Abstract
The performance of regionalization methods used for regional flood frequency analysis is affected considerably by the features used to identify the homogeneous regions (e.g., climatological, meteorological, geomorphological, and physiographic characteristics of the watersheds). In this study, a regionalization method is proposed that takes advantage of the two widely used techniques in regionalization of watersheds: canonical correlation analysis and cluster analysis. In the proposed method, the canonical correlation analysis is utilized to select or weight features that then will be used by a hybrid clustering algorithm for regionalization of watersheds. The proposed method is applied to Sefidrud basin, located in the north of Iran, to implement regionalization with two, three, four, and five regions. Performance assessment of the proposed method shows that all the options of the proposed method can be effective alternatives to some common regionalization methods to improve the homogeneity of the regions. The results indicate that the method can satisfy the homogeneity conditions approximately for all the regions which were identified in the study area.
NOTATION
- FFA
Flood frequency analysis
- RFFA
Regional flood frequency analysis
- CCA
Canonical correlation analysis
- WAKM
A hybrid clustering algorithm consisting of Ward's algorithm and K-means
- ASW
Average silhouette width
- CCA-WAKM
A set of four implementation options of the proposed regionalization method including the combination of CCA and WAKM in which the feature vectors consisting of the canonical variables of the watershed features are used in clustering by WAKM
- CCA-WAKM1
An option for implementing CCA-WAKM in which the values of the first canonical variable of the watershed features are used as the feature vectors of sites
- CCA-WAKM1,2
An option for implementing CCA-WAKM in which the values of the first and second canonical variables of the watershed features are used as the features of the feature vectors of sites
- CCA-WAKM1,3
An option for implementing CCA-WAKM in which the values of the first and third canonical variables of the watershed features are used as the features of the feature vectors of sites
- CCA-WAKM1,2,3
An option for implementing CCA-WAKM in which the values of the first, second and third canonical variables of the watershed features are used as the features of the feature vectors of sites
- CCA-wWAKM
An implementation option of the proposed regionalization method in which the coefficients of the watershed features in the first canonical variables of the watershed features are used as the weights of the features in clustering by WAKM
INTRODUCTION
Flood frequency analysis (FFA) is used to estimate the magnitude of a flood with a specified return period or estimate the return period of a flood with a specified magnitude. Flood quantiles can be estimated by at-site FFA using only flood data recorded at the site of interest. However, in many cases, the length of flood data records in sites of interest are not appropriate to provide reliable flood estimates. In such situations, regional flood frequency analysis (RFFA) is an efficient approach to compensate for the temporal shortage of flood data records by pooling flood data over a number of sites with similar flood generation mechanisms.
The objective of the regionalization is to identify homogeneous regions, i.e., groups of sites with similar flood generation mechanisms (Hosking & Wallis 1997). In a homogeneous region, the flood frequency distribution varies from site to site with only a site-specific factor named the index-flood. RFFA based on index-flood was first introduced by Dalrymple (1960).
The identification of homogeneous regions for RFFA is required to find an appropriate regional flood frequency distribution. However, identifying a group of sites satisfying homogeneity conditions is not always possible easily. When the regionalization features not directly related to flood data records (e.g., climatological, meteorological, geomorphological, and physiographic characteristics of the watersheds), it is difficult to assign all the watersheds to the regions where all of them satisfy homogeneity conditions.
One of the most widely used methods for regionalization is cluster analysis. The cluster analysis methods are multivariate statistical analysis methods which have been utilized by many researchers in several hydrological studies especially for the regionalization (e.g., Acreman & Sinclair 1986; Burn 1989; Hall & Minns 1999; Jingyi & Hall 2004; Lin & Chen 2006; Ramachandra Rao & Srinivas 2006a, 2006b; Srinivas et al. 2008; Chen & Hong 2012; Toth 2013). While in traditional methods of regionalization, regions were identified based on administrative borders or geographical contiguity, in the methods like cluster analysis, different types of features effective on flood generation mechanism, such as physiographic features or meteorological attributes, can be used as regionalization features. By applying the traditional methods, it is very difficult to identify homogeneous regions, because geographical contiguity of sites does not result in a similarity in their flood generation mechanism. On the other hand, in cluster analysis methods, the regions may be identified based on similarity of sites in terms of various features such as physiographic attributes, meteorological characteristics, plant cover, land use, etc. Thus, in the new methods, regions are not identified essentially based on geographical contiguity.
