News Column

MCMC-Based Assessment of the Error Characteristics of a Surface-Based Combined Radar-Passive Microwave Cloud Property Retrieval

August 1, 2014

Mace, Gerald G


Collocated active and passive remote sensing measurements collected at U.S. Department of Energy Atmospheric Radiation Measurement Program sites enable simultaneous retrieval of cloud and precipitation properties and air motion. Previous studies indicate the parameters of a bimodal cloud particle size distribution can be effectively constrained using a combination of passive microwave radiometer and radar observations; however, aspects of the particle size distribution and particle shape are typically assumed to be known. In addition, many retrievals assume the observation and retrieval error statistics have Gaussian distributions and use least squares minimization techniques to find a solution. In truth, the retrieval error characteristics are largely unknown. Markov chain Monte Carlo (MCMC) methods can be used to produce a robust estimate of the probability distribution of a retrieved quantity that is nonlinearly related to the measurements and that has non-Gaussian error statistics. In this work, an MCMC algorithm is used to explore the error characteristics of cloud property retrievals from surface-based W-band radar and low-frequency microwave radiometer observations for a case of orographic snowfall. In this particular case, it is found that a combination of passive microwave radiometer measurements with radar reflectivity and Doppler velocity is sufficient to constrain the liquid and ice particle size distributions, but only if the width parameter of the assumed gamma particle size distribution and mass-dimensional relationships are specified. If the width parameter and mass-dimensional relationships are allowed to vary realistically, a unique retrieval of the liquid and ice particle size distribution for this orographic snowfall case is rendered far more problematic.

(ProQuest: ... denotes formulae omitted.)

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 Monte Carlo (MCMC) algorithms (Metropolis et al. 1953; Hastings 1970; Posselt and Vukicevic 2010, hereinafter PV10; Posselt and Bishop 2012, hereinafter PB12; Posselt 2013) have recently been used to examine the information content of remote sensing retrievals (Tamminen and KyrÖlÄ 2001; Posselt et al. 2008) and in rare cases to perform the retrieval itself (Tamminen 2004). MCMC methods produce a Bayesian estimate of the probability distribution of a set of parameters given a set of observations, prior knowledge, and a model that relates parameters and observations. There is no requirement that the probability distributions be Gaussian, and additional constraints (e.g., positive definite PSD parameters) can be easily and flexibly accommodated. As such, MCMC algorithms can be used to obtain an accurate assessment of observation information content, to assess the effect of uncertainty in the forward model, and determine which observations might be used to constrain the retrieval should the solution prove to be nonunique. The cost of this flexibility is the computational burden required to implement MCMC algorithms, as will become clear below.

In this paper, a Markov chain Monte Carlo algorithm is used to retrieve cloud and precipitation properties in a case of orographic snowfall for the purpose of illustrating the sensitivity of such retrievals to assumptions and prior information. The results 1) indicate whether there is sufficient information in a combined microwave radiometer and radar retrieval to constrain a bimodal (liquid and ice) cloud particle size distribution, and 2) determine, via computation of the probability distribution of retrieved liquid and ice, the error characteristics of the retrieval and the dominant sources of uncertainty. It should be noted at the outset that errors in any retrieval of cloud PSD properties tend to be strongly regime dependent, and retrieval outcomes in general are sensitive to the specific forward model(s) used and measurement errors assumed. Our experiments highlight the uncertainty in mixed-phase PSD retrievals from surface-based measurements: W-band cloud radar and low-frequency microwave radiometer observations. The results are specific to our combination of forward models and were obtained from a single orographic snowfall case. As such, though the MCMC methodology is robust, and our forward models well tested, our results should be viewed as applicable to a relatively narrow range of conditions.We will return to this point in the conclusions, and offer a preview of how we are currently increasing the generality of our experiments.

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 B. Section 3 contains an analysis of how forward-modeled measurements change in response to changes in cloud PSD-the cloud property response functions. Results fromidealized and real-data retrievals using MCMC are presented in section 4, and our major conclusions are summarized in section 5.

