• Ackerman, A. S., O. B. Toon, and P. V. Hobbs, 1995: A model for particle microphysics, turbulent mixing, and radiative transfer in the stratocumulus-topped marine boundary layer and comparisons with measurements. J. Atmos. Sci., 52 , 12041236.

    • Search Google Scholar
    • Export Citation
  • Barker, H., and J. A. Davies, 1992: Solar radiative fluxes for stochastic, scale-invariant broken cloud fields. J. Atmos. Sci., 49 , 11151126.

    • Search Google Scholar
    • Export Citation
  • Cahalan, R. F., 1994: Bounded cascade clouds: Albedo and effective thickness. Nonlinear Processes Geophys., 1 , 156167.

  • Cahalan, R. F., and Coauthors, 2005: The I3RC: Bringing together the most advanced radiative transfer tools for cloudy atmospheres. Bull. Amer. Meteor. Soc., 86 , 12751293.

    • Search Google Scholar
    • Export Citation
  • Di Giuseppe, F., and A. M. Tompkins, 2003: Effect of spatial organization on solar radiative transfer in three-dimensional idealized stratocumulus cloud fields. J. Atmos. Sci., 60 , 17741794.

    • Search Google Scholar
    • Export Citation
  • Evans, K. F., 1993: A general solution for stochastic radiative transfer. Geophys. Res. Lett., 20 , 20752078.

  • Evans, K. F., and W. J. Wiscombe, 2004: An algorithm for generating stochastic cloud fields from radar profile statistics. Atmos. Res., 72 , 263289.

    • Search Google Scholar
    • Export Citation
  • Evans, K. F., A. Marshak, and T. Várnai, 2008: The potential for improved cloud optical depth retrievals from the multiple directions of MISR. J. Atmos. Sci., 65 , 31793196.

    • Search Google Scholar
    • Export Citation
  • Gentle, J. E., 2003: Random Number Generation and Monte Carlo Methods. 2nd ed. Springer, 381 pp.

  • Gikhman, I. I., and A. V. Skorokhod, 1977: An Introduction to the Theory of Random Processes. (in Russian). Nauka, 567 pp.

  • Hogan, R. J., and S. F. Kew, 2005: A 3D stochastic cloud model for investigating the radiative properties of inhomogeneous cirrus clouds. Quart. J. Roy. Meteor. Soc., 131 , 25852608.

    • Search Google Scholar
    • Export Citation
  • Kargin, B. A., and S. M. Prigarin, 1988: Modelling of stochastic fields of heap clouds and a study of their radiative properties with the Monte Carlo method (in Russian). Preprint 817, Siberian Division, USSR Academy of Sciences, Novosibirsk, 18 pp.

  • Marshak, A., A. Davis, R. F. Cahalan, and W. J. Wiscombe, 1994: Bounded cascade models as nonstationary multifractals. Phys. Rev. E, 49 , 5569.

    • Search Google Scholar
    • Export Citation
  • Marshak, A., A. Davis, W. J. Wiscombe, W. Ridgway, and R. F. Cahalan, 1998: Biases in shortwave column absorption in the presence of fractal clouds. J. Climate, 11 , 431446.

    • Search Google Scholar
    • Export Citation
  • Marshak, A., S. Platnick, T. Várnai, G. Wen, and R. F. Cahalan, 2006: Impact of 3D radiative effects on satellite retrievals of cloud droplet sizes. J. Geophys. Res., 111 .D09207, doi:10.1029/2005JD006686.

    • Search Google Scholar
    • Export Citation
  • Ogorodnikov, V. A., and S. M. Prigarin, 1996: Numerical Modelling of Random Processes and Fields: Algorithms and Application. s. VSP, 240 pp.

    • Search Google Scholar
    • Export Citation
  • Platnick, S., M. D. King, S. A. Ackerman, W. P. Menzel, B. A. Baum, J. C. Riedi, and R. A. Frey, 2003: The MODIS cloud products: Algorithms and examples from Terra. IEEE Trans. Geosci. Remote Sens., 41 , 459473.

    • Search Google Scholar
    • Export Citation
  • Prigarin, S. M., 1995: Spectral models of random fields and their applications. Adv. Modell. Analysis, 29A , 3951.

  • Prigarin, S. M., 2001: Spectral Models of Random Fields in Monte Carlo Methods. VSP, 198 pp.

  • Prigarin, S. M., and G. A. Titov, 1996: Spectral methods of numerical modeling of geophysical fields. Atmos. Oceanic Opt., 9 , 629635.

    • Search Google Scholar
    • Export Citation
  • Prigarin, S. M., and A. Marshak, 2005: Numerical model of broken clouds adapted to observations. Atmos. Oceanic Opt., 18 , 256263.

  • Prigarin, S. M., and A. Marshak, 2008: Simulation of vector semibinary homogeneous random fields and modeling of broken clouds. Numeric. Anal. Appl., 3 , 285292.

    • Search Google Scholar
    • Export Citation
  • Prigarin, S. M., B. A. Kargin, and U. G. Oppel, 1998: Random fields of broken clouds and their associated direct solar radiation, scattered transmission, and albedo. Pure Appl. Opt., 7A , 13891402.

    • Search Google Scholar
    • Export Citation
  • Prigarin, S. M., A. Martin, and G. Winkler, 2004: Numerical models of binary random fields on the basis of thresholds of Gaussian functions. Siberian J. Numer. Math., 7 , 165175.

    • Search Google Scholar
    • Export Citation
  • Scheirer, R., and S. Schmidt, 2005: CLABAUTAIR: A new algorithm for retrieving three-dimensional cloud structure from airborne microphysical measurements. Atmos. Chem. Phys., 5 , 23332340.

    • Search Google Scholar
    • Export Citation
  • Schertzer, D., and S. Lovejoy, 1987: Physical modeling and analysis of rain and clouds by anisotropic scaling multiplicative processes. J. Geophys. Res., 92 , 96939714.

    • Search Google Scholar
    • Export Citation
  • Schmidt, K. S., V. Venema, F. Di Giuseppe, R. Scheirer, M. Wendisch, and P. Pilewski, 2007: Reproducing cloud microphysics and irradiance measurements using three 3D cloud generators. Quart. J. Roy. Meteor. Soc., 133 , 765780.

    • Search Google Scholar
    • Export Citation
  • Sveshnikov, A. A., 1968: Applied Methods in Random Function Theory. (in Russian). Nauka, 463 pp.

  • Várnai, T., 2000: Influence of three-dimensional radiative effects on the spatial distribution of shortwave cloud reflection. J. Atmos. Sci., 57 , 216229.