Region of influence (ROI) is another widely used approach for RFFA that was developed by Burn (1990). In ROI, a region of influence, which is a hydrologically homogeneous neighborhood, is formed for each watershed in a study area. ROI has been used in several regional frequency analysis studies and its performance evaluated in different case studies (e.g., Zrinji & Burn 1994; Burn 1997; Castellarin et al. 2001).
Clustering algorithms can be divided into hierarchical and partitional clustering algorithms (Ramachandra Rao & Srinivas 2008). Hierarchical algorithms include agglomerative algorithms and divisive algorithms where the agglomerative hierarchical algorithms have been used for regionalization of watersheds in several RFFA studies (e.g., Mosley 1981; Tasker 1982; Acreman & Sinclair 1986; Nathan & McMahon 1990; Burn et al. 1997; Hosking & Wallis 1997; Ramachandra Rao & Srinivas 2006a). One of the most important advantages of hierarchical algorithms is that they often do not require the determination of initial conditions (such as the determination of initial cluster centers). On the other hand, a noticeable limitation of hierarchical algorithms is that after assigning a data point to a cluster, it is not possible to move it between clusters. Partitional algorithms, which often are based on the minimization of an objective function, require the determination of initial conditions, such as the initial cluster centers, but these algorithms often provide the benefits of the possibility of moving data points between different clusters in different iterations of the algorithm. One of the most widely used partitional clustering algorithms in regional frequency analysis studies is the K-means algorithm (e.g., Wiltshire 1986; Burn 1989; Bhaskar & O'Connor 1989; Burn & Goel 2000; Ramachandra Rao & Srinivas 2006a; Jin et al. 2017; Xie et al. 2018).
Ramachandra Rao & Srinivas (2006a) investigated the performances of combinations of the three hierarchical algorithms with one partitional algorithm for regionalization of 245 watersheds in Indiana State, USA. They used single linkage, complete linkage, and Ward's algorithm as hierarchical algorithms to specify initial cluster centers for K-means as a partitional algorithm. The quality of the clusters formed by each of the proposed hybrid algorithms was evaluated according to values of four cluster validity indices: cophenetic correlation coefficient (CPCC), silhouette width, Dun's index, and Davies–Bouldin index. Also, the homogeneity of the regions identified by each algorithm was assessed based on the values of the heterogeneity measures proposed by Hosking & Wallis (1993). The proposed hybrid algorithms showed better performances in comparison with the hierarchical and partitional clustering algorithms. In addition, among the proposed algorithms, the combinations of Ward's and K-means algorithms (WAKM) provided the best regionalization results for RFFA in the study area.
From another aspect, clustering methods can be divided into hard and fuzzy clustering. In hard clustering, each data point belongs to only one cluster and does not belong to any other cluster. On the other hand, in fuzzy clustering, each data point can be assigned to all clusters simultaneously with specified degrees of membership between 0 and 1. The sum of the degrees of membership of each data point in all clusters is equal to 1. The most popular fuzzy clustering algorithm used in regional frequency analysis studies to implement regionalization is the fuzzy C-means (FCM) clustering algorithm (e.g., Hall & Minns 1999; Ramachandra Rao & Srinivas 2006b; Srinivasa Raju & Nagesh Kumar 2008; Sadri & Burn 2011; Asong et al. 2015; Basu & Srinivas 2015).
