Advances in Meteorology

Advances in Meteorology / 2017 / Article
Special Issue

Extreme Precipitation and Floods: Monitoring, Modelling, and Forecasting

View this Special Issue

Research Article | Open Access

Volume 2017 |Article ID 6954902 | https://doi.org/10.1155/2017/6954902

Said Arab Khan, Ijaz Hussain, Tajammal Hussain, Muhammad Faisal, Yousaf Shad Muhammad, Alaa Mohamd Shoukry, "Regional Frequency Analysis of Extremes Precipitation Using L-Moments and Partial L-Moments", Advances in Meteorology, vol. 2017, Article ID 6954902, 20 pages, 2017. https://doi.org/10.1155/2017/6954902

Regional Frequency Analysis of Extremes Precipitation Using L-Moments and Partial L-Moments

Academic Editor: Mouleong Tan
Received21 Dec 2016
Revised26 Jan 2017
Accepted07 Feb 2017
Published07 Mar 2017

Abstract

Extremes precipitation may cause a series of social, environmental, and ecological problems. Estimation of frequency of extreme precipitations and its magnitude is vital for making decisions about hydraulic structures such as dams, spillways, and dikes. In this study, we focus on regional frequency analysis of extreme precipitation based on monthly precipitation records (1999–2012) at 17 stations of Northern areas and Khyber Pakhtunkhwa, Pakistan. We develop regional frequency methods based on L-moment and partial L-moments (L- and PL-moments). The L- and PL-moments are derived for generalized extreme value (GEV), generalized logistic (GLO), generalized normal (GNO), and generalized Pareto (GPA) distributions. The -statistics and L- and PL-moments ratio diagrams of GNO, GEV, and GPA distributions were identified to represent the statistical properties of extreme precipitation in Northern areas and Khyber Pakhtunkhwa, Pakistan. We also perform a Monte Carlo simulation study to examine the sampling properties of L- and PL-moments. The results show that PL-moments perform better than L-moments for estimating large return period events.

1. Introduction

Hydraulic and hydrologic designs are key steps in planning of any water project. Any problem pitched at designing stage will result in the failure of design irrespective of the fact how correctly the other steps are taken. Hydrologists deal with water-related issues, problems of quantity, quality, and availability, in the society that known as hydrologic events. Stochastic methods are often used to understand sources of uncertainties in physical processes that give rise to observed hydrologic events, as precipitation and stream flow estimates depend on the past or future events. Several statistical methods offered to minimize and summarize the uncertainties of observed data and frequency analysis is one of them. It determines that how often a particular event will occur by estimating the quantile for return period of , where is the magnitude of the event that occurs at a given time and location.

Dalrymple [1] proposed regional frequency analysis (RFA) method for pooling various data samples. It is index-flood procedure in hydrology. Hosking et al. [2] studied the properties of probability-weighted moments (PWMs) method based on RFA method. This method is first used by Greis and Wood [3] and Wallis [4]. Cunnane [5] reviewed twelve methods of RFA and related regional PWMs algorithm.

Initially, PWMs are considered as an alternative parameter estimation method; however, it was difficult to interpret directly as measures of the shape and scale parameters of distribution. RFA can forecast the flood flow using empirical formula and unit-hydrograph procedure Subramanya [6], and it can also estimate the quantiles of extreme precipitation.

Hosking and Wallis [7] showed that RFA method based on L-moments is used to detect homogeneous regions, to select suitable regional frequency distribution, and to predict extreme precipitation quantiles at region of interest.

Whilst L-moment methodology is effective in estimating parameters, it may not valid for predicting high return period events. Wang [8] suggested that relatively small floods might create disturbance in the analysis. To overcome such situation, a censored sample can be used as by Cunnane [9].

Wang [8] proposed partial probability-weighted moments (PPWMs) for fitting the probability distribution function to the censored sample. Partial L-moments (PL-moments) are variants of the L-moments and similar to PPWMs. PL-moments method has used in fitting generalized extreme value (GEV) distribution for censored flood samples (see Wang [8, 10, 11] and Bhattarai [12]). Bhattarai [12] found that censoring flood samples are nearly thirty percent of basic L-moments.

Shabri et al. [13] used Trim L-moments (TL-moments) for the RFA and compared its performance with L-moments. Saf [14] determined hydrologically homogeneous region and regional flood frequency estimates by using index-flood technique along with L-moments for the West Mediterranean River, Turkey. L-moments method is also used to assign a suitable regional distribution for the individual subregions and to assess their homogeneity; see Abolverdi and Khalili [15]. Hussain and Pasha [16] suggested the regional flood frequency analysis based on L-moments. They used discordancy measure for data screening and used the four-parameter Kappa distribution with 500 simulations for the heterogeneity analysis. Zakaria et al. [17] used the PL-moments technique and found another link for the homogeneity analysis. Shahzadi et al. [18] showed that the generalized normal (GNO) distribution is suitable for the regional quantile estimation at maximum return period and the GEV distribution for the overall regions at low return period based on relative RMSE and relative absolute bias. Most commonly used statistical distributions for high climate modeling are as follows: the logistic distribution with three parameters, lognormal distribution with three parameters, Log Pearson type III, GEV, and generalized Pareto (GPA) (Coles [19]; Katz et al. [20]; Abida and Ellouze [21]; Feng et al. [22]; Yang et al. [23]; Villarini et al. [24]; Zakaria et al. [17]; She et al. [25]). Moreover, GEV and GPA distributions are suitable if data contains extremes values. Over the years, the GNO, GEV, generalized logistic (GLO), and GPA distributions have been widely employed in the extreme value estimation of annual flood peaks. In this study, we aim to develop RFA method based on L- and PL-moments approach. Our proposed method can be used at all levels of regional analysis such as identification of homogeneous regions, identification, and also testing of the suitable probability regional frequency distribution based on -statistic and the L- and PL-moment ratio diagram and estimation of the flood quantile at site of interest. We use 17 sites of Northern areas and Khyber Pakhtunkhwa as a case study to perform the analysis.

We explain the methodology of the L-moments and PL-moments in Section 2.2. Section 2.3 shows the application of the RFA where we choose the appropriate distribution for the regional analysis. Section 3.3 provides the estimation of quantile for both, the small and large return period. The results of our simulations study will be presented in Section 4.