    • Search Google Scholar
    • Export Citation
  • Venema, V., S. Bachner, H. W. Rust, and C. Simmer, 2006a: Statistical characteristics of surrogate data based on geophysical measurements. Nonlinear Processes Geophys., 13 , 449466.

    • Search Google Scholar
    • Export Citation
  • Venema, V., and Coauthors, 2006b: Surrogate cloud fields generated with the iterative amplitude adapted Fourier transform algorithm. Tellus, 58A , 104120.

    • Search Google Scholar
    • Export Citation
  • Voss, R., 1985: Random fractal forgeries. Fundamental Algorithms in Computer Graphics, R. A. Earnshaw, Ed., Springer-Verlag, 805–835.

  • Wen, G., A. Marshak, R. F. Cahalan, L. A. Remer, and R. G. Kleidman, 2007: 3-D aerosol–cloud radiative interaction observed in collocated MODIS and ASTER images of cumulus cloud fields. J. Geophys. Res., 112 .D13204, doi:10.1029/2006JD008267.

    • Search Google Scholar
    • Export Citation
  • Yamaguchi, Y., A. B. Kahle, H. Tsu, T. Kawakami, and M. Pniel, 1998: Overview of Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER). IEEE Trans. Geosci. Remote Sens., 36 , 10621071.

    • Search Google Scholar
    • Export Citation
  • Zuev, V. E., and G. A. Titov, 1995: Radiative transfer in cloud fields with random geometry. J. Atmos. Sci., 52 , 176190.

  • View in gallery
    Fig. 1.

    A schematic illustration of models A and B; H0 is a constant cloud base and d is a cutting threshold level.

  • View in gallery
    Fig. 2.

    Number of clouds per unit area mc (k = 1), as a function of cloud fraction Ac for models A and B.

  • View in gallery
    Fig. 3.

    Configuration of cloud fields (300 km × 300 km) for models (left) A and (right) B on the basis of a Gaussian random field with a correlation function J0 for the same cloud fraction Ac = 0.58 (ρ = 0.5; d = −0.20 for model A and 0.55 for model B). To simulate a Gaussian random field, a spectral model from Prigarin (2001, section 1.1.4) was used; see also appendix B.

  • View in gallery
    Fig. 4.

    A 68 km × 68 km region in Brazil centered at 17°S and 42°W collected on 9 Aug 2001 at 1015 LT. The solar zenith angle θ0 = 41°; the solar azimuth angle φ0 = 23° (from the top). (a) MODIS true color RGB (red–green–blue) 1-km resolution; (b) ASTER RGB 15-m resolution; (c) retrieved cloud optical thickness; (d) retrieved cloud-top height (in km).

  • View in gallery
    Fig. 5.

    Indicator functions I(x, y) of a cloud field: white is cloud (I = 1) and black is a cloud-free area (I = 0). Cloud fraction Ac = 0.4. (left) A 68 km × 68 km MODIS image centered at (17.1°S, 42.16°W) acquired on 9 Aug 2001. (right) A realization of a simulated field.

  • View in gallery
    Fig. 6.

    Four realizations of cloud optical depth; all of them have the same covariance function of the cloud mask KI(x, y) and histogram g(τ) as the one in Fig. 4c. Color scale is as in Fig. 4c. The size of the images is 68 km × 68 km, as in Fig. 5.

  • View in gallery
    Fig. 7.

    Estimates of the covariance function KI(x, y) = KI(r) (r = ) of the indicator function I(x, y) for the observed (solid line) and simulated (dotted line) cloud fields from Fig. 5.

  • View in gallery
    Fig. 8.

    The original histogram g(τ) and four other histograms that correspond to four realizations of cloud optical depth shown in Fig. 6.

  • View in gallery
    Fig. 9.

    Joint distribution of the given cloud optical depth and cloud-top height fields.

  • View in gallery
    Fig. 10.

    The conditional distribution function F(h|τ) used to simulate cloud-top height h; F(h|τ) is shown for two values of optical depth τ.

  • View in gallery
    Fig. 11.

    (a) A realization of cloud optical depth from Fig. 6b. (b)–(d) Three additional realizations of cloud-top height distribution; they correspond to the cloud optical depth field shown in (a). All realizations have the same conditional distribution F(h|τ).

  • View in gallery
    Fig. 12.

    A pdf of three realizations of cloud-top height shown in Fig. 11. The original pdf of cloud-top height from Fig. 4d is also shown.

All Time Past Year Past 30 Days
Abstract Views 0 0 0
Full Text Views 118 49 1
PDF Downloads 78 30 3

A Simple Stochastic Model for Generating Broken Cloud Optical Depth and Cloud-Top Height Fields

Sergei M. PrigarinNovosibirsk State University, and Institute of Computational Mathematics and Mathematical Geophysics, Russian Academy of Sciences (Siberian Branch), Novosibirsk, Russia

Search for other papers by Sergei M. Prigarin in
Current site
Google Scholar
PubMed
Close
and
Alexander MarshakNASA Goddard Space Flight Center, Climate and Radiation Branch, Greenbelt, Maryland

Search for other papers by Alexander Marshak in
Current site
Google Scholar
PubMed
Close
Full access

Abstract

A simple and fast algorithm for generating two correlated stochastic two-dimensional (2D) cloud fields is described. The algorithm is illustrated with two broken cumulus cloud fields: cloud optical depth and cloud-top height retrieved from the Moderate Resolution Imaging Spectroradiometer (MODIS). Only two 2D fields are required as an input. The algorithm output is statistical realizations of these two fields with approximately the same correlation and joint distribution functions as the original ones. The major assumption of the algorithm is statistical isotropy of the fields. In contrast to fractals and the Fourier filtering methods frequently used for stochastic cloud modeling, the proposed method is based on spectral models of homogeneous random fields. To retain the same probability density function as the (first) original field, the method of inverse distribution function is used. When the spatial distribution of the first field has been generated, a realization of the correlated second field is simulated using a conditional distribution matrix. This paper serves as a theoretical justification of the publicly available software “Simulation of a two-component cloud field,” which has been recently released. Although 2D rather than full 3D, the stochastic realizations of two correlated cloud fields that mimic statistics of given fields have proven to be very useful to study 3D radiative transfer features of broken cumulus clouds for a better understanding of shortwave radiation and the interpretation of remote sensing retrievals.