In some studies, combinations of hard and fuzzy clustering algorithms have been studied for regionalization. Srinivas et al. (2008) used a combination of self-organizing maps (SOM) and FCM to implement regionalization of Indiana State watersheds. The results showed that the method was effective in forming homogeneous regions. Farsadnia et al. (2014) also carried out a similar study for regionalization of the watersheds of Mazandaran province in northern Iran and obtained similar results. Ahani & Mousavi Nadoushani (2016) studied the fuzzy development of the hybrid clustering algorithms proposed by Ramachandra Rao & Srinivas (2006a) by combining single linkage, complete linkage, average linkage hierarchical algorithms, and Ward's algorithm, as well as SOM with C-means fuzzy clustering algorithm for performing regionalization of the watersheds in the Sefidrud watershed. The study of the size of the formed regions, the values of the clustering validity indices, and also the $H$ heterogeneity indices showed that, in general, the combination of the Ward and SOM algorithms with FCM algorithm provide the best results for regionalization of the studied region in order to analyze the flood regional frequency.
The ability of cluster analysis methods in dealing with multivariate analysis problems and reducing the need for visual judgments and time-consuming assessments are the benefits of these methods for regionalization of watersheds. However, there are some issues that may affect the efficiency of cluster analysis methods for regionalization of watersheds. In regionalization by cluster analysis methods, each watershed is represented by a vector that includes values of a set of features affecting flood generation mechanism. The feature vectors are used to evaluate similarity of the watersheds. The identification of the feature vectors to be used in clustering is one of the most challenging issues in regionalization studies (e.g., Nezhad et al. 2010; Di Prinzio et al. 2011; Razavi & Coulibaly 2013; Ahani et al. 2018).
One of the useful methods to identify and select the effective features on the flood generation mechanism of watersheds is canonical correlation analysis (CCA). CCA (Hotelling 1936) is a method for describing the correlation between two sets of variables (Cavadias 1990). Cavadias (1990) developed a method based on CCA to determine the hydrological neighborhoods and estimate flood quantile. Also, Cavadias et al. (2001) proposed a method based on the use of CCA in order to determine homogeneous regions or hydrological neighborhoods for flood estimation in both gauged and ungauged sites. The proposed method was useful to identify effective watershed features on flood generation mechanism. However, the features useful for identification of homogeneous regions were selected based on visual judgments on the similarities between patterns of data points in original feature space and canonical space. The application of CCA in RFFA was studied by several researchers (e.g., Ribeiro-Correa et al. 1995; GREHYS 1996a, 1996b; Ouarda et al. 2001, 2008) and the results indicated the desirable effects of CCA on RFFA.
In a method introduced by Ilorme & Griffis (2013), CCA was initially used along with some other multivariate analysis methods to identify the watershed features influencing the flood generation mechanism. Then, the selected features were used to perform regionalization by Ward's clustering algorithm. The method reduced the need for visual judgment to identify homogeneous regions and select regionalization features. Also, it overcame the visual judgment limitation. However, the different effects of the different features on the final regionalization was not considered in the proposed method. In addition, skipping a number of features with relatively lower values of correlation coefficient might significantly reduce homogeneity of regions and accuracy of flood quantile estimation (Basu & Srinivas 2014).
In general, determining the relationship between the watershed features (such as geographical location characteristics, physiographic attributes, geological features, land-use, plant cover, etc.) and the flood-related features (such as flood statistics) can be considered as an important advantage of CCA-based RFFA methods. However, most CCA-based regionalization methods depend on visual judgments to some extent, and in some cases, they are theoretically limited to two-dimensional space.
The main objective of the current study is to propose an efficient regionalization method focusing on feature selection and feature weighting to improve the homogeneity of the regions. To this aim, a new hybrid method is proposed by combining CCA and cluster analysis in order to take the advantages of both of them and overcome their limitations in regionalization of watersheds for RFFA. After describing the proposed method, some implementation options of the method are presented for regionalization of watersheds in Sefidrud basin located in the north of Iran. Then the performance of implementation options of the method is compared with that of a common regionalization method in the study area.
