1. Introduction

Rain is a complex and highly intermittent phenomenon and therefore difficult to model and even to measure. During the past decades, thanks to advances in the turbulence theory and in the description of multiplicative cascades, various models based on scale invariance and multifractals have been proposed (see, e.g., Veneziano et al. 1996; Over and Gupta 1996; Mazzarella 1999; Deidda 2000). These models open the way to a realistic description of rain, since it is not possible to dissociate rain from atmospheric turbulence. In this paper, we focus on the universal multifractals framework and more specifically on the fractionally integrated flux (FIF) model developed by Schertzer and Lovejoy (1987, 1997). This model was chosen because of its strong links to turbulence theory and because most multiplicative processes tend to resort to it. It has been applied successfully to a wide variety of phenomena, from cloud radiance (Lovejoy and Schertzer 2006) to ocean color (Lovejoy et al. 2001), or to financial assets (Schmitt et al. 1999).

As far as rain is concerned, this model has been tested against experimental data measured by rain gauges, meteorological radars [see Lilley et al. (2006) for a review], or—more recently—remote sensing (Lovejoy et al. 2008). Surprisingly, although this type of data is useful in general climatologic or weather statistical studies, very little is known in scientific literature about high-resolution rain-rate time series retrieved from disdrometer measurements. The latter provide individual drop data, thus leading to a large number of explored time scales and overcoming some limitations of others’ instruments—namely, the uncertainty affecting rain-rate retrieval in satellite and radar measurements and the variability of the sampling frequency of standard rain gauges due to the integrating effect of the tipping bucket (de Lima 1998). In this paper, measurements from the dual-beam spectropluviometer (DBS) developed at the Centre d’étude des Environnements Terrestre et Planétaires (CETP) are analyzed. The datasets include rain-rate time series covering several months from three different locations. The experiments are described in section 4, and the multifractal analysis techniques are presented in section 5.

Rain-rate time series have already been extensively studied within the universal multifractals framework, and many authors have reported similar results [typically H ≈ 0, α ≈ 0.5–0.7, and C₁ ≈ 0.4–0.6, using daily accumulations in Tessier et al. (1993, 1996), Hubert et al. (1993), Ladoy et al. (1993), and Fraedrich and Larnder (1993), or using accumulations during about 10 min in de Lima and Grasman (1999), de Lima (1998), and Hubert et al. (2002)]. However, when looking at high-resolution DBS data, a break in the scaling is observed at higher frequencies. This deviation from the model [previously pointed out by Fraedrich and Larnder (1993), Pathirana et al. (2004), and Onof et al. (1996)] leads to the following question: do smaller time scales have different scaling properties from larger time scales? In this paper, we investigate the possibility that this problem may not be due to physical reasons but to an artifact related to rain–no rain intermittency. Rain-rate time series include many no-rain periods with zero rain rates, which universal multifractals are not designed to model, and as suggested by Schmitt et al. (1998), the effect of this rain–no rain intermittency on the estimation of the universal multifractals model parameters has to be taken into account. This effect will indeed be shown to possibly lead to an artificial scaling at low resolutions and thus to biased multifractal parameters (section 3). The data are, therefore, analyzed on an event-by-event basis to avoid this difficulty (section 6). Long rain-rate time series that include rain–no rain intermittency (section 7) will also be analyzed to better characterize the effect of rain–no rain intermittency. Using the new parameters found and on a simple threshold, a new model that simulates high-resolution rain-rate time series will be presented (section 8) as well as the properties of the corresponding synthetic time series (section 9).

2. The fractionally integrated flux model

The FIF model is based on the concept of multiplicative cascades that had been suggested by Richardson (1922) in the context of turbulence and notably developed by Kolmogorov (1941), Yaglom (1966), and Mandelbrot (1974). According to this model, the kinetic energy injected into the system at a large scale is transferred by a conservative multiplicative process to smaller and smaller scales and finally dissipated as heat. More generally, this type of process can be modeled by (i) uniformly distributing a given quantity during a given interval; (ii) splitting this interval into subintervals; (iii) assigning these subintervals the original given quantity multiplied by a random variable whose law does not depend on scale; and (iv) repeating steps (ii) and (iii) iteratively. By construction, such processes exhibit interesting scale invariance properties and to describe them, it is necessary to introduce the concept of fractal dimension.

a. The fractal dimension and codimension

The definition of the fractal dimension based on the box-counting method will be recalled here [see Falconer (2003) for more details and the classical example of Cantor dust and other definitions of the fractal dimension]. Assume an observation period of duration T including rain periods and drought periods. Then split this original interval into contiguous subintervals of length Δt (here and throughout, the scale is given by Δt and the resolution by the number of subintervals λ = T/Δt). Next, count the number of subintervals n_rain(λ) where rain is present. If this number is a power law of resolution, then the rain process is fractal and the power law exponent D_f is its fractal dimension (here and throughout, ∼ stands for equality within slowly varying and constant factors):

The fractal dimension and the fractal codimension are related to each other by the following equation in a one-dimensional embedding space:

The fractal codimension is an interesting concept, because it characterizes the probability that a randomly chosen subinterval I_λ contains rain:

For the rain-rate process D_f ≈ 0.8, Lavergnat and Golé (1998) found 0.82 for time scales ranging from 10⁻² to 10⁵ min, and Hubert and Carbonnel (1989) found 0.8 for time scales ranging from 1 day to 45 yr.

b. Singularities and multifractals

The concept of fractal dimension is restricted to ensembles and therefore to the binary case rain or no rain. To study the scale invariance properties of a process whose intensities vary across a whole range of possible values, the multifractal formalism must be introduced. The idea is that the fractal dimension is not unique but rather depends on the intensity that is used to define the ensemble. Moreover, the scale-dependent notion of intensity has to be left aside, and the scale invariant concept of singularity (denoted γ) has to be used instead. The intensity related to a given singularity is the γ^th power of resolution λ, so that when the scale varies, the intensity is modified but not the singularity. The basic equation of the “codimension multifractal formalism” is as follows [see Schertzer et al. (2002) for a detailed introduction to multifractal processes and multiplicative cascades]:

where Φ_λ is the process at resolution λ and c(γ) the fractal codimension related to the singularity γ. Since the probability density is related to the statistical moments of Φ_λ by the Mellin transform, Eq. (4) can equivalently be written (Schertzer et al. 2002, section 4.2) as

where <·> is the mean operator, q the order of the moment, and K(q) the so-called moment scaling function. Here, K(q) and c(γ) are linked to each other via the Legendre transform, which is a one-to-one relation (Parisi and Frisch 1985):

Here, K(q) is a convex function, and it can easily be shown that K(0) = 0 and K(1) = 0. The latter property is a consequence of the assumption of conservation, namely, that the mean of the process is a constant M regardless of resolution,

The notion of conservative process is derived from turbulence theory, according to which the energy flux density between smaller and smaller eddies is assumed to be constant down to the dissipation scale. Multiplicative cascades with a constant mean are, therefore, called conservative cascades.

c. Universal multifractals and fractional integration

By densifying the cascade, that is, by increasing the number of multiplicative steps between the inner and the outer scales of the cascade, it has been shown (Schertzer and Lovejoy 1997) that multifractal cascades generally tend toward a universal class of processes called universal multifractals, whose moment scaling function K(q) can be described by only two parameters (α and C₁),

Here, C₁ is the fractal codimension of the level providing the dominant contribution to the mean of the process; α describes how rapidly the fractal codimension varies with the singularity (α varies from 0 to 2; α = 0 is the monofractal case; and α = 2 is the lognormal case).

In general, geophysical processes are not stationary and an additional fractional integration is required, which leads to the following expression for the rain-rate process (see the appendix for a formal definition of fractional integration):

where ΔR(Δt) is the rain-rate gradient over time lags of length Δt and where parameter H is related to the degree of smoothness. When Φ_λ is a conservative cascade, the process is referred to as a fractionally integrated flux [see Gagnon et al. (2006) for a comparison with classical models, such as the fractional Brownian motion or fractional Lévy motion]. Combining Eqs. (9) and (5), the q^th-order structure function of the rain-rate process can be written as

where ζ(q) is known as the structure function exponent and has the following form:

3. The effect of rain–no rain intermittency

By definition, the FIF model cannot generate zero-value periods, because it is a fractional integration of a multiplicative cascade with strictly positive multiplicative increments. Therefore, it is not clear whether it would be possible to perform a multifractal analysis of time series that include rain–no rain intermittency. This section shows that the effect of rain–no rain intermittency may strongly affect the estimation of the universal multifractals model parameters.

As for the estimation of H, the effect of the rain–no rain intermittency can be understood by considering that, for Δt sufficiently long, if R(t) ≠ 0, the probability of having R(t + Δt) = 0 is close to 1 and therefore

Thus, for Δt sufficiently long, the first-order structure function does not depend on Δt, and H is found to be 0.

Concerning the estimation parameters of the stationary cascade (α and C₁) with the moment scaling function K(q), a simplified case may be considered. Since rain events have a limited extension in space and time, assume that, at a sufficiently long time scale Δt, the successive averaging operations performed to reconstruct the conservative cascade are roughly similar to the degeneracy of a rectangular function (a function equal to zero everywhere except over an interval of duration Δt). Let Θ_λmax be the rectangular function at resolution λ_max = T/Δt and θ_λmax the level of the rectangle itself. When Θ_λmax is degraded to resolution λ_max/2 by averaging pairs of contiguous values, since the rectangle is surrounded by zeros, its level falls to

If this operation is repeated n times, Eq. (13) can be generalized to

Then, the qth-order moments of the rectangular function at resolution λ_max/2ⁿ are easily calculated because, at all scales, only one value is not zero. This value is the level of the rectangle θ_max/2ⁿ raised to the qth power and divided by the resolution λ_max/2ⁿ:

Modifying Eq. (15) to make the scaling explicitly appear leads to

Then, by taking the logarithm of both sides and since λ ≡ λ_max/2ⁿ,

Consequently, a scaling is found and the corresponding moment scaling function is

A comparison between Eq. (18) and the theoretical form of K(q) given by Eq. (8) leads to α = 0 and C₁ = 1. This simplified example shows that, at low resolutions, rain–no rain intermittency may result in an artificial scaling, with α close to 0 and C₁ close to 1. To avoid this artifact, the multifractal analysis of the spectropluviometer data will be performed on an event-by-event basis in section 6.

4. Measurements

The rain rate has been measured by means of a DBS (Delahaye et al. 2006). The DBS is more accurate than a previously used instrument, the optical spectropluviometer (OSP; Hauser et al. 1984), because it comprises two closely spaced flat optical beams and an increased catchment surface (100 against 50 cm² for commercial instruments). With this instrument, false detections due to air turbulence can be reduced and raindrops as small as 0.3 mm can be detected unambiguously by verifying that each drop actually crosses both beams. In the laboratory, the instrument’s accuracy for particle diameters is 1% for bias and 3% for standard deviation. This accuracy has also been tested in the outside turbulent atmosphere with beads larger than 1 mm. For smaller diameters, calibration can only be performed indoors. The DBS also provides an accurate estimation of the vertical velocity of raindrops. Its outputs are raindrop diameters, velocities, and times of arrival (in milliseconds). These outputs were transformed into rain-rate time series sampled at 1 s. However, the retrieved rain rate is not accurate with a 1-s integration time, and the time series have to be resampled at a longer sampling period, typically 60 s (see sections 5a and 6 for more details). The standard deviations of the rain-rate estimation error are 31% and 4% for sampling periods of 1 and 60 s, respectively, but the error is much larger for smaller rain rates (Mallet and Barthès 2009).

The experimental dataset (Table 1) comprises three subsets collected in different climatic areas: Paris [Site Instrumental de Recherche par Télédétection Atmosphérique (SIRTA) experimental site, France], Iowa City [Disdrometer Evaluation Experiment (DEVEX) campaign, United States; Krajewski et al. 2006], and Djougou [African Monsoon Multidisciplinary Analyses (AMMA) campaign, Benin]. The latter corresponds to the African monsoon period and is characterized by strong rain events.

5. Multifractal analysis techniques

This section describes the analysis tools needed to perform the multifractal analysis. Each technique is illustrated based on an example derived from a strong rain event observed on 12 July 2002 during the DEVEX experiment in Iowa (Fig. 1).

a. Power spectrum analysis

The power spectrum is a useful tool to decide whether the process is scaling or not. If scaling is present, the power spectrum, plotted in a log–log graph, should be a straight line. Then, it has to be decided whether the process is a stationary cascade or a FIF. This can be easily done by considering the spectral slope β: if the process is a stationary cascade, then, as a consequence of the Wiener–Khintchine theorem, the theoretical slope is

Therefore, if β >1, the process cannot be stationary. An additional fractional integration is required, which leads to the FIF model with the following theoretical slope:

The power spectrum of the 12 July 2002 event is presented in Fig. 1. For this particular event, β is found to be 1.72, and the process is therefore a FIF.

The limits of the power spectrum scaling are also important, because they indicate the maximum and minimum scales over which it is meaningful to perform a multifractal analysis. For the 12 July 2002 event, the spectrum is clearly leveling off at frequencies above 1/64 Hz (Fig. 1) and therefore the time series has to be averaged over 32-s time lags to keep only the scaling part of the power spectrum. This is a fundamental feature because, in the case of a FIF, the multifractal analysis depends on the gradient of the process at the finest time scale [see Eq. (21)], which therefore would have no physical meaning with a very noisy process.

c. Moment scaling analysis

To check the scaling of the statistical moments of Φ_λ [Eq. (5)] and to estimate the values of α and C₁, the multiplicative cascade Φ_λ has to be retrieved over a wide range of scales. This is done first by deducing Φ_λ at the finest scale (denoted Φ_λmax) from the rain-rate time series, and then by reconstituting the multiplicative cascade at longer time scales [see Gagnon et al. (2006) for an instructive example]. Theoretically, the first step requires an H-order fractional differentiation of the rain-rate process. However, since H ≈ 0.5, an adequate approximation (Lavallée et al. 1993) is to take the absolute value of the gradient at the finest time scale Δt_min (i.e., the highest resolution λ_max), which is normalized to have M = 1:

The multiplicative cascade is then reconstituted at longer time scales by consecutive averagings of pairs of contiguous values. Finally, the statistical moments are calculated at each scale and plotted in a log–log graph as functions of the scale. According to Eq. (5), these functions should be straight lines with slopes equal to K(q), where q is the order of the moment. Such straight lines are indeed found for the 12 July 2002 event (Fig. 3, top), so that the corresponding moment scaling function K(q) can be estimated (Fig. 3, bottom).

d. Estimation of α and C₁

The multifractal parameters α an C₁ can be estimated using the moment scaling function K(q) given in Eq. (8). However, the analysis of synthetic time series created with known multifractal parameters shows that α is not retrieved correctly by directly fitting K(q) (the simulation technique is described in the appendix). A more sophisticated technique known as double trace moment (DTM) can improve the accuracy of the estimation (see Lavallée et al. 1993 for details). For that purpose, the η^th power of Φ_λmax is first computed and the multiplicative cascade is then reconstituted at larger time scales. The statistical moments of such a process, M_λ(η,q), are scaling, and their scaling exponent is denoted K(η,q):

Exponents K(η,q) have the following property in the framework of universal multifractals:

Therefore, when q has a fixed value (different from the special case 0 or 1), a log–log plot of K(η, q) as a function of η should give a straight line with a slope of α. The scaling of the double trace moments and the corresponding scaling function K(η,q) are shown in Fig. 4 for the 12 July 2002 event. The fit of this scaling function to its linear part leads to α = 1.69. The classical moment scaling function K(q) is then fitted according to Eq. (8) to retrieve C₁, which is found to be 0.12.

6. Event-by-event analysis

Since the presence of rain–no rain intermittency may bias the estimated parameters, it was chosen to avoid this difficulty by performing the multifractal analysis of spectropluviometer data on an event-by-event basis. In this section, 30 rain events were analyzed separately (10 for each of the three datasets, see Tables 2 –4). These rain events were extracted in such a way that all zero values were excluded. The rain-rate process is thus said to be apparently uninterrupted, because although there are no zero values at the resolution at which they are analyzed, such values would appear if the resolution was increased.

The power spectrum was calculated for each rain event, and the highest resolution at which the scaling is valid was determined (so as to avoid flattening the power spectrum at the highest frequencies). For most events, a 64-s averaging time was used. However, for certain events, a 32-s averaging time was sufficient. The leveling off of the power spectrum occurring at higher frequencies could be because the sampling frequency exceeds the inner scale of the multiplicative cascade (i.e., the dissipative scale in the turbulence framework). However, Lovejoy and Schertzer (2008) have shown that rain becomes decoupled from atmospheric turbulence at a spatial scale of 1 m. Since most of the drops are falling faster than 1 m s⁻¹ (Gunn and Kinzer 1949), the rain-rate process should at least scale up to the 1-s time scale. A more realistic explanation could be that, when the rain rate is calculated with an integration time that is too short, the number of raindrops detected by the instrument is too small and the measurement is not representative of the process (although the DBS has been designed with a relatively large catchment surface of 100 cm² to counter this limitation). Another explanation could be that the vertical velocity of the drops is affected at higher frequencies by the turbulence due to the proximity to the earth’s surface. However this point remains to be investigated.

The event-by-event multifractal analysis led to the multifractal parameters presented in Table 5. Instead of the values H ≈ 0, α ≈ 0.5–0.7, and C₁ ≈ 0.4–0.6 generally found in scientific literature (see references cited in section 1), significantly different values were found, namely, H ≈ 0.53, α ≈ 1.7, and C₁ ≈ 0.13. The fact that H is not 0 indicates that the process is not a stationary cascade but rather a nonstationary FIF. Although this result is consistent with those of Harris et al. (1996) who used high-resolution rain-rate time series (15-s integration time), this contradicts most other studies in which the cascade is found to be stationary. This difference can be explained by the fact that these studies used low-resolution data and therefore were restricted to the flat part of the spectrum observed at the lower frequencies. The standard deviation of H is large (27% of its mean value). Therefore, it is not possible to conclude that H is constant for all rain events. However, H was difficult to estimate because of its sensitivity to the noise level and the sample size, which is small in an event-by-event analysis. Thus, the variability of H might be due to instrumental reasons and, to decide whether H is constant or not, it would be interesting to increase the catchment surface so that finer time scales could be explored.

As to the other parameters, α is large, which means that the multiplicative cascade is close to the lognormal case. As expected, the value of C₁ is smaller than the one found in scientific literature, since no rain intermittency artificially increases the mean level of intermittency. Note that the values of α and C₁ are remarkably similar regardless of the dataset and that their variability is small, although the data have been collected in very different climatic areas (the climate is continental for Iowa, temperate–oceanic for Paris, and tropical with monsoon for Djougou).

7. Analysis of long rain-rate time series

In this section, a multifractal analysis of the three rain-rate time series is performed without extracting individual events. Since these time series last several months, they inherently include zero-value periods, so that rain–no rain intermittency should have an effect.

Figure 5 shows the power spectrum of each time series. At time scales longer than 1 min, a scaling is observed up to the 1-h scale (corresponding to 1/7200 Hz in Fig. 5) and then the spectrum becomes flat. Between the 1-min and 1-h time scales, the spectral slope was found to be 1.76 for AMMA data, 1.79 for DEVEX data, and 1.77 for SIRTA data. Figure 6 shows the first-order structure function for each dataset and corresponding fits that were performed only for time lags shorter than 15 min. It was found that H = 0.55 for AMMA data, H = 0.52 for DEVEX data, and H = 0.5 for SIRTA data. At larger time intervals, a flattening is also observed. These results are consistent with the assessment of the effect of rain–no rain intermittency performed in section 3, because above a time interval that roughly corresponds to the typical duration of a rain event (with data sampled every minute), the first-order structure function flattens and yields H ≈ 0.

Figure 7 shows the scaling of moments and the corresponding moment scaling function K(q) for the AMMA dataset. This scaling also exhibits a break and, if only low resolutions are considered (time scales longer than 1 h), the parameters are found to be α = 0.24 and C₁= 0.63 [note that these values are the same whether the process is assumed to be a stationary cascade (the multiplicative cascade is reconstructed directly from the rain rate) or a nonstationary FIF (the multiplicative cascade is reconstructed from the absolute value of the gradient)]. These parameters are close to those suggested by other authors who used low-resolution data (α ≈ 0.5–0.7; C₁ ≈ 0.4–0.6) and consistent with section 3, where it is shown that rain–no rain intermittency may lower α and increase C₁.

8. Simulation of the rain–no rain intermittency

A new model called “thresholded FIF” is proposed, which simulates high-resolution rain-rate time series based on the new parameters and a simple threshold. A FIF process is first generated with H ≈ 0.53, α ≈ 1.7, and C₁ ≈ 0.13. Then lower values lying below a given threshold are set to 0. Those above the threshold are shifted downward by subtracting the threshold value. The threshold is chosen so that rain–no rain intermittency reproduces the fractal dimension of the rain “support.” Contrary to stationary cascades, the values of FIF processes do not have an absolute meaning. Therefore, it is possible to shift them up or down without breaking the scaling (this operation does not modify the differentiated process). Note that in this model, the threshold corresponds to an intensity of the fractionally integrated cascade, which means that it does not correspond to a unique singularity of the cascade itself.