Corresponding author address: Alexander Marshak, NASA GSFC, Code 613.2, Greenbelt, MD 20771. Email: alexander.marshak@nasa.gov

Abstract

A simple and fast algorithm for generating two correlated stochastic two-dimensional (2D) cloud fields is described. The algorithm is illustrated with two broken cumulus cloud fields: cloud optical depth and cloud-top height retrieved from the Moderate Resolution Imaging Spectroradiometer (MODIS). Only two 2D fields are required as an input. The algorithm output is statistical realizations of these two fields with approximately the same correlation and joint distribution functions as the original ones. The major assumption of the algorithm is statistical isotropy of the fields. In contrast to fractals and the Fourier filtering methods frequently used for stochastic cloud modeling, the proposed method is based on spectral models of homogeneous random fields. To retain the same probability density function as the (first) original field, the method of inverse distribution function is used. When the spatial distribution of the first field has been generated, a realization of the correlated second field is simulated using a conditional distribution matrix. This paper serves as a theoretical justification of the publicly available software “Simulation of a two-component cloud field,” which has been recently released. Although 2D rather than full 3D, the stochastic realizations of two correlated cloud fields that mimic statistics of given fields have proven to be very useful to study 3D radiative transfer features of broken cumulus clouds for a better understanding of shortwave radiation and the interpretation of remote sensing retrievals.

Corresponding author address: Alexander Marshak, NASA GSFC, Code 613.2, Greenbelt, MD 20771. Email: alexander.marshak@nasa.gov

1. Introduction

To better understand and predict shortwave radiation in realistic cloudy atmospheres, we need to specify the 3D distribution of cloud liquid water. Also, statistical cloud retrievals that include 3D radiative transfer need to be trained on a large number of 3D cloud fields (Evans et al. 2008). Realistic cloud fields and spatial distributions of cloud liquid water can be obtained from either dynamical or stochastic cloud models. Based on cloud dynamics, physical (or dynamical) cloud models such as a large-eddy simulation (LES) or a cloud-resolving model (e.g., Ackerman et al. 1995) are physically consistent but require specification of a lot of atmospheric parameters and often are computationally expensive. On the other hand, stochastic cloud models based on aircraft, satellite, or ground measurements of cloud structure are computationally inexpensive and can output a much larger range of scales than dynamical models. Stochastic cloud models are mostly 2D because currently there are no techniques to measure a full 3D cloud structure. (To get a 3D stochastic model, one can assume the same statistics for both horizontal directions; see Evans and Wiscombe 2004; Hogan and Kew 2005.)

Over the last two decades, many different cloud stochastic models have been developed. We break them into two classes. The first class of cloud models uses only a few parameters to simulate the main aspects of the realistic cloud fields like the mean, standard deviation, and correlation (often assumed to be a power law). These models are very simple and are generally used to test hypotheses and better understand cloud–radiation interaction; they include the fractionally integrated cascade model (Schertzer and Lovejoy 1987), the bounded cascades (Cahalan 1994; Marshak et al. 1994), the fractional Brownian motion (Voss 1985), the Fourier filtering of Gaussian noise (Barker and Davies 1992; Evans 1993; Várnai 2000), and the Poisson distribution of cloud elements (Zuev and Titov 1995), to name a few. These models generally produce an unbroken (overcast) 2D xy field of cloud optical depth or cloud liquid water path. To obtain the desired cloud fraction, a simple threshold can be used (e.g., Barker and Davies 1992; Marshak et al. 1998).

The second class of cloud stochastic models provides a statistical reconstruction of an observed field and generates the detailed cloud structure. They are also called statistical cloud generators (Venema et al. 2006a; Schmidt et al. 2007). These cloud models are usually 3D rather than 2D. For cumulus clouds, Evans and Wiscombe (2004) used time–height radar data to generate 2D realizations of cloud liquid water that are then generalized to 3D fields assuming statistical homogeneity and horizontal isotropy. For stratocumulus clouds, Di Giuseppe and Tompkins (2003) generated 3D cloud liquid water fields by combining stochastic horizontal models based on a power spectrum and Fourier filtering (at each height) with realistic vertical profiles of total water and temperature. From radar time–height series, using a Fourier transform technique, Hogan and Kew (2005) generated realistic 3D cirrus clouds with a fallstreak structure vertically changing the slope of the power spectrum. Venema et al. (2006a, b) generated a surrogate cloud field with a liquid water distribution and spatial correlation (through a power spectrum) statistically similar to the observed one. Venema et al. (2006b) also compared radiative properties of LES clouds with their surrogate fields and Venema et al. (2006a) provided an excellent review of different cloud generators. Finally, we mention the Scheirer and Schmidt (2005) generator, which reproduces cloud fields of liquid water and effective radius using aircraft data. Schmidt et al. (2007) used cloud fields simulated by the last three generators as input to a 3D radiative transfer model to compare its output with the radiative flux measurements.

The current paper describes a simple stochastic model that belongs to the second class of cloud stochastic models. For given 2D fields of cloud optical depth and cloud-top height, the model generates realizations of these two fields with the same covariance of the cloud mask and the joint distribution as the original fields. In contrast to Evans and Wiscombe (2004), it does not generate 3D cloud liquid water fields but rather provides the xy fields of cloud optical and geometrical thicknesses (assuming a constant cloud base). To simulate the required autocorrelation function, it uses spectral models of homogeneous random fields (Prigarin 1995, 2001) rather than commonly used Fourier filtering (e.g., Evans 1993; Di Giuseppe and Tompkins 2003; Evans and Wiscombe 2004; Hogan and Kew 2005; Venema et al. 2006b). Another distinguishable feature of this paper is that it provides a theoretical background for the publicly available software that has been recently developed and released by the authors.

The plan of the paper is as follows: The next section briefly discusses two stochastic models of broken cloudiness that are based on a truncated Gaussian homogeneous field. The (auto)correlation function of a 2D field defines its structure. Section 3 describes how to generate a quasi-Gaussian field with a given correlation function that is retrieved from the covariance of the indicator function of the original field. Section 4 then explains how to modify the field to reproduce the observed distribution. In section 5 we generate the second field using the joint distribution of the given fields of cloud optical depth and cloud-top height. Section 6 illustrates the theory with Moderate Resolution Imaging Spectroradiometer (MODIS) data; section 7 summarizes the main steps of the proposed algorithm and discusses its applications. Section 8 gives a brief summary of the results. At the end, appendix A demonstrates the relations between the covariance functions of a Gaussian random filed and its indicators, and appendix B illustrates spectral models of Gaussian isotropic homogeneous random fields on the xy plane.