STUDY AREA AND DATA
Sefidrud basin in the north of Iran with a total area about 59,200 km2 was chosen as a study area to evaluate the performance of proposed methods for regionalization of watersheds. Sefidrud River is formed in the confluence of two rivers, named Shahrud and Ghezel-Ozan, and flows into the Caspian Sea. Thirty-nine gauged sites with unregulated flow in Sefidrud basin were selected for this study and their watershed features were extracted for regionalization (Figure 1). The annual maximum flood data records of the sites of interest were obtained from the database of the Iran Water Resources Management Company. The flood data records in the selected sites cover the time period from 1967 to 2012 and the average length of flood data records is about 23 years. The total number of flood data is equal to 898 station-years and record length in the sites varies between 10 and 39 years.
Location of Sefidrud Basin and the hydrometric stations.
Longitude, latitude, elevation from the sea level, drainage area, mean annual precipitation, and the runoff coefficient were selected as the watershed features contributing to the regionalization procedure (Table 1). It is worth noting that four sites have runoff coefficients greater than 1 because the watersheds of these sites are located in areas with karst geologic structures.
Descriptive statistics of the regionalization features
| Feature | Range | Mean | Standard deviation |
|---|---|---|---|
| Longitude (dd) | 47.05–51.07 | 48.74 | 1.30 |
| Latitude (dd) | 35.18–37.53 | 36.55 | 0.73 |
| Elevation (m a.s.l.) | 40–2,800 | 1,376.22 | 649.20 |
| Drainage area (km2) | 29–49,300 | 5,591.38 | 11,569.86 |
| Mean annual precipitation (mm) | 184–1,400 | 467.70 | 323.62 |
| Runoff coefficient | 0.03–1.32 | 0.49 | 0.34 |
| Feature | Range | Mean | Standard deviation |
|---|---|---|---|
| Longitude (dd) | 47.05–51.07 | 48.74 | 1.30 |
| Latitude (dd) | 35.18–37.53 | 36.55 | 0.73 |
| Elevation (m a.s.l.) | 40–2,800 | 1,376.22 | 649.20 |
| Drainage area (km2) | 29–49,300 | 5,591.38 | 11,569.86 |
| Mean annual precipitation (mm) | 184–1,400 | 467.70 | 323.62 |
| Runoff coefficient | 0.03–1.32 | 0.49 | 0.34 |
The features were selected based on the availability of the relevant data and their potential role in the flood generation mechanism. The longitude, latitude, and elevation from the sea level were selected, because of the special geographical situation of the study area. In fact, the noticeable variablity in values of longitude, latitude, and elevation from the sea level in the study area may affect climatological and meteorological conditions of the sites considerably. The considerable variation of these features in the study area can affect the flood generation mechanisms of the watersheds noticeably. The precipitation and the drainage area are considered for regionalization in several RFFA studies due to their pivotal role in flood generation (e.g., Ramachandra Rao & Srinivas 2006a, 2006b; Srinivas et al. 2008; Farsadnia et al. 2014). Precipitation is often the main factor generating floods (Ramachandra Rao & Srinivas 2008; Srinivas et al. 2008) and so, mean annual precipitation of watersheds was selected as one of the watershed features. Additionally, in hydrological models, the drainage area often is considered as one of the most important factors in estimating the flood magnitudes (Hosking & Wallis 1997; Ramachandra Rao & Srinivas 2008). Thus, it is logical to use the drainage area as one of the watershed features in the regionalization procedure. Also, the runoff coefficient was selected as it determines the ratio of precipitation transformed to runoff and, therefore, it may be useful to identify homogeneous regions (e.g., Ramachandra Rao & Srinivas 2006a, 2006b; Srinivas et al. 2008; Basu & Srinivas 2015).
is the value of the feature j in the data point i; 
is the mean value of the feature j over the dataset, and 
is the standard deviation of the feature j over the dataset. In addition, 
is the standardized value of the feature j for the data point i.METHODS
Discordancy evaluation
Prior to the feature selection or feature weighting steps, the flood data records were evaluated by using the discordancy measure D proposed by Hosking & Wallis (1993) in terms of the L-moments of flood data. A site is identified as discordant if D exceeds the critical value. When the number of sites is greater than 14, the critical value of D is equal to 3. This screening procedure can be performed either before regionalization for all the sites as one group, or after regionalization for the sites belonging to each region. Among the 39 sites, two sites were identified as discordant and were excluded from the regionalization process as suggested by Hosking & Wallis (1997). Therefore, the data related to the 37 remaining sites was used in the next stages of the study.