2. Material and Methods

2.1. Study Area and Data Sources

The data was collected for this study from Karachi Data Processing Center through Pakistan Meteorological (PMD) Islamabad. Monthly precipitation data has been recorded from 1999 to 2012. There are 17 meteorological stations of Northern areas and Khyber Pakhtunkhwa, Pakistan. These stations are full precipitation regions that affect areas in Pakistan, where water is essential for hydropower and flood plains. Figure 1 shows the location of the study area and geographic distribution of precipitation stations. There is no missing value in this data set.

2.2. Statistical Methods
2.2.1. The L-Moments

Conventional moments method may be used for estimating the parameters of probability distribution. However, this approach has some serious drawbacks. Ratio of moment estimators is biased and often assumption of being normally distributed is violated, Wallis et al. [26]. Furthermore, it is sensitive to outliers, Pearson [27]. Therefore, it is unreliable for skewed distributions.

Hosking [28] proposed L-moments approach to overcome the above problems. The L-moments may precisely describe the statistical properties of hydrological information, and it can write as a linear function of the PWMs. The PWMs of order were properly described by Greenwood et al. [29] as where , and are the real numbers and is the cumulative distribution function of . A functional case is Therefore, where is the cumulative distribution function of a random variable and is a nonnegative integer of the real number that is . Therefore the first four L-moments, which are the linear combinations of the PWMs, are The L-moments have no units of measurement, which are called the L-moments ratios. The L-moments ratios, proposed by Hosking [28], are computed as where represents the L-coefficient of variation , represents the L-coefficient of skewness , and represents the L-coefficient of kurtosis .

The arranged sample is given as . Wang [8] stated that the statistic is an unbiased estimator of .

So where , and are the first four L-moments of the sample. And similarly are the sample L-moments ratios.

2.2.2. The Partial L-Moments

Wang [8, 10, 11] introduced a concept of partial probability-weighted moments (PPWMs) that will estimate the higher quantiles of flood flows. Data can be censored to the right tail or left tail.

Initially, PPWMs were to take out the smaller values from the process of distribution fitting because such values have slight influence on the frequency analysis and are nuisance to the fitting process.

The left tail PPWMs are defined by Wang [8, 11] as where which is the lower limit of the censored observations and is the censoring threshold value.

The PPWMs elongated form described by Wang [10] are to be given a censored sample as If the value of is starting from the zero, then the result of PPWMs will be the same as the usual PWMs. As is the arranged sample, Wang [10] describes the unbiased estimator of as where The censoring level, , is the prior selection of the number of censored sample data. The procedure that determines the number of sample data points are to be censored: where and are the lengths of sample which are to be censored and uncensored, respectively. Similarly, is the highest value of the censored sample. The first four PL-moments for the PPWMs are Similarly, the PL-moments ratios can be written as where , and denote the partial L-coefficient of variation , partial L-coefficient of skewness , and partial L-coefficient of kurtosis , respectively. Therefore, the first four sample PL-moments can be computed as And the first four sample PL-moments ratios can be computed as where , , and represent the sample partial L-moments ratios of the , , and , respectively. The derivation is L-moment and PL-moments are given in Appendix. In the present study, different level of censoring threshold is selected.

2.3. Regional Frequency Analysis

Hosking and Wallis [7, 30] identified the following four steps to explain the procedure of the RFA:(1)Data screening(2)Designing of the homogeneous region(3)Selection of an appropriate probability distribution(4)Parameters estimation of the appropriate probability distribution

2.3.1. Data Screening

We screened data anomalies before applying any statistical analysis.

2.3.2. Discordance Test

Hosking and Wallis [7] suggested a discordancy measure test that recognizes the locations where sample L-moments are marked contrarily from the most other locations. Locations with the large flaws in the data will be flagged as discordant.

The discordancy test for a region containing locations, for site , is proposed by Hosking and Wallis [7] as follows:where is the vector containing the three sample L-moments ratios for the site expressed as is the average vector of for the overall region that is and is the covariance matrix for the sample that can be expressed as Broadly speaking, a location or a site is considered to be discordant from the whole region or group if the value of is larger than the critical value.

2.3.3. Heterogeneity Test

The homogeneity measure identifies homogenous regions, Hosking and Wallis [7]. It is also useful to tag the locations if they are plausible to handle as a homogeneous region. It estimates the amount of heterogeneity in the overall region. The heterogeneity test is computed as where and are representing the mean and standard deviation of the simulated values. Also, Here, , and are region average L-moments or PL-moments ratios. We assessed the heterogeneity of a region as suggested by Hosking and Wallis [7]:Region is acceptably homogeneous if .Region is possibly heterogeneous if .Region is definitely heterogeneous if .

2.4. Selection of the Appropriate Probability Distribution

Hosking and Wallis [7] proposed two approaches to select the distribution that fitted best the data: the L-moment ratios diagram and the -test. The L-moment ratios diagram is using the unbiased estimators, Hosking [28], Stedinger et al. [31], Vogel and Fennessey [32], and Hosking [33]. The L-moments ratio diagram is a plot of the computed values and the observed values of the distribution function. The curves indicate the hypothetical connections between and of the candidate distribution. The L-moment ratio diagrams have been proposed for discriminating between the candidate probability distributions in describing the regional information (Hosking [28]; Stedinger et al. [31]; Hosking and Wallis [7]). The L-moments ratio diagrams have been used as a component of probability distribution process for regional information (Schaefer [34]; Pearson [35]; Vogel and Fennessey [32], Vogel et al. [36]; Chow and Watt [37]; ÖnÖz and Bayazit [38]; Vogel and Wilson [39]; Peel et al. [40]).

Hosking and Wallis [7] suggested a measure to see how well the and of the fitted probability distribution match the regional average and of the observed information.

The measure goodness of fit for every single selected probability distribution is computed as follows: where represents the value of the of the fitted distribution, represents the weighted regional average , and represents standard deviation of the , which is obtained from the simulation of the Kappa probability distribution.

If the computed value of is equal to zero, the probability distribution will be the most suitable fit. If the computed value of -statistic is less than 1.64 at 90% confidence level (i.e., ), it will indicate that the distribution qualifies the goodness of fit criteria. If there are more than one distribution that qualify the criteria, the most suitable distribution has the minimum value.

3. Estimation

The sites’ information and statistic by using L-moments for the present study are presented in Table 1. In Table 1, mean represents the first sample L/PL-moments and , and are the sample L/PL-moments ratios of the , , and , respectively. The lower level censoring threshold is selected from 10 to 23%. Table 2 expresses the feasible threshold values according to the percentile technique along with Average Annual Occurrence Number (AAON). Jiang et al. [41] and Yuguo [42] suggested that the optimal threshold can be obtained if the values of AAON lie between and . Table 2 shows that 90th percentile observations are suitable for the optimum threshold selection of most of areas in the present study. We have 168 values in each station; according to the above table, Astore station has 3.2 threshold values due to which 16 values are being censored in 168. According to censored level, 10.2% censored level was selected. Similarly, 10.5% was selected for Balakot and Muzaffarabad. According to this process, maximum threshold level 22.3% was selected for Gupis. By using the above process for each station, 17 different censoring levels were selected. So we decide that the range from 10 to 23% of censoring level should be kept for selecting threshold values.


SitesL-momentsPL-moments
MeanMean

Astore37.6240.5180.3660.193 41.752 0.467 0.377 0.183
Balakot120.4500.4780.3220.192 133.259 0.428 0.335 0.198
Bunji14.2970.6090.4560.259 16.339 0.553 0.450 0.257
Chilas 15.0910.5850.3920.194 17.603 0.516 0.375 0.200
Cherat 49.0380.5800.3750.162 55.289 0.526 0.355 0.161
DI khan 28.8850.6890.5160.262 32.351 0.651 0.487 0.250
Dir 105.6940.4320.2160.098 116.741 0.378 0.222 0.097
Drosh 45.2840.5250.3170.136 50.316 0.474 0.307 0.137
Garhi Dupatta 111.3520.4540.2270.111 123.374 0.398 0.229 0.120
Gilgit 12.6770.5910.4120.228 14.290 0.539 0.399 0.239
Gupis 23.4460.6780.5290.331 30.299 0.583 0.506 0.350
Kakul 99.4310.4530.2470.102 110.068 0.399 0.254 0.093
Kotli 95.0700.5420.3320.136 106.903 0.487 0.314 0.138
Muzaffarabad 118.4390.4590.2840.182 131.122 0.405 0.301 0.194
Peshawar 46.2160.5720.3980.216 51.761 0.520 0.392 0.222
Saidu Sharif 85.6670.4550.2650.181 94.978 0.399 0.283 0.200
Skardu 20.4380.6160.45640.238 22.885 0.570 0.444 0.231


Station 80th 90th 95th 97.5th 99th

Astore
 Threshold 9.8 3.2 0.2 0 0
 AAON 2.43 1.21 0.64 0.43 0.43
Balakot
 Threshold 32.5 14.8 5.1 3 0
 AAON 2.43 1.29 0.57 0.29 0.14
Muzaffarabad
 Threshold 30.4 11.8 5 1 0
 AAON2.5 1.21 0.57 0.29 0.21
Garhi Dupatta
 Threshold 28.8 10 3.8 0 0
 AAON 2.43 1.21 0.57 0.29 0.29
Dir
 Threshold 31 20 5 0.05 0
 AAON2.43 1.21 0.570.290.21
Kakul
 Threshold 28.9 13.4 2.1 0.5 0
 AAON1.641.29 0.57 0.29 0.21
Kotli
 Threshold13 60.05 0 0
 AAON 2.43 1.35 0.57 0.36 0.36
Saidu Sharif
 Threshold 25.6 8.3 2 0 0
 AAON2.431.21 0.57 0.36 0.36
cherat
 Threshold 50.05 0 0 0
 AAON 2.43 1.43 0.93 0.93 0.93
Peshawar
 Threshold6 0.05 00 0
 AAON2.43 1.29 0.57 0.57 0.57
Drosh
 Threshold7.6 1.50.2 0.05 0
 AAON 2.43 1.21 0.57 0.5 0.21
D I Khan
 Threshold 0.3 00 00
 AAON 2.57 1.29 1.29 1.29 1.29
Gupis
 Threshold0 000 0
 AAON 2.71 2.71 2.71 2.71 2.71
Skardu
 Threshold 20.3 0 0 0
 AAON2.43 0.71 0.71 0.71 0.71
Chilas
 Threshold1 0.5 0 0 0
 AAON1.71 1.07 1.07 1.07 1.07
Bunji
 Threshold1.3 0 0 0 0
 AAON 1.5 1.5 1.5 1.5 1.5
Gilgit
 Threshold 1.2 0.05 0 0 0
 AAON2.43 1.35 0.71 0.71 0.71

3.1. Regional Frequency Analysis

The following four steps are considered as prerequisite for frequency analysis, Hosking and Wallis [7, 30]:(1)Data screening(2)Designing of the homogeneous region(3)Selection of an appropriate probability distribution(4)Parameters estimation of the appropriate probability distribution

3.1.1. Data Screening

In this study, we use secondary data after carefully examining all locations for abnormalities and missing observations. Therefore, we use 14 years of data for RFA that were obtained from seventeen locations.

3.1.2. Discordance Test

Table 3 shows result of (18) for 17 locations of this study region. It can be observed from the results of L/PL-moments in Table 3 that the value of -statistic varies from 0.07 to 2.44. If is greater than 3, the location is considered to be discordant from the rest of the regional data, Hosking and Wallis [7]. In this study region, no location is diagnosed as discordant . Therefore, we use all data for the development of the RFA based on L-moments and PL-moments.


Name of site L-moments ()PL-moments ()

Astore 2.05 2.44
Balakot 1.42 0.85
Bunji 0.55 0.60
Chilas 0.29 0.07
Cherat 0.79 0.64
DI khan 1.29 1.90
Dir 0.82 0.86
Drosh 0.51 0.43
Garhi Dupatta 1.43 1.33
Gilgit 0.26 0.59
Gupis 2.37 2.31
Kakul 0.82 1.22
Kotli 0.77 0.59
Muzaffarabad 0.99 0.85
Peshawar 0.07 0.08
Saidu Sharif 1.91 1.57
Skardu 0.67 0.65

3.1.3. Regional Heterogeneity Measure

The next step is the formation of the homogeneous region, which is conventionally tougher and needs the higher number of subjective judgments. The homogeneity conditions are defined as the locations that have the same frequency distributions.

For the present study area, realization of the Kappa probability distribution is used to conduct the heterogeneity test based on the L-moments and PL-moments.

Number of simulations are 10,000 for computing the heterogeneity. We computed the regional average L-moments ratios, the regional PL-moments ratios, and the corresponding parameter values of the fitted Kappa probability distribution (see Table 4). Table 4 shows the results of the heterogeneity measure using L-moments and PL-moments methods. It can be observed from Table 4 that the different values for the -statistic are −0.41, −1.60, and −2.89 based on L-moments and −0.06, −1.8, and −3.17 based on PL-moments. Therefore, we concluded that, by comparing these results and the heterogeneity conditions, study region is acceptably homogeneous for L-moments and PL-moments. No further subdivisions of the present study are necessary.


Heterogeneity measures L-momentPL-moment

Heterogeneity measure
Observed standard deviation of group L-Cv 0.078 0.077
Simulated mean of standard deviation of group L-Cv 0.084 0.077
Simulated standard deviation of standard deviation of group L-Cv 0.015 0.014
Value of the heterogeneity measure −0.410 −0.060

Heterogeneity measure
Observed average of L-Cv/L-skewness distance 0.105 0.139
Simulated mean of average L-Cv/L-skewness distance 0.141 0.139
Simulated standard deviation of average L-Cv/L-skewness distance 0.023 0.022
Value of the heterogeneity measure −1.600 −1.800

Heterogeneity measure
Observed average of L-skewness/L-kurtosis distance 0.094 0.088
Simulated mean of average L-skewness/L-kurtosis distance 0.1750.176
Simulated standard deviation of average L-skewness/L-kurtosis distance 0.028 0.028
Value of the heterogeneity measure −2.890−3.170

3.2. Fitting Appropriate Probability Distribution

After homogeneity analysis of the study area, a suitable probability distribution is required for the RFA. The objective is not only to recognize a suitable probability distribution for RFA but also to observe a probability distribution that will provide robust quantile estimate for each location and for the regional growth cure. List of candidate probability distributions for RFA is GLO, GEV, GPA, and GNO.

We plotted L-moments and PL-moments diagrams for preliminary evaluation of the probability distribution for the study area.

Figure 2 illustrates an analogy of the observed and hypothetical relationships of the probability distribution. Figure 2 shows that GLO distribution is not a suitable candidate for the L-moments and PL-moments.

Interestingly, both analyses of the L-moments and the PL-moments diagram show that the sample average values are appropriately distinguished by the hypothetical L-moments and PL-moments for GPA and GNO distributions.

However, it is hard to find a suitable probability distribution that fits most of the regional observed data. Table 5 shows the goodness of fit test results for candidate probability distributions.


Method GLO GEV GNO GPA

L-moments 1.96 1.38 0.61−0.42
PL-moments 1.77 1.19 0.43−0.61

Table 5 shows that GLO distribution failed the goodness of fit test for both L-moments and for PL-moments methods as the calculated value of the -test for the GLO distribution is larger than the critical value of 1.64 (at 90% confidence level).

It has been observed that the computed values of are less than 1.64 (at 90% confidence level), namely, GEV, GNO, and GPA distributions. However, GEV, GNO, and GPA distributions are suitable for regional distribution based on L-moments and PL-moments methods and for obtaining the future estimates of the quantile.

Further, it can be noted that GPA distribution is suitable for L-moments method (lowest critical value). Similarly, GNO distribution is suitable for PL-moments method (lowest critical value). Table 6 shows the estimates of the regional parameters for L-moments and PL-moments for the suitable probability distribution.


MethodDistribution
Parameters

L-moments GEV0.46600.5644 −0.2751
GNO0.66710.7565 −0.7594
GPA−0.05510.9944 −0.0575

PL-moments GEV 0.52200.5116 −0.2686
GNO0.70460.6837 −0.7487
GPA0.0478 0.9073 −0.0472

3.3. Estimation of the Quantiles

The regional quantile estimates , with varying nonexceedance probability for the GNO, GEV, and GPA distributions, are presented in Table 7 based on L-moments and PL-moments. Quantile function is normally represented as for fitted regional frequency distribution. The quantile estimate at location is established by joining the estimate of and .


MethodDistribution
0.10000.50000.8000 0.9000 0.9500 0.9800 0.99000.99750.998750.9990

L-momentsGEV 0.0454 0.6836 1.5139 2.2247 3.0591 4.4161 5.6871 9.0748 11.3165 12.1338
GNO 0.0473 0.6671 1.5585 2.3072 3.1448 4.4098 5.4997 8.0676 9.5666 10.0823
GPA 0.0500 0.6481 1.6217 2.3931 3.1958 4.3073 5.1878 7.0578 8.0502 8.3782

PL-momentsGEV 0.1397 0.7190 1.4670 2.1034 2.8469 4.0497 5.1703 8.1363 10.086 10.7950
GNO 0.1412 0.7046 1.5062 2.1752 2.9203 4.0410 5.0032 7.2608 8.5739 9.0249
GPA 0.1436 0.6871 1.5649 2.2547 2.9674 3.9460 4.7149 6.3304 7.1786 7.4577

Mathematical form of the quantile estimate with nonexceedance probability isThe regional growth curves for the GEV, GNO, and GPA distributions are shown in Figure 3.

Figure 3 shows the regional growth curves of each candidate distribution for L-moments and PL-moments. GEV, GNO, and GPA distributions are approximately identical up until -year return period ( for both L-moments and PL-moments. However, afterward the growth curves of the GPA distribution lie below the GEV and GNO distributions.

Therefore, it is necessary to assess the performance of regional quantile estimates.

4. Accuracy of the Estimated Quantiles and the Regional Growth Curve

A Monte Carlo simulation is designed to assess accuracy of the regional quantile estimates that are obtained by the RFA. We use logical L-moments algorithm that has been reported by Hosking and Wallis [7] in Section . This algorithm takes samples from a region that has comparable characteristics as of the actual region, such as having the same record length, same number of locations, and the regional L-moments ratios. The area used for simulation should report the plausible heterogeneity in the area and intersite dependency if exist (Hosking and Wallis [7]). In the repeated sampling procedure, the quantile estimates are computed for the different nonexceedance probabilities. Suppose that, at th repetition and location , quantile estimate can be written as for the nonexceedance probability . The relative error for this estimator is The bias and the RMSE of the above quantity over all repetition are Also, for the estimated quantile, the regional average bias and the relative RMSE are We use empirical quantities of quantile distribution for the assessment analysis that can be computed by taking the ratio of estimated to true values, for the quantile, and for the regional growth curves. Therefore, 90% of the regional growth curve lie in between the interval: Inverting the expression for , we have The 90% confidence interval limits show the measure of variation between the estimated and the true quantiles. These limits provide the expected magnitude of errors in the estimated quantiles and the regional growth curves.