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 Park Range Mountains near Steamboat Springs, Colorado, as the principal component of the Storm Peak Laboratory Cloud Property Validation Experiment (StormVEx; Mace et al. 2010). The purpose of StormVEx was to document the properties of orographic mixed-phase snow clouds in conjunction with the collection of correlative data at Storm Peak Laboratory (SPL) located at the summit of Mount Werner (Borys and Wetzel 1997; Hallar et al. 2011a,b). The remote sensors used for this study were located near the Thunderhead Lodge at an elevation of 2759m approximately 2.4km to the west and 400m below SPL and collected data continuously from late November 2010 until early April 2011. The Scanning W-Band Cloud Radar (SWACR) and a 2-channel (23 and 31 GHz) microwave radiometer, as well as ancillary data from a laser ceilometer, are the primary instruments used in this study.

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 Colorado Rocky Mountains with cloud cover exceeding 60%. Measureable snow was recorded on more than 75 days of the campaign. Major snowfall exceeding 10 cmday21 was recorded on approximately 40 of those days. Total measureable snowfall exceeded 800 cm over the entire winter season.

The case we examine in this paper was recorded on 4 February 2011 when a quiescent northerly flow existed over the Park Range. This case was selected because both liquid and ice particles were known to exist within the cloud, and because snowfall was light and nearly continuous over a period of several hours. Snow began to fall during the early local afternoon from a mixed-phase cloud layer with bases (determined from ceilometer data) that varied from near to just below the SPL level where temperatures were in the 2108C range. Cloud tops from the SWACR were observed to be approximately 2 km above the radar where temperatures from an afternoon radiosonde were 2158C. Operators at SPL reported rime buildup on the instruments during the afternoon with the light snowfall that consisted primarily of rimed dendritic ice crystals. Snowfall was light enough that automated snowfall instruments recorded no measureable accumulation. Figure 1 shows the radar Doppler moments that were collected during the 1-h period centered on the profile that we evaluate in this study. It is notable that a bimodal (and in select layers, trimodal) spectrum is evident in the observations of Doppler spectral width. It is not uncommon to observe a population of precipitating ice hydrometeors collocated with smaller ice particles falling more slowly, and it is likely that this was the case during a portion of the observed time period. On this particular afternoon, the SWACR did not conduct routine RHI and PPI scans, and only zenith-pointing data were collected.

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)5am,iDbm,i in cgs units, and G is the gamma function. The derivations of additional microphysical parameters are presented in appendix A.

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 Monte Carlo algorithm

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 Monte Carlo methods provide a solution to (3) that does not require assumptions as to the specific form of the probability distributions (e.g., Gaussian, as is the case in optimal estimation) and utilizes the full nonlinear forward model. MCMC algorithms sample the posterior probability distribution via a random walk that consists of multiple sequential runs of the forward model. In each iteration, candidate values of all retrieved parameters are randomly drawn from a proposal distribution centered on the current parameter estimate. The forward model is then used to produce simulated observations from the candidate parameter values and the resulting forward observations and prior are compared with the observations via a cost function. The form of this cost function depends on what the user has chosen as a likelihood function P(y j x). The proposed set of parameter values x (and the corresponding set of forward-modeled observations y) are accepted as a sample of the posterior probability distribution with probability:

... (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 MCMC algorithm uses an uncorrelated Gaussian proposal distribution centered on the current set of parameter values, with variance that is adaptively tuned during an initial set of forward-model calculations (Haario et al. 1999; Tamminen and KyrÖlÄ 2001; Posselt 2013). Parameter sets generated during the tuning process are not included in the posterior solution.