Canonical correlation analysis
and 
and 
, then the two sets of the canonical variables 
and 
are formed as in Equation (2), such that the correlation between pairs of the corresponding canonical variables is maximized and the correlation between other pairs are minimized. The highest correlation is related to the canonical variables 
and 
, and the lowest correlation is related to the canonical variables 
and 
. For more details on CCA, see Hotelling (1936) and Cavadias (1990).
In the present study, the six watershed features and three L-moment (Hosking 1990) ratios of flood data are used as the two sets of original variables for CCA. The three selected L-moment ratios are the linear coefficient of variation (L-CV), linear skewness (L-skewness) and linear kurtosis (L-kurtosis). They are chosen because the three H heterogeneity measures proposed by Hosking & Wallis (1997) are calculated based on the values of these three L-moment ratios. Therefore, the use of canonical variables of watershed features which are highly correlated with the canonical variables of the L-moment ratios may increase the homogeneity of the regions identified in the regionalization.
After standardization of watershed features, L-moment ratios were calculated for each site and the standardization technique was applied to the values of L-CV, L-skewness, and L-kurtosis because the standardization of both original datasets is recommended before implementing CCA (Ribeiro-Correa et al. 1995). The standardized watershed features, longitude, latitude, elevation from the sea level, drainage area, mean annual precipitation and the runoff coefficient may be represented by 
, 
, 
, 
, 
, and 
, respectively, hereafter. Also, 
, 
, and 
denote the standardized L-moment ratios, L-CV, L-skewness, and L-kurtosis, in this order. Then CCA was performed on the standardized dataset of the six watershed features and the standardized dataset of the three L-moment ratios. Consequently, three pairs of canonical variables were calculated. The canonical variables related to the watershed feature space are represented by 
, 
, and 
, and the canonical variables related to the L-moment ratios space are denoted by 
, 
, and 
.
WAKM clustering algorithm
Regarding the advantages of the hybrid clustering algorithms (Ramachandra Rao & Srinivas 2006a; Srinivas et al. 2008; Farsadnia et al. 2014; Ahani & Mousavi Nadoushani 2016), the WAKM algorithm (Ramachandra Rao & Srinivas 2006a) is used as the clustering algorithm for the regionalization of watersheds in this study. The name WAKM represents the combination of Ward's algorithm (Ward Jr 1963) and K-means algorithm (Hartigan & Wong 1979). In this algorithm, first, by applying Ward's algorithm to the data points, a desired number of clusters are provided. Then, the cluster centers are used as initial cluster centers for clustering the data points by the K-means algorithm (Ramachandra Rao & Srinivas 2008). More details on Ward's and K-means algorithms are available in Ramachandra Rao & Srinivas (2008).
Cluster validity index
To compare the quality of different clusterings performed on the same dataset, cluster validity indices are used. The clustering quality is improved as the distances between the data points belonging to each cluster decrease (smaller intra-cluster distances) and the distances between the data points belonging to different clusters increase (greater inter-cluster distances) (Ramachandra Rao & Srinivas 2008).
Ramachandra Rao & Srinivas (2006a) evaluated the performances of a number of cluster validity indices to determine the optimal number of clusters in order to perform regionalization of watersheds. They concluded that the average silhouette width (ASW) is an effective measure for this purpose.
