513-1535, doi:10.1175/ 2009MWR3094.1.

_____, and

_____, T. S. L'Ecuyer, and

_____,D.Hodyss, and

Pruppacher, H. R., and J. D. Klett, 1978: Microphysics of Clouds and Precipitation. Reidel, 714 pp.

Rambukkange, M. P.,

Rodgers, C. D., 2000: Inverse Methods for Atmospheric Sounding, Theory and Practice. Series on Atmospheric, Oceanic and Planetary Physics, Vol. 2, World Scientific, 240 pp.

Sadiku, M. N. O., 1985: Refractive index of snow at microwave frequencies. Appl. Opt., 24, 572-575, doi:10.1364/AO.24.000572.

Schmitt, C. G., and

Szyrmer, W., and I. Zawadzki, 2010: Snow studies. Part II: Average relationship between mass of snowflakes and their terminal velocity. J.

Tamminen, J., 2004: Validation of nonlinear inverse algorithms with Markov chain

_____,and E. KyrÖlÄ, 2001: Bayesian solution for nonlinear and non- Gaussian inverse problems by Markov chain

Tarantola, A., 2005: Inverse Problem Theory and Methods for Model Parameter Estimation. SIAM, 342 pp.

Wood, N. B., 2011: Estimation of snow microphysical properties with application to millimeter-wavelength radar retrievals for snowfall rate. Ph.D. dissertation,

Zhang, Y., and

Zhao, Y.,

(Manuscript received

Corresponding author address:

E-mail: dposselt@umich.edu

(