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 MCMC algorithm, and ensures that randomly chosen parameter sets with higher probability (better fit to the observations) are automatically accepted, while parameter values that provide a similar (but perhaps poorer) fit to observations are considered. Parameter sets that produce forward observations far from the measurements are rejected. In the process, the algorithm preferentially samples high-probability regions of the posterior parameter space, avoids very low-probability regions, and appropriately samples the parameter space in between. The probabilistic accept/ reject procedure allows the algorithm to move away from local probability maxima and is the primary reason MCMC is not simply an optimization algorithm. Considered as a whole, the sequence of randomly generated parameter vectors accepted as samples of the posterior distribution is the Markov chain. At the user's discretion, an MCMC algorithm can be constructed so that it uses a single Markov chain, or multiple chains simultaneously exploring the probability distribution associated with the same set of observations (Posselt 2013).

The sample of the posterior PDF generated by an MCMC algorithm represents the most complete characterization of the retrieval solution for a given forward model, prior, and set of measurements. Sources of uncertainty can be easily identified, observation information content can be quantified, and the impact of changes to observation uncertainty as well as the introduction of new observations can be tested in a straightforward manner. The implementation of the MCMC algorithm in this paper is based on the algorithm described in Posselt et al. (2008), and PB12, and illustrated in the flowchart presented in Fig. 2. In each of the aforementioned papers, the set of parameter values was fixed in time and space.

It should be noted that the implementation of MCMC used in this paper has been modified to account for the fact that, as opposed to a set of scalar quantities, vertical profiles of cloud properties are estimated. There are a number of ways to estimate vertical profile variables in an MCMC algorithm. One can assume parameters are spatially uncorrelated, and randomly generate proposed parameter profiles by perturbing a single layer at a time or perturbing all layers simultaneously and randomly. In practice, the former option leads to very slow convergence, while the latter introduces a significant amount of noise in the profiles. An alternative option is to smooth perturbations in the vertical, either by perturbing a single layer at a time and spreading the influence using a correlation function with specified vertical decorrelation length, or by perturbing all layers simultaneously, then adjusting the proposed profile using a specified smoothing function. Tests with each option indicate the former technique is most effective in producing a robust estimate of the posterior PDF in the most computationally efficient manner. The vertical decorrelation length scale was estimated from the vertical decorrelation of the radar reflectivity profiles, and was determined to be approximately 400m. The results were relatively insensitive to the choice of decorrelation length, though of course large decorrelation length essentially causes the entire profile to be perturbed simultaneously, unrealistically constraining the vertical variability in the profile. As mentioned above, very small lengths lead to effectively independent perturbation of each layer. This is not a problem per se, but it leads to much slower convergence in the MCMC algorithm.

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 Mitchell (1996), Szyrmer and Zawadzki (2010), and Heymsfield et al. (2010) and are appropriate for the aggregates and dendritic particles observed on this case day. The PSD parameters were determined using data collected in situ during StormVEx from the Colorado Airborne Mixed Phase Study (CAMPS; Dorsi 2013) that was funded by theNational Science Foundation. Weperformtwo sets of experiments: one in which the true values of each PSD parameter are predetermined and used to generate simulated radar and microwave radiometer observations, and another in which MCMC is used to retrieve PSD properties from observations taken from the 4 February 2011 StormVEx case. The idealized experiment is used to examine the functional relationships between changes to PSD parameters and forward-model output (the forward-model response function). Following idealized tests, MCMC experiments are used to retrieve the vertical profile of liquid and ice PSD parameters for a single representative cloud profile. The information content of the observations is assessed and the effect of uncertainty in the cloud PSD is examined.

3. Idealized retrieval experiments

Prior to conducting MCMC experiments, we explore the performance of the combined radar and microwave radiometer retrieval for an idealized case. We first analyze the response of the forward-modeled observations to changes in particle size distribution parameters with the intent of 1) exploring the inherent sensitivity in the retrieval and 2) characterizing the (univariate) parameter-observation functional relationship. We then specify a (vertically homogeneous) vertical profile of liquid and ice content based on the specified default parameters listed in Table 1, and generate simulated observations by using these ''true'' profiles in the radar and microwave radiometer forward models. The resulting simulated observations are then used to retrieve the cloud particle size distribution parameters, using observation uncertainty consistent with the real-world StormVEx observations. The results provide information as to whether the observations provide sufficient constraint on the retrieval.

