Download as pdf
Download as pdf
You are on page 1of 17
1526 JOURNAL OF APPLIED METEOROLOGY Voume 33 Statistical Analysis of Mesoscale Rainfall: Dependence of a Random Cascade Generator on Large-Scale Forcing THOMAS M. OVER Center forthe Study of Earth from Space/CIRES and Geophysics Program, Univesity of Colorado, Boulder, Colorado ‘way K. GUPTA Center forthe Study of Earth from Space /CIRES and Department of Geological Sciences, University of Colorado, Boulder, Colorado (Manuscript received 16 August 1993, in final form 30 April 1994) ‘ABSTRACT Under the theory of independent and identically distributed random cascades, the probability distribution of the cascade generator determines the spatial and the ensemble properties of spatial rainfall. Three sts of, ‘adar-erive rainfll data in space and time are analyze to estimate the probability distribution ofthe generator. ‘detailed comparison between instantaneous scans of spatial rainfall and simulated cascades using te sealing properties of the marginal moments is carried out. This comparison highlights important similarities end dif- Ferences between the data and the random cascade theory. Dferences are quantified and measured for the three datasets. Evidence is presented to show tht the scaling properties ofthe rainfall can be captured to the frst order by a random cascade with a single parameter. The dependence of this parameter on forcing by the large-scale meteorological conditions, as measured by the large-scale spatial average rain rate, is investigated for ‘these three datasets. The data show that this dependence can be captured by a one-to-one function. ince the large-sale average rain rate can be diagnosed from the large-scale dynamics, this relationship demonstrates an ‘important linkage between the large-scale atmospheric dynamics andthe statistical cascade theory of mesoscale ‘ainfal, Potential application of this research to parameterization of runoff from the land surface and regional flood frequency analysis i briely discussed, and open problems for further research are presented. 1. Introduction Applications of recent mathematical developments in the theory of random cascades to mesoscale rainfall have underscored at least three important problems (Gupta and Waymire 1993), The first problem con- cerns estimation of the probability distribution of the cascade generator from an analysis of instantaneous radar scans of spatial rainfall. This problem arises be- to the random cascade theory, the spatial rainfall averages cannot be equated to the en- semble averages (Holley and Waymire 1992). More- over, an important objective of this paper is to inves- tigate how the structure of rainfall at the mesoscale varies with time according to large-scale dynamical forcing. Therefore, computing ensemble averages by averaging spatial data in the time domain is not ap- propriate here. We address the problem of estimating the cascade generator in a scan by scan analysis of space-time radar rainfall data at the mesoscale by computing quantities whose scaling properties converge to nonrandom constants in the high-resolution limit fora single scan of data. The second problem concerns Corresponding author address: Vijay K. Gupta, CSES/CIRES, ‘Campus Box 216, University of Colorado, Boulder, CO 80309. © 1994 American Meteorological Society the development of the dynamical basis of the random cascade theory. The generic problem of establishing dynamical interpretations of parameters in statistical theories has long been recognized to be extremely im- portant. Unfortunately little progress has been made ‘on this front in the last decade. As a first step in this, direction, in this paper we present an empirical rela- tionship between the random cascade parameters and the forcing due to the large-scale meteorological con- ditions. Throughout this paper, large-scale forcing will be synonymous with the large-scale average rain rate as measured by radar. However, because the large-scale average rain rate can be diagnosed from large-scale dy- ‘namics, this relationship constitutes a real linkage be- ‘tween the physics and the statistics of rainfall. The third problem is to extend the spatial theory of random cas- ccades to incorporate space-time evolution under large- scale forcing. Even though an answer to the third prob- lem awaits future work, we expect that the investiga tions reported in this article will guide future research. on this problem. A hierarchical spatial structure of rainfall intensity in which clusters of high-intensity rainfall cells are imbedded within regions of tow rainfall intensity, which in tum are clustered within still larger regions of even lower intensity, is widely observed in mesoscale rainfall. Decenaer 1994 ‘Some of the early observational evidence for this kind of structure came from Austin and Houze (1972), who observed hierarchical structure within storms of a va- riety of synoptic types occurring in New England. In