2. Quasi-Gaussian model of broken cloudiness

Let us assume that our cloud field has a constant cloud base at height H0 and a variable cloud top described by
i1520-0469-66-1-92-e1a
Here, υ(x, y) is a homogeneous Gaussian field with zero mean, unit variance, and a correlation function K(x, y) [with K(0, 0) = 1]. The point (x, y, z) belongs to a cloud if H0zw(x, y). A value w(x, y) = H0 simply means that there is a cloud gap in the horizontal point (x, y). The cloud-top field w has two parameters: a and d. Parameter a > 0 stretches the cloud-top field vertically and parameter d defines the truncation level (cf. Marshak et al. 1998).
Simultaneously with Eq. (1a), we consider another model (Kargin and Prigarin 1988) of cloud top:
i1520-0469-66-1-92-e1b
We will call Eqs. (1a) and (1b) models A and B, respectively. Figure 1 illustrates the difference between the two models. We can see that for model A, a positive d corresponds to a broken cloud field while a negative d corresponds rather to a more “overcast” cloud with a few gaps. Based on Fig. 1, one can say that model B better represents the cellular structure of stratocumulus whereas model A is for more broken cumulus clouds.
It can be shown that cloud fraction Ac has a different value for both models, namely:
i1520-0469-66-1-92-e2a
for model A and
i1520-0469-66-1-92-e2b
for model B, where
i1520-0469-66-1-92-eq1
is the standard normal cumulative distribution function, which ranges between 0 and 0.5 for negative d and between 0.5 to 1 for positive d.
For both models, we can determine the average number of clouds mc per unit area as a limit of the number of clouds in a convex domain of area S divided by S when S tends to infinity. Note that mc is dimensionless. Obviously mc will depend on d and the autocorrelation function K. For isotropic fields, one obtains [e.g., Sveshnikov 1968, Eq. (45.51), p. 441]
i1520-0469-66-1-92-e3
where i = 1 for model A and i = 2 for model B. Here, k > 0 is the second derivative of −K(x, y) with respect to x or y taken at x = y = 0 (with a negative sign). Note that for isotropic fields, the derivatives are equal [see Eq. (5) below].

Figure 2 shows the dependence of mc on cloud fraction Ac for both models and isotropic fields. We see that the number of clouds first increases with cloud fraction and then decreases. This is because the cloud fraction Ac itself monotonically decreases with the truncation level d whereas the number of clouds mc first increases with d and then decreases (see Fig. 1). Note that for model A, the number of clouds mc for Ac > 0.5 (d < 0) is not defined. The generalization of Eq. (3) to anisotropic fields is straightforward (Prigarin and Marshak 2005).

To summarize, both cloud models A and B are uniquely determined by parameters a and d and a correlation function K. To simulate a cloud field with a given cloud fraction Ac, we first solve Eq. (2) for the truncation level d. Then it is necessary to generate the correlation function K(x, y) based on some additional information about correlation in a real cloud field. Parameter a is determined from a simple one-point statistic describing the cloud-top height field. The most difficult part of such an approach is the choice (or generation) of the correlation function K; it will be discussed in the next section.

Finally, we emphasize that the cloud-top height field w(x, y) is simulated in two steps: (i) simulation of the Gaussian field υ(x, y) with mean zero and correlation function K(x, y) and (ii) calculation of w(x, y) with parameters a and d using (1a) or (1b). The problem of numerical simulation of Gaussian random fields has been well studied [e.g., chapters 1, 2, and 4 in Ogorodnikov and Prigarin (1996), as well as Prigarin (2001) and references therein] and will not be discussed here. Appendix B illustrates the approximation of a Gaussian homogeneous random field used in the numerical examples in section 6.

3. Correlation function

The correlation function defines the geometrical structure of a cloud field, including the size and distribution of individual clouds and the space between them. Perhaps the simplest and/or the most deterministic isotropic cloud field used in the first stochastic models can be defined by a Bessel function of the first kind, J0 (see Gikhman and Skorokhod 1977, p. 87). In this case, the correlation function is
i1520-0469-66-1-92-e4
where parameter ρ is responsible for cloud sizes (the larger the value of ρ, the smaller an average cloud is). It is easy to see that
i1520-0469-66-1-92-e5
Thus, to define ρ, one uses Eq. (3), which relates the average number of clouds per unit area mc and the second derivative k. Because ρ is fixed, the use of the correlation function (4) is very limited and the cloud fields based on it are unrealistic (see Fig. 3 as an example).
To generalize (4), Prigarin et al. (1998) used the radial spectral density of a Gaussian field z(ρ) and a representation of the normalized correlation function as an integral over all cloud sizes ρ of a product between z(ρ) and J0:
i1520-0469-66-1-92-e6
Here, z(ρ) ≥ 0 and ∫0 z(ρ)dρ = 1. Varying z(ρ), in general, one can get any correlation function of a random isotropic field on the plane.

Below we briefly describe a general procedure for generating the correlation function K(x, y) based on observations (leaving the details for appendix A). As an illustration, in section 6 we apply our algorithm to a broken cloud scene retrieved from MODIS.

Let I(x, y) be an indicator function (a binary cloud mask) that takes value 1 if there is a cloud above point (x, y) and 0 otherwise. Based on observations, we can estimate the mathematical expectation of I, which is a cloud fraction Ac:
i1520-0469-66-1-92-e7
and its covariance function,
i1520-0469-66-1-92-e8
It is known (Ogorodnikov and Prigarin 1996, p. 65) that the covariance function KI(x, y) of the indicator field I(x, y) and the correlation function K(x, y) of a Gaussian field υ(x, y) are nontrivially related. This relationship allows us to retrieve the correlation function K(x, y) from the measured covariance function KI(x, y). The main steps of the retrieval are described in appendix A. Note that the truncation level d is uniquely defined from Eqs. (2) and (7).

4. Geometrical thickness with a given density

Equations (1a) and (1b) alone do not allow us to control the distribution of cloud geometrical thickness w(x, y) − H0. Determined by Eqs. (1a) and (1b), this distribution is a scaled-up (a > 1) or -down (a < 1) truncated (by a parameter d) Gaussian distribution; its density is defined by
i1520-0469-66-1-92-e9
where φ(x) = 1/2π exp(−x2/2) is a standard Gaussian density. However, the observed distribution of cloud thickness does not necessarily satisfy Eq. (9). In general, one has to modify Eqs. (1a) and (1b) to reproduce the observed distribution. We describe below a modification of a Gaussian model that allows reproduction of any given distribution. This modification is based on the method of inverse distribution function widely used in statistical modeling (e.g., Ogorodnikov and Prigarin 1996, 65–71).
Let g(h) (h > 0) be a density of the observed distribution of cloud thickness. We denote its distribution function by
i1520-0469-66-1-92-e10
It is easy to see that if
i1520-0469-66-1-92-e11
is the distribution function [with density f1(h) defined in Eq. (9) where a = 1] and xi is a random variable distributed with the density f1, then G−1F(xi) will have a density of the observed distribution of cloud geometrical thickness. Indeed, there is a general statement (e.g., Gentle 2003, p. 42): if F is a distribution function of a random variable xi then F(xi) is uniformly distributed on the interval (0, 1). Therefore, the random variable G−1F(xi) has the probability density g. Note that a method of inverse distribution function similar to the above has been also used by Evans and Wiscombe (2004) to generate lookup tables for cloud liquid water content and droplet effective radius [see their Eq. (A.3)].
This leads us to the following modification of Eqs. (1a) and (1b):
i1520-0469-66-1-92-e12a
for model A, and
i1520-0469-66-1-92-e12b
for model B.