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 MCMC experiments, we allow liquid modal diameter a much larger range to accommodate the diverse conditions encountered in real clouds.

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 MCMC. As our primary goal is an exploration of the sensitivity of the retrieval, we do not need to retain the full vertical resolution in the observations or estimated parameters. Estimation of cloud PSD properties for every radar range bin would also add significantly to the computational expense of implementingMCMC. For this reason, we significantly reduce the vertical resolution of the SWACR data from 42-m range bins to just three layers. These layers are evenly divided in the vertical, where the lowest layer (142-741m above the surface) captures primarily the subcloud layer where mostly precipitation is found, the midlayer (741- 1341m above the surface) is the mixed-phase region where snow is falling through liquid cloud, and the upper layer (1341-1983m above the surface) is also mixed phase but the ice particles are smaller and the liquid phase contributes more significantly to the radar observables. The vertical averaging is accomplished by summing the radar reflectivity Z over each layer and dividing by the number of range bins in the layer, while the Doppler velocity is computed as a Z-weighted average in the layer. A single representative profile is chosen for analysis (Table 2), and two different experiments are conducted; one in which the ice mass-dimensional parameters and liquid and ice gamma distribution width parameters are set equal to default values, and one in which they are allowed to vary. In each experiment, observations of reflectivity, Doppler velocity, and microwave radiometer Tb are used to constrain the retrieved particle size distribution. The vertical column is perturbed one layer at a time, with perturbations spread in the vertical via a Gaussian function with 400-m decorrelation length.AsingleMarkov chain is used, and the MCMC algorithm is run for increasing numbers of iterations of the forward model as the dimensionality of the problem is increased (Table 3). Convergence of the Markov chains is assessed via intercomparison of multiple randomly drawn subsamples of 105 iterations, as well as via visual inspection of the chains themselves (not shown in the interest of brevity). In each case, the integrated absolute differences between the subsampled PDFs were less than 2% and the chains were observed to be mixing readily in the solution space.

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 MCMC. By definition, the optimal fit to the available observations is the maximum likelihood point in the PDF, the point at which the posterior PDF has a maximum in value. If there are multiple maxima, this reflects the fact that multiple different sets of parameter values produce forward-modeled values that are consistent with the real observations. If there is no unique maximum likelihood point, or the probability maximum is broad in the vicinity of such a point, then by definition the retrieval is not well constrained.

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 (Hammonds et al. 2014). Forward-modeled Doppler velocity is very sensitive to changes in both the mass-dimensional parameters, while microwave radiometer Tb exhibits little sensitivity (Figs. 7k,l,o,p). Note that there is an inverse functional relationship between the Doppler velocity and the coefficient and exponent in the mass-dimensional relationship, indicating that increases in the mass coefficient can be effectively compensated by decreases in the mass exponent (and vice versa).

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 Monte Carlo methods produce a Bayesian sample of the retrieval solution space, and are more efficient than multivariate brute force perturbation and more complete than one-at-a-time response function analysis.

This paper focused specifically on an MCMC-based exploration of the solution space associated with a surface-based combined W-band radar-low-frequency microwave radiometer brightness temperature retrieval of cloud properties for a case of orographic snowfall observed during the StormVEx field campaign. Threelayer liquid and snow water content and number were retrieved from 94-GHz radar and 23- and 31-GHz microwave radiometer Tb observations, and the successive addition of observational constraints provided information as to whether a radar-only retrieval is capable of constraining the liquid and ice particle size distributions. The effect of variability in PSD parameters and in the ice crystal shape on the retrieval solution was also assessed.

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 U.S. Department of Energy'sAtmospheric System Research, anOffice of Science, Office of Biological and Environmental Research program, under Grants DESC0007059 (DP), ER65324-1038673-0017514 (GM), and ER65235-1038449-0017520 (GM). Data were obtained from the Atmospheric Radiation Measurements Program sponsored by the U. S. Department of Energy Office of Science, Office of Biological and Environmental Research, Environmental Science Division. We also thank members of StormVEx science team for their tireless work in making the program a success: Gannet Hallar, Ian McCubbin, Matt Shupe, Roger Marchand, Sergey Matrosov, and Chuck Long. Sergey Matrosov provided T-matrix calculations of the W-band backscatter cross sections. Stephanie Avey and Qiuqing Zhang at the University of Utah provided programming assistance.


Bohren, C. F., and D. R. Huffman, 1983: Absorption and Scattering of Light by Small Particles. Wiley-Interscience, 530 pp.

Borys, R. D., and M. A. Wetzel, 1997: Storm Peak Laboratory: A research, teaching and service facility for the atmospheric sciences. Bull. Amer. Meteor. Soc., 78, 2115-2123, doi:10.1175/ 1520-0477(1997)078,2115:SPLART.2.0.CO;2.

DelanoË, J., and R. J. Hogan, 2008: A variational scheme for retrieving ice cloud properties from combined radar, lidar, and infrared radiometer. J. Geophys. Res., 113, D07204, doi:10.1029/ 2007JD009000.

Deng, M., and G. Mace, 2006: Cirrus microphysical properties and air motion statistics using cloud radar Doppler moments. Part I: Algorithm description. J. Appl. Meteor. Climatol., 45, 1690- 1709, doi:10.1175/JAM2433.1.

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 southern Rockies. Ph.D. dissertation, University of Colorado, 204 pp.

Ebell, K., U. LÖhnert, S. Crewell, and D. D. Turner, 2010: On characterizing the error in a remotely sensed liquid water content profile. Atmos. Res., 98, 57-68, doi:10.1016/ j.atmosres.2010.06.002.

Frisch, A. S., G. Feingold, C. W. Fairall, T. Uttal, and J. B. Snider, 1998: On cloud radar and microwave radiometer measurements of stratus cloud liquid water profiles. J. Geophys. Res., 103, 23 195-23 197, doi:10.1029/98JD01827.

Gossard, E. E., 1994: Measurement of cloud droplet size spectra by Doppler radar. J. Atmos. Oceanic Technol., 11, 712-726, doi:10.1175/1520-0426(1994)011,0712:MOCDSS.2.0.CO;2.

Haario, H., E. Saksman, and J. Tamminen, 1999: Adaptive proposal distribution for random walk Metropolis algorithm. Comput. Stat., 14, 375-395, doi:10.1007/s001800050022.

Hale, G. M., and M. R. Querry, 1973: Optical constants of water in the 200 nm to 200 mm wavelength region. Appl. Opt., 12, 555- 563, doi:10.1364/AO.12.000555.

Hallar, A. G., G. Chirokova, I. B. McCubbin, T. H. Painter, C. Wiedinmyer, and C. Dodson, 2011a: Atmospheric bioaerosols transported via dust storms in western United States. Geophys. Res. Lett., 38, L17801, doi:10.1029/2011GL048166.

_____, D. H. Lowenthal, G. Chirokova, C. Wiedinmyer, and R. D. Borys, 2011b: Persistent daily new particle formation at a mountain-top location. Atmos. Environ., 45, 4111-4115, doi:10.1016/j.atmosenv.2011.04.044.

Hammonds, K. D., G. G. Mace, and S. Y. Matrosov, 2014: Characterizing the radar backscatter cross-sectional sensitivities of ice-phase hydrometeor size distributions via a simple scaling of the Clausius-Mossotti factor. J. Appl. Meteor. Climatol., doi:10.1175/JAMC-D-13-0280.1, in press.

Hansen, J. E., and L. D. Travis, 1974: Light scattering in planetary atmospheres. Space Sci. Rev., 16, 527-610, doi:10.1007/ BF00168069.

Hastings, W. K., 1970: Monte Carlo sampling methods using Markov chains and their applications. Biometrika, 57, 97-109, doi:10.1093/biomet/57.1.97.

Hewison, T. J., and S. J. English, 1999: Airborne retrievals of snow and ice surface emissivity at millimeter wavelengths. IEEE Trans. Geosci. Remote Sens., 37, 1871-1879, doi:10.1109/ 36.774700.

Heymsfield, A. J., C. Schmitt, A. Bansemer, and C. H. Twohy, 2010: Improved representation of ice particle masses based on observations in natural clouds. J. Atmos. Sci., 67, 3303-3318, doi:10.1175/2010JAS3507.1.

Kummerow, C., W. S. Olson, and L. Giglio, 1996: A simplified scheme for obtaining precipitation and vertical hydrometeor profiles from passive microwave sensors. IEEE Trans. Geosci. Remote Sens., 34, 1213-1232, doi:10.1109/36.536538.

Lebsock, M.D., T. S. L'Ecuyer, andG. L. Stephens, 2011:Detecting the ratio of rain and cloud water in low-latitude shallow marine clouds. J. Appl. Meteor. Climatol., 50, 419-432, doi:10.1175/ 2010JAMC2494.1.

Locatelli, J. D., and P. V. Hobbs, 1974: Fall speeds and masses of solid precipitation particles. J. Geophys. Res., 79, 2185-2197, doi:10.1029/JC079i015p02185.

LÖhnert, U., S. Crewell, C. Simmer, and A. Macke, 2001: Profiling cloud liquid water by combining active and passive microwave measurements with cloud model statistics. J. Atmos. Oceanic Technol., 18, 1354-1366, doi:10.1175/1520-0426(2001)018,1354: PCLWBC.2.0.CO;2.

_____, G. Feingold, T. Uttal, A. S. Frisch, and M. D. Shupe, 2003: Analysis of two independent methods for retrieving liquid water profiles in spring and summer Arctic boundary clouds. J. Geophys. Res., 108, 4219, doi:10.1029/2002JD002861.

Luke, E. P., and P. Kollias, 2013: Separating cloud and drizzle radar moments during precipitation onset using Doppler spectra. J. Atmos. Oceanic Technol., 30, 1656-1671, doi:10.1175/ JTECH-D-11-00195.1.

Mace, G. G., T. P. Ackerman, P. Minnis and D. F. Young, 1998: Cirrus layer microphysical properties derived from surfacebased millimeter radar and infrared interferometer data. J. Geophys. Res., 103, 23 207-23 216, doi:10.1029/98JD02117.

Mace, J., and Coauthors, 2010:STORMVEX: The Storm Peak Lab Cloud Property Validation Experiment science and operations plan. U.S. Department of Energy Tech Rep. DOE/ SC-ARM-10-021, 45 pp.

Mather, J. H., and J. W. Voyles, 2013: TheARMClimate Research Facility: A review of structure and capabilities. Bull. Amer. Meteor. Soc., 94, 377-392, doi:10.1175/BAMS-D-11-00218.1.

Matrosov, S. Y., B. W. Orr, R. A. Kropfli, and J. B. Snider, 1994: Retrieval of vertical profiles of cirrus cloud microphysical parameters from Doppler radar and infrared radiometer measurements. J. Appl. Meteor., 33, 617-626, doi:10.1175/ 1520-0450(1994)033,0617:ROVPOC.2.0.CO;2.

_____, M. D. Shupe, and I. V. Djalalova, 2008: Snowfall retrievals using millimeter-wavelength cloud radars. J. Appl. Meteor. Climatol., 47, 769-777, doi:10.1175/2007JAMC1768.1.

McFarlane, S. A., K. F. Evans, and A. S. Ackerman, 2002: A Bayesian algorithm for the retrieval of liquid water cloud properties from microwave radiometer and millimeter radar data. J. Geophys. Res., 107, 4317, doi:10.1029/ 2001JD001011.

Metropolis, N., A. W. Rosenbluth, M. N. Rosenbluth, A. H. Teller, and E. Teller, 1953: Equation of state calculations by fast computingmachines. J. Chem. Phys., 21, 1087-1092, doi:10.1063/ 1.1699114.

Mitchell, D. L., 1996: Use of mass- and area-dimensional power laws for determining precipitation particle terminal velocities. J. Atmos. Sci., 53, 1710-1723, doi:10.1175/1520-0469(1996)053,1710: UOMAAD.2.0.CO;2.

_____, and A. J. Heymsfield, 2005: Refinements in the treatment of ice particle terminal velocities, highlighting aggregates. J. Atmos. Sci., 62, 1637-1644, doi:10.1175/JAS3413.1.

Posselt, D. J., 2013: Markov chain Monte Carlo methods: Theory and applications. Data Assimilation for Atmospheric, Oceanic and Hydrologic Applications, 2nd ed. S. K. Park and L. Xu, Eds., Springer 59-87.

_____, 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 C. H. Bishop, 2012: Nonlinear parameter estimation: Comparison of an ensemble Kalman smoother with a Markov chain Monte Carlo algorithm. Mon. Wea. Rev., 140, 1957- 1974, doi:10.1175/MWR-D-11-00242.1.

_____, T. S. L'Ecuyer, and G. L. Stephens, 2008: Exploring the error characteristics of thin ice cloud property retrievals using a Markov chain Monte Carlo algorithm. J. Geophys. Res., 113, D24206, doi:10.1029/2008JD010832.

_____,D.Hodyss, and C.H. Bishop, 2014: Errors in ensemble Kalman smoother estimates of cloud microphysical parameters. Mon. Wea. Rev., 142, 1631-1654, doi:10.1175/MWR-D-13-00290.1.

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

Rambukkange, M. P., J. Verlinde, E. W. Eloranta, C. J. Flynn, and E. E. Clothiaux, 2011: Using Doppler spectra to separate hydrometeor populations and analyze ice precipitation in multilayered mixed-phase clouds. IEEE Geosci. Remote Sens. Lett., 8, 108-112, doi:10.1109/LGRS.2010.2052781.

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 A. J. Heymsfield, 2010: The dimensional characteristics of ice crystal aggregates from fractal geometry. J. Atmos. Sci., 67, 1605-1616, doi:10.1175/2009JAS3187.1.

Szyrmer, W., and I. Zawadzki, 2010: Snow studies. Part II: Average relationship between mass of snowflakes and their terminal velocity. J. Atmos. Sci., 67, 3319-3335, doi:10.1175/ 2010JAS3390.1.

Tamminen, J., 2004: Validation of nonlinear inverse algorithms with Markov chain Monte Carlo method. J. Geophys. Res., 109, D19303, doi:10.1029/2004JD004927.

_____,and E. KyrÖlÄ, 2001: Bayesian solution for nonlinear and non- Gaussian inverse problems by Markov chain Monte Carlo method. J. Geophys. Res., 106, 14 377-14 390, doi:10.1029/ 2001JD900007.

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, Colorado State University, 231 pp. [Available online at 10217/48170.]

Zhang, Y., and G. G. Mace, 2006: Retrieval of cirrus microphysical properties with a suite of algorithms for airborne and spaceborne lidar, radar, and radiometer data. J. Appl. Meteor. Climatol., 45, 1665-1689, doi:10.1175/JAM2427.1.

Zhao, Y., G. G. Mace, and J. Comstock, 2011: The occurrence of particle size distribution bimodality in midlatitude cirrus as inferred from ground-based remote sensing data. J. Atmos. Sci., 68, 1162-1179, doi:10.1175/2010JAS3354.1.


University of Michigan, Ann Arbor, Michigan


University of Utah, Salt Lake City, Utah

(Manuscript received 16 July 2013, in final form 2 April 2014)

Corresponding author address: Derek J. Posselt, Department of Atmospheric, Oceanic, and Space Sciences, University of Michigan, 2455 Hayward Street, Ann Arbor, MI 48109-2143.


(ProQuest: Appendix omitted.)

For more stories covering the world of technology, please see HispanicBusiness' Tech Channel

Source: Journal of Applied Meteorology and Climatology

Story Tools Facebook Linkedin Twitter RSS Feed Email Alerts & Newsletters