the Global Atmospheric Research Program (GARP) Atlantic Tropical Experiment (GATE), precipitation was observed to follow a life cycle that at its peak in- cluded clusters of convective cells with associated me- soscale anvils of stratiform rainfall (Houze and Betts 1981). Groups of thunderstorms imbedded within a mesoscale system, termed mesoscale convective com- plexes (MCSs), were shown to be major contributors, of summertime rainfall over the central United States by Maddox (1980). A recent review of the observa- tional evidence for such systems and their effects is given by Houze (1989). Attempts to model this hierarchical spatial structure hhave taken one of two general approaches. The first set of approaches, largely developed within the mete- orology literature, involve the numerical solution of a set of equations describing the dynamics and ther- ‘modynamics of the atmosphere. When the model res- olution is sufficiently fine to resolve the convective scalle, the proper mesoscale structures can be simulated and predicted over short timescales without parame- terization of convection (Lilly 1990; Molinari and Du- deck 1992). However, in many applications, most no- tably in general circulation models (GCMS), predic- tions are needed over long timescales, and it is not possible to obtain a fine spatial resolution. Therefore, the problem of convective parameterization arises. If clouds are modeled in a convective parameterization scheme, they are generally modeled as a statistically independent set of individuals without interactions (Arakawa and Schubert 1974). We are not aware of any convective parameterization schemes that include the effects of mesoscale clustering on convection, al- though there have been diagnostic analyses that have taken it into account (Johnson 1984; Houze 1989). The second set of approaches involves the use of stochastic methods to describe the hierarchical spatial organization in the rainfall intensity at the mesoscale. These have been mostly developed within the hydrol- ogy literature. Work on the use of the mathematical theory of cluster point processes to describe this type of rainfall variability began with LeCam (1961). The point process models, for example, the WGR model (Waymire et al. 1984), have provided useful insights into the connections between the empirically observed timescale of the breakdown of Taylor's hypothesis of uid turbulence (Zawadzki 1973; Crane 1990)and the space-time organization of mesoscale rainfall. The point process models still continue to be refined and {ested on different datasets (Phelan and Goodall 1990; Valdes et al. 1990). However, a major difficulty with the point process approach has been its parameteriza- tion, because each spatial scale is treated separately and thus has its own set of parameters. Overall, the OVER AND GUPTA 1527 large number of parameters in these models, say up to 10, immensely complicates the problems of parameter estimation and physical interpretation. ‘The problem of parameterizing each scale separately ‘can be circumvented by entertaining some kind of @ scaling invariance hypothesis in the statistical structure of mesoscale rainfall. Scaling invariance means that there are no preferred scales in the mesoscale range at which particular “structures” occur. Instead, structures remain similar to each other across scales, and this similarity can be described mathematically by scaling invariance. The general question of the presence or the absence of characteristic scales of atmospheric phe- nomena at the mesoscale is discussed by Emanuel (1986). He concludes that for the most part the me- soscale lacks preferred scales. The carly pioneering ‘work in application of scaling theories to mesoscale rainfall is by Lovejoy (1981), who employed the fractal formalism to describe the scaling behavior of the ge- ‘ometry of rainy regions. Current efforts to describe rain rates as well as the geometry of rainy and nonrainy regions at the mesoscale are based on the theory of random cascades and multfractal spatial mass distri- butions (Gupta and Waymire 1993; Lovejoy and Scherizer 1990). The idea behind the random cascade construct, as a conceptual basis of turbulence theory, is that a cascade of turbulent eddies transfers kinetic energy from a large energy-input scale to a small dissipation scale. This conceptual framework is expressed in Richardson's fa- mous ditty: “big whirls have little whirls that feed on their velocity, and little whirls have lesser whirls and so on to viscosity—in the molecular sense” (Richard- son 1965, p. 66; first published in 1922). It also un- derlies the “two-thirds law” relating the expectation of the square of velocity differences at two points in ho- mogeneous and isotropic turbulence to the energy dis- sipation rate and the separation distance (Kolmogorov 1941), The remark by Landau (see, e.g., Kolmogorov 1962) that the energy dissipation would be subject to random fluctuations, thus