In contrast to (1a) and (1b), distributions of cloud thickness w(x, y) − H0 defined by either (12a) or (12b) match the observed probability distribution G(h). In addition, we recall that υ(x, y) is a homogeneous Gaussian field with zero mean and unit variance; its correlation function K(x, y) is retrieved from the covariance function KI(x, y) of the observed cloud mask field I(x, y). For both models, parameter d is uniquely determined from the average value of I(x, y), that is, cloud fraction Ac [see (7)].

5. Joint distribution of optical and geometrical thicknesses

We assume here that we have two random variables: cloud optical depth τ(x, y) and cloud geometrical thickness h(x, y). Then a pair (τ, h) will be a two-dimensional variable and P(τ1 < τ < τ2, h1 < h < h2) will be the probability that the values of τ and h fall in the intervals (τ1, τ2) and (h1, h2), respectively.

Practically (see next section), when two matrixes τ and h are given from observations, we first subdivide all their values into M and N bins, respectively. Then we calculate a conditional distribution matrix:
i1520-0469-66-1-92-e13
Now, if we assume that we have a realization of one variable, say h, then by using the conditional distribution matrix 𝗣 we can simulate a distribution of a second variable, τ. This is a straightforward procedure similar to a simulation of random number with a given distribution. As a result for each point (x, y) we will get both τ(x, y) and h(x, y) preserving conditional distribution (13) as well as the distribution of the random vector [h(x, y), τ(x, y)]. The order of simulation (first τ and then h or first h and then τ) is irrelevant for the reproduction of the joint distribution of the components of this two-dimensional vector.

Note that realizations of the second component generated using a conditional distribution matrix (13) are usually more stochastic (or noisy) than the given one. This is especially well pronounced if the original field has a strong spatial heterogeneity (e.g., the highest values are localized in several neighboring pixels). In a simulated field, these high values are not necessarily well localized and sometimes can be distributed through the whole scene, making it much noisier. This problem has been discussed in more detail in Prigarin and Marshak (2008).

6. Illustration with MODIS data

To illustrate the above theory with observations, we have selected a 1-km spatial resolution MODIS 68 km × 68 km broken cumulus cloud scene (Fig. 4a) from a less polluted region in Brazil, centered at 17°S and 42°W. The data were acquired on 9 August 2001 at 1015 LT. The solar zenith angle θ0 = 41°. This scene is part of phase 3 of the International Intercomparison of 3D Radiative Transfer Codes (I3RC; Cahalan et al. 2005) and has been used for the analysis of the retrieved droplet size by Marshak et al. (2006) and for the radiative effects of broken clouds on aerosol retrievals by Wen et al. (2007). The cloud fraction in the scene is Ac = 0.4. The MODIS image is collocated with a high–spatial resolution (15 m) Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) image (Yamaguchi et al. 1998) plotted in Fig. 4b. The solar azimuth angle φ0 = 23° (from upper right corner), as can be confirmed from the casting of the shadows.

In Figs. 4c and 4d we have also added the retrieved cloud optical depth and cloud-top height at a 1 km × 1 km resolution. Unlike the retrieved 1 km × 1 km cloud optical depth from the operational MODIS product, the operational cloud-top height retrievals have a 5 km × 5 km resolution (Platnick et al. 2003). To estimate the 1 km × 1 km resolution of cloud-top height, we used the brightness temperature at 11 μm (MODIS band 31) (see Wen et al. 2007 for details). As a result, Figs. 4c and 4d will serve as the basic scenes for our illustration.

First, Fig. 5a shows the indicator function I(x, y) of the cloud optical depth field from panel 4c. The cloud fraction, as a mathematical expectation of I defined in (7), is 0.4. Figure 5b is the indicator function of a realization of a simulated field that has the same covariance function KI(x, y) as the measured one. As shown in appendix A, to get K(x, y) we first estimated the covariance function KI defined in (8) and then retrieved K(x, y) with the help of the Owen function (Prigarin et al. 2004).

Next we illustrate how the distribution of cloud optical depth can be reproduced using Eq. (12a). This is done for model A, which perhaps agrees better with the observations than model B (Prigarin and Marshak 2005). Four realizations of cloud optical depth distribution are plotted in Fig. 6. All of them have approximately the same covariance function KI(x, y) of the indicator field (see Fig. 7) and probability density function (pdf) g(τ) as the original cloud optical depth field shown in Fig. 5c. Figure 8 illustrates these five pdfs: the original one and the four realizations of cloud optical depth from Fig. 6.

Now we illustrate the joint distribution of optical depth and cloud-top height. Figure 9 shows a joint distribution function; Fig. 10 shows an example of two conditional distributions F(h|τ) for τ = 3.5 ± 0.5 and τ = 10 ± 1. The conditional distributions of cloud-top height h are obviously different. Finally, Fig. 11 (for the realization of the cloud optical depth plotted in Fig. 6b) shows three realizations of cloud-top height. As we can see from Fig. 12, their pdfs match (approximately) the original pdf of cloud-top height from Fig. 4d.

7. Main steps of the model

Based on the above description, the software titled “Simulation of a two-component cloud field,” which generates realizations of cloud optical depth and cloud-top height from given observations, has been developed and is freely available for download from http://i3rc.gsfc.nasa.gov/Public_codes_clouds.htm (click on “PDF-based stochastic cloud model”).