9. Analysis of synthetic rain-rate time series

The length of the synthetic time series is set at 32 768 samples. Since the finest scale is 1 min, these time series simulate the rain-rate process over a period of three weeks, which may be considered to be the average lifetime of large-scale structures in the atmosphere. The parameters are set in accordance with the results given in section 6 (H = 0.53; α = 1.7; C₁ = 0.13). The threshold is set to obtain a fractal dimension of the rain “support” equal to 0.82 [this value has been reported by Lavergnat and Golé (1998), who also used a spectropluviometer]. It may be noticed that, with these parameters and this fractal dimension, only 5% of the process is not set to 0, which corresponds to the rain time percentage observed in the experimental data (at a 1-min time scale). Figure 8 shows the effect of thresholding on the power spectrum and on the first-order structure function. The effect of rain–no rain intermittency is clearly to flatten both curves at low resolutions. Figure 9 shows that the effect of the rain–no rain intermittency is to straighten out K(q) and to cause an artificial break in the scaling of moments. When K(q) is only estimated on the basis of the scaling at the lower resolutions, this leads to 0.1 < α < 0.5 and 0.4 < C₁ < 0.6, depending on the synthetic time series realization. This is consistent with the parameters found for long time series (section 7), with the parameters previously proposed in scientific literature (section 1) and with the expected effect of rain–no rain intermittency (section 3).

Moreover, the scale at which the break in the scaling is occurring can be roughly predicted. If there is no rain–no rain intermittency, the number of intervals during which rain is present decreases as 2⁻ⁿ, where n is the number of consecutive time averagings of pairs of contiguous intervals (0 corresponds to the highest resolution). If the rain support is assumed to have a fractal dimension D_f , then this number decreases as 2^−nDf. Therefore, 2^−nDf − 2⁻ⁿ gives the approximate value of the number of time intervals that will be averaged with zero values at step n during the reconstitution of the cascade. If it is assumed that the effect of rain–no rain intermittency becomes the major contributor to the scaling when the number of time intervals that are averaged with zero values is equal to the theoretical number of time intervals without rain–no rain intermittency, then the scaling should break at step n such that (2^−nDf − 2⁻ⁿ) = 2⁻ⁿ. This condition is actually equivalent to n = 1/C_f , which provides an estimation of the scale at which the break is expected to occur. In the case of rain, if we assume D_f to be 0.8, this leads to n = 5. Since the finest time scale is 1 min, five consecutive time averagings correspond to a time scale of 2⁵ = 32 min. It is in agreement with the breaks observed in Figs. 7 and 9 that occur between 32- and 64-min time scales.

The alternative model based on the parameters previously proposed in scientific literature (H ≈ 0; α ≈ 0.5–0.7; C₁ ≈ 0.4–0.6) has also been investigated through simulations. It consists in generating a stationary bounded cascade (H = 0; α < 1) and then in setting to 0 small values that are below a given singularity. It was found that with α = 0.6 and C₁=0.5, the singularity corresponding to D_f = 0.8 was very small and that setting the values below it to 0 had almost no effect on the scaling. This explains why the effect of rain–no rain intermittency was generally not considered. Conversely, in the model proposed here based on high-resolution data, the effect on rain–no rain intermittency is significant because, as previously mentioned, the threshold does not correspond to a unique small singularity of the cascade but to a given intensity of the fractionally integrated cascade. In other words, a differentiated “thresholded FIF” is close to a stationary cascade multiplied by an independent fractal support. This means that all kinds of singularities may be set to 0 by the threshold, not only the smallest singularities.

10. Conclusions

The presence of zero-value intervals in analyzed time series may cause a spurious break in the scaling and a change in the universal multifractals model parameters. The multifractal analysis was, therefore, performed on an event-by-event basis, so that the analyzed process is apparently uninterrupted. This approach led to new parameter estimates (H ≈ 0.53; α ≈ 1.7; C₁ ≈ 0.13) significantly different from those previously suggested in scientific literature (H ≈ 0; α ≈ 0.5–0.7; C₁ ≈ 0.4–0.6). In particular, the rain process is found to be a FIF, meaning that the multifractal analysis should not be performed directly on the rain rate but on the absolute value of its gradient.

A new model has been proposed to simulate high-resolution rain-rate time series including rain-no rain intermittency based on these new parameters and on a simple threshold. The resulting time series correctly reproduce the parameters that were found for individual events as well as the break of the scaling due to rain–no rain intermittency and the fractal dimension of the rain “support.”

This study underlines the need for improved analysis techniques that would be capable of modeling the degeneracy of the multiplicative cascade due to the averaging with zero values during the reconstitution. Moreover, although the parameters proposed in this study were estimated from uninterrupted rain periods, they are also affected by rain–no rain intermittency because the process is only apparently uninterrupted—that is, the measurement is actually an average of rain periods and no rain periods. As a consequence, future analysis techniques should also take into account the degeneracy of the fractionally integrated process itself due to the measurement and relate it to the resolution of the data.