is the average distance of the data point i from the data points with which it is placed in the same cluster; and 
is the minimum average distance from the data point i from the data points in a cluster different from the cluster that the data point i belongs to. The value of 
can be in the range 
, where values close to 1 indicate the allocation of the data point i to an appropriate cluster, and values close to −1 represents the assignment of the data point i to an inappropriate cluster. The average silhouette width (ASW) criterion is obtained by averaging on the values of the silhouette width of all the clustered data points and hence it varies over the range 
as well (Ramachandra Rao & Srinivas 2008). Considering the acceptable performance of ASW in evaluating the quality of clusters (Ramachandra Rao & Srinivas 2006a), it was used for clustering evaluation in this study.A hybrid method for regionalization
In the present study, a new hybrid method is proposed for regionalization of watersheds. To implement the regionalization method, four options are presented based on feature selection and one option is presented based on feature weighting. In the four feature selection-based options, after implementing CCA, the canonical variables consisting of the watershed features highly correlated with the canonical variables of the flood statistics (L-CV, L-skewness, L-kurtosis) were used as input features of the WAKM clustering algorithm. These four options are represented by CCA-WAKMfv in the remainder of the article. In CCA-WAKMfv, the subscript fv is an abbreviation for feature vector and can be replaced by one of the options 1, 1,2, 1,3, and 1,2,3. The four CCA-WAKM options differ in defining the feature vectors corresponding to the sites used in clustering by WAKM.
The input feature vectors of clustering for the CCA-WAKM1, CCA-WAKM1,2, CCA-WAKM1,3, and CCA-WAKM1,2,3 are described in Table 2. In the first option, denoted by CCA-WAKM1, only the value of the first canonical variable of the watershed features is used as the feature vector of each site. In the second option, represented by CCA-WAKM1,2, the feature vector of each site includes the values of the first and second canonical variables of the watershed features. In the third option, introduced by CCA-WAKM1,3, the feature vector of each site contains the values of the first and third canonical variables of the watershed features. Finally, in the fourth option, denoted by CCA-WAKM1,2,3, the values of all the three canonical variables consisting of the watershed features, are used to form the feature vector of each site. Since in the space of canonical variables, the highest correlation exists between the first pair of canonical variables of the watershed features and the L-moment ratios, the first canonical variable of the watershed features is used in all the four options. Also, the second and third canonical variables of the watershed features are added to the feature vectors in different options in order to investigate their effects on the regionalization results.
Variables included in the input feature vector of clustering for the CCA-WAKM options and CCA-wWAKM
| Option | Variables of the feature vector |
|---|---|
| CCA-WAKM1 | V1 |
| CCA-WAKM1,2 | V1, V2 |
| CCA-WAKM1,3 | V1, V3 |
| CCA-WAKM1,2,3 | V1, V2, V3 |
| CCA-wWAKM | a11A1, a12A2, a13A3, a14A4, a15A5, a16A6 |
| Option | Variables of the feature vector |
|---|---|
| CCA-WAKM1 | V1 |
| CCA-WAKM1,2 | V1, V2 |
| CCA-WAKM1,3 | V1, V3 |
| CCA-WAKM1,2,3 | V1, V2, V3 |
| CCA-wWAKM | a11A1, a12A2, a13A3, a14A4, a15A5, a16A6 |
In the feature weighting-based option, weights of the watershed features in the linear combination of the first canonical variable of watershed features (i.e., a11, a12, a13, a14, a15, a16), are used as the weights of the original watershed features (i.e., A1, A2, A3, A4, A5, A6) in the regionalization by WAKM. Thus, the regionalization feature vector used this option is [a11A1, a12A2, a13A3, a14A4, a15A5, a16A6]. To facilitate referring to this option, the acronym CCA-wWAKM is used in the rest of the article, in which wWAKM represents weighting WAKM. The input feature vector of clustering for the CCA-wWAKM is determined in Table 2.
For implementing the feature weighting-based option CCA-wWAKM, the coefficients of the watershed features in the first canonical variable are applied to the feature vectors of the standardized watershed features as the weights. Then, WAKM is used to perform clustering based on these weighted feature vectors.
Homogeneity assessment
Identification of homogeneous regions by regionalization methods is an important and challenging part of RFFA. Hosking & Wallis (1993) proposed three heterogeneity measures, H1, H2, and H3, based on the L-moment ratios. These measures were used in several RFFA studies and were approved by the researchers (e.g., Viglione et al. 2007). For a given region, if 
, the region is ‘acceptably homogeneous’, if 
, the region is identified as ‘possibly heterogeneous’, and if 
, the region is regarded as ‘definitely heterogeneous’ (Hosking & Wallis 1997).