We computed L-moments ratios to find the most suitable distribution and the precision of original growth curves. The correlation between the study region sites varies from −0.05 to 0.86 with an average of 0.40. Therefore, we use algorithm from Table of Hosking and Wallis [7].

We held out the analysis for recurrence of different years. We run 10,000 simulations with sample size of 30, 60, and 90 in each case. The whole process is repeated for GEV, GNO, and GPA distributions. From these repetitions, we computed several performance measures, such as the regional average relative bias, regional average RMSE, regional average relative RMSE, and the error bounds for the estimated regional growth curves for the selected nonexceedance probability . Overall results for the suitable probability distribution for both methods are presented in Tables 8, 9, and 10 for sample size of 30, 60, and 90, respectively.


MethodDistribution0.10.50.80.90.950.980.990.9950.9980.999

L-momentsGEV−0.5400−19.7674−19.5328−19.3602−19.1579−18.8659−18.6376−18.4097−18.1165−17.9056
0.09000.06950.09630.19520.33340.60380.90571.32612.13373.0048
1.98730.10190.06350.08760.10900.13690.15950.18390.21910.2478
−2.09370.88290.92960.88420.84110.78620.74600.70650.65570.6177
1.87141.19491.12081.15801.18841.22441.25161.28181.32851.3683

L-momentsGNO0.8124−19.6279−19.3846−19.2837−19.1816−19.0506−18.9573−18.8699−18.7636−18.6899
0.09860.06960.09170.20510.35200.60270.84211.13121.60082.0312
2.07660.10450.05870.08890.11200.13680.15330.16860.18780.2016
−2.13310.86830.92710.87800.83790.79580.76910.74530.71720.6978
1.65691.19741.10771.15691.19541.23551.26151.28511.31491.3369

L-momentsGPA3.8654−19.4351−19.3564−19.3226−19.2616−19.1706−19.1060−19.0506−18.9966−18.9727
0.12420.06690.09980.21360.35520.59680.83131.11881.59412.0368
2.47850.10330.06140.08920.11110.13860.16020.18330.21620.2431
−2.51920.85910.92150.88170.84440.79790.76460.73270.69190.6623
1.42721.19111.11151.15881.19761.24591.28411.32571.38591.4367

PL-momentsGEV−0.2662−19.7059−19.4770−19.3340−19.1600−18.8996−18.6909−18.4797−18.2052−18.0066
0.53790.06870.09470.20230.34670.61710.90871.30582.05512.854
3.81170.09480.06490.09600.12100.15080.17360.19770.23240.2609
−7.78140.88920.92520.87250.82590.76970.72960.69110.64170.6050
4.55161.18271.12071.17161.21171.25541.28811.32111.36931.4104

PL-momentsGNO3.5142−19.5922−19.3485−19.2704−19.1894−19.0806−19.0003−18.9235−18.8284−18.7617
0.62030.06890.09390.21410.36440.61430.84831.12721.57471.9809
4.35220.09710.06260.09800.12380.15040.16750.18310.20240.2162
−10.6180.87720.92040.86550.82330.78000.75300.72900.70120.6822
4.63531.18491.11221.17341.21891.26531.29491.32181.35461.3780

PL-momentsGPA12.123−19.4300−19.3238−19.3113−19.2716−19.2046−19.1555−19.1141−19.0768−19.0649
0.85130.06660.10410.22500.36840.60210.82151.08531.51441.9094
5.88110.09620.06690.09960.12330.15090.17200.19410.22580.2518
−17.5070.86980.91300.86800.82990.78460.75240.72200.68310.6552
4.73071.17911.11891.17741.22151.27391.3141.35671.41951.4722


MethodDistribution0.10.50.80.90.950.980.990.9950.9980.999

L-momentsGEV−1.3271−19.4230−19.3551−19.3241−19.2649−19.1612−19.0736−18.9839−18.8690−18.7888
0.08460.06240.08510.18730.32500.58410.86521.25051.98392.7721
1.91550.08330.06120.08550.10240.12130.13600.15230.17700.1983
−1.95930.87560.92540.88510.84910.80740.77960.75010.71180.6834
1.62421.16751.10151.14721.18791.23091.26781.30381.36591.4141

L-momentsGNO0.8417−19.3335−19.2795−19.2984−19.2999−19.2879−19.2742−19.2593−19.2399−19.2264
0.09680.06340.08580.20000.34390.58500.81231.08451.52351.9239
2.10550.08810.05820.08610.10500.12350.13540.14630.15990.1699
−2.28870.86470.92440.87940.84660.81410.79570.77980.75970.7443
1.58631.17141.0961.15111.19341.24051.26661.29391.32531.3505

L-momentsGPA4.6371−19.2512−19.2631−19.3160−19.3417−19.3606−19.3745−19.3933−19.4306−19.4708
0.12800.06240.09500.20940.34680.57400.79031.05361.48781.8928
2.64420.09150.06140.08820.10640.12700.14290.15990.18490.2059
−3.80880.86280.91860.88150.85100.81670.79240.77030.74220.7178
1.46991.17211.10201.15341.19251.24541.27911.32361.38691.4405

PL-momentsGEV−1.6968−19.3984−19.3601−19.3623−19.3319−19.2576−19.1861 −19.1081−19.0026−18.9255
0.54310.05960.08880.19780.33660.58440.84171.18381.81792.4868
3.86560.07830.06340.0950.11660.13870.15410.17000.19280.2120
−9.29210.89070.92010.87500.84030.79840.77140.74340.70430.6783
5.35531.15211.10991.16641.20981.25721.28691.32161.36971.4105

PL-momentsGNO2.3785−19.3306−19.2979−19.3417−19.3628−19.3678−19.3617−19.3509−19.3331−19.3185
0.62780.06010.09110.21000.35230.58040.7881.03031.41131.7516
4.37430.08120.06230.09650.11860.13910.15130.16180.17420.1830
−12.0830.88560.91880.87090.83890.80610.78700.77060.75170.7380
5.27071.15611.10591.17041.21521.26131.28981.31381.33611.356

PL-momentsGPA11.187−19.2590−19.2910−19.3681−19.4084−19.4322−19.4404−19.4463−19.4560−19.4673
0.85960.05970.10090.22290.35810.56150.73910.94181.25581.5341
5.83100.08400.06600.09910.11950.13910.15220.16500.18280.1971
−18.9770.88000.91270.86990.84130.80980.79120.76820.74470.7255
5.12811.15511.11171.17691.21731.26151.28871.32271.36091.3899


MethodDistribution0.10.50.80.90.950.980.990.9950.9980.999

L-momentsGEV−0.0117−19.4226−19.3109−19.2926−19.2542−19.1809−19.1151−19.0448−18.9497−18.8790
0.07870.05560.08230.18500.31510.54590.78401.09891.67932.2883
1.73290.08140.05440.08320.10310.12380.13810.15250.17270.1892
−1.93030.89350.92870.88850.85870.82460.80060.77720.74630.7229
1.57171.14991.09731.14611.18041.21561.23931.26351.29661.3242

L-momentsGNO1.5851−19.3606−19.2517−19.2695−19.2747−19.2678−19.2569−19.2437−19.2245−19.2098
0.09170.05670.08460.19670.33050.54460.73920.96601.32231.6400
1.93280.08500.05430.08530.10520.12370.13470.14420.15540.1632
−2.32680.88480.92610.88440.85680.83010.81390.80000.78330.7722
1.51681.15421.09421.14891.18511.21971.24011.25751.27791.2925

L-momentsGPA4.9081−19.2995−19.2417−19.2865−19.3078−19.3175−19.3195−19.3209−19.3256−19.3330
0.12340.05710.09350.20810.33650.53220.70540.90531.21801.4975
2.46880.08820.05760.08700.10540.12390.13640.14890.16600.1798
−4.53530.87830.92100.88380.85940.83260.81410.79610.77310.7557
1.48331.15741.09961.15321.18701.22231.24671.27111.30441.3312

PL-momentsGEV−0.0877−19.4145−19.3060−19.3062−19.2847−19.2278−19.1698−19.1038−19.0100−18.9374
0.54580.05650.08550.19630.33380.57120.80901.11561.66502.2276
3.90790.07840.05850.09340.11720.14060.15580.17030.18970.2050
−10.1900.89830.92360.87650.84310.80670.78180.75790.72710.7041
5.87301.14481.10361.16521.20811.25041.27711.30261.33641.3623

PL-momentsGNO3.8414−19.3665−19.2536−19.2859−19.3044−19.3102−19.3057−19.2966−19.2804−19.2665
0.63080.05780.08960.20880.34880.56900.76590.99261.34341.6524
4.46510.08190.05960.09610.11930.14050.15260.16280.17440.1822
−13.0110.89100.91920.87160.84110.81220.79490.78010.76340.7519
5.67561.14911.10301.16931.21281.25391.27751.29741.31971.3347

PL-momentsGPA12.6398−19.3142−19.2484−19.3077−19.3404−19.3589−19.3637−19.3655−19.3680−19.3720
0.86410.05880.09990.22200.35580.55350.72270.91251.20091.4521
6.01160.08550.06400.09850.11990.14000.15290.16510.18120.1937
−19.8890.88420.91300.87000.84350.81560.79700.77890.75660.7399
5.37431.15281.11031.17551.21581.25521.28071.30551.33851.3636

Figures 4, 5, and 6 show estimated regional growth curves for sample sizes 30, 60, and 90, respectively, and also GEV, GNO, and GPA distributions with the 90% error bounds.

Tables 8, 9, and 10 show that increase in the sample size such as 30 to 90 improved the performance particularly in the prediction of the large nonexceedance probability . L-moments method provides similar performance for the GNO and GPA distributions in terms of relative bias that is presented in Tables 8, 9, and 10. We found that the GPA distribution produced the lowest relative bias compared to GEV and GNO distribution for the PL-moment for the various values of the nonexceedance probability . However, the GPA distribution performs better in terms of RMSE than the GEV and GNO distribution for both methods (L-moments and PL-moments). Furthermore, RMSE is lowest for PL-moments compared to L-moments. In addition, the error bounds for the GPA distribution of regional quantiles are narrow compared to GEV and GNO distributions. It shows that the estimation of censored sample improves the prediction of extreme precipitation explicitly at large nonexceedance probability .

5. Discussion and Conclusion

This study provides a comprehensive evaluation of the L-moments and the PL-moments. First, revisiting RFA on L-moments by Hosking and Wallis [7], we aimed to develop similar connections of regional homogeneity for PL-moments. The L-moments and the PL-moments for candidate probability distributions (GLO, GEV, GNO, and GPA) are also developed for presenting the corresponding L-ratio and PL-ratio diagrams with the goodness of fit test results. The regional growth curves for the selected distribution have been shown in Figures 4, 5, and 6. At the lower tail, GEV, GPA, and GNO distributions are approximately the same, but, at the upper tail, there is variation between the regional quantiles. The regional homogeneity analysis starts by assuming 17 locations of Northern areas and Khyber Pakhtunkhwa, Pakistan, as one homogeneous region, based on L-moments and PL-moments at censoring level ranging from 10 to 23%. This assumption is statistically accepted after applying the heterogeneity and discordancy tests. The -statistic provides appropriate distribution for modeling the monthly extreme precipitation in Northern areas and Khyber Pakhtunkhwa, Pakistan. We found that GPA distribution is suitable for the L-moments and GNO distribution for the PL-moments.

Finally, Monte Carlo simulation used for performance evaluation by commonly used error functions. Several accuracy measures such as relative bias, RMSE, relative RMSE, and error function bounds for the regional quantiles are computed with 10,000 runs of Monte Carlo simulations. We found that GPA distribution produced robust quantile estimates for both return periods and methods (L-moments and PL-moments). Our results support the finding of previous study (e.g., Cunnane [9]; Bhattarai [12]) for censored sample analysis where PL-moments method outperformed the L-moments method for the estimation of large return periods events.

Appendix

The partial L-moments (PL-moments) for generalized logistic (GLO), generalized Pareto (GPA), generalized normal (GNO), and generalized extreme value (GEV) distributions were derived based on the formula defined by Wang (Water Resour Res 32:1767Ö 1771, 1996 (In references lines from 442 to 444 mentioned that in study)). The summary of the derived distributions and parameters estimation for these distributions is as follows.

PL-moments for the GEV Distribution [10] are as follows.

The CDF and quantile function of the GEV are given byand quantile function

Wang [10] developed the partial probability-weighted moments (PWMs) of the GEV aswhere The first four PL-moments of the GEV are defined as

In previous equation, is an Incomplete Gamma function:Then the first four PL-moments are computed to develop the PL-moment ratios (PL-Cv, PL-Cs, and PL-Ck) for the GEV distribution.

The PL-moments for the GLO Distribution are as follows.

The CDF and quantile function of the GLO are given byand quantile function The partial PWMs of the GLO are developed as follows: where is an Incomplete Beta function

The first four PL-moments of the GLO are defined as

Then the first four PL-moments are computed to develop the PL-moment ratios (PL-Cv, PL-Cs, and PL-Ck) for the GLO distribution.

The PL-moments for the GPA Distribution are as follows.

The CDF and quantile function of the GPA are given byand quantile function

The partial PWMs of the GPA are developed as follows:

The first four PL-moments of the GPA are defined aswhere

Then the first four PL-moments are computed to develop the PL-moment ratios (PL-Cv, PL-Cs, and PL-Ck) for the GPA distribution.

Ethical Approval

The manuscript is prepared in accordance with the ethical standards of the responsible committee on human experimentation and with the latest (2008) version of Helsinki Declaration of 1975.

Conflicts of Interest

The manuscript is prepared by using secondary data and authors declared that there are no conflicts of interest.

Acknowledgments

The authors extend their appreciation to the Deanship of Scientific Research at King Saud University for funding this work through Research Group no. RG-1437-027.

References

  1. T. Dalrymple, “Flood-frequency analyses,” Water Supply Paper 1543-A, U.S.Geological Survey, Reston, Va, USA, 1960. View at: Google Scholar
  2. J. R. M. Hosking, J. R. Wallis, and E. F. Wood, “An appraisal of the regional flood frequency procedure in the UK flood studies report,” Hydrological Sciences Journal, vol. 30, no. 1, pp. 85–109, 1985. View at: Publisher Site | Google Scholar
  3. N. P. Greis and E. F. Wood, “Regional flood frequency estimation and network design,” Water Resources Research, vol. 17, no. 4, pp. 1167–1177, 1981. View at: Publisher Site | Google Scholar
  4. J. R. Wallis, “Hydrologic problems associated with oilshale devel opment,” in Environmental Systems and Management, S. Rinaldi, Ed., pp. 85–102, 1982. View at: Google Scholar
  5. C. Cunnane, “Methods and merits of regional flood frequency analysis,” Journal of Hydrology, vol. 100, no. 1–3, pp. 269–290, 1988. View at: Publisher Site | Google Scholar
  6. K. Subramanya, Engineering Hydrology, McGraw-Hill, Singapore, 2nd edition, 2007.
  7. J. R. Hosking and J. R. Wallis, Regional Frequency Analysis: An Approach Based on L-Moments, Cambridge University Press, New York, NY, USA, 1997. View at: Publisher Site
  8. Q. J. Wang, “Estimation of the GEV distribution from censored samples by method of partial probability weighted moments,” Journal of Hydrology, vol. 120, no. 1–4, pp. 103–114, 1990. View at: Publisher Site | Google Scholar
  9. C. Cunnane, “Statistical distributions for flood frequency analysis,” World Meteorological Organization Operational Hydrology Report 33, 1987. View at: Google Scholar
  10. Q. J. Wang, “Using partial probability weighted moments to fit the extreme value distributions to censored samples,” Water Resources Research, vol. 32, no. 6, pp. 1767–1771, 1996. View at: Publisher Site | Google Scholar
  11. Q. J. Wang, “Unbiased estimation of probability weighted moments and partial probability weighted moments from systematic and historical flood information and their application to estimating the GEV distribution,” Journal of Hydrology, vol. 120, no. 1-4, pp. 115–124, 1990. View at: Publisher Site | Google Scholar
  12. K. P. Bhattarai, “Partial L-moments for the analysis of censored flood samples,” Hydrological Sciences Journal, vol. 49, no. 5, pp. 855–868, 2004. View at: Google Scholar
  13. A. B. Shabri, Z. M. Daud, and N. M. Ariff, “Regional analysis of annual maximum rainfall using TL-moments method,” Theoretical and Applied Climatology, vol. 104, no. 3-4, pp. 561–570, 2011. View at: Publisher Site | Google Scholar
  14. B. Saf, “Regional flood frequency analysis using L-moments for the West Mediterranean region of Turkey,” Water Resources Management, vol. 23, no. 3, pp. 531–551, 2009. View at: Publisher Site | Google Scholar
  15. J. Abolverdi and D. Khalili, “Development of regional rainfall annual maxima for Southwestern Iran by L-moments,” Water Resources Management, vol. 24, no. 11, pp. 2501–2526, 2010. View at: Publisher Site | Google Scholar
  16. Z. Hussain and G. R. Pasha, “Regional flood frequency analysis of the seven sites of Punjab, Pakistan, using L-moments,” Water Resources Management, vol. 23, no. 10, pp. 1917–1933, 2009. View at: Publisher Site | Google Scholar
  17. Z. A. Zakaria, A. Shabri, and U. N. Ahmad, “Regional Frequency Analysis of Extreme Rainfalls in the West Coast of Peninsular Malaysia using Partial L-Moments,” Water Resources Management, vol. 26, no. 15, pp. 4417–4433, 2012. View at: Publisher Site | Google Scholar
  18. A. Shahzadi, A. S. Akhter, and B. Saf, “Regional frequency analysis of annual maximum rainfall in monsoon region of pakistan using L-moments,” Pakistan Journal of Statistics and Operation Research, vol. 9, no. 1, pp. 111–136, 2013. View at: Publisher Site | Google Scholar
  19. S. Coles, An Introduction to Statistical Modeling of Extreme Values, Springer Series in Statistics, Springer-Verlag, London, England, 2001. View at: Publisher Site | MathSciNet
  20. R. W. Katz, M. B. Parlange, and P. Naveau, “Statistics of extremes in hydrology,” Advances in Water Resources, vol. 25, no. 8–12, pp. 1287–1304, 2002. View at: Publisher Site | Google Scholar
  21. H. Abida and M. Ellouze, “Probability distribution of flood flows in Tunisia,” Hydrology and Earth System Sciences, vol. 12, no. 3, pp. 703–714, 2008. View at: Publisher Site | Google Scholar
  22. S. Feng, S. Nadarajah, and Q. Hu, “Modeling annual extreme precipitation in China using the generalized extreme value distribution,” Journal of the Meteorological Society of Japan, vol. 85, no. 5, pp. 599–613, 2007. View at: Publisher Site | Google Scholar
  23. T. Yang, Q. Shao, Z.-C. Hao et al., “Regional frequency analysis and spatio-temporal pattern characterization of rainfall extremes in the Pearl River Basin, China,” Journal of Hydrology, vol. 380, no. 3-4, pp. 386–405, 2010. View at: Publisher Site | Google Scholar
  24. G. Villarini, J. A. Smith, M. L. Baeck, R. Vitolo, D. B. Stephenson, and W. F. Krajewski, “On the frequency of heavy rainfall for the Midwest of the United States,” Journal of Hydrology, vol. 400, no. 1-2, pp. 103–120, 2011. View at: Publisher Site | Google Scholar
  25. D. She, J. Xia, J. Song, H. Du, J. Chen, and L. Wan, “Spatio-temporal variation and statistical characteristic of extreme dry spell in Yellow River Basin, China,” Theoretical and Applied Climatology, vol. 112, no. 1-2, pp. 201–213, 2013. View at: Publisher Site | Google Scholar
  26. J. R. Wallis, N. C. Matalas, and J. R. Slack, “Just a moment!,” Water Resources Research, vol. 10, no. 2, pp. 211–219, 1974. View at: Publisher Site | Google Scholar
  27. C. P. Pearson, “New Zealand regional flood frequency analysis using L-moments,” Journal of Hydrology, vol. 30, no. 2, pp. 53–64, 1991. View at: Google Scholar
  28. J. R. Hosking, “L-moments: analysis and estimation of distributions using linear combinations of order statistics,” Journal of the Royal Statistical Society. Series B. Methodological, vol. 52, no. 1, pp. 105–124, 1990. View at: Google Scholar | MathSciNet
  29. J. A. Greenwood, J. M. Landwehr, N. C. Matalas, and J. R. Wallis, “Probability weighted moments: definition and relation to parameters of several distributions expressable in inverse form,” Water Resources Research, vol. 15, no. 5, pp. 1049–1054, 1979. View at: Publisher Site | Google Scholar
  30. J. R. M. Hosking and J. R. Wallis, “Some statistics useful in regional frequency analysis,” Water Resources Research, vol. 29, no. 2, pp. 271–281, 1993. View at: Publisher Site | Google Scholar
  31. J. R. Stedinger, R. M. Vogel, and E. Foufoula-Georgiou, “Frequency analysis of extreme events,” in Hand Book of Hydrology, D. R. Maidment, Ed., McGraw-Hill, New York, NY, USA, 1993. View at: Google Scholar
  32. R. M. Vogel and N. M. Fennessey, “L-moment diagrams should replace product moment diagrams,” Water Resources Research, vol. 29, no. 6, pp. 1745–1752, 1993. View at: Publisher Site | Google Scholar
  33. J. R. M. Hosking, “The use of L-moments in the analysis of censored data,” in Recent Advances in Life-Testing and Reliability, N. Balakrishnan, Ed., pp. 545–564, CRC Press, Boca Raton, Fla, USA, 1995. View at: Google Scholar
  34. M. G. Schaefer, “Regional analyses of precipitation annual maxima in Washington State,” Water Resources Research, vol. 26, no. 1, pp. 119–131, 1990. View at: Publisher Site | Google Scholar
  35. C. Pearson, “Application of l-moments to maximum river ows,” The New Zealand Statistician, vol. 28, no. 1, pp. 2–10, 1993. View at: Google Scholar
  36. R. M. Vogel, W. O. Thomas, and T. A. McMahon, “Flood-flow frequency model selection in southwestern united states,” Journal of Water Resources Planning and Management, vol. 119, no. 3, pp. 353–366, 1993. View at: Publisher Site | Google Scholar
  37. K. C. A. Chow and W. E. Watt, “Practical use of the L-moments,” Stochastic and Statistical Methods in Hydrology and Environmental Engineering, vol. 1, no. 3, pp. 55–69, 1994. View at: Google Scholar
  38. B. ÖnÖz and M. Bayazit, “Best-fit distributions of largest available flood samples,” Journal of Hydrology, vol. 167, no. 1-4, pp. 195–208, 1995. View at: Publisher Site | Google Scholar
  39. R. M. Vogel and I. Wilson, “Probability distribution of annual maximum, mean, and minimum streamflows in the United States,” Journal of Hydrologic Engineering, vol. 1, no. 2, pp. 69–76, 1996. View at: Publisher Site | Google Scholar
  40. M. C. Peel, Q. J. Wang, R. M. Vogel, and T. A. McMahon, “The utility of L-moment ratio diagrams for selecting a regional probability distribution,” Hydrological Sciences Journal, vol. 46, no. 1, pp. 147–155, 2001. View at: Publisher Site | Google Scholar
  41. Z. Jiang, Y. Ding, L. Zhu, and J. Zhang, “Extreme precipitation experimentation over eastern china based on generalized pareto distribution,” Plateau Meteorology, vol. 28, no. 3, pp. 573–580, 2009. View at: Google Scholar
  42. S. B. Y. J. D. Yuguo, “Research on extreme value distribution of short-duration heavy precipitation in the Sichuan basin,” Journal of the Meteorological Sciences, vol. 4, article 007, 2012. View at: Google Scholar

Copyright © 2017 Said Arab Khan et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.


More related articles

 PDF Download Citation Citation
 Download other formatsMore
 Order printed copiesOrder
Views3564
Downloads1169
Citations

Related articles

We are committed to sharing findings related to COVID-19 as quickly as possible. We will be providing unlimited waivers of publication charges for accepted research articles as well as case reports and case series related to COVID-19. Review articles are excluded from this waiver policy. Sign up here as a reviewer to help fast-track new submissions.