Let us summarize here the main steps of the simulation procedure. There are only two input files: the cloud optical depth τ(x, y) and cloud geometrical thickness h(x, y) (as shown in Figs. 4c and 4d). The main 11 steps are the following:

  1. Read input file τ(x, y);

  2. Estimate cloud fraction Ac [see (7)];

  3. Find the truncation level d from (2);

  4. Estimate the covariance function of the indicator field KI [see (8)];

  5. Compute the correlation function K (see appendix A);

  6. Generate a Gaussian homogeneous random field υ(x, y) with mean zero and correlation function K (see appendix B);

  7. Simulate τ*(x, y), modifying the Gaussian field according to (12);

  8. Read input file h(x, y);

  9. Calculate joint distribution of τ and h fields (see Fig. 9);

  10. Calculate a conditional distribution matrix (13);

  11. Using the conditional distribution matrix, simulate the realization h*(x, y) that corresponds to realization τ*(x, y) generated at step 7 (see Fig. 11).

In the software, the first seven steps are accomplished by the executable file 0x-sp-a-s.exe. The input file is matrix τ(x, y). The output files are a realization τ*(x, y) of cloud optical depth field, the estimated covariance function of the indicator field, the computed autocorrelation function of the Gaussian field, and histograms of the input and output optical depth fields. The executable file DISTR-M2.exe estimates a joint distribution function of two random fields, τ(x, y) and h(x, y). The output files of this program are the joint and conditional distributions (steps 8–10). For the last step, the executable file X_Ysim.exe is used. It provides a realization h*(x, y) of the cloud-top height field. In that way, the realizations τ*(x, y) and h*(x, y) imitate the input fields τ(x, y) and h(x, y) by reproducing the covariance function of the indicator field and joint distribution of τ and h components for the two-dimensional vectors [τ(x, y), h(x, y)]. Note that here the random fields are assumed to be statistically homogeneous and isotropic.

8. Summary

Cloud stochastic models have proved to be an important tool to study 3D radiative effects in clouds, especially in broken cumulus clouds (e.g., Barker and Davies 1992; Evans 1993; Marshak et al. 1998; Várnai 2000; Evans and Wiscombe 2004; Schmidt et al. 2007). Here we have provided a theoretical description of a simple algorithm that generates realizations of the two correlated stochastic two-dimensional cloud fields that have statistical characteristics similar to given cloud fields. Each step of the algorithm has been illustrated with two broken cumulus cloud fields: cloud optical depth and cloud-top height retrieved from MODIS. Whereas most stochastic cloud models use Fourier filtering of a Gaussian signal to generate the required correlation (e.g., Schertzer and Lovejoy 1987; Evans 1993; Di Giuseppe and Tompkins 2003; Evans and Wiscombe 2004; Hogan and Kew 2005; Venema et al. 2006b), our algorithm is based on spectral models of homogeneous random fields (Prigarin 1995, 2001). A nonlinear transformation of Gaussian functions (the method of inverse distribution function) allows us to keep the distribution function similar to the one of the first original field. Realizations of the correlated second field are generated using a conditional distribution matrix.

This paper is accompanied by the software “Simulation of a two-component cloud field,” which has been recently released and can be freely downloaded from http://i3rc.gsfc.nasa.gov/Public_codes_clouds.htm. The software generates a two-component cloud field and provides programs to simulate two-dimensional distributions. The software contains a program (0x-sp-a-s) that generates realizations of a broken cloud field (X) with statistical characteristics (autocorrelation, density, and indicator functions) similar to the first given sample, a program (DISTR-M2) that estimates joint and conditional distributions for the two given samples, and a program (X_Ysim) that simulates a sample Y when a sample X is given. At present, the software runs only on Windows PCs but will be later extended to other platforms.

Finally we note that this model is a 2D stochastic model rather than 3D. To extend it to a 3D cloud model, we need to assume a vertical profile of cloud liquid water (see Di Giuseppe and Tompkins 2003). A simple linear increase of liquid water with height can be easily implemented in the frame of this model in its next stage.

Acknowledgments

Both S. Prigarin and A. Marshak have been supported by an NSF Collaboration in Basic Science and Engineering (COBASE) travel grant. S. Prigarin has also been supported by INTAS (Grant 05-1000008-8024), RFBR (Grant 06-05-64484), and the President’s program Leading Scientific Schools (Grant NSH-4774.2006.1). A. Marshak has been supported by the Department of Energy (under Grant DE-A105-90ER61069 to NASA’s GSFC) as part of the Atmospheric Radiation Measurement (ARM) program and by NASA’s Radiation Program Office (under Grants 621-30-86 and 622-42-57). We thank Dr. Tamas Várnai for his help with testing the software and Dr. Warren Wiscombe for stimulating the development of stochastic models of broken cloud fields. We are grateful to two anonymous reviewers and to Dr. Sebastian Schmidt for the excellent job they did reviewing the paper.

REFERENCES

  • Ackerman, A. S., O. B. Toon, and P. V. Hobbs, 1995: A model for particle microphysics, turbulent mixing, and radiative transfer in the stratocumulus-topped marine boundary layer and comparisons with measurements. J. Atmos. Sci., 52 , 12041236.

    • Search Google Scholar
    • Export Citation
  • Barker, H., and J. A. Davies, 1992: Solar radiative fluxes for stochastic, scale-invariant broken cloud fields. J. Atmos. Sci., 49 , 11151126.

    • Search Google Scholar
    • Export Citation
  • Cahalan, R. F., 1994: Bounded cascade clouds: Albedo and effective thickness. Nonlinear Processes Geophys., 1 , 156167.

  • Cahalan, R. F., and Coauthors, 2005: The I3RC: Bringing together the most advanced radiative transfer tools for cloudy atmospheres. Bull. Amer. Meteor. Soc., 86 , 12751293.

    • Search Google Scholar
    • Export Citation
  • Di Giuseppe, F., and A. M. Tompkins, 2003: Effect of spatial organization on solar radiative transfer in three-dimensional idealized stratocumulus cloud fields. J. Atmos. Sci., 60 , 17741794.

    • Search Google Scholar
    • Export Citation
  • Evans, K. F., 1993: A general solution for stochastic radiative transfer. Geophys. Res. Lett., 20 , 20752078.

  • Evans, K. F., and W. J. Wiscombe, 2004: An algorithm for generating stochastic cloud fields from radar profile statistics. Atmos. Res., 72 , 263289.

    • Search Google Scholar
    • Export Citation
  • Evans, K. F., A. Marshak, and T. Várnai, 2008: The potential for improved cloud optical depth retrievals from the multiple directions of MISR. J. Atmos. Sci., 65 , 31793196.

    • Search Google Scholar
    • Export Citation
  • Gentle, J. E., 2003: Random Number Generation and Monte Carlo Methods. 2nd ed. Springer, 381 pp.

  • Gikhman, I. I., and A. V. Skorokhod, 1977: An Introduction to the Theory of Random Processes. (in Russian). Nauka, 567 pp.

  • Hogan, R. J., and S. F. Kew, 2005: A 3D stochastic cloud model for investigating the radiative properties of inhomogeneous cirrus clouds. Quart. J. Roy. Meteor. Soc., 131 , 25852608.

    • Search Google Scholar
    • Export Citation
  • Kargin, B. A., and S. M. Prigarin, 1988: Modelling of stochastic fields of heap clouds and a study of their radiative properties with the Monte Carlo method (in Russian). Preprint 817, Siberian Division, USSR Academy of Sciences, Novosibirsk, 18 pp.

  • Marshak, A., A. Davis, R. F. Cahalan, and W. J. Wiscombe, 1994: Bounded cascade models as nonstationary multifractals. Phys. Rev. E, 49 , 5569.

    • Search Google Scholar
    • Export Citation
  • Marshak, A., A. Davis, W. J. Wiscombe, W. Ridgway, and R. F. Cahalan, 1998: Biases in shortwave column absorption in the presence of fractal clouds. J. Climate, 11 , 431446.

    • Search Google Scholar
    • Export Citation
  • Marshak, A., S. Platnick, T. Várnai, G. Wen, and R. F. Cahalan, 2006: Impact of 3D radiative effects on satellite retrievals of cloud droplet sizes. J. Geophys. Res., 111 .D09207, doi:10.1029/2005JD006686.

    • Search Google Scholar
    • Export Citation
  • Ogorodnikov, V. A., and S. M. Prigarin, 1996: Numerical Modelling of Random Processes and Fields: Algorithms and Application. s. VSP, 240 pp.

    • Search Google Scholar
    • Export Citation
  • Platnick, S., M. D. King, S. A. Ackerman, W. P. Menzel, B. A. Baum, J. C. Riedi, and R. A. Frey, 2003: The MODIS cloud products: Algorithms and examples from Terra. IEEE Trans. Geosci. Remote Sens., 41 , 459473.

    • Search Google Scholar
    • Export Citation
  • Prigarin, S. M., 1995: Spectral models of random fields and their applications. Adv. Modell. Analysis, 29A , 3951.

  • Prigarin, S. M., 2001: Spectral Models of Random Fields in Monte Carlo Methods. VSP, 198 pp.

  • Prigarin, S. M., and G. A. Titov, 1996: Spectral methods of numerical modeling of geophysical fields. Atmos. Oceanic Opt., 9 , 629635.

    • Search Google Scholar
    • Export Citation
  • Prigarin, S. M., and A. Marshak, 2005: Numerical model of broken clouds adapted to observations. Atmos. Oceanic Opt., 18 , 256263.

  • Prigarin, S. M., and A. Marshak, 2008: Simulation of vector semibinary homogeneous random fields and modeling of broken clouds. Numeric. Anal. Appl., 3 , 285292.

    • Search Google Scholar
    • Export Citation
  • Prigarin, S. M., B. A. Kargin, and U. G. Oppel, 1998: Random fields of broken clouds and their associated direct solar radiation, scattered transmission, and albedo. Pure Appl. Opt., 7A , 13891402.

    • Search Google Scholar
    • Export Citation
  • Prigarin, S. M., A. Martin, and G. Winkler, 2004: Numerical models of binary random fields on the basis of thresholds of Gaussian functions. Siberian J. Numer. Math., 7 , 165175.

    • Search Google Scholar
    • Export Citation
  • Scheirer, R., and S. Schmidt, 2005: CLABAUTAIR: A new algorithm for retrieving three-dimensional cloud structure from airborne microphysical measurements. Atmos. Chem. Phys., 5 , 23332340.

    • Search Google Scholar
    • Export Citation
  • Schertzer, D., and S. Lovejoy, 1987: Physical modeling and analysis of rain and clouds by anisotropic scaling multiplicative processes. J. Geophys. Res., 92 , 96939714.

    • Search Google Scholar
    • Export Citation
  • Schmidt, K. S., V. Venema, F. Di Giuseppe, R. Scheirer, M. Wendisch, and P. Pilewski, 2007: Reproducing cloud microphysics and irradiance measurements using three 3D cloud generators. Quart. J. Roy. Meteor. Soc., 133 , 765780.

    • Search Google Scholar
    • Export Citation
  • Sveshnikov, A. A., 1968: Applied Methods in Random Function Theory. (in Russian). Nauka, 463 pp.

  • Várnai, T., 2000: Influence of three-dimensional radiative effects on the spatial distribution of shortwave cloud reflection. J. Atmos. Sci., 57 , 216229.

    • Search Google Scholar
    • Export Citation
  • Venema, V., S. Bachner, H. W. Rust, and C. Simmer, 2006a: Statistical characteristics of surrogate data based on geophysical measurements. Nonlinear Processes Geophys., 13 , 449466.

    • Search Google Scholar
    • Export Citation
  • Venema, V., and Coauthors, 2006b: Surrogate cloud fields generated with the iterative amplitude adapted Fourier transform algorithm. Tellus, 58A , 104120.

    • Search Google Scholar
    • Export Citation
  • Voss, R., 1985: Random fractal forgeries. Fundamental Algorithms in Computer Graphics, R. A. Earnshaw, Ed., Springer-Verlag, 805–835.

  • Wen, G., A. Marshak, R. F. Cahalan, L. A. Remer, and R. G. Kleidman, 2007: 3-D aerosol–cloud radiative interaction observed in collocated MODIS and ASTER images of cumulus cloud fields. J. Geophys. Res., 112 .D13204, doi:10.1029/2006JD008267.

    • Search Google Scholar
    • Export Citation
  • Yamaguchi, Y., A. B. Kahle, H. Tsu, T. Kawakami, and M. Pniel, 1998: Overview of Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER). IEEE Trans. Geosci. Remote Sens., 36 , 10621071.

    • Search Google Scholar
    • Export Citation
  • Zuev, V. E., and G. A. Titov, 1995: Radiative transfer in cloud fields with random geometry. J. Atmos. Sci., 52 , 176190.