introducing a correction to the two-thirds law, spurred the beginning of the theory of intermittent turbulence. A primary objective of re- search in intermittent turbulence has been to find sta- tistical models ofthe energy dissipation field that match experimental measurements and dynamical consid- erations. Toward that end, Kolmogorov (1962) and Oboukhiov (1962) hypothesized that the energy dissi- pation was marginally lognormally distributed, and Yaglom (1966) presented an apparent proof of this hypothesis for a spatial statistical model that is essen- tially a three-dimensional discrete random cascade of the type used in this paper. However, as shown by Mandelbrot ( 1974), the mathematics of such cascades is rather subtle. In fact a variety of marginal distribu- tional behaviors may be obtained, ranging from those that are nearly lognormal to those vastly different from lognormal. Mandelbrot (1974) introduced a variant of | 1528 JOURNAL ‘Yaglom’s model, which he termed a “canonical” cas- cade, in which energy dissipation is conserved only on average. By contrast, in “microcanonical” cascades, energy is conserved pathwise. As Mandelbrot points out, complicated and probably unphysical dependen- cies are required between the multipliers to obtain pathwise energy conservation in a random cascade, ‘whereas the canonical cascade is obtained with simple independent and identically distributed (i.i.) multi- pliers, Recent mathematical developments in the the- ory of random cascades are given in Holley and Way- mire (1992). The present context of the spatial properties of rain- fall fields can be distinguished from the turbulence the- ory context in at least two respects. First, itis not known, at present what constraints should be put on rainfall ‘models due to physical considerations. Unlike turbu- lent kinetic energy in the inertial range of fluid tur- bulence, it is not clear that conservation of rainfall needs to hold. Nor is it known what, if anything, plays the role of velocity in the rainfall context. These con- nections between rainfall dynamics and its statistical behavior remain one of the most critical areas of re- search, The application of random cascade theory is, based simply on two empirical observations: first, that rainfall fields appear to have scaling structure that can be described by random cascades, and second, rainfall fields are clustered in space. As a second point of con- ‘rast between turbulence and rainfall, we show in this, article that no universal model with fixed parameter values can describe rainfall as it varies through time, Rather, even if a single model form is appropriate, its parameter values will vary as a function of the large- scale forcing. This paper is divided into five sections. In section 2 following this introduction an empirical relationship between the large-scale average rain rate and a cascade parameter p is given. This relationship describes how the statistical spatial properties of the rainfall fields vary with large-scale dynamics. Section 2 is meant to be accessible to readers who lack an in-depth understand- ing of and familiarity with the random cascade theory. Section 3 contains a brief review of some terminology and results from the random cascade theory. We have chosen to include only those results that are applied to spatial rainfall in this paper. These results focus on the issue of estimating the parameters of a random cascade, especially the parameter p, and demonstrate precisely its role in spatial modeling using random cas- cades. In section 4, the estimation techniques described in section 3 are applied to three sets of radar-derived instantaneous spatial rainfall data, including a detailed comparison of two scans of rainfall with two corre- sponding cascade simulations. Even though the ran- dom cascade generator analyzed here docs not exhibit completely the behavior of the scans of rainfall data, it shows that, as a first-order approximation, the rainfall fields can be described by a particularly simple gen- OF APPLIED METEOROLOGY Vouume 33 ‘erator that is completely determined by the parameter P. This finding justifies the use of p in section 2 to characterize the spatial properties of the rainfall fields ‘The paper concludes with a summary, a brief discussion cof potential areas of application of the results presented here, and a description of some open problems for fur- ther research. 