In the current study, the heterogeneity measures H are used to assess the homogeneity of the regions and a region is considered as homogeneous if 
and 
and 
.
RESULTS AND DISCUSSION
The coefficients of standardized watershed features and standardized L-moment ratios in the linear combinations related to the canonical variables of the watershed features and the L-moment ratios are presented in Table 3. The canonical variables of the watershed feature space are represented by V1, V2, and V3, and the canonical variables of the L-moment ratios space are denoted by W1, W2, and W3.
The coefficients of the standardized watershed features and L-moment ratios in linear combinations of their canonical variables
| Standardized variable | V1 | V2 | V3 | W3 | W | W3 |
|---|---|---|---|---|---|---|
| A1 (Longitude) | −0.590 | −0.021 | −1.380 | – | – | – |
| A2 (Latitude) | −0.264 | −0.373 | 0.435 | – | – | – |
| A3 (Elevation from the sea level) | 0.272 | −1.491 | 0.412 | – | – | – |
| A4 (Drainage area) | 0.044 | −0.588 | −0.287 | – | – | – |
| A5 (Mean annual precipitation) | −0.126 | −1.579 | 0.538 | – | – | – |
| A6 (Runoff coefficient) | −0.264 | 0.614 | 0.727 | – | –- | – |
| B1 (L-CV) | – | – | – | 1.004 | 0.431 | −1.644 |
| B2 (L-skewness) | – | – | – | 0.034 | −0.707 | 2.827 |
| B3 (L-kurtosis) | – | – | – | −0.261 | 1.407 | −1.523 |
| Standardized variable | V1 | V2 | V3 | W3 | W | W3 |
|---|---|---|---|---|---|---|
| A1 (Longitude) | −0.590 | −0.021 | −1.380 | – | – | – |
| A2 (Latitude) | −0.264 | −0.373 | 0.435 | – | – | – |
| A3 (Elevation from the sea level) | 0.272 | −1.491 | 0.412 | – | – | – |
| A4 (Drainage area) | 0.044 | −0.588 | −0.287 | – | – | – |
| A5 (Mean annual precipitation) | −0.126 | −1.579 | 0.538 | – | – | – |
| A6 (Runoff coefficient) | −0.264 | 0.614 | 0.727 | – | –- | – |
| B1 (L-CV) | – | – | – | 1.004 | 0.431 | −1.644 |
| B2 (L-skewness) | – | – | – | 0.034 | −0.707 | 2.827 |
| B3 (L-kurtosis) | – | – | – | −0.261 | 1.407 | −1.523 |
As shown in Table 4, the correlation coefficient between the first pair of canonical variables is considerably greater than the values of the correlation coefficient between the second pair of canonical variables as well as the third pair of canonical variables. The values of the correlation coefficients of the second pair of canonical variables and the third pair of canonical variables are nearly equal to each other and their difference is lower than 0.04. Therefore, it seems that the first canonical variable of the watershed features and its coefficients can play a more important role than two other canonical variables in identifying regions with better homogeneity. It is also important to note that among the coefficients of the first canonical variable of the L-moment ratios, the largest coefficient belongs to L-CV, which is the basis for calculating the heterogeneity measure H1, which is more effective in identifying homogeneous regions than H2 and H3 heterogeneity measures according to Hosking & Wallis (1997) and Viglione et al. (2007).
The correlation coefficient values between the canonical variable pairs
| Canonical variable pair | W1, V1 | W2, V2 | W3, V3 |
|---|---|---|---|
| Correlation coefficient | 0.855 | 0.383 | 0.347 |
| Canonical variable pair | W1, V1 | W2, V2 | W3, V3 |
|---|---|---|---|
| Correlation coefficient | 0.855 | 0.383 | 0.347 |
In Table 5, the values of the linear correlation coefficients between the original variables and canonical variables are presented. Among watershed features, the drainage area and elevation from the sea level show the greatest positive correlations with the first canonical variables (V1 and W1), respectively. On the other hand, the longitude and the runoff coefficient have the largest magnitudes of negative correlations with the first canonical variables, respectively. Concerning the second canonical variables (V2 and W2), the highest linear correlations are those of the drainage area and the runoff coefficient, respectively, and the largest inverse correlation values are related to the elevation and the mean annual precipitation. The two features of latitude and mean annual precipitation show the highest positive correlation with the third canonical variables (V3 and W3), while the largest negative correlation values with these canonical variables are respectively related to the longitude and the drainage area.
