Veneziano 12

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 16

WATER RESOURCES RESEARCH, VOL. 48, W08516, doi:10.

1029/2012WR012105, 2012

The scaling of temporal rainfall


Daniele Veneziano1 and Chiara Lepore1
Received 9 March 2012; revised 30 May 2012; accepted 16 June 2012; published 22 August 2012.

[1] The scaling analysis of temporal rainfall produces very different results if one uses the
entire rainfall record inclusive of rainstorms and dry interstorm periods (continuous
analysis) or only intrastorm data (within-storm analysis). We show that the continuous
results reflect mainly the alternation of dry and wet periods and that the rain support is not
fractal. Therefore, the continuous results are spurious. This conclusion is significant
since continuous analysis is the most popular scaling analysis method for rainfall.
The literature on within-storm analysis is much more limited, and standard methods do
not exist. We develop such methods and show how their bias can be corrected and the
accuracy maximized. The results from within-storm analysis show higher intermittency
(higher-intensity fluctuations) than continuous analysis. This result has important
implications on downscaling and the evaluation of rainfall extremes. Frequently used
multifractal models for rainfall are of the log-Levy (“universal”) type. A key parameter of
those models is the stability index 0 < a ≤ 2, with a = 2 corresponding to lognormal
models. To account for the alternation of dry and wet periods (also inside the storms),
one can add a “beta component,” obtaining beta-log-Levy and beta-lognormal models.
By using simulations with a = 2, we show that standard estimators of a are negatively
biased and that the hypothesis of beta-lognormal multifractality inside the storms is
statistically acceptable.
Citation: Veneziano, D., and C. Lepore (2012), The scaling of temporal rainfall, Water Resour. Res., 48, W08516, doi:10.1029/
2012WR012105.

1. Introduction Verrier et al., 2011]. We refer to these as continuous anal-


ysis and within-storm analysis, respectively. The scaling
[2] The discovery that rainfall has renormalization prop- properties one needs to estimate are the fractal dimension of
erties has produced new statistical models, new forecasting
the rain support and the scaling of the intensity fluctuations
and downscaling methods, and new approaches to rainfall
when it rains. Compared to within-storm analysis, continu-
extremes [Gupta and Waymire, 1993; Hubert et al., 1993; ous analysis produces much smaller fractal dimensions
Lovejoy and Schertzer, 1995; Over and Gupta, 1996;
(more lacunarity) and much weaker intensity fluctuations
Deidda, 2000; Veneziano and Furcolo, 2002; Molnar and
(less intermittency when it rains). The difference in the
Burlando, 2005; Veneziano et al., 2006a; Langousis and
estimated fractal dimension could be easily explained by the
Veneziano, 2007]. Yet the scaling properties of temporal
fact that the rain support is more compact inside the storms
rainfall remain somewhat elusive. Published results vary
than in the continuous record, whereas the difference in the
widely, calling into question whether rainfall indeed obeys
estimated intermittency is more intriguing because the same
scaling laws, what those laws are, and whether they have
nonzero rain values are used in both analyses. As we shall
some degree of universality [Tessier et al., 1993; Lovejoy and
see, all differences may be imputed to the fact that rainfall
Schertzer, 1995; Nykanen and Harris, 2003; Veneziano et al.,
scales within storms but does not scale as a continuous
2006b; Molnar and Burlando, 2008; Molini et al., 2009; process. Providing support to this conclusion is one main
Serinaldi, 2010; Verrier et al., 2010, 2011]. This study aims
objective of the present study.
at understanding the sources of these differences and estab-
[4] Continuous analysis methods are well codified [see,
lishing proper rainfall scaling analysis procedures.
e.g., Lavallée et al., 1991; Gupta and Waymire, 1993;
[3] The main factor on which the estimates of rainfall
Lovejoy and Schertzer, 1995; Veneziano and Furcolo, 1999],
scaling depend is whether one analyzes the continuous
but those for within-storm analysis are not as established
record inclusive of rainy and nonrainy periods or only the
[Langousis and Veneziano, 2007; Verrier et al., 2011].
portion of the record within storms [de Montera et al., 2009;
Challenges in determining within-storm scaling arise from
the variability of the average storm intensity, the short and
1
Department of Civil and Environmental Engineering, MIT, Cambridge, different duration of the storms, and the fact that the pattern
Massachusetts, USA. of dry intervals inside the storms depends on the storm
Corresponding author: C. Lepore, Department of Civil and identification method and differs from the “true” pattern.
Environmental Engineering, MIT, Room 1-348, Cambridge, MA 02139, Hence, a second objective of this paper is to develop robust
USA. (chlepore@mit.edu) procedures for within-storm scaling analysis.
©2012. American Geophysical Union. All Rights Reserved. [5] Several factors in addition to using continuous or
0043-1397/12/2012WR012105 within-storm analysis affect the scaling results. These include
W08516 1 of 16
W08516 VENEZIANO AND LEPORE: SCALING OF TEMPORAL RAINFALL W08516

the type of fitted multifractal model (e.g., beta, lognormal, where Z is a unit mean random variable called the dressing
beta-lognormal, log-Levy, log-Poisson, etc.), whether the factor [Schertzer and Lovejoy, 1987; Veneziano and Furcolo,
model is conserved in the mean or not (whether it is of the 2003].
multiplicative cascade type or includes fractional integra- [9] The distribution of Ar in equation (1) gives a complete
tion), the quantities used to analyze scaling (average inten- characterization of the scaling properties of the measure Id,
sities, absolute wavelet coefficients, wavelet modulus but in practice it is more convenient to use the moments
maxima, partition coefficients, power spectra), the range of of Ar and more specifically the moment scaling function
scales and moment orders considered, and artifacts intro- K(q) = logr(E[Aqr ]). This function does not depend on r. It is
duced by the recording device or by data preprocessing called the moment-scaling function because, for all q such
(filling in missing values, deaggregation of cumulative that E[Iqd] exists,
measurements, setting to zero the values below some  
threshold, etc.). Here we focus on the scale range from 1 h to E Idq ∝ d K ðqÞ ð3Þ
one or very few days. In this range, a conserved multiplica-
tive model usually suffices [Gupta and Waymire, 1993; Equation (3) continues to hold if one replaces the average
Deidda, 2000; Langousis and Veneziano, 2007] and the intensity Id with a wavelet coefficient at scale d. These
analysis may be conducted using average intensities or par- moment scaling relations are at the basis of most scaling
tition coefficients. Many past analyses have used log-Levy inference methods: one makes log-log plots of the empirical
multifractal models and have reported different values of the q moments of Id against scale d, determines whether and in
Levy index 0 < a ≤ 2 [Tessier et al., 1993; Sun and Barros, what range the plots are straight, and estimates K(q) as the
2010; Verrier et al., 2010, 2011]. We examine the bias in slope of the plots in the scaling range. Variants of this
the estimation of a in continuous and within-storm analysis method are sometimes used to fit specific multifractal
and provide support for the hypothesis that inside the storms models to data. One example is the so-called double trace
a = 2 (lognormal model). moment method for fitting log-Levy models [Lavallée et al.,
[6] Section 2 describes the multifractal models we employ. 1991; Veneziano and Furcolo, 1999].
Section 3 deals with continuous scaling analysis. This [10] The most general K(q) function we consider here is
includes a description of the time series used in the analysis, the so-called beta-log-Levy model, which corresponds to
the estimation procedure, and results. Section 4 develops assuming that Ar in equation (1) has a probability mass
a methodology for within-storm analysis. This includes P[Ar = 0] = 1  rCb at zero and ln(A+r ) = [ln(Ar)∣Ar > 0]
methods for the estimation of the multifractal parameters and has maximally skewed Levy distribution with stability
the assessment of estimation bias and accuracy. Section 5 index 0 < a ≤ 2 and dispersion parameter 2CLL ln(r). In
applies the within-storm methodology to historical data. this case [Schertzer and Lovejoy, 1987]
A summary of the main findings and an outlook of future
CLL;b ab
research are given in section 6. KbLL ðqÞ ¼ Cb ðq  1Þ þ ðq  qÞ ð4Þ
ab  1