2. Relationship between spatial r large-scale forcing fall patterns and In this section we show how the spatial patterns of a set of radar-derived mesoscale rainfall scans depend (on the large-scale forcing, As mentioned previously, the spatial patterns and large-scale forcing are each measured by a single parameter. We use the radar- 1, b can be written b = Th; &,, where b, is the number of subdivisions in each step in the ith dimension, Here we will assume that all the b; are equal, that is, b, = b'4, i= 1, 2, +, d. The ith subcube after n subdivisions is denoted by 1. The sidelength of Aj at level m represents the spatial scale and is denoted by Ly. A dimensionless spatial-scale parameter , is given by GB.) “Mass” is distributed on the cube as follows. First the cube is assigned a nonrandom initial mass RoLé. The subcubes 4 after the fist subdivision are assigned the masses RoLAW'1/b, i= 1,2, ++, b. We assume that the H’s, i= 1, 2, =++, b” are nonnegative and mutually independent fandom variables. Their com- ‘mon probability distribution is the same as that of IV, which is called the cascade generator. Because the gen- erator random variables 1 are i.id., the cascade they generate can be called an iid, random cascade. Ran- dom cascades whose generators are not i.i.d, will be briefly mentioned at the end of this paper, but the work presented in the body of the paper considers only ran- dom cascades with i.d. generators As the cascade proceeds downward in scale, let the mass associated with a subdivision A, at the nth level be denoted by (4). This mass isthe product of the initial mass and all the intervening JV i along the path tothe subcube 4, divided by the number of subcubes, JOURNAL OF APPLIED METEOROLOGY Vouume 33 Goneraion number Spatial seaie and Waymire 1993) 'b” at this scale. Mathematically this can be expressed, ma(A4) = Roto" T] Wj (2) ‘The limiting mass 2. is obtained by letting n > co It satisfies an important recursion equation given by Gupta and Waymire (1993, p. 255), Hoo Sn) = br AW\Zocli), i= 12, + (33) ‘where Z..(i) is statistically independent of y,(\,). The Z.,(i) are independent and identically distributed as Hz. (0, Lo]*) Role a, Zell G4) for all i. Equation (3.4) can be derived using (3.3) and the definition Ho( Ao) = Ho([0, Lol) = RoLd We will use Z,, to designate any random variable with this distribution, Equation (3.3) says that the cascade mass 1,, at any given scale , consists of a low-fre- quency, large-scale component 4», and a high-fre- Guency, small-scale component Z,.. Under certain ‘mathematical conditions it can be shown that Z,,, exists and is nondegenerate; see Holley and Waymire (1992) for details. The most important of these conditions is that the mathematical expectation or the ensemble mean E(W] = 1, which we assume is true throughout Decenuer 1994 this presentation, An important physical interpretation of this condition is that the mass on the average is conserved as it cascades down to smaller spatial scales, However, this condition does not guarantee the co servation of mass for each realization of the cascade. As discussed in the introduction, cascades having con- servation of mass on average are called canonical cas- cades, while those with pathwise conservation are called ‘microcanonical cascades. The assumption E{W] ~ | in conjunction with (3.2) and (3.3) also implies that Elda] = 1. In applying the theory of random cascades to rainfall, we take d = 2 and b = 4 as in Fig. 4. Here, Lo is the size of the spatial domain under consideration, for ex- ample, for the middle 64? 4-km pixels of the GATE region, Lo = 256 km. Here Ry is the expected average rain rate over the domain and thus has units of depth per time. So that Z,, is dimensionless, ,, has units of volume per time by (3.4). Thus u.,.( Aj,) represents the volume of water falling on the ith box of size L, in a unit time. The total mass 4..,([0, Lo]*) thus represents, the volume of water falling on the whole domai a unit time. Therefore, in view of (3.4), the distribution of the spatial average rain rate R is given by oo({0, Lol”) Li = RL Gs) By (3.5), since E[Z.] = 1, E{R] = Ro. This result verifies the interpretation of Ry as the expected average rain rate. 2) SCALING OF THE MARGINAL SPATIAL, MOMENTS: THE 7(q) FUNCTION For the purpose of estimating the probability distri- bution of the cascade generator HV from data, we now define the scaling of the marginal spatial moments of a multifractal spatial mass distribution. This definition also applies to spatial mass distributions generated by a random cascade. The marginal spatial moments of Hee are defined as Mg) = & wi (An), 20 0. In view of (3.1), the scaling of the moments is defined as, a tim EMaL@ — joy HOBMat) wa [Oghe mace (nt logb)/d? eo a) = 1, and if aah <>. (38) then with probability | 1(q) = dXo(q), (3.9) where Xolg) = 1owEW=(9= 1). B.10) Furthermore, fim = BMAD) = tim Med) ¥(q), (3.11) mow (RoLA)B™E(Z 4] where ¥(q) is a random variable for each q The presence of the random variable Y(g)in (3.11) shows that the spatial moments M,(q) fail to converge to their expectation in the high-resolution limit, This demonstrates the nonergodicity of the measures gen- erated by random cascades. Despite this, (3.9) and (G.10) show that the r(q) function, computed from a single scan of data in the limit of high resolution, can be used to infer the probability distribution of W. Note also that since X,(q) is a convex function of (Holley and Waymire 1992, p. 832), it follows