APPENDIX A

Relations for Covariance Functions of a Gaussian Random Field and Its Indicators

Assume that υ(x, y) is a homogeneous Gaussian random field on the plane with mean zero and correlation function K(x, y) = E[υ(x, y) υ(0, 0)]. Let us consider two indicator fields with respect to a fixed level d:
i1520-0469-66-1-92-eqa1
These indicators correspond to model A (1a) and model B (1b) introduced in section 2. In this appendix we present the basic relations between covariance functions of the random field υ(x, y) and its indicators (for details, see Prigarin et al. 2004). For the covariance functions KI(n)(x, y) = E[I(n)(x, y) I(n)(0, 0)] = P[I(n)(x, y) = 1, I(n)(0, 0) = 1] we have
i1520-0469-66-1-92-ea1
where
i1520-0469-66-1-92-ea2
and
i1520-0469-66-1-92-ea3
is the probability density of a two-dimensional Gaussian random vector with zero mean, unit variance of the components, and a correlation coefficient r between the components.
To find the correlation function K(x, y) of the Gaussian random field for a quasi-Gaussian model of broken clouds, it is necessary to estimate function KI(n)(x, y)—i.e., the covariance function of the cloud indicator field—and to solve numerically Eq. (A1) (n = 1 for model A and n = 2 for model B). For computations it is reasonable to use the following representations of (A2) in terms of Owen’s function:
i1520-0469-66-1-92-ea4
where Φ is the standard normal cumulative distribution function, a = (1 − r)/(1 + r), and
i1520-0469-66-1-92-ea5
is Owen’s function.

APPENDIX B

On the Simulation of a Gaussian Homogeneous Isotropic Field with a Given Correlation Function

Here we briefly present a numerical method used in the software “Simulation of a two-component cloud field” for modeling Gaussian random fields on a plane [for details, see chapter 2 in Ogorodnikov and Prigarin (1996); chapter 1, and particularly section 1.1.4, in Prigarin (2001); and Prigarin and Titov (1996)]. To simulate a Gaussian homogeneous isotropic random field υ(x, y) with mean zero and correlation function K(x, y), we use an approximation of the following type:
i1520-0469-66-1-92-eqb1
where
i1520-0469-66-1-92-eqb2
αjm, βjm, and γj are independent random variables uniformly distributed in (0, 1), and the same symbol K is used for the correlation function (of the isotropic field) depending on a point on the plane (x, y) and on the distance r = x2 + y2: K(r) = K(x, y). Function z(ρ) is the radial spectral density [see (6) in section 3]. Such numerical models are called spectral models because they approximate stochastic integrals in the spectral decomposition of the random field
i1520-0469-66-1-92-eqb3
by finite sums. Here ξ(dλdν) and η(dλdν) are the orthogonal stochastic measures. [For the details on the spectral decompositions see, e.g., Gikhman and Skorokhod (1977, p. 273).] The spectral model is a sum of J × M random harmonics and it depends on three parameters J, M, and B, where B is an upper boundary of the radial spectrum of the model. (In the accompanying software “Simulation of a two-component cloud field,” parameters J and M are chosen manually, whereas B is specified automatically). Additional information on the construction, properties, errors, and convergence of spectral models can be found in Prigarin (2001).

Fig. 1.
Fig. 1.

A schematic illustration of models A and B; H0 is a constant cloud base and d is a cutting threshold level.

Citation: Journal of the Atmospheric Sciences 66, 1; 10.1175/2008JAS2699.1

Fig. 2.
Fig. 2.

Number of clouds per unit area mc (k = 1), as a function of cloud fraction Ac for models A and B.

Citation: Journal of the Atmospheric Sciences 66, 1; 10.1175/2008JAS2699.1

Fig. 3.
Fig. 3.

Configuration of cloud fields (300 km × 300 km) for models (left) A and (right) B on the basis of a Gaussian random field with a correlation function J0 for the same cloud fraction Ac = 0.58 (ρ = 0.5; d = −0.20 for model A and 0.55 for model B). To simulate a Gaussian random field, a spectral model from Prigarin (2001, section 1.1.4) was used; see also appendix B.

Citation: Journal of the Atmospheric Sciences 66, 1; 10.1175/2008JAS2699.1

Fig. 4.
Fig. 4.

A 68 km × 68 km region in Brazil centered at 17°S and 42°W collected on 9 Aug 2001 at 1015 LT. The solar zenith angle θ0 = 41°; the solar azimuth angle φ0 = 23° (from the top). (a) MODIS true color RGB (red–green–blue) 1-km resolution; (b) ASTER RGB 15-m resolution; (c) retrieved cloud optical thickness; (d) retrieved cloud-top height (in km).

Citation: Journal of the Atmospheric Sciences 66, 1; 10.1175/2008JAS2699.1

Fig. 5.
Fig. 5.

Indicator functions I(x, y) of a cloud field: white is cloud (I = 1) and black is a cloud-free area (I = 0). Cloud fraction Ac = 0.4. (left) A 68 km × 68 km MODIS image centered at (17.1°S, 42.16°W) acquired on 9 Aug 2001. (right) A realization of a simulated field.

Citation: Journal of the Atmospheric Sciences 66, 1; 10.1175/2008JAS2699.1

Fig. 6.
Fig. 6.

Four realizations of cloud optical depth; all of them have the same covariance function of the cloud mask KI(x, y) and histogram g(τ) as the one in Fig. 4c. Color scale is as in Fig. 4c. The size of the images is 68 km × 68 km, as in Fig. 5.

Citation: Journal of the Atmospheric Sciences 66, 1; 10.1175/2008JAS2699.1

Fig. 7.
Fig. 7.

Estimates of the covariance function KI(x, y) = KI(r) (r = ) of the indicator function I(x, y) for the observed (solid line) and simulated (dotted line) cloud fields from Fig. 5.

Citation: Journal of the Atmospheric Sciences 66, 1; 10.1175/2008JAS2699.1

Fig. 8.
Fig. 8.

The original histogram g(τ) and four other histograms that correspond to four realizations of cloud optical depth shown in Fig. 6.

Citation: Journal of the Atmospheric Sciences 66, 1; 10.1175/2008JAS2699.1

Fig. 9.
Fig. 9.

Joint distribution of the given cloud optical depth and cloud-top height fields.

Citation: Journal of the Atmospheric Sciences 66, 1; 10.1175/2008JAS2699.1

Fig. 10.
Fig. 10.

The conditional distribution function F(h|τ) used to simulate cloud-top height h; F(h|τ) is shown for two values of optical depth τ.

Citation: Journal of the Atmospheric Sciences 66, 1; 10.1175/2008JAS2699.1

Fig. 11.
Fig. 11.

(a) A realization of cloud optical depth from Fig. 6b. (b)–(d) Three additional realizations of cloud-top height distribution; they correspond to the cloud optical depth field shown in (a). All realizations have the same conditional distribution F(h|τ).

Citation: Journal of the Atmospheric Sciences 66, 1; 10.1175/2008JAS2699.1

Fig. 12.
Fig. 12.

A pdf of three realizations of cloud-top height shown in Fig. 11. The original pdf of cloud-top height from Fig. 4d is also shown.

Citation: Journal of the Atmospheric Sciences 66, 1; 10.1175/2008JAS2699.1

Save