The subscripts b and LL in ab, Cb and CLL,b indicate that


2. Scaling Properties of Rainfall this is a log-Levy model with a beta component. The
[7] As mentioned above, many studies have found that, in parameters Cb and CLL,b are nonnegative and satisfy
the scale range from 1 h to 1 day or longer, rainfall behaves Cb + CLL,b < 1. Cb controls the fractal dimension of the
like a stationary multifractal measure. This means that, if Id rain support, Df = 1  Cb; hence, we refer to Cb as the
is the average rainfall intensity in an interval of duration d, lacunarity parameter. CLL,b determines the erraticity of
then for any d and rd within the scaling range, the fluctuations of rainfall intensity when it rains and is
referred to as the intermittency parameter.
Id ¼ Ar Ird ð1Þ [11] In the important special case when a = 2, equation (4)
reduces to the beta-lognormal model
where Ar is a nonnegative random variable with mean value  
1 (if the mean value of Id exists) and r > 1 is a scale change KbLN ðqÞ ¼ Cb ðq  1Þ þ CLN ;b q2  q ð5Þ
or relative resolution factor. The distribution of Ar depends
on r. Random measures that satisfy equation (1) are most for which ln(A+r ) has normal distribution with mean value
simply obtained as multiplicative cascades, i.e., as products (Cb  CLN,b)ln(r) and variance 2CLN,b ln(r). Note that in this
of time-contracted independent copies of a nonnegative case we use LN in place of LL as a subscript. For Cb there is
stationary process. Discrete cascades [Mandelbrot, 1974; no need to use different symbols in equations (4) and (5)
Schertzer and Lovejoy, 1987; Gupta and Waymire, 1993] are because this parameter is the same in the two models.
good pedagogical examples of such constructions. [12] For Cb = 0 (no beta component and compact rain
[8] An important distinction is that between bare and support), equation (4) reduces to the simpler log-Levy
dressed rainfall intensities: bare intensities I′d result from model,
terminating the multiplicative cascade construction at scale d,
CLL
whereas dressed intensities Id are averages over d of the KLL ðqÞ ¼ ðqa  qÞ ð6Þ
cascade developed to infinitesimal scales. The two intensities a1
are related as
In what follows we use the models in equations (4), (5),
Id ¼ I′d  Z ð2Þ and (6). We do not use purely lognormal models obtained
by setting Cb = 0 in equation (5) or a = 2 in equation (6)

2 of 16
W08516 VENEZIANO AND LEPORE: SCALING OF TEMPORAL RAINFALL W08516

Figure 1. Illustration of the upper scaling limit dmax and the duration d′max.

because rainfall invariably displays some level of lacunarity. annual precipitation, wet fraction, and mean value and
The reason for including the purely log-Levy model in coefficient of variation of the positive rainfall intensity. The
equation (6) is that this model has been popular in previous six U.S. records are from NOAA (http://www.esrl.noaa.gov/
scaling analyses of temporal rainfall [e.g., Tessier et al., precip/hpd/). The Portuguese records are from the Portu-
1993; Olsson, 1995; de Lima and de Lima, 2009; Sun and guese Geophysical Institute, the Heathrow airport data set is
Barros, 2010; de Montera et al., 2009; Verrier et al., 2011]. from the UK Meteorological Office, and the Florence record
[13] In addition to the moment scaling function K(q), one is from Becchi and Castelli [1989]. The records vary in
must estimate the limits of the scaling range [dmin, dmax]. length and represent a wide variety of climates. The mean
The lower limit dmin is usually below 1 h [Olsson et al., annual precipitation (MAP) varies between 100 mm for Las
1993; Veneziano and Iacobellis, 2002; Mandapaka et al., Vegas (Nevada) and 1600 mm for Mobile (Alabama). The
2009; Verrier et al., 2011] and cannot be estimated from wet fraction (WF) ranges from 1% in very dry climates to
the hourly data used in the present study. For the upper limit, 13% in very wet regions. Taken together, the mean positive
one should distinguish between dmax and the duration d′max intensity and the wet fraction can be used to characterize the
such that, for d in the scaling range, Id ≈ mAd′max/dZ, where m rainfall climate. For example, Seattle has a very high WF of
is the mean rainfall intensity and Z is the dressing factor in 11% but an MI of only 1 mm/hr, suggesting persistent rain
equation (2). In an ideal multiplicative cascade, d′max = dmax, with low average intensity, whereas Phoenix has a WF of
but in practice dmax may be much smaller than d′max; see only 1% and MI = 1.48 mm/hr, indicating rare but intense
Figure 1 for an illustration. In within-storm analysis, the rainfall.
moments E[Iqd] can be reliably estimated only for short [17] (b) Time series with the same storm characteristics
durations d (due to the limited duration of the storms), (occurrence, duration, average intensity) as the historical
making it difficult to determine dmax, whereas it is possible records, but simulated intensity fluctuations inside the
to estimate d′max with good accuracy. Hence, here we deal storms. Simulation uses beta-lognormal cascades with param-
with the estimation of d′max not dmax. In what follows we eters extracted from the historical records as explained in
refer to d′max as simply dmax. section 4.2. If a storm has duration D < dmax, only the initial
D portion of the dmax long cascade is used. For storms of
duration D > dmax, we use the first D period of a series of
3. Continuous Scaling Analysis
independent contiguous cascades, each of length dmax. The
[14] We start by considering the scaling of continuous historical pattern of storms is repeated (with different intra-
rainfall records, including storms and dry interstorm periods. storm intensity simulations) until a total length of 1000 years
To illustrate and interpret the results, we use a combination is obtained.
of historical rainfall records and synthetic time series [18] (c) Time series (c) is the same as time series (b), but
designed to quantify the effects of different factors. All time the rainfall intensity inside each storm is set to the historical
series have hourly resolution. The various time series are average for that storm, resulting in no intensity fluctuation or
represented in Figure 2 in a schematic illustration. dry periods inside the storms. These synthetic time series
have the same length as the historical records they are
3.1. Time Series Used in the Analysis based on.
[15] The time series we use are (see Figure 2 for a sche- [19] (d) Time series (d) is the same as time series (c), but
matic illustration). with unit storm intensities.
[16] (a) A set of 10 historical records representative of [20] (e) Time series (e) is the same as time series (d), but
different rainfall climates. Table 1 gives the length of the with widely separated storms, such that not more than one
records and basic meteorological parameters: the mean storm contributes rainfall in a d interval.
3 of 16
W08516 VENEZIANO AND LEPORE: SCALING OF TEMPORAL RAINFALL W08516

Figure 2. Type of temporal series used in the scaling analysis: (a) historical records, (b) historical
records with simulated multifractal within-storm intensity fluctuations (1000 yr), (c) same as (b) with uni-
form within-storm intensity equal to the historical value (length of historical record), (d) same as (c) with
unit within-storm intensity (length of historical record), (e) same as (d) with widely separated storms (same
number of storms as in the historical record).

[21] Model (b) produces realistic rainfall simulations. et al., 2011] and one of our objectives is to investigate
Comparison of the scaling results from these simulations with whether the estimates of a reported in the literature are con-
their known within-storm properties highlights the effects of sistent with an underlying beta-lognormal process (a = 2).
including the dry interstorm periods in the analysis. Time [24] For the range of durations used to infer the moment-
series (c)–(e) progressively suppress various sources of rainfall scaling function K(q), we consider two cases: a “local anal-
variability and allow one to assess how those sources of vari- ysis” in which one uses only two durations d/2 and d, with
ability influence the scaling results. d = 2, 4, 8, … hours, and a “scale-range analysis” in which
one uses ranges of duration from 2 hours to d = 4, 8, … hours.
3.2. Continuous Scaling Estimation
[22] The parameters of the beta-lognormal model in
Table 1. Meteorological Parameters of the Historical Rainfall
equation (5) are estimated as Cb = K(0) and CLN ;b ¼ Recorda
2 K ð2Þ  Cb . Other common procedures, for example least
1
MAP Years WF MI VI
squares regression of K(q) over a range of q, produce similar
results. For a multifractal process, the moment scaling Florence, Italy 757.99 23 0.074 1.17 1.76
function K(q) is convex and CLN,b is nonnegative. However, Heathrow, England 598.76 51 0.087 0.78 1.52
if the process lacks scale invariance, the empirical K(q) Seattle, Washington 939.07 45 0.107 1.01 1.04
Phoenix, Arizona 183.85 55 0.014 1.48 1.56
function may be concave and produce negative CLN,b esti- Boston, Massachusetts 1090.42 61 0.082 1.52 1.43
mates. Rather than constraining CLN,b to be positive, we Las Vegas, Nevada 103.41 58 0.010 1.14 1.48
report negative values, as these estimates are useful to Mobile, Alabama 1601.49 53 0.056 3.26 1.78
interpret the results obtained from the historical time series. New Orleans, Alabama 1560.89 54 0.053 3.35 1.78
Coimbra, Portugal 954.28 22 0.105 1.04 1.57
[23] Cb is estimated as K(0) also in the beta-log-Levy Lisbon, Portugal 691.39 22 0.068 1.17 1.66
model, whereas the parameters (a, CLL) and (ab, CLL,b) are
a
obtained through least squares regression of K(q) in the MAP: mean annual precipitation (mm); WF: wet fraction (number of
rainy hours divided by total number of hours); MI: mean rainfall intensity
range 0.5 ≤ q ≤ 3. The reason is that regression-based when it rains (mm/h); VI: coefficient of variation of hourly rainfall
methods are most often used in practice [see, e.g., Verrier intensity when it rains.

4 of 16
W08516 VENEZIANO AND LEPORE: SCALING OF TEMPORAL RAINFALL W08516

Figure 3. Continuous analysis of the Florence record: (top) the moment plot and (bottom) the moment
scaling functions K(q) from local and scale-range analysis. In Figure 3 (bottom left) the arrows and circle
indicate the K(q) function for certain ranges of duration d.

The first analysis produces local moment scaling functions analysis are shown in the left column and scale-range results
Kd,l(q) and the second analysis produces scale-range are in the right column. In both cases, the duration on the
functions Kd,r(q). The functions one uses in practice are of horizontal axis is the highest value of d used to estimate the
the scale-range type, but local scaling is useful to interpret the K(q) function. The records are grouped based on the mean
scale-range results. Significant variation of Kd,l(q) with d annual precipitation: less than 400 mm (Las Vegas and
indicates lack of scale invariance. Phoenix), between 400 and 800 mm (Florence, Heathrow,
[25] As an example, Figure 3 shows moment plots and the Lisbon), and higher than 800 mm (Seattle, Boston, Mobile,
local and scale-range K(q) functions for the Florence record. New Orleans, Coimbra). To a first approximation, the scale-
The solid lines in Figure 3 (top) are least squares fits for range parameters for a given d are averages of the local
durations d from 2 to 8 h. While the moment plots seem to estimates for durations up to d. In the following comments
have good scaling, the local K(q) curves reveal significant we refer to the local scaling analysis.
variation with d. As one would expect, the variation is [28] A striking feature of Figure 4 is the high value of Cb,
damped in range-scale analysis. especially in dry climates. Recall that Cb = 1 corresponds to
[26] For the estimation of dmax, we use the relationship a zero fractal dimension and therefore to isolated rain epi-
Id = mAdmax/dZ, where m is the mean rainfall intensity and Z is sodes. In arid climates this condition is approached for d
the dressing factor. This relationship implies around 64 h, meaning that in those climates rainstorms are
  typically of duration less than 64 h and are separated by dry
  d max K ðqÞ periods longer than about 64 h. At smaller scales, rainfall is
E Idq ¼ mq E ½Z q  ð7Þ
d more persistent, producing lower values of Cb. In wetter
climates, persistency extends to longer durations and the
where the moments of Z can be calculated using the method peak value of Cb is lower. At long averaging durations,
of Veneziano and Furcolo [2003]. Here we estimate dmax Cb decreases and eventually approaches zero, which corre-
 K ð2Þ sponds to a fractal dimension 1 of the rain support. This
such that, on log paper, m2 d max
d E½Z 2  produces the least
squares fit to the empirical second moments E[I2d]. The range happens because it is rare for very long periods to be com-
of d is the same as that used for K(q) estimation (see above). pletely dry. The rate at which Cb → 0 is faster in the wetter
Estimates of dmax are obtained only for the beta-lognormal climates.
model. [29] The intermittency parameter CLN,b displays signifi-
cant fluctuations from record to record. Recall  that this
parameter is obtained as CLN;b ¼ 12 K ð2Þ  Cb ; hence, CLN,b
3.3. Continuous Scaling Results
is used to explain the scaling exponent K(2) beyond the value
[27] Figure 4 shows the multifractal parameters estimated expected from the beta component alone. In some cases Cb
from the 10 historical records. Results from local scaling
5 of 16
W08516 VENEZIANO AND LEPORE: SCALING OF TEMPORAL RAINFALL W08516

Figure 4. Continuous analysis of historical records. Comparison of results from local and scale-range
analysis. The duration d on the horizontal axis is the largest value used in the analysis. Symbols and colors
distinguish stations in variously wet climates; see text for details.

exceeds K(2). This condition implies nonconvexity of K(q) plays a dominant role in explaining the fluctuations of
and negative values of CLN,b. As we show later when dis- rainfall intensity and as a → 0 the log-Levy model becomes
cussing model d results, the source of these (mathematically a pure beta model. Similar results have been reported in the
impossible) negative values is lack of scaling of the rain sup- literature [Tessier et al., 1993; Olsson, 1995; de Lima and
port. In general, the local values of CLN,b increase as d de Lima, 2009].
increases, in a pattern that is the opposite of that of Cb. In wet [32] When a beta component is added to the log-Levy
climates, as d becomes large, Cb decreases to zero and the model, the a and CLL parameters change drastically, because
observed scaling exponent K(2) is explained through large the log-Levy component is used to explain not the entire
values of CLN,b. Dry climates display the same behavior at K(q) function but the difference between K(q) and the value
durations longer than those considered in Figure 4. Cb(q  1) from the beta component. In general ab is close to
[30] The local estimates of dmax reflect mainly the scaling 2 (in the analysis, ab is constrained to be between 0 and 2),
exponent Kd,l(2) (dmax increases as the exponent decreases). with a tendency to decrease as d increases, and CLL,b is close
For example, at small scales both Cb and CLN,b are relatively to CLN,b. The local estimates of ab are erratic, especially at
small; hence, Kd,l(2) is small and dmax is large. For very long short durations. This variability is due to the fact that the log-
durations, Cb in wet climates is very small and again dmax is Levy component, CLL,b, is small and ab is sensitive to small
large. variations of the empirical moments.
[31] Log-Levy analysis without the beta component gives [33] For a deeper understanding of Figure 4, Figures 5a–5c
stability indices a close to zero and intermittency parameters compare results from the different rainfall time series in
CLL close to Cb in the beta-lognormal model (Figure 4, first Figure 2. The time series are based on the Florence record
row). These results are explained by the fact that lacunarity (Figure 5a), the Boston record (Figure 5b), and the Phoenix
record (Figure 5c), which have been chosen as representative
6 of 16
W08516 VENEZIANO AND LEPORE: SCALING OF TEMPORAL RAINFALL W08516

Figure 5a. Continuous scaling analysis of the time series shown in Figure 2 based on the Florence
record. The first two rows refer to fitting of the beta-lognormal model. The last two rows refer to fitting
of the log-Levy model, with (subscript b) and without a beta component. Solid lines are results from using
the historical record.

of variously wet climates, with a mean annual precipitation of reproduces also the historical local intermittencies CLN,b for
756, 1090, and 183 mm/yr, respectively. The solid lines large d, confirming that the historical estimates reflect the
(labeled a) are results from the historical records and are the pattern of storms and their average intensities more than the
same as in Figure 4. within-storm fluctuations. At short durations, CLN,b is close
[34] Model b preserves the pattern and average intensity to zero because in model c the rainfall intensity inside the
of the historical storms but replaces the observed rainfall storms is uniform.
fluctuations inside the storms with simulations from beta- [36] In model d, setting to 1 the rainfall intensity inside the
lognormal cascades. The cascade parameters [(Cb, CLL,b) = storms does not affect Cb but reduces CLN,b, especially at
(0.09, 0.14), (0.04, 0.11), and (0.08, 0.14) for Florence, short durations where CLN,b becomes highly negative. This
Boston and Phoenix, respectively] are estimated from the negativity indicates that model d is not scaling, i.e., that the
historical records as explained in section 4. The fact that the rainfall support is not fractal. The increase of CLN,b with
(a) and (b) parameters in Figures 5a–5c are very similar and duration explains the general trend of this parameter in the
differ significantly from those used to simulate the storm analysis of the historical records: for short durations the
interiors indicates that the results in Figure 4 reflect the historical estimates are essentially the same as the model d
pattern and intensity of the storms more than the fluctuations values incremented by the within-storm CLN,b. The net effect
of rainfall intensity inside the storms. As the label in is that, at small scales, the historical estimates of CLN,b are
Figures 5a–5c increases from (c) to (e), the time series significantly lower than the within-storm values. At scales
becomes more schematic and resembles less and less the that exceed the average storm duration, the effect of the
historical record; see Figure 2. As we discuss next, these within-storm fluctuations on CLN,b is smaller and the model
time series help understanding which rainfall features con- d values become closer to the historical estimates.
trol the scaling results in Figure 4. [37] For model e, if one further spaces the storms such that
[35] In model c, suppression of the interior storm fluc- not more than one storm contributes rainfall to any d inter-
tuations does not alter the lacunarity parameter Cb, except at val, then at large scales the fractal dimension of the support
very short durations due to short dry periods inside the his- drops to 0, the local lacunarity Cb → 1, and the local inter-
torical storms that in model c are considered wet. Model c mittency CLN,b → 0. It is interesting that this extremely
7 of 16
W08516 VENEZIANO AND LEPORE: SCALING OF TEMPORAL RAINFALL W08516