The correlation coefficient values between the original variables and canonical variables
| Standardized variable | V1 | V2 | V3 | W1 | W2 | W3 |
|---|---|---|---|---|---|---|
| A1 (Longitude) | −0.805 | −0.218 | −0.426 | −0.688 | −0.083 | −0.148 |
| A2 (Latitude) | −0.369 | 0.113 | 0.499 | −0.316 | 0.043 | 0.173 |
| A3 (Elevation from the sea level) | 0.387 | −0.440 | −0.188 | 0.331 | −0.169 | −0.065 |
| A4 (Drainage area) | 0.480 | 0.307 | −0.260 | 0.410 | 0.118 | −0.090 |
| A5 (Mean annual precipitation) | −0.752 | −0.310 | 0.208 | −0.643 | −0.119 | 0.072 |
| A6 (Runoff coefficient) | −0.783 | 0.120 | 0.118 | −0.669 | 0.046 | 0.041 |
| B1 (L-CV) | 0.972 | 0.230 | 0.046 | 0.831 | 0.088 | 0.016 |
| B2 (L-skewness) | 0.555 | 0.656 | 0.511 | 0.474 | 0.252 | 0.177 |
| B3 (L-kurtosis) | −0.019 | 0.970 | 0.243 | −0.016 | 0.372 | 0.084 |
| Standardized variable | V1 | V2 | V3 | W1 | W2 | W3 |
|---|---|---|---|---|---|---|
| A1 (Longitude) | −0.805 | −0.218 | −0.426 | −0.688 | −0.083 | −0.148 |
| A2 (Latitude) | −0.369 | 0.113 | 0.499 | −0.316 | 0.043 | 0.173 |
| A3 (Elevation from the sea level) | 0.387 | −0.440 | −0.188 | 0.331 | −0.169 | −0.065 |
| A4 (Drainage area) | 0.480 | 0.307 | −0.260 | 0.410 | 0.118 | −0.090 |
| A5 (Mean annual precipitation) | −0.752 | −0.310 | 0.208 | −0.643 | −0.119 | 0.072 |
| A6 (Runoff coefficient) | −0.783 | 0.120 | 0.118 | −0.669 | 0.046 | 0.041 |
| B1 (L-CV) | 0.972 | 0.230 | 0.046 | 0.831 | 0.088 | 0.016 |
| B2 (L-skewness) | 0.555 | 0.656 | 0.511 | 0.474 | 0.252 | 0.177 |
| B3 (L-kurtosis) | −0.019 | 0.970 | 0.243 | −0.016 | 0.372 | 0.084 |
All the values of correlation coefficient between the L-moment ratios and canonical variables are positive. Greatest values of correlations with first, second, and third canonical variables are related to L-CV, L-kurtosis, and L-skewness, respectively.
Since only the canonical variables of watershed features (i.e., V1, V2, and V3) are used in the next step of the proposed regionalization method, the values of correlation coefficient between watershed features and their canonical variables are more useful to identify the watershed features that may be more effective on results of regionalization.
According to the number of selected sites in the study area and the length of their flood data records and also regarding the 5T rule (Reed et al. 1999), the regionalization was implemented by changing the number of regions from two to five. In this study, it was considered as a constraint that the smallest region (in terms of station-years) in each regionalization includes about 50 station-years flood data in order to provide flood quantiles corresponding to a ten-year return period. To evaluate the effect of the proposed methods on the homogeneity of the regions, the results of applying CCA-WAKM and CCA-wWAKM were compared with the results of applying the single WAKM clustering algorithm to feature vectors consisting of six standