Simulation of the Stokes vector in inhomogeneous precipitation



[1] The large absorption and scattering cross sections of liquid and frozen hydrometeors, respectively, introduce appreciable signatures to measured polarized brightness temperatures, degrading the retrieval of other geophysical parameters such as near-surface ocean winds. In particular, the retrieval of wind direction requires precise knowledge of polarization. This study investigates the fully polarized atmospheric contribution of precipitation and compares these effects with the current sensitivities of passive wind vector retrieval algorithms. A realistic microphysical cloud model supplies atmospheric parameters, including hydrometeor water contents, which are input into a vector radiative transfer model. Scattering is handled using a reverse Monte Carlo method. Radiances are simulated for three frequencies of interest to microwave polarimetry, 10.7, 18.7, and 37.0 GHz, for the four elements of the Stokes vector. The simulations show that the dichroic nature of precipitating media has a significant impact on passive wind vector retrievals.

1. Introduction

[2] Precipitation is a dominating quantity in passive microwave remote sensing. Because of the complexities of tropospheric cloud structures, the retrieval of precipitation is particularly difficult. The large microwave signal of rain also interferes with the measurement of other oceanic quantities. In particular, the polarimetric signatures of ocean surface roughness, which are essential to radiometric wind direction retrievals, are highly sensitive to interference. While the polarized surface emission is immune to most atmospheric effects, precipitation adversely affects the ability to obtain accurate ocean vector wind measurements [Adams et al., 2006].

[3] Since the ocean surface is a boundary interface, a calm ocean with negligible wind speed results in specular reflection where the Fresnel formulae describe the reflection and emission. As wind flows over the fluid ocean surface, roughness of the boundary increases, which changes the reflection and emission characteristics; therefore, the total intensity of the Stokes vector increases with increasing roughness (wind speed). Also, the ocean surface roughness is preferential to wind direction. This results in separate harmonic dependencies of the Stokes vector elements,

equation image

with respect to the difference between wind direction and viewing angle [Johnson, 2006]. I, Q, U, and V are the four components of the Stokes vector; E0v and E0h are the parallel (vertical) and perpendicular (horizontal) electric fields, respectively; and ε and μ are the permittivity and permeability, respectively, of the medium.

[4] The retrieval of wind direction requires highly accurate brightness temperature measurements, with nominal effective noise equivalent differential temperatures at or below 0.1 K. Retrieval sensitivities in the Naval Research Laboratory's WindSat geophysical retrieval algorithm [Bettenhausen et al., 2006] for Q are 0.21, 0.29, and 0.7 K for 10.7, 18.7, and 37.0 GHz, respectively, while retrieval sensitivities for U and V are on the order of 0.05–0.1 K for those same frequencies (M. H. Bettenhausen, private communication, 2007). Thus, low levels of polarized interference will impact the brightness temperature inversion process.

[5] Until recently, the research of polarization effects of precipitation in passive microwave radiometry has primarily focused on how Q, the difference between vertical and horizontal polarizations, is related to particle type and shape. Haferman [2000] lists research efforts for both simulating and observing microwave radiances, specifically polarization, for precipitation. Haferman mentions a study by Roberti and Kummerow [1999] that looks at simulations of cloud structure radiances for nonspherical hydrometeors, primarily the vertical and horizontal polarization. While the study does not look directly at the full Stokes vector, it does present some useful insights and methods.

[6] Roberti and Kummerow [1999] perform Monte Carlo radiative transfer simulations for a set of one-dimensional precipitation profiles for 19.35, 37, and 85.6 GHz at an incidence angle of 50°. The authors model rain and snow as horizontally aligned oblate spheroids, while graupel is modeled as spheres. The aspect ratio of the rain is a function of drop size, while the aspect ratio of snow is uniform over all particle sizes. As part of the analysis, they increase the snow and reduce the graupel concentrations to determine the effects of nonspherical ice particles and to match observations. Results show polarization differences up to 15 K, depending on frequency, but the authors state that the third Stokes values are small and are included only for calculations, not analysis. Also, a conversion of 50% of the graupel to snow gives simulated results that match observation.

[7] Battaglia and Simmer [2007] detail an in-depth study of differences between one-dimensional and three-dimensional radiative transfer simulations in order to explain a strong negative Q in downwelling measurements that has not been replicated in 1-D simulations. Simulations are performed at 6.6, 19.4, and 85.6 GHz for an idealized rain column of 25 mm/h. The study shows strong disparity in 1-D and 3-D downwelling simulations, emphasized at lower frequencies, while the differences in simulations are much less apparent for upwelling scenarios.

[8] One of the first studies of the presence of a third Stokes component in precipitation originated from a group of scientists from Germany and Russia. The most relevant study to this research is by Kutuza et al. [1998]. While the study is limited in scope, model calculations assert an appreciable third Stokes parameter in precipitation. The simulations include a canting angle which accounts for the horizontal drag component, i.e., horizontal wind in the rain volume. This results in a nonzero emission component for the third Stokes parameter. However, the simulation only uses a uniform vertical profile and is only performed at nadir for upwelling simulations and at zenith for downwelling simulations. Other simplifying assumptions are made, such that scattering by rain is ignored and emission and backscattering by ice are ignored. Results show U values on the order of 0.9 K of upwelling radiation at 35 GHz and 0.1 K at 13.3 and 20 GHz.

[9] More recently, Battaglia et al. [2007] compare the three-dimensional results from different Monte Carlo methods and also compare these results to a one-dimensional discrete ordinate method. The study analyzes 19.4 GHz downwelling simulations for a 25 mm/h idealized rain column over a range of viewing angles from zenith (0°) to 90°. As in the work by Battaglia and Simmer [2007], the authors observe a strong negative Q in 3-D simulations for a range of angles. More pertinent to this study is the resolution of third and fourth Stokes measurements in 3-D simulations, which are not only dependent on incidence angle but show a strong correlation with the azimuthal symmetry of the precipitation with respect to the position of the sensor.

[10] This study presents a unique analysis of the atmospheric contribution of inhomogeneous precipitation to the Stokes vector. By simulating the transfer of microwave radiation through realistic cloud structures, brightness temperatures are generated for three frequencies matching the polarimetric channels of the WindSat polarimetric radiometer (10.7, 18.7, and 37 GHz) under various precipitating conditions. Scenarios of one-dimensional and three-dimensional inhomogeneity are considered. The shapes of rain and ice hydrometeors are modeled so as to simulate the effect of particle shape on polarization. Section 2 explains the modeling of hydrometeors for radiative transfer simulations. Section 3 describes the radiative transfer model and scattering calculations. Section 4 lists the methodology and considerations for the simulations. In section 5, simulation results are given. Section 6 concludes the study.

2. Precipitation Modeling

[11] Understanding the effect of precipitation on microwave radiation requires a proper model of cloud microphysics. Knowledge of liquid and ice water profiles is necessary to generate particle size distributions and shapes for generating and scaling rain and ice (snow and graupel) scattering properties. The Goddard Cumulus Ensemble (GCE) model [Tao and Simpson, 1993] is a tool used to describe the complex environment of convective systems in four dimensions. GCE simulations have proven useful in providing the vertical cloud structure required for detailed radiative transfer calculations and retrieval inversions [Simpson and Tao, 1993; Kummerow et al., 1996].

2.1. Cloud Structure

[12] Data from a GCE simulation of a tropical squall line that developed near the Tropical Ocean–Global Atmosphere–Coupled Ocean-Atmosphere Response Experiment (TOGA-COARE) observational array [Webster and Lukas, 1992] in the western Pacific Ocean on 22 February 1993 provide liquid and ice profiles for radiative transfer calculations. The available simulation data file has a horizontal resolution of 2 km on a 140 × 140 pixel grid. The vertical profile consists of 28 layers plus the surface. From the surface to an altitude of 10 km, each layer is resolved at 0.5 km, and above this altitude to a bounding height of 18 km, layers are resolved at 1 km.

[13] Since falling hydrometeors are the primary mechanism for scattering due to the large size (on the order of a few millimeters or more), especially at 37 GHz, these quantities are modeled carefully. The simulations also include cloud ice and liquid since these also affect the simulated radiances; however, scattering and, therefore, polarization effects are negligible because of the small particle size (on the order of tens of microns). With ice and liquid densities, particle size distributions can then be calculated. Determining the dielectric properties of the particles and calculating gas absorption both require the temperature profile. Additionally, gas absorption calculations depend on the pressure and humidity profiles.

2.2. Particle Size Distributions

[14] Accurate calculations of particle absorption, scattering, and emission require knowledge of both particle sizes and the distribution of those particle sizes. Marshall and Palmer [1948] describe the distribution of raindrop sizes as an inverse exponential distribution of the form

equation image

where D is drop diameter, N(D) is the number density over the range D + dD, N0 is intercept parameter, and λ is the slope of the distribution. Marshall and Palmer find an intercept of 0.08 cm−4 to be consistent with observations, and the slope typically relates to rain rate via a power law fit. The inverse exponential distribution extends to both snow and graupel. The intercept values vary with rain type and geographical location. The intercepts used in these simulations fall within typical value ranges: the intercept for rain and graupel is that of Marshall-Palmer, and the intercept used for snow is 0.17 cm−4. Since water contents, not precipitation rates, are available from the GCE data set, λ must match the water content [Kummerow et al., 1996]. To determine water content, the masses of the particles are integrated assuming spherical particles [Liou, 2002]:

equation image

where ρ is the particle density, 1 g/cm3 for rain, 0.1 g/cm3 for snow, and 0.4 g/cm3 for graupel, and N(D) is the particle number distribution, in this case the inverse exponential. By solving (3) and then inverting,

equation image

[15] A modified gamma distribution represents the cloud liquid water distribution (C. Davis, PyArts user guide, algorithm description and theoretical basis,∼cdavis/PyARTS/userguide.pdf, 2006). McFarquhar and Heymsfield [1997] represent gamma as cloud ice, as this parameterization is indicative of tropical cloud ice and is therefore compatible with the tropical squall line used for the simulation.

2.3. Particle Shape

[16] Besides particle size, particle shape is an important characteristic when considering the polarizing effects of hydrometeors. As rain falls, aerodynamic drag flattens the spherical shape of the drops, with the large dimension perpendicular to the drop direction. Snow crystals form as a hexagonal prism, from which dendritic arms grow. Snow particles “rock” back and forth as they fall, but the large dimension also tends to be perpendicular to the fall direction. At the frequencies of interest, the intricacies of the hydrometeor shapes are inconsequential; however, the general shape is of great importance. Standard practice is to estimate the shape of rain and snow as horizontally aligned oblate spheroids [Oguchi, 1983]. An oblate spheroid is an ellipse that is rotated about its minor axis. The oblateness of both rain and snow increases with the size of the hydrometeor. Oblateness, quantified by the aspect ratio, is the ratio of the major to minor axes of the defining ellipse. Since the distribution properties of precipitation depend on the volume of a sphere, as in (3), many polynomial expansions relate the aspect ratio RA to the radius (or diameter) of an equivalent volume sphere. The expansion chosen for this study is [Andsager et al., 1999]

equation image

[17] Graupel forms when supercooled water droplets accrete on snow crystals and tends to be either spherical or conical. Graupel is assumed to be spherical for the purposes of these simulations. Cloud droplets are also spherical, while the shape of cloud ice varies. Since cloud ice particles are much smaller than the smallest wavelength (about 20 μm versus an 8 mm wavelength for 37 GHz), a precise model for cloud ice is unnecessary; therefore, cloud ice is estimated as randomly oriented oblate spheroids with a constant aspect ratio of 2.

3. Radiative Transfer

[18] All gas absorption and radiative transfer calculations are performed using the Atmospheric Radiative Transfer Simulator (ARTS) [Buehler et al., 2005]. ARTS is a flexible radiative transfer model capable of modeling diverse atmospheric conditions for a variety of sensor configurations and has been validated for frequencies below 1 THz. The original implementation of ARTS (version 1.0) is a one-dimensional tool capable of generating atmospheric absorption coefficients for trace gases such as water vapor, oxygen, and nitrogen. It also calculates scalar radiative transfer. The more recent implementation of ARTS (version 1.1) extends radiative transfer calculations to up to three atmospheric dimensions and computes the full Stokes vector, which allows for the consideration of scattering. ARTS version 1.1.1095 (hereinafter referred to as ARTS) is the primary tool used for this study. Since this version does not generate atmospheric absorption coefficients internally, ARTS version 1.0.195 (hereinafter referred to as ARTS 1.0) generates these coefficients.

3.1. Gas Absorption

[19] At 22.235 GHz, there is a weakly absorbing water vapor spectral line broadened by molecular collisions. These collisions are dependent on pressure, temperature, and water vapor [Ulaby et al., 1981]. There is a strong oxygen absorption band at 60 GHz resulting from changes in the orientation of electron spin with respect to molecular rotation. Of the frequencies of interest, gas absorption affects 10.7 GHz the least, with a small contribution from water vapor and a negligible contribution from oxygen. Simulations at 18.7 GHz are sensitive to water vapor and are only slightly more sensitive to oxygen than those of 10.7 GHz. At 37 GHz there is an almost equal contribution of gas absorption from water vapor and oxygen; however, the total absorption is on par with that of 18.7 GHz [Liou, 1992]. Thus, water vapor and oxygen absorption must be considered when performing accurate atmospheric radiative transfer simulations.

[20] ARTS 1.0 can compute absorption coefficients for both water vapor and oxygen using a number of popular models. This study utilizes the PWR98 model to generate water vapor absorption coefficients [Rosenkranz, 1998] and utilizes PWR93 for oxygen absorption [Rosenkranz, 1993]. ARTS 1.0 calculates gas absorption coefficients using volume mixing ratios. The oxygen volume mixing ratio remains constant in the atmosphere, while volume mixing ratios for water vapor are calculated with vapor saturation pressure from Flatau et al. [1992].

3.2. Scattering Calculations

[21] Scattering, emission, and absorption calculations are performed external to ARTS and are stored in extensible markup language (XML) files for use by the ARTS environment. Michael Mishchenko's FORTRAN T-matrix codes are utilized for scattering calculations. These codes have been modified and implemented in PyARTS (Davis, 2006). PyARTS is a tool capable of controlling ARTS simulations, writing data files in XML for interaction with ARTS, and generating atmospheric data files and scattering parameters.

[22] The amplitude scattering matrix gives the dependence of a scattered spherical wave to the field incident on a particle [Mishchenko et al., 2005]:

equation image

where Ev is the electric field parallel to the plane of incidence (vertical); Eh is perpendicular to the plane of incidence (horizontal); i and s denote incident and scattered fields, respectively; k is the wave number; r is the radial distance from the particle; equation image defines the direction of propagation; and S11, S12, S21, and S22 are the elements of the amplitude scattering matrix. From the amplitude-scattering matrix, the phase (scattering) and extinction matrices, as well as the absorption vector, can be derived.

[23] For individual horizontally aligned oblate spheroids, scattering is independent of azimuth viewing direction [Mishchenko et al., 2005]. S12 and S21 are both zero; therefore, the extinction matrix K is represented by a block diagonal configuration:

equation image

where θ is the difference between the incident and scattered angles. Here there are only three independent terms. Also, this configuration of the extinction matrix separates the dependence of I and Q from U and V; that is, extinction of unpolarized or vertically/horizontally polarized radiation will not result in third or fourth Stokes components. Additionally, the absorption vector, which also governs emission, consists of only I and Q components. However, all of the elements of the 4 × 4 scattering matrix must be considered, resulting in nonzero third Stokes components for energy that is scattered in the direction of incidence from other directions in the case of multiple scattering in a rotationally nonsymmetric medium.

[24] To calculate the scattering data for snow and rain, the double-precision T-matrix code for nonspherical particles in fixed orientation [Mishchenko, 2000] is used. First, the T-matrix is calculated once for each particle size and corresponding aspect ratio, incident wavelength, and complex index of refraction. Then the amplitude scattering matrix can be calculated for all incident and scattered directions.

[25] Large collections of particles with random orientations are considered to be isotropic and symmetric; therefore, the ensemble-averaged extinction matrix is nondirectional and diagonal [Mishchenko et al., 2005]. The scattering matrix simplifies to a block diagonal structure for the general case, and it becomes diagonal for forward and backward scattering. Emission is unpolarized. For the special case of spherical particles, the T-matrix reduces to Mie theory.

[26] For graupel, cloud liquid, and cloud ice, the double-precision T-matrix code for nonspherical particles in random orientation [Mishchenko and Travis, 1998] is used. The aspect ratios for graupel and cloud liquid are set at 1.000001 to avoid convergence issues of using an aspect ratio of exactly 1. Only one set of calculations is required, unlike the case of horizontally aligned particles. For a given particle size and corresponding aspect ratio, incident wavelength, and complex index of refraction, only the extinction and scattering cross section, and the block diagonal elements of the phase matrix, are calculated as a function of scattering angle.

[27] In addition to particle shape and orientation, the absorption, emission, and scattering quantities depend on the dielectric properties of hydrometeors. The large imaginary component of the permittivity of liquid water at microwave frequencies over that of ice demonstrates the strongly absorptive properties of rain over those of snow or graupel, while both rain and snow/graupel have similar scattering cross sections. Still, computing the permittivity of rain is trivial when compared to snow or graupel, as raindrops are considered pure liquid water. Graupel contains many air pockets within the ice structure. Although the small-scale structure of snow is ignored for the scattering calculations, the spacing between the dendritic arms affects the density and dielectric properties of the uniformly estimated particle. Thus, pockets and spacings are considered air inclusions when computing the dielectric properties of snow and graupel. Accurate T-matrix calculations require physically reasonable approximations of the dielectric properties of rain, snow, and graupel. Most of the dielectric models are empirical fits of permittivity to frequency with a temperature dependence included.

[28] Liebe et al. [1989] develop the frequency and temperature-dependent dielectric properties for liquid water and pure ice. For liquid water, the double-Debye equation determines the permittivity. For ice, the permittivity is based on a slightly simpler fit. For snow and graupel, the Maxwell Garnett mixing scheme [Maxwell Garnett, 1904] introduces air inclusions and calculates an effective permittivity.

3.3. Monte Carlo Simulation

[29] Within ARTS, there are two available methods for performing radiative transfer calculations for scattering atmospheres: a discrete ordinate iterative (DOIT) method [Emde et al., 2004] and a reverse Monte Carlo method [Davis et al., 2005]. While the DOIT method would be more expedient for one-dimensional radiative transfer calculations, the Monte Carlo algorithm is utilized for consistency with the three-dimensional simulations. The reverse Monte Carlo method [Davis et al., 2005] follows a prescribed number of arbitrary, fixed increments of energy, termed photons, from the measurement point, backward through the scattering medium. The extinction contribution (including gas absorption) is calculated for a chosen propagation path length. By using a random number, a decision is made at the propagation path step as to whether a photon is scattered or absorbed. If the photon is absorbed, the propagation path ends, and the emission contribution is included. Otherwise, a new incident direction is chosen, and the scattering contribution is logged. Once all photon tracing is complete, the contributions are combined, and the radiance at the measurement location and the simulation error are calculated, which gives a measure of accuracy or convergence. The accuracy of the simulation is controlled by the number of photons used in the scattering calculations; however, increasing the number of photons, and accuracy, increases run time.

4. Simulation Methodology

4.1. Simulation Time Versus Accuracy

[30] The simulations conducted for this study require an accurate representation of the particle number densities of highly inhomogeneous precipitation fields. The exponential dependence of the Marshall-Palmer distribution on liquid water content precludes using Gauss-Laguerre integration, as in the work by Battaglia et al. [2007], over a large range of liquid or ice water contents; therefore, a large number of bins (200 each for rain, snow, and graupel) is required for integrating the scattering properties over the particle size distributions. Such a large collection of particle sizes results in long Monte Carlo simulation times, so a trade must be made between desired accuracy and simulation time. By performing independent simulations for a range of photons, a profile with large amounts of rain, snow, and graupel gives an upper bound to the accuracy and run time expected from subsequent simulations. Starting at 25,000 photons, each run doubles the number of the photons from the previous run, up to 1.6 million photons.

[31] Figure 1 is the Monte Carlo simulation error (standard deviation), which is inversely proportional to the square root of the number of photons, for six independent simulations for a single benchmark profile consisting of large amounts of rain, snow, and graupel. This single profile has been expanded in both horizontal dimensions to give the effect of a one-dimensional profile, and the instrument geometry is configured such that the medium is rotationally symmetric about a vertical axis at the point where the instrument line-of-sight vector impinges the surface of the model environment. The errors for Q and U are equivalent, while the errors for I are about 1.5–2.5 times that of Q or U, depending on frequency. The error in V is many orders of magnitude lower than those of the other Stokes elements.

Figure 1.

Stokes error versus number of photons. Solid line is 10.7 GHz, dashed line is 18.7 GHz, and dotted line is 37.0 GHz.

[32] Figure 2 shows Monte Carlo run times for the benchmark profile. The simulations are performed on a dual-processor 1.8 GHz Power Mac G5. Over the range of 100,000–200,000 photons, the simulation errors are on the order of the retrieval sensitivities for Q, and simulation times are on the order of 1 h. Thus, one-dimensional simulations use 120,000 photons to keep Monte Carlo simulation time short. To acquire sufficient accuracy to resolve a third Stokes component, 2 million photons are used for three-dimensional simulations. Simulation times for the three-dimensional cases range from 12 to 16 h, depending on frequency and microphysics.

Figure 2.

Monte Carlo run time versus number of photons. Solid line is 10.7 GHz, dashed line is 18.7 GHz, and dotted line is 37.0 GHz.

4.2. Surface Consideration

[33] Unfortunately, current surface models are not mature enough to accurately represent fully polarimetric surfaces seen in nature. Models, such as the two-scale model [Johnson, 2006] or FASTEM3 [Saunders, 2006], must be empirically tuned to represent satellite observations [Bettenhausen et al., 2006]. Thus, a quantitative error analysis of rain-contaminated wind vector retrievals is not possible. Additionally, polarimetric surface models are not compatible with three-dimensional vector radiative transfer models because of the lack of cross-polarization terms for U and V, and proper handling of reflections over a full 2π sr is computationally intractable given the already intensive atmospheric calculations. Therefore, only highly simplified surfaces are considered for this work, and modulation of the surface Stokes vector is not analyzed. This allows for the isolation of the atmospheric contribution from precipitation to the Stokes vector. For one-dimensional simulations, the surfaces are Lambertian with 0.9 emissivity, and the surface temperatures are approximately 300 K but are dependent on the profile. For all three-dimensional simulations, blackbody surfaces are used, instead of Lambertian ones, to reduce run time by eliminating surface reflections. Again, surface temperatures are approximately 300 K but are dependent on location.

4.3. One-Dimensional Simulations

[34] The images presented in Figure 3 are slices of the GCE profiles that are used to perform one-dimensional radiative transfer simulations, similar to the work by Roberti and Kummerow [1999]. Each vertical profile is considered to be an individual one-dimensional profile. The slice of profiles gives a line of spatially correlated cloud structures, perpendicular to the line of convection, that vary slowly with position.

Figure 3.

Slices of precipitation profiles used for one-dimensional simulations.

[35] This collection of precipitation data allows for a close inspection of the effect of a large number of combinations of rain, snow, and graupel. All simulations are calculated for an incidence angle of 50°. One-dimensional profiles are prepared in the same manner as described for the time-and-accuracy trade study. U and V are ignored.

4.4. Three-Dimensional Simulations

[36] The images in Figure 4 are the integrated rain, snow, and graupel contents used for three-dimensional simulations. The slices used for the 1-D simulations are taken at y = 0 km. The data region is partitioned into eight overlapping 40 km × 40 km regions, centered at 32, 52, 72, 82, 92, 102, 112, and 132 km. The increased sampling between 72 and 112 km is to include more data with high snow and graupel contents. The regions are handled as separate atmospheres to reduce the large amounts of memory needed to store the Marshall-Palmer-derived particle number density fields. The sensor geometry, shown in Figure 5, is configured so that the line of sight for each simulation results in a 50° incidence angle with the satellite viewing the -x direction. The line-of-site vector intersects the surface at the previously stated center point for each region.

Figure 4.

Images of the integrated precipitation profiles used for three-dimensional simulations.

Figure 5.

Sensor configuration for three-dimensional simulations.

5. Simulation Results

[37] The results of both one- and three-dimensional simulations, detailed in sections 5.1 and 5.2, support the findings in the studies listed in section 1. However, the three-dimensional simulations offer a unique analysis of atmospheric contributions to the full Stokes vector for realistic precipitating scenarios.

[38] When calibrating a polarized radiometer channel, the measured power is linearly regressed to the measured target temperature. If both horizontal and vertical horns view a blackbody of 300 K, each will be calibrated to that temperature, resulting in an effective I of 600 K. To compensate for this calibration artifact, all WindSat retrieval sensitivities presented in this paper have been divided by a factor of 2.

5.1. One-Dimensional Results

[39] The results displayed in Figures 6, 7, and 8show a definite frequency dependence on the effect that absorption, emission, and scattering have on polarization. Q shows a large polarization signal, both from scattering and from absorption/emission. The surface temperature of 300 K and an emissivity of 0.9 result in a high background brightness temperature; therefore, scattering and absorption will be the mechanisms apparent when examining I. U and V are ignored since these elements of the Stokes vector are irrelevant to one-dimensional radiative transfer scenarios.

Figure 6.

One-dimensional simulation results for 10.7 GHz.

Figure 7.

One-dimensional simulation results for 18.7 GHz.

Figure 8.

One-dimensional simulation results for 37 GHz.

[40] For 10.7 GHz (Figure 6) the polarization signal is purely from emission and absorption. Q (VH) shows an opposing signature to that of I (V + H). The largest changes in I and Q correspond to the profiles with the greatest amounts of rain in the 32–40 km range, while the snow and graupel between 76 and 88 km do not seem to add any contribution beyond that of rain.

[41] At 18.7 GHz, given in Figure 7, absorption and emission are still the dominant mechanisms; however, the effects of scattering are also noticeable. Unlike at 10.7 GHz, the polarization effects in the regions with high snow and graupel are of similar magnitude to the region with large amounts of rain and negligible snow or graupel. Absorption effects are still the strongest mechanism, however. Also, from about 100 to 112 km the snow is polarizing the simulated radiation.

[42] Figure 8 presents the results for 37 GHz, which is much more sensitive to scattering. This effect is quite apparent in I. The greatest dip in brightness temperature occurs where there is a large amount of snow and graupel. While rain absorption also results in lower intensities, the negligible absorption cross section and large scattering cross section of frozen water guarantee lower intensities.

[43] Unfortunately, a quantitative analysis of the impact of atmospheric polarization contributions from precipitation is beyond the scope of this paper and is an extensive research subject on its own. However, the simulation results can be put in context with operational retrieval algorithms. Given the geophysical inversion sensitivities to Q listed in section 1, ranging from 0.2 K at 10.7 GHz to 0.7 K at 37 GHz, the atmospheric contributions to the polarization, which are well above these sensitivities, will have a noticeable effect on geophysical retrievals. Adams et al. [2006] have shown that even low levels of precipitation can adversely affect regression-based inversion algorithms, artificially increasing wind speed because of interference in I and Q. Similar results have been seen in physically based retrievals that do not include precipitation in the inversion process (Bettenhausen, private communication, 2007).

5.2. Three-Dimensional Results

[44] Figures 9, 10, and 11 display the results of the three-dimensional simulations. The three-dimensional results for I and Q agree well with the one-dimensional simulations with one obvious exception detailed in the following paragraph. For Q, the parity between the one-dimensional and three-dimensional simulations matches the results of Battaglia and Simmer [2007], which show close agreement in the polarization differences for upwelling 1-D and 3-D simulations. Any differences are likely to be more pronounced in a less realistic scenario of uniform heavy rain than in the realistic profiles used in this study. The results for U support the idealized simulations performed by Battaglia et al. [2007]; however, V is about 5–8 orders of magnitude lower than U in realistic precipitating environments.

Figure 9.

Three-dimensional simulation results for 10.7 GHz.

Figure 10.

Three-dimensional simulation results for 18.7 GHz.

Figure 11.

Three-dimensional simulation results for 37 GHz.

[45] While the precipitation features are not as finely resolved, because of the small number of 3-D simulations, the polarization effects of heavy precipitation features, visible across all frequencies, have shifted along the x direction of the simulation space with respect to the 1-D simulations. Both the region of heavy rain near 32 km and the region of heavy multiphase precipitation between 76 and 88 km display this effect. For the case that consists of primarily rain, the effects are more apparent in I at the lower frequencies, where attenuation of the high background signal is prevalent. With respect to the 1-D simulation, there is roughly a 2–4 km shift in the signal generated by this precipitation feature. The slant effect is more obvious in the multiphase precipitation region, as more 3-D samples have been simulated. The peak response for the 1-D simulations occurs at about 82 km and is most evident in I at 37 GHz. Again, while the 3-D simulations are not as finely sampled as the 1-D cases, the additional simulations at 82 km show that this feature has moved approximately 10 km, to close to the 72 km surface distance mark. These shifts are due to a combination of a slanted cloud structure (obvious in the rain between 66 and 86 km) and the 50° instrument line-of-sight angle. Both Roberti et al. [1994] and Turk et al. [2006] explain this phenomenon in much further detail.

[46] Besides the perceived shift with respect to surface geolocation of precipitation features observed in Q, spatial signatures are also apparent in U and, to a lesser extent, V. Off-angle scattering is the only mechanism by which precipitation generates these elements of the Stokes vector; therefore, the appreciable positive and negative values of U offer a measure of the spatial inhomogeneity. Figure 9 shows little third or fourth Stokes signal, as is expected since scattering from precipitation is minimal at 10.7 GHz. The signs of U and V in Figures 10 and 11 (and to a much lesser extent in Figure 9) correlate particularly well to the position of the rain features in Figure 4. When the rain features are positioned in the −y region, U exhibits a strong negative signal as in the results at 72 km. In contrast, positive U correlates strongly with the presence of rain features in the +y region, apparent in the results at 32 km.

[47] Three-dimensional simulations are particularly well suited for studying how the contribution to the U and V components of the Stokes vector can affect geophysical retrievals. V results for all frequencies are well below the lower limit of 0.05 K of the polarimetric retrieval sensitivities. U, on the other hand, is appreciable at 18.7 and 37 GHz, where scattering is significant. Even though V does not show an atmospheric contribution from precipitation, Adams et al. [2006] report errors in directional retrievals, primarily because the surface contribution of U is much greater than that of V. Similar issues with directional errors can be expected when including Q to resolve wind direction (Bettenhausen, private communication, 2007).

6. Conclusion

[48] The results from both sets of simulations show a large atmospheric contribution in I, Q, and U for realistic precipitation scenarios. While the simulation of a polarized surface is beyond the scope of this paper, the results for unpolarized backgrounds can be used to elucidate more complex scenarios. The effects of hydrometeor phase are largely dependent on frequency since scattering is prominent when the size of the particle is on the order of the incident wavelength. In both the scattering and absorption cases, the I and Q effects result from a greater extinction for the horizontally oriented electric field since the horizontal field component has the same orientation as the large dimension of horizontally aligned oblate hydrometeors. For frequencies where scattering becomes appreciable, the atmospheric contribution to U also becomes measurable. At 37 GHz, where scattering is significant, a large amount of spherical graupel can result in a depolarization of effects introduced by the nonspherical hydrometeors. For all frequencies, the atmospheric Q and U from precipitation, when compared with the retrieval sensitivities, will result in erroneous wind vector solutions. This is aggravated by the high incidence angles required by conically scanning sensors, like WindSat, to give adequate Earth coverage and wind direction response. While the atmospheric interference for the fourth Stokes parameter is negligible, this does not preclude distortion of the V component of a polarized surface signal (not considered here) directly because of scattering or more indirectly from extinction due to coupling with U.

[49] While the simulations aim to model realistic conditions, wind-induced canting angles of raindrops and a wind-roughened ocean surface are ignored. To investigate all of the effects of precipitation on the surface signal, future research requires the inclusion of both the hydrometeor canting angle and an advanced surface wind model. Regardless, the simulations show that there is still useful surface information at 10.7 GHz, even at moderate rain rates, with some surface information present at lower rain rates at 18.7 GHz, where scattering is negligible.


[50] We would like to thank Chris Kummerow and Jody Crook at Colorado State University for giving us access to Goddard Cumulus Ensemble data and to their radiative transfer code, which facilitated proper use of the GCE data. We would also like to thank the developers of ARTS for making their code available to the scientific community and for their support in using and understanding ARTS.