Figure 5b. Same as Figure 5a with time series based on the Boston record.

simplified model already captures the general shape of the Since the same behavior is seen in model c, these features
historical plots of CLN,b against d, confirming that that shape have nothing to do with the within-storm fluctuations.
originates from the distribution of storm duration.
[38] In summary, the intermittency parameter Cb and the 4. Within-Storm Scaling Analysis
variation of CLN,b with duration from historical time series
are determined by the pattern of storm arrivals and durations. [40] The fact that the continuous scaling results are
In the absence of within-storm intensity fluctuations, CLN,b insensitive to the within-storm rainfall fluctuations is a
is negative at small scales and close to zero at large scales. concern because operations like downscaling and critical
The within-storm intensity fluctuations increase CLN,b at quantities like rainfall extremes are controlled by such
small scales, but have a small effect at large scales. The net fluctuations. The continuous results do not even reflect the
result is that CLN,b from continuous analysis significantly scaling of the storm occurrence process (except in a best
underestimates the intermittency within storms. For exam- fitting sense), because the rain support is nonfractal. There-
ple, for Florence, Boston and Phoenix one obtains CLN,b fore, we are compelled to examine whether rainfall has
values around 0.05, 0, and 0.01, when the within-storm multifractal properties within the storms. Section 4.1 dis-
values are around 0.14, 0.11, and 0.14, respectively. cusses procedures to identify the storms, section 4.2 devel-
[39] Concerning the fitting of log-Levy and beta-log-Levy ops methods to estimate scaling, and section 4.3 assesses the
models, Figures 5a–5c show that also the estimates of a and bias and variance of the estimators.
ab are largely controlled by the pattern and average intensity 4.1. Storm Identification
of the storms (see the good correspondence between histor-
ical and model c results). If one does not include a beta [41] A common storm identification method is to associate
component (as is typically the case when fitting log-Levy storms with uninterrupted rainy periods [Dixon, 1994;
models), a is close to zero and CLL is large, effectively Lakshmanan et al., 2003; Gomes and Held, 2005]. Problems
mimicking a purely beta process. When a beta component is with this approach are the sensitivity to the temporal reso-
added, ab becomes close to 2 (optimal fits to K(q) are often lution of the data and the fact that even short dry periods
obtained with values larger than 2, which are inadmissible), may split into multiple events what meteorologically should
with a tendency to slightly decrease as duration increases. be considered a single rainstorm. To avoid this to occur, one
must allow storms to include short dry periods.
8 of 16
W08516 VENEZIANO AND LEPORE: SCALING OF TEMPORAL RAINFALL W08516

Figure 5c. Same as Figure 5a with time series based on the Phoenix record.

[42] One popular strategy is to consider as storm separa- [44] Both the traditional and fractal methods produce
tors dry periods that exceed some threshold duration D*. storms that start and end with nonzero values. This feature
The choice of D* may be judgmental [Huff, 1967; Menabde affects the distribution of storm duration and more impor-
and Sivapalan, 2000; Upton, 2002] or based on objective tantly the lacunarity within storms. In section 4.3 we shall
criteria, such as passing a statistical test of independence of introduce a bias correction for the latter effect.
the resulting storm events [Koutsoyiannis and Foufoula- [45] In analyzing the historical data in Table 1, we use the
Georgiou, 1993]. We call this the traditional storm identi- traditional method with D* = 6 h and the fractal method with
fication method. 1 and 6 h thresholds. The number of identified storms in the
[43] The threshold D* should ideally depend on the fractal method is about 30% higher than in the traditional
duration of the adjacent storms. For example, a dry period of method, with large increases in the number of short storms
2 h should not be used to split a 24 h storm, but could and a reduction in the number of long storms. However, for
plausibly separate shorter storms. To implement this con- the development of within-storm estimation procedures, the
cept, Langousis and Veneziano [2007] proposed a fractal storm identification method is less important; for that pur-
storm identification method. This method partitions the pose we use only the fractal method.
record into continuously wet and continuously dry periods
and then examines the dry periods in order of increasing 4.2. Within-Storm Scaling Estimation
duration. If a dry period separates rainy intervals of equal or [46] The analysis within the storms includes the inference of
longer duration, the dry period and the adjacent rainy inter- the lacunarity and intermittency parameters in equations (4),
vals are considered part of the same rainstorm. Convergence (5), and (6) (Cb and CLN,b for the beta-lognormal case,
is usually attained after very few iterations. This procedure is CLL and a for the log-Levy case, and CLL,b and ab for the
applied with two exceptions: dry periods of duration not beta-log-Levy case) collectively referred to as fluctuation
exceeding a given lower limit are always considered part of parameters, and the outer scale of multifractality dmax. A
a storm interior and dry periods of duration equal to or above property of rainfall that simplifies the within-storm analysis
a given upper limit are always considered to be storm is that the rainfall intensity in a wet intrastorm period of
separators. Langousis and Veneziano [2007] found that a duration d, Id+, does not significantly depend on the duration
lower limit of 1 h and an upper limit of 6 h produce satis- D of the host storm [Langousis and Veneziano, 2007]. This
factory results. Models b–e in Figures 2 and 5 are based on property excludes any variation of the multifractal parameters
this fractal identification method. with D and allows one to combine data from storms of dif-
ferent duration.
9 of 16
W08516 VENEZIANO AND LEPORE: SCALING OF TEMPORAL RAINFALL W08516

[47] Still, several issues remain to be addressed. Specifi- coefficient C2 under the condition that both d subintervals
cally, the lacunarity and intermittency parameters could be are wet; hence, 0 < C+2 < 2. The 2
second-moment scaling
inferred from the moments of Id (absolute scale) or Ir = Id=D/r exponent is K(2) = log2E[C+2 ] and CLN,b is found as
(relative scale), the storms may or may not be normalized to 1
CLN ;b ¼ K ð2Þ. The reason for using the positive rather than
have unit intensity (normalization gives equal weight to all 2
the storms, irrespective of their average intensity), and ordinary partition coefficients is that the former decouple the
storms could be partitioned into substorms before normali- estimation of Cb and CLN,b and do not let the bias of Cb
zation. In addition, Cb and the intermittency parameters contribute bias to CLN,b. In accordance with Veneziano and
(CLN, CLL, a) may be estimated jointly or from different Furcolo [2009], we have found that the present selection of
statistics. Also for the outer scale of multifractality dmax low r and d values minimizes the estimation variance, but
there are choices to be made. We briefly discuss various increases the bias relative to using larger r and d values. The
possibilities below, first for the fluctuation parameters and higher bias is of no great concern because it can be corrected
then for dmax. as we show in section 4.3.
4.2.1. Estimation of the Fluctuation Parameters [54] The log-Levy and beta-log-Levy intermittency param-
[48] For the fluctuation parameters, one might consider eters are found as in the continuous case, i.e., by using
the following estimation strategies. method 1 above (the moments of Id) and least squares fitting
[49] 1. Use of Id without storm normalization. The most of the empirical K(q) functions for q between 0.5 and 3. The
straightforward way to estimate the fluctuation parameters is durations used are d = 2, 4 and 8 h. While suboptimal, this
q
to use the moments md,q = E[Id ] of the average rainfall method is close to procedures followed in practice to fit log-
intensities Id for different d and q. Since the intensities Id are Levy models; hence, its use facilitates determining whether
extracted from storms with duration D ≥ d, the mix of storm past results are consistent with a true stability index a = 2.
durations that contribute to md,q depends on d. As the prob- 4.2.2. Estimation of dmax
ability P[Id = 0] that a d interval is dry depends on D, this [55] The parameter dmax cannot be found from normalized
mixture variation introduces bias in the estimation of the data or partition coefficients. Rather, one can use the
fluctuation parameters. The storm identification procedure unnormalized positive average intensities I+d = (Id∣Id > 0) as
may contribute additional bias by differentially removing follows.
dry interstorm periods of different length. Another, more [56] 1. Calculate the nonnormalized moments E[(I+d )q] for
important drawback of this method is that the estimators some q and a range of d. For the small values of Cb found
have a high variance because the moments of order q > 1 inside the storms, these moments scale approximately as
depend mainly on a few intense storms. E[(I+d )q] = cdK0(q), where c is some constant and K0(q) is the
[50] 2. Use of Id after storm normalization. The variance moment scaling function for Cb = 0.
of the estimators can be reduced by normalizing the storms [57] 2. Find the value of c such that cdK0(q) best fits the
to have unit average intensity. In this case one uses the calculated moments E[(I+d )q] (in log space).
moments m′d,q = E[I′qd] where the prime sign denotes quan- [58] 3. Find dmax from the condition cdK 0(q)
max = mqE[(Z+)q],
+
tities calculated after storm normalization. A drawback is where m is the mean storm intensity and Z = (Z∣Z > 0) is the
that the resulting estimators are highly biased because after positive dressing factor.
normalization the distribution of I′d has a strong dependence [59] Since the moments E[(I+d )q] for different d are highly
on D. Hence, this estimation strategy should be discarded. correlated, one may use the above procedure with just one
[51] 3. Use of partition coefficients Cr. A better alternative value of d. This gives
is to use the moments m′r,q = E[Cqr ], where Cr = Id/Ird is the     
partition coefficient between scales rd and d (we use a prime d max ln E Idþ q  lnðmq E½Z þq Þ
ln ¼ ð8Þ
sign on the moments because the partition coefficients d K0 ðqÞ
effectively normalize the rainfall intensity at scale rd). The
use of partition coefficients is known to bias the estimation Equation (8) has the advantage that in the beta-lognormal
of the fluctuation parameters because, unlike the moment case one can theoretically calculate the mean value and
q q variance of ln(dmax); see Appendix A. In what follows, we use
ratios E[Ir ]/E[Ird], the moments E[Cqr ] = E[(Id/Ird)q] scale
only in approximation with r [Veneziano and Furcolo, equation (8) with d = 2 h.
2009]. However, the bias is much smaller than that of
method 2 and can be corrected (see below). Since rd can 4.3. Bias and Variance of Within-Storm Estimators
have any value up to the duration D of a storm, one can [60] To assess the accuracy of the above estimators, we
extract a large number of partition coefficients Cr from also use model b in Figure 2 to generate a large number of syn-
short rainfall records, producing estimates with small error thetic records with beta-lognormal scaling within storms.
variance. The simulations are based on the historical storm pattern for
[52] With these considerations in mind, we have devised Florence produced by the fractal storm identification method
the following parameter estimation strategy. The within- (using other historical rainstorm patterns or other storm-
storm lacunarity parameter Cb is obtained as Cb = K(0) identification procedures produces very similar results). One
where K(0) is estimated using the moments m′r,0 = E[C0r ] hundred 25 year simulations have been generated for each of
for r = 1, 2 and d = 1 h. This corresponds to setting the following 32 combinations of within-storm parameters:
Cb = log2(P+) where P+ is the fraction of rainy hours dmax = 16, 32 h, Cb = 0, 0.025, 0.05, 0.08, and CLN,b = 0.05,
within wet 2 h intervals inside the storms. 0.10, 0.15, 0.20. These parameter values span the range of
[53] CLN,b is estimated using the second moments of real data applications; see section 4.4. Each synthetic record
positive partition coefficients C+2 at relative resolution r = 2, has been analyzed under the assumption that the storm
for d = 2 or 4 h. More precisely, C+2 is the partition interiors have beta-lognormal, log-Levy, or beta-log-Levy
10 of 16
W08516 VENEZIANO AND LEPORE: SCALING OF TEMPORAL RAINFALL W08516

Figure 6. Mean value and coefficient of variation of the beta-lognormal parameters, using the proposed
within-storm estimators. The red lines correspond to zero bias. Different line patterns are used for different
storm identification cases: solid lines for no storm reidentification (ideal case when the start and end times
of the storms are assumed known), solid lines with circles for the traditional method, and dashed lines for
the fractal method. All results are based on one hundred 25 year simulations using the temporal pattern and
average storm intensities from the Florence record. The intensity fluctuations inside the storms are simu-
lated using a beta-lognormal multifractal model with parameters Cb and CLN,b as indicated.

multifractal scaling. In addition, each analysis has been fractal method (because of the misclassification of some
repeated three times: (1) assuming that the time of occurrence short dry intervals as storm separators and because the
and duration of the storms are known (no storm reidentifi- method does not allow a storm to start or end with zeros).
cation), (2) reidentifying the storms using the traditional [63] The coefficient of variation of CLN,b is around 0.03; it
method with D* = 6 h, and (3) reidentifying the storms using is independent of Cb and decreases somewhat as CLN,b
the fractal method with critical durations set to 1 and 6 h. increases. The coefficient of variation of Cb is around 0.05
[61] From these runs, we have assessed the bias and when the storm pattern is known or the storms are identified
coefficient of variation of the estimators of all multifractal by the traditional method and around 0.10 when the fractal
parameters. The results are shown in Figure 6 (fluctuation storm identification criterion is used. The larger value for
parameters of the beta-lognormal model), Figure 7 (fluctu- the fractal method is due to the smaller number of wet
ation parameters of the log-Levy model), and Figure 8 (dmax 2 h intervals from which Cb is estimated and to the smaller
for the beta-lognormal model). Each column refers to a dif- estimator mean.
ferent value of Cb, whereas CLN,b varies along the horizontal [64] One can use the simulation results in Figure 6 to
axis of each panel. obtain approximate analytical expressions of the bias and
[62] Consider first the fluctuation parameters of the beta- accuracy of the Cb and CLN,b estimators. Bias-corrected
lognormal model (Figure 6). The first two rows show the estimates are given by (for Cb, we use the results from the
expected values of the two parameters and the bottom rows fractal storm identification method)
give their coefficients of variation, calculated as the standard
deviation divided by the estimator mean. Solid, dot-dashed ^ b ¼ 1:38  Cb
C
and dashed lines are for no storm reidentification, traditional qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð9Þ
 
storm identification, and fractal storm identification, ^ LN ;b ¼ 0:44 
C 1:19 þ 0:8  0:26Cb  CLN ;b
respectively. The bias of CLN,b is small (compare with the
no-bias 45 line) and is the same for the different storm where Cb and CLN,b on the right side of the equation are the
identification methods. As one would expect, the storm empirical biased estimates. The coefficients of variation
identification procedure is influential on Cb: the bias of Cb is depend on the length of the record and more precisely on the
null when the storm pattern is known, is positive when using number n+2 of wet 2 h intervals within storms for Cb
the traditional method (because that method leaves too many and the number n+C of partition coefficients C+2 for CLN,b.
short dry interval inside the storms), and is negative for the
11 of 16
W08516 VENEZIANO AND LEPORE: SCALING OF TEMPORAL RAINFALL W08516

Figure 7. Same as Figure 6 for the parameters of the log-Levy model, with and without a beta
component.
Considering that for the Florence record n+2 ≈ 4700 and lines refer to (a, CLL). The other lines are for (ab, CLL,b)
n+C ≈ 2500, the coefficients of variation in Figure 6 are fitted using the estimated biased Cb values (dashed lines) or the
well by theoretical Cb values (circle with solid lines). All results are
obtained under no reidentification of the storms.
3:4 [65] When Cb = 0, there is no lacunarity inside the storms
VCb ¼ pffiffiffiffiffi ffi
nþ2 and the correct values of the log-Levy parameters are a = 2
ð10Þ
1:5 and CLL = CLN. The first column in Figure 7 shows that the
VCLN ;b ¼ pffiffiffiffiffi ffi estimate of a has a negative bias that increases as CLN
nþ C
increases. This negative bias is compensated for by a posi-
Figure 7 shows results for the log-Levy parameters (a, CLL) tive bias of CLL. As Cb increases, in an analysis without beta
and (ab, CLL,b), using the same format as in Figure 6. Solid component (solid lines) the dry within-storm intervals cause

Figure 8. Same as Figure 6 for the parameter dmax. Lines are theoretical results from Appendix A, and
symbols are simulation results. Solid lines and circles are for dmax = 32 h; dashed lines and squares are for
dmax = 16 h. Black symbols and lines are from using moments of order q = 2; gray symbols and lines are
from using moments of order q = 0.5.
12 of 16
W08516 VENEZIANO AND LEPORE: SCALING OF TEMPORAL RAINFALL W08516

results to find for what moment order q the estimator in


equation (8) is most accurate. The accuracy increases as q
decreases (compare black and gray lines in Figure 8).
However, there are reasons for using moments of higher
order: low-order moments are more sensitive to the low
rainfall intensities, which in actual records are often unreli-
able. Moreover, equation (8) uses dmax to reproduce empir-
ical scaling exponents of order q. In practice, interest is often
in the high rainfall intensities, which are controlled by the
higher moments. All considered, we regard the choice q = 2
as about optimal.

5. Within-Storm Analysis of Historical Records


and Discussion
[69] Having developed an analysis method for within-
storm scaling, we apply the method to the historical time
series in Table 1. Results are shown in Figure 9. The biased
and bias-corrected estimates of the beta-lognormal param-
eters are shown as dashed and solid lines, respectively.
Log-Levy model parameters obtained using Cb estimates
before and after bias correction are given by the dashed line
and solid line with circles, respectively. In all cases, storms
are identified by the fractal method.
[70] In Figure 9, the stations are ordered by increasing
mean positive intensity (MI, bottom plot). MI is positively
correlated with CLN and CLL and negatively correlated with
the other multifractal parameters. In particular, CLN is around
0.10–0.15 in dry and moderately wet climates and exceeds
0.2 along the wet Gulf of Mexico coast (New Orleans and
Mobile).
[71] When the internal lacunarity is accounted for by
using a beta-log-Levy model, the stability index a is gen-
Figure 9. Within-storm scaling results for the historical erally close to 2. This is consistent with estimates reported
records. The dashed and solid lines in the top plot give by de Montera et al. [2009] and Verrier et al. [2010, 2011]
parameter estimates before and after bias correction, respec- for continuously wet segments of rain records. The only
tively. Values of the beta-log-Levy model obtained using Cb exceptions are Mobile and New Orleans, where a is around
estimates before and after bias correction are given by the 1.5. Notice that at these sites CLN is around 0.22. Figure 7
dashed line and solid line with circles, respectively. The sta- shows that, under the assumption that a = 2, the estimator
tions are ordered according to increasing mean intensity of a for CLN = 0.22 has a mean value around 1.7 and a
when it rains (MI, bottom plot). standard deviation of about 0.3. Hence, the assumption that
also at these sites a = 2 is statistically acceptable.
a to slightly decrease and CLN to increase. When the beta
component is included, the biases are smaller. The coeffi-
cients of variation of both parameters increase as CLN 6. Conclusions
increases, with representative values around 0.1. [72] This study pursued three main objectives related to
[66] Finally, Figure 8 concerns the estimation of dmax. The the scaling of temporal rainfall: (1) develop methods for the
symbols are simulation results, and the lines are theoretical analysis of within-storm scaling, (2) explain the difference in
results from Appendix A. The circles and solid lines are for a the scaling results between continuous and within-storm
true dmax = 32 h, whereas the squares and dashed lines are analysis (and determine in which way rainfall scales), and
for a true dmax = 16 h. The true values are indicated by red (3) examine whether scaling follows a beta-lognormal model
horizontal lines in the top row. The symbols and black lines or a more general log-Levy (“universal”) or beta-log-Levy
are for the case when one uses equation (8) with q = 2 and model. Our main findings are as follows.
the gray lines are theoretical results for q = 0.5. [73] By analyzing historical records and synthetic time
[67] As one can see from the top row of Figure 8, the series generated by different models, we have found that the
estimator of ln(dmax) is essentially unbiased. Its standard continuous scaling results are dominated by the pattern of
deviation is insensitive to Cb, but increases as CLN increases. storms/interstorm periods (essentially by the rain support).
The coefficient of variation for CLN = 0.15 is about 0.07. This This is especially true for the lacunarity (beta component)
coefficient of variation is essentially proportional to n1/2, parameter Cb, the stability index a and intermittency
where n is the number of years on record (here n = 25). parameter CLL of the log-Levy model without a beta com-
[68] Appendix A makes a theoretical analysis of the bias ponent, and the outer limit of scaling dmax. The intermittency
and accuracy of ln(dmax) estimation (the lines in Figure 8 parameters of the beta-lognormal and beta-log-Levy models,
show those theoretical values). One can use the theoretical CLN and CLL, are affected also by the variability of the
13 of 16
W08516 VENEZIANO AND LEPORE: SCALING OF TEMPORAL RAINFALL W08516

average storm intensity and the fluctuations of hourly lognormal intensity fluctuations. First we derive the rela-
intensity within the storms. We have found that the esti- tionship between the moments mq,d = E[I+q d ] and m1,d = E[Id ]
+

mated multifractal parameters vary in a systematic way with of order q and 1 of the positive intensities at scale d (no zero
the range of durations used in the analysis. This variation is rainfall values is included in this analysis, therefore Cb is set
caused by lack of scaling of the rain support. In essence, to zero). We notice that
rainfall does not scale in continuous time and the results of
continuous analysis are to be treated as spurious. mq;d ¼ mq rqCb rKLN ðqÞ E½Z þq 
[74] One of the models we have used to investigate the  q qC K ðqÞ
¼ mrCb rKLN ðqÞ rZ b rZ LN
issue of continuous-time scaling (model b in Figure 2) pre-

C q K ðqÞ
serves the historical occurrence and average intensity of the ¼ mrd b rKLN ðqÞ rZ LN
storms but replaces the historical intensity fluctuations inside ðqÞ
¼ mq1;d rKLN ðqÞ rZ LN
K
the storms with simulated multifractal beta-lognormal cas- ðA1Þ
cades. The continuous scaling results are virtually identical
to those from the historical records and differ greatly from where r = dmax/d, the moments of Z are approximated as
the within-storm multifractal properties used in the simula- E[Z q] = rK(q)
Z with rZ to match the second moment of Z
tion. Therefore, continuous analysis should not be used to [Veneziano and Langousis, 2005; Veneziano and Furcolo,
characterize scaling inside the storms. 2003], rd = rrZ, and KLN(q) = CLN(q2  q) is the moment
[75] Having found that rain does not scale as a continuous scaling function associated with the lognormal component of
process, we turned our attention to within-storm scaling. In multifractal scaling. Replacing the exact moments mq,d and
spite of increasing evidence that temporal rainfall scales only m1,d in equation (A1) with the sample moments m  q;d ¼
inside storms [Langousis and Veneziano, 2007; de Montera
q
þ þ
et al., 2009; Verrier et al., 2010, 2011], robust within-storm nr ∑i¼1 Ir;i
1 nr
 1;d ¼ n1r ∑i¼1
and m nr
Ir;i and solving for ln(dmax) =
analysis procedures do not exist. We have investigated sev- ln(r) + ln(d) gives the following estimator of ln(dmax):
eral strategies and concluded that the best approach is to use

the moments of partition coefficients (see section 4 for  



ln m q;d  q ln m 1;d
details). Theoretically, these moments do not scale and the ln d^ max ¼ lnðd Þ þ  lnðrZ Þ ðA2Þ
resulting multifractal parameter estimates are biased KLN ðqÞ
[Veneziano and Furcolo, 2009]. However, the bias can be
corrected using simulation and the estimates have very small Equation (A2) finds dmax to reproduce the observed moment
error variance. Therefore, it is possible to construct within- mq,d at a single scale
d.
Our objective is to quantify the bias
storm scaling estimators that are both unbiased and very and variance of ln d^max in equation (A2) and optimize the
accurate. This is important especially for the analysis of choice of q and d. Notice that equation (A1) assumes
short historical records or records from arid climates. knowledge of the KLN(q) function, when in reality that
[76] In the literature, much has been said about the devia- function must be estimated from data. However, if one uses
tions of the stability index a from the value 2 that corre- low-variance bias-corrected estimators of CLN like the estima-
sponds to lognormal or beta-lognormal models. For example, tor in section 4, one may assume that KLN(q) in equation (A2)
values around 0.5 are obtained when a log-Levy model is is known.
used to describe the scaling of rainfall in continuous time, [79] The mean value and variance of mq, d are
whereas values between 1.5 and 2 are found when one ana-
lyzes periods of compact rainfall, in essence rainfall within h i   qCb þCLN ðq2 qÞ
E mq;d ¼ E Idþ q ¼ mq rd
storms. We have used simulation to determine the bias and h i n 
1    2 o
variance of the estimator of a. For all 10 historical records, Var mq;d ¼ E Idþ 2q  E Idþ q ðA3Þ
the estimates of a are less than one standard deviation away neq ðd; qÞ
1 h
2qCb þ2CLN ðq2 qÞ 2CLN q2
i
from their expected values under a = 2. We conclude that ¼ m2q rd rd 1
within-storm fluctuations conform to a beta-lognormal neq ðd; qÞ
model. This is an important finding because the estimation of
a is notoriously difficult and lognormal distributions are where neq(d, q) is the equivalent independent sample size
much easier to work with than log-Levy distributions. that considers the effect of correlations among the Id+q values
[77] In future extensions of this work we plan to system- (see later).
atically analyze all the NOAA precipitation data in the [80] Notice that ln(Id+) and ln(Id+q) = q ln(Id+) have a (sin-
continental U.S. (several thousand hourly records). Issues of gular) joint normal distribution with correlation coefficient 1.
interests are the spatial and temporal variation of the multi- It follows [e.g., Thomopoulos and Johnson, 2004] that Id+
fractal parameters within storms and how these variations and Id+q have (singular) joint lognormal distribution with
may relate to climatic variables and correlate with rainfall correlation coefficient
extremes. This analysis may possibly lead to revisions of 2
eqs  1
classical extreme rainfall estimates [e.g., Hershfield, 1961] r ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ðA4Þ
and an assessment of their possible future variation due to ðes  1Þðeq2 s2  1Þ
2

climate changes.
where s2 = 2CLN ln(rd) is the variance of ln(I+d ). In good
approximation, r in equation (A4) is also the correlation
Appendix A: Accuracy of Estimation of dmax  1;d and m
coefficient between m  q;d .
[78] We examine the bias and variance of the estimator [81] Finally, assuming that m 1;d and m
 q;d have joint log-
of ln(dmax) in the case of storms with multifractal beta- normal distribution, their natural logarithms are jointly
14 of 16
W08516 VENEZIANO AND LEPORE: SCALING OF TEMPORAL RAINFALL W08516

normal, with mean values, variances and covariance given [83] Consider now a single storm of duration D = nd. For
by D ≤ dmax, the variance s2(D) = Var[∑ni=1Iqd,i] is
h
i
qCb þCLN ðq2 qÞ " #
E ln m q;d ¼ ln mq rd 2CLN q2
  2CLN ðq2 qÞ X
n1
r
h i s ðDÞ ¼ nVar Idq þ 2rd
2
ðn  j Þ 1
1 1 2 j¼1
jþ1
 ln 1 þ rd2CLN q  1
2 neq ðd; qÞ ðA9Þ
h
i h i
1 2
Var ln m  q;d ¼ ln 1 þ rd2CLN q  1
neq ðd; qÞ where r = dmax/d. For D ≤ dmax, the storm includes s(D) =
h  
i
Cov ln m  1;d ; ln m
 q;d int[D/dmax] independent substorms of length dmax and a
8 rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi “residual” substorm of length Dres = D  s(D)dmax. Hence,
h i h i9
>
> r Var m  1;d Var m  q;d > > in this case the variance s2(D) is
< 1 =
¼ ln 1 þ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi h i h i
>
>
: neq ðd; 1Þneq ðd; qÞ E m j
 1;d E m  q;d j >
>
; s2 ðDÞ ¼ sðDÞs2 ðd max Þ þ s2 ðDres Þ ðA10Þ

ðA5Þ
where s2(dmax) and s2(Dres) are found from equation (A9).

[84] The value of g(d,q) for a single storm of duration D is


Using equation (A5), ln d^max in equation (A2) has normal D
Var½I q 
distribution with mean value and variance given by g ðd; q∣DÞ ¼ d s2 ðDÞd . For D random, one obtains g(d,q) by
replacing numerator and denominator in this expression by
0
1 h 2
i1 their expected values with respect to D. This gives
h
i 1þ r2CLN q  1
1 B neq ðd; qÞ d C  q
E ln d^max ¼ lnðd max Þ  lnB C E ½D 
2½CLN ðq2  qÞ @ 1  2CLN A d Var Id
1þ rd  1 g ðd; qÞ ¼ ðA11Þ
neq ðd; 1Þ E ½ s 2 ðD Þ
h
i    h  
i
h
i Var ln m q;d þ q2 Var ln m
 1;d  2qCov ln m
 1;d ; ln m
 q;d
^
Var ln dmax ¼
Figure 8 compares these analytical findings with empirical
½CLN ðq2  qÞ2 results.
ðA6Þ
[85] Acknowledgments. This work was sponsored by the National
Science Foundation under grant EAR-0910721. The UK Meteorological
Office provided the Heathrow data set, which was used as part of collabo-
Now we turn to the calculation of the equivalent inde- rative work with Christian Onof. The Portuguese data were provided by
pendent sample size neq(d,q), which we express as M.F.E.S. Coelho of the Institute of Meteorology in Lisbon. We gratefully
neq(d,q) = g(d,q)n(d) where n(d) is the actual sample size. acknowledge Tanvir Ahmed for the processing and quality control of the
The factor g(d,q) theoretically depends on d, q, the distri- NOAA data.
bution of storm duration D, and the multifractal parameters
(Cb, CLN, dmax). Below we derive this factor for the case References
when Cb = 0. Setting Cb = 0 simplifies the analysis and Becchi, I., and F. Castelli (1989), Lettura ed archiviazione digitali di regis-
produces factors that are accurate for the low values of Cb trazioni pluviografiche ad alta risoluzione temporale, Tech Rep. 1/89,
found inside storms. We also set m = 1 without affecting Dep. of Civ. Eng., Univ. of Florence, Florence, Italy.
g(d,q). de Lima, M. I. P., and J. L. M. P. de Lima (2009), Investigating the multi-
fractality of point precipitation in the Madeira archipelago, Nonlinear
[82] Consider two nonoverlapping intervals of length Processes Geophys., 16, 299–311, doi:10.5194/npg-16-299-2009.
d = dmax/r inside a storm of duration D and let do = dmax/ de Montera, L., L. Barthès, C. Mallet, and P. Golé (2009), The effect of
ro ≥ 2d be the length of the smallest interval that includes rain–no rain intermittency on the estimation of the universal multifractals
both d intervals. The values of Id in the two d intervals are model parameters, J. Hydrometeorol., 10, 493–506, doi:10.1175/
2008JHM1040.1.
denoted by Id, 1 and Id, 2. First we note that Deidda, R. (2000), Rainfall downscaling in a space-time multifractal
h i  2KLN ðqÞ framework, Water Resour. Res., 36(7), 1779–1794, doi:10.1029/
r 2 2CLN ðq qÞ
2
2000WR900038.
q q
E Id;1 Id;2 ¼ roKLN ð2qÞ ðE½Z q Þ2 ¼ ro2CLN q rd Dixon, M. (1994), Automated storm identification, tracking and forecasting
ro
—A radar-based method, PhD thesis, Univ. of Colo., Boulder.
ðA7Þ Gomes, A. M., and G. Held (2005), Characterization of storm properties
during the TroCCiBras experiment, paper presented at 11th Conference
Therefore, the covariance between Id,1 and Id,2 is on Mesoscale Processes and 32nd Conference on Radar Meteorology,
Am. Meteorol. Soc., Albuquerque, N. M.
h i h i    Gupta, V. K., and E. C. Waymire (1993), A statistical analysis of meso-
q q q q 2 scale rainfall as a random cascade, J. Appl. Meteorol., 32, 251–267,
Cov Id;1 Id;2 ¼ E Id;1 Id;2  E Idq
doi:10.1175/1520-0450(1993)032<0251:ASAOMR>2.0.CO;2.
2 2CLN ðq2 qÞ 2CLN ðq2 qÞ Hershfield, D. M. (1961), Rainfall frequency atlas of the United States for
¼ ro2CLN q rd  rd durations from 30 minutes to 24 hours and return periods from 1 to
2CLN ðq2 qÞ
h 2
i 100 years, Weather Bur. Tech. Pap. 40, U.S. Dep. of Commer.,
¼ rd ro2CLN q  1 ðA8Þ Washington, D. C.
Hubert, P., Y. Tessier, S. Lovejoy, D. Schertzer, F. Schmitt, P. Ladoy,
J. P. Carbonnel, and S. Violette (1993), Multifractals and extreme rainfall
LN(q q) events, Geophys. Res. Lett., 20(10), 931–934, doi:10.1029/93GL01245.
2
whereas the variance of Iqd is Var[Iqd] = r2C
d Huff, F. A. (1967), Time distribution of rainfall in heavy storms, Water
2CLNq2
[rd  1]. Resour. Res., 3(4), 1007–1019, doi:10.1029/WR003i004p01007.

15 of 16
W08516 VENEZIANO AND LEPORE: SCALING OF TEMPORAL RAINFALL W08516

Koutsoyiannis, D., and E. Foufoula-Georgiou (1993), A scaling model for a Serinaldi, F. (2010), Multifractality, imperfect scaling and hydrological
storm hyetograph, Water Resour. Res., 29(7), 2345–2361, doi:10.1029/ properties of rainfall time series simulated by continuous universal multi-
93WR00395. fractal and discrete random cascade models, Nonlinear Processes
Lakshmanan, V., R. Rabin, and V. DeBrunner (2003), Multiscale storm iden- Geophys., 17, 697–714, doi:10.5194/npg-17-697-2010.
tification and forecast, J. Atmos. Res., 67-68, 367–380, doi:10.1016/ Sun, X., and A. P. Barros (2010), An evaluation of the statistics of rainfall
S0169-8095(03)00068-1. extremes in rain gauge observations, and satellite-based and reanalysis
Langousis, A., and D. Veneziano (2007), Intensity-duration-frequency products using universal multifractals, J. Hydrometeorol., 11, 388–404,
curves from scaling representations of rainfall, Water Resour. Res., 43, doi:10.1175/2009JHM1142.1.
W02422, doi:10.1029/2006WR005245. Tessier, Y., S. Lovejoy, and D. Schertzer (1993), Universal multifractals:
Lavallée, D., D. Schertzer, and S. Lovejoy (1991), On the determination Theory and observations for rain and clouds, J. Appl. Meteorol., 32,
of the codimension function, in Non-linear Variability in Geophysics: 223–250, doi:10.1175/1520-0450(1993)032<0223:UMTAOF>2.0.CO;2.
Scaling and Fractals, edited by D. Schertzer and S. Lovejoy, pp. 99–109, Thomopoulos, N. T., and A. C. Johnson (2004), Some measures on the
Kluwer Acad., Dordrecht, Netherlands, doi:10.1007/978-94-009-2147-4_7. standard bivariate log-normal distribution, in Proc. Decision Sci. Inst.,
Lovejoy, S., and D. Schertzer (1995), Multifractals and rain, in New Uncer- 2004, pp. 1721–1726, Decision Sciences Institute, Atlanta, Ga.
tainty Concepts in Hydrology and Hydrological Modelling, edited by Upton, G. J. J. (2002), A correlation-regression method for tracking rain-
A. W. Kundzewicz, pp. 61–103, Cambridge University Press, Cambridge, storms using rain-gauge data, J. Hydrol., 261, 60–73, doi:10.1016/
U. K., doi:10.1017/CBO9780511564482.009. S0022-1694(01)00618-7.
Mandapaka, P. V., P. Lewandowski, W. E. Eichinger, and W. F. Krajewski Veneziano, D., and P. Furcolo (1999), A modified double trace moment
(2009), Multiscaling analysis of high resolution space-time lidar-rainfall, method of multifractal analysis, Fractals, 7(2), 181–195, doi:10.1142/
Nonlinear Processes Geophys., 16, 579–586, doi:10.5194/npg-16-579- S0218348X99000207.
2009. Veneziano, D., and P. Furcolo (2002), Multifractality of rainfall and
Mandelbrot, B. B. (1974), Intermittent turbulence in self-similar cascades: intensity-duration-frequency curves, Water Resour. Res., 38(12), 1306,
Divergence of high moments and dimension of the carrier, J. Fluid 12 pp., doi:10.1029/2001WR000372.
Mech., 62, 331–358, doi:10.1017/S0022112074000711. Veneziano, D., and P. Furcolo (2003), Marginal distribution of stationary
Menabde, M., and M. Sivapalan (2000), Modeling of rainfall time series and multifractal measures and their Haar wavelet coefficients, Fractals,
extremes using bounded random cascades and levy-stable distributions, 11(3), 253–270, doi:10.1142/S0218348X03002051.
Water Resour. Res., 36(11), 3293–3300, doi:10.1029/2000WR900197. Veneziano, D., and P. Furcolo (2009), Improved moment scaling estimation
Molini, A., G. G. Katul, and A. Porporato (2009), Revisiting rainfall clus- for multifractal signals, Nonlinear Processes Geophys., 16, 641–653,
tering and intermittency across different climatic regimes, Water Resour. doi:10.5194/npg-16-641-2009.
Res., 45, W11403, doi:10.1029/2008WR007352. Veneziano, D., and P. Iacobellis (2002), Multiscaling pulse representation
Molnar, P., and P. Burlando (2005), Preservation of rainfall properties in of temporal rainfall, Water Resour. Res., 38(8), 1138, doi:10.1029/
stochastic disaggregation by a simple random cascade model, Atmos. 2001WR000522.
Res., 77, 137–151, doi:10.1016/j.atmosres.2004.10.024. Veneziano, D., and A. Langousis (2005), The maximum of multifractal
Molnar, P., and P. Burlando (2008), Variability in the scale properties of cascades: Exact distribution and approximations, Fractals, 13, 311–324.
high-resolution precipitation data in the Alpine climate of Switzerland, Veneziano, D., A. Langousis, and P. Furcolo (2006a), Multifractality
Water Resour. Res., 44, W10404, doi:10.1029/2007WR006142. and rainfall extremes: A review, Water Resour. Res., 42, W06D15,
Nykanen, D., and D. Harris (2003), Orographic influences on the multiscale doi:10.1029/2005WR004716.
statistical properties of precipitation, J. Geophys. Res., 108(D8), 8381, Veneziano, D., P. Furcolo, and P. Iacobellis (2006b), Imperfect scaling of
doi:10.1029/2001JD001518. time and space-time rainfall, J. Hydrol., 322(1–4), 105–119.
Olsson, J. (1995), Limits and characteristics of the multifractal behaviour of Verrier, S., L. De Montera, L. Barthès, and C. Mallet (2010), Multifractal
a high-resolution rainfall time series, Nonlinear Processes Geophys., 2, analysis of African monsoon rain fields, taking into account the zero
23–29, doi:10.5194/npg-2-23-1995. rain-rate problem, J. Hydrol., 389, 111–120, doi:10.1016/j.jhydrol.
Olsson, J., J. Niemczynowicz, and R. Berndtsson (1993), Fractal analysis 2010.05.035.
of high-resolution rainfall time series, J. Geophys. Res., 98(D12), Verrier, S., C. Mallet, and L. Barthès (2011), Multiscaling properties of rain
23,265–23,274, doi:10.1029/93JD02658. in the time domain, taking into account rain support biases, J. Geophys.
Over, T. M., and V. K. Gupta (1996), A space-time theory of mesoscale Res., 116, D20119, doi:10.1029/2011JD015719.
rainfall using random cascades, J. Geophys. Res., 101, 26,319–26,331,
doi:10.1029/96JD02033.
Schertzer, D., and S. Lovejoy (1987), Physical modeling and analysis of rain
and clouds by anisotropic scaling multiplicative processes, J. Geophys.
Res., 92(D8), 9693–9714, doi:10.1029/JD092iD08p09693.

16 of 16

You might also like