from (3.9) that 1(q) is also convex. ‘A method for estimating 1(q) = dXs(q) from data with finite resolution can be also obtained from (3.11). ‘Using (3.1), for large but finite n, (3.11) can be writen TogMy(4) = logY (a) + @ log(RoL 4) + logE[Z%} ~ (4) logrs. (3.12) ‘Thus by linear regression of log Mf,(4) versus —log,, wwe can obtain an estimate #(q) of 7(q) from the regres- sion slope and an estimate of a realization of the ran- dom quantity log (a) + @ log(RoL$) + log£1Z 4] from the regression intercept. The log-log linearity of log.M,(q) versus —log), also provides atest of the scal- ing hypothesis for the moment of order g. 3) CROSS MOMENTS In many applications of random measures—for ex- ample, the statistical theory of turbulence—the spatial correlation or cross-moment properties are very im- portant, In turbulence theory, two-point cross-moment properties have been computed by a number of authors 1534 for mulifractal or cascadelike constructions using heuristic arguments (Yaglom 1966; Cates and Deutsch 1987; Siebesma and Pietronero 1988; Lee and Halsey 1990; Meneveau and Chhabra 1990). For the random cascades discussed here, we have performed a rigorous calculation of the two-point cross moment of arbitrary order for a one-dimensional random cascade. For the sake of brevity, we will not present the result in detail, The basic idea is that the two-point cross moment is approximately a power law in the separation distance, with an exponent that depends in a simple way on the moments of the generator W¥. Since x(q) is also a function of the moments of W’, the cross- ‘moment exponent can also be written in terms of 7(q). Comparing the empirical value of the cross-moment exponent with a prediction of its value using the em- pirical 7(q) thus provides a consistency check on the hypothesis that a measure comes from an j.i,d. random cascade. b. Special cases of random cascades arising in the ‘modeling of spatial rain rates 1) CASCADE GENERATORS WITH AN ATOM AT ZERO A very important feature of the cascade theory is that it can model regions of zero rain rate. This is ac- complished by allowing the generator W to have an. atom at zero with probability p; that is, fie, where 1V'* is the positive part of 1”. When Wis given by (3.13), (3.9) and (3.10) provide a method for es- timating the parameter p. It should first be noted that W given by (3.13) fails the condition Pr(W’ > a) for some a > 0 required in Theorem 3.1. However, while it has not been proven, Theorem 3.1 is believed to st hold in this case (E. C. Waymire 1993, personal com- munication), and simulations such as those shown here support this view. To estimate p, take g = 0 in (3.10) to obtain xs(0) = logs(1 — p) + 1. Substituting this into (3.9) yields 7(0)/d = logs( 1 — p) + 1. Solving for paives p= | ~ 6 "4, which suggests the estimate Bat - pion G14) with probability p with probability 1 w 3.13) An alternate but equivalent form of the estimator ‘can be obtained by considering the scaling of the frac- tional rainy area f(X,) = b-"M,(0). Now, setting ¢ = 0 in (3.12) gives log f(n) ~ log (0) + logPr(Z.. > 0) ~ dlogs(1 ~ p) log. (3.15) d loge(1 — p) for p gives the Solving the slope s alternate form JOURNAL OF APPLIED METEOROLOGY Vou 33 B ane (3.16) where sis the estimate of s obtained by linear regression of log /(,) versus log). 2) CASCADE GENERATORS COMPLETELY SPECIFIED BY p A cascade generator with an atom at zero that is completely specified by the parameter p can be con- structed by using the E[W’] = | criterion to assign a single value to the positive part of W; that is, ft a-py, This model was designated in the turbulence literature as the 6 model by Frisch et al. (1978) Since the distribution of Win this case is completely determined by p, it is easy to see from (3.9) and (3.10) that the corresponding 7(q) is also parameterized by p in a simple way. In particular, with probability p 3.17. with probability 1 — p) F! 1a) d From (3.18) itis casy to see that 7'(q) in this case is a constant and is given by 1 = g)flogy(1 ~p) + 1]. (3.18) r@) 7 ~tlogs( ~ p) + 1 (G19) In fact, as claimed by the following theorem, this property is unique to generators of the form (3.17) This fact is important because, as will be shown in section 4, the empirical r(q) functions of the scans of rainfall considered here are nearly straight lines. THEOREM 3.2. If the generator W of a random cas- cade has an atom at zero and satisfies the appropriate conditions so that (3.9) holds, then x(q) is a straight line if and only if *=(-py, (3.20) where Lapa perenn G21) A proof of Theorem 3.2 is given in appendix B. Since (3.21) relates the constant derivative 7’(q) of the 7(q) function to the value ofp, it provides an ad- ditional means of estimating p. In particular, if we de- fine an estimate of the slope /(q) by # (qq: = OD | 3.22) aa ee 4) 0), where ¥(0) approximates the random variable ¥ (0) given by (3.11). Here Pr(Z.. > 0) depends on p and can be computed from a formula given in Gupta and Waymire ( 1993, p. 262). This formula with b = 4 gives Pr(Zq, > 0) = 0.9980 for p = 0.21 and Pr(Z., > 0) = 0.983 for p = 0.35. Thus logPr(Z.., > 0) is negative bbut small in magnitude (0.0020 and ~0.017, respec- tively). Since fy) = 6-"M,(0), using (3.11), ¥(0) can also be written = tim Lode) YO)~ BP EVO Thus when f(\,) = EL/(x)], log¥ (0) = 0, so random fluctuations should appear above and below this value. (4) JOURNAL OF APPLIED METEOROLOGY scan 1201. (b) The A-model simulation with ‘The solid line was computed by weight least-squares repression Vouume 33 } feteescepenerenyeerseeny a A oer.) (8) eo = 021. (c) GATE This is what we see for the simulations, where in Fig. 6b evidently log { f(Ay)/ELf(x)]} < 0, while in Fig, 6d, log{f(X,)/EL/(X,)]} > 0. In other words, the simulation with p = 0.21 has an unusually small frac- tional wetted area, whereas the p = 0.35 simulation has an unusually large fractional wetted area. By this measure, both the scans of rainfall have un- usually large fractional wetted areas. For a sample of just two scans, this would by no means be statistically significant. However, it turns out that this is typically true of the scans in the datasets considered here. Gupta and Waymire (1993) suggest that this may be the result of a “break” in scaling at about 100 km, meaning that, above this scale the random cascade theory may not apply. We are examining this issue further for other possible explanations. ‘The scaling of the marginal moments Mz(q) for q = 0,05, 1, «++, 4 is shown in Figs. 7a-d. The log- log linearity is reasonably good at small scales for both the data and the simulations, For GATE-II scan 1201 (Fig. 7a), deviations from log-fog linearity even at large scales are small, though not as small as for the corre- Decemmen 1994 = © OVER AND GUPTA wea, Fic, 7. Sealing ofthe marginal moments with g = 010,015, 1.0, = (b) The @-model simulation with computed by weight least-squares regression, sponding simulation (Fig. 7b). GATE-I scan 176 has larger deviations (Fig. 7c), especially for larger values of g. Part of this increase in deviations is due to sam- pling variability, which grows with p. This is because a large p produces a large number of zeroes at each scale, and therefore fewer boxes contribute to the sums. ‘The 7(q) functions arising from the scaling of the marginal moments are pictured in Figs. 8a-d. GATE- It scan 1201 shows pronounced convexity near q = 2; it was chosen to illustrate this feature of some of the data. As discussed below in subsection 4b, convexity in the 7(q) is common for rainfall seans with p < 0.3. When p > 0.3, the r(q) curves tend to be straight lines. GATE-I scan 176, with # = 0.35, is typical in this re- gard. The 7(q) functions of the simulations are quite straight, as they should be according to Theorem 3.2. We note in passing that a slight concavity appears in the r(q) functions at q = 0.5 in the two scans of data. This is often observed in the data. We are exploring an explanation of this feature 4.0. from the bottom ofthe plots up. (a) GATE phase I sean 1201 021. (c) GATE phase I scan 176. (d) The s-model simulation with p = 0.38. The solid line was b. Results of application to the complete dataset The results of an analysis of # versus R for the com- plete dataset were presented in section 2. Here we con- sider to what extent the predictions of the 6 model deviate from the observed scans for the entire dataset. Insubsection a above, three such deviations were iden- tified: 1 ) the intercept of the log-log plots of fractional wetted area versus scale is typically positive, whereas, according to theory (Eq. (3.15)], the intercept should be slightly negative on average; 2) for p < 0.3, 1(q) curves tend to be convex, whereas for -model cascades, the 7(q) curves are straight lines; 3) there is often a slight concavity in the empirical 7(q) curves near q 5. In this subsection, deviation 2 will be considered and the remainder will be left as open problems for future research, To investigate the extent of the deviations of the empirical 7(q) functions from straight-line behavior, we created plots of (0:4) versus f for all three of the JOURNAL OF APPLIED METEOROLOGY Vouume 33 ) 4 — —— “| | a Fic. 8. Estimated #(g) curves (a) GATE phase II scan 1201. (b) The f-model simulation with p = 0.21. (€) GATE phase I scan 176. (4) The B-mode! simulation with p= 0.35. datasets, They are presented in Fig. 9. It is clear from these plots that (0:4) versus fis for the most part evenly scattered about the line (0:4) = 5. However, for the smaller values of f, the points tend to lie above the line (0:4) = 6. To interpret these results, we frst note that by (3.23), (0:4) > p implies that #(q) is striclly convex over the range 0 0. However, application of the cross-mo- ment consistency check briefly described in section 3 suggests that this alternative does not adequately ex- plain the rainfall data. The second alternative is that a rainfall field can be modeled by a dependent mixture of two random cas- cades, both having generators of the form (3.17) but with different values of p. Dependency is introduced by allowing the generator W, at a location at level n to depend on the generator W,.. at level n ~ 1. Such cascade construction is also suggested by the two types of rain, stratiform and convective, that are commonly observed. However, the mathematical framework of DECEMBER 1994 (@ 08 os| 0:4) 9. Plot of (0-4), estimated from slope of #(g) versus A, ‘estimated from sealing of fractional rainy area. (a) GATE phase (b) GATE phase I, (¢) Elbow radar, 6 July 1991, The sold curve isthe line (0:4) = p the iid. cascades given in Holley and Waymire (1992), and applied here to rainfall, cannot accommodate de- pendent mixtures of two or more random cascades. ‘The mathematical theory of such mixtures is currently being developed, and some basic results are given in ‘Waymire and Williams (1994). 5. Summary, potential applications, and open problems a. Summary ‘The above results can be summarized as follows. We described a random cascade model for rain rates in two dimensions. The model depends on three pa- rameters: 1) Lo, the size of the spatial domain under consideration; 2) Ro, the expected average rain rate OVER AND GUPTA 1539 over the spatial domain; and 3) W’, a random variable that multiplicatively “generates” the cascade mass f. from an initial condition defined by Lo and Ro. Meth- ods for estimation of the parameters from data were presented. Because E[R] = Ro, Ro is estimated by spa~ tial average rain rate R. The distribution of W’can be ‘obtained from the empirically computable multifractal function 7(q). For the three sets of data analyzed here it was seen that 1¥ can be approximated by a simple form that depends on a single parameter p. Therefore, can be viewed as the descriptor of the spatial rainfall patterns analyzed here. Two methods of estimation of 1p were presented. The first one was derived from the scaling of the fractional rainy area. The second was derived from the slope of the estimated r(q) when the slope is constant. Deviations from the theoretical be- havior are discussed. b. Applications We consider two potential applications of this work to important problems in hydrology. The first appli- cation derives from the parameterization of runoff from the land surface during the operation of a GCM. The grid squares of a GCM ate typically of the order of a few hundred kilometers on a side, and the GCM pre- dicts the convective precipitation to be distributed uni- formly over the grid square. The average convective rain rate over the whole grid square is typically quite small, but some of the local rain rates can be quite large. Since rainfall generally covers only a portion of the grid square, failing to adjust the rainfall distribution for partial coverage results in severe underestimation of runoff (Johnson et al. 1993). For the purpose of illustrating the idea, we assume that the relationship between the large-scale forcing and the cascade theory presented in section 2 is valid for the convective precipitation. Moreover, suppose that the values of the empirical parameters Riss, and k in (2.1) can be compiled globally from empirical obser- vations or predicted from physical considerations. Now the application would proceed as follows. First, set the domain scale Lo to be equal to the size of the grid square of a GCM, assuming that no break in scaling. ‘occurs up to this scale. Second, select the parameters Rmx and K for the season and the spatial location under consideration, Third, equate the expected average rain rate Ro to the rain rate predicted by the GCM. Use (2.1) with Ro in place of to predict p. With Lo, Ro, and p given, the -model random cascade is completely specified. Therefore, a realization of the random cas- cade with the given parameters can be produced and Used as input to the rainfall-runoff component of the Jand-surface parameterization. Alternatively, ensemble average properties of the rainfield with the given pa- rameters can be computed. For example, the expected fractional rainy area or the CV of the rain field at some spatial scale of interest can be determined. The specifics 1540 of the information that one needs would depend on. the specifics of the land-surface parameterization. Nevertheless, the type of results presented here can be used to obtain detailed information about rainfall vari- ability on the subgrid scales. The second application concerns the multiscaling spatial theory of floods within a geographic region (Gupta et al. 1994). Itis observed that the annual peaks of rainfall-generated floods within a homogeneous re- gion exhibit multiscaling spatial variability. Since mul- tiscaling is intimately connected to random cascades (Gupta and Waymire 1993), itis reasonable to hy- pothesize that spatial floods inherit this property from spatial rainfall. A precise demonstration of this feature using a space-time rainfall model remains open. ¢. Some open problems The research presented here suggests open problems for future work in several areas, The frst area is the deviations from iid. random cascade behavior, par- ticularly the @ model, described in section 4. It was suggested there that the theory of dependent cascade generators should be investigated to explain the con- vexity of the 7(q) function for small p. We suspect that the other two deviations, the typical presence of large positive intercepts in the scaling of fractional rainy area and the tendency of the estimated 7(q) function to be concave near q = 0.5, have a common origin, These features can be reproduced in iid. random cascade simulations by selecting a randomly located subset of the simulated field that lies between the subcubes ‘Ay, much as a radar samples a rainfall field. The full implications of this procedure are still being investi- gated. The second area is to explain the relation between large-scale average rain rate R and the cascade param- eter p given by (2.1) on physical grounds. Three prob- Jems fall in this area. The first and most basic is an explanation of the dynamic origin ofthe observed scal- ing invariance that allows estimation of the parameter _p. Empirical investigations of the dependence of scaling ‘behavior on physical forcing such as reported here pro- vide a first step toward a solution of this problem. The second is 0 explain the relation (2.1) itself. As dis- cussed in section 2, it is reasonable to“Suppose for a Single type of forcing that & and p will be related by a one-to-one funetion. However, that does not explain, the form of the function or help predict the values of the parameters. The third is to understand the fluctu- ations around the mean 2 versus p relation given by (2.1). Can they best be understood statistically, that is, as arising from the sensitivity ofthe small-scale pro- cesses to slight differences in the large-scale forcing? Orare they atleast partly due to differences inthe type of convection? For example, it seems reasonable to suppose that a given 2 in the growing stage of convec- tion would have a different associated value of p than JOURNAL OF APPLIED METEOROLOGY Voue 33 the same R in the dissipating stage. We are proceeding ‘with investigations of these questions. ‘The third area concerns an extension of this spatial theory to model the space-time evolution of a rainfall field. Such an extension is important for modeling purposes and for the investigation of basic theoretical questions about the temporal evolution of rainfall ficlds, The present work demonstrates that for a certain ‘lass of rainfall fields the space-time evolution depends ‘on the large-scale forcing. Over a short time interval, during which the large-scale forcing changes slightly, construction of an entirely new field with properties depending on the new forcing is obviously inappro- riate. A correct description would require that an ad- justment of an initial field be made such that its prop- erties respond to the new forcing. The new spatial field after a short time should not look too much different from the initial field. We are currently pursuing this line of research by letting each cascade multiplier Wj, evolve asa continuous-time stochastic process that depends on the large-scale forcing, Acknowledgments. The authors wish to thank Dave Short and Tom Bell of NASA /Goddard Space Flight Center and Rick Lawford of the Atmospheric Envi- ronment Service, Canada, for providing the GATE and the Elbow radar data, respectively. Helpful discussions with Tom Bell, Rick Lawford, Dave Short, Brent Troutman, Skip Vecchia, and Ed Waymire during the course of this research are gratefully acknowledged, as are the comments of Witold Krajewski and two anon- ymous reviewers. This research was supported by NASA Grant NAGW-2731 and NSF Grant EAR- 9220047. The first author was also supported by the USS. Geological Survey, Water Resources Division. APPENDIX A. Random Cascade Simulation Algorithm It is more convenient to simulate the densities of a cascade. In view of Eq. (3.1), the densities for d can be expressed a8 Py = n/N = nb”. We take RoLj = | and simulate on the two-dimensional “cube” [0, Lo]?. To make the two-dimensional na- ture of the b” subdivisions at level explicit, we de- note them by A‘/, where i= 0,1, +++, b%/?~1and J=0,1, +++, 6° ~ 1, We let denote the number of levels in the low-frequency component and M the number of levels in the high-frequency component (see step 6) 1) Initialize the density array at level m= 0, po(8) = 1.0. 2) Increment n= n+ 1 3) Generate a 6” x b"!? array of iid. generator random variables W(Ai!), where i= 0, 1, += +, 5" =Vandj=0,1, ++, 5" = 1 4) For i= 0,1, «+s, 6"? - 1, and j=0, 1, + br 1, set DeceMner 1994 Bol M5!) = pr AEWA), (AL) where [x] denotes the greatest integer less than or equal tox 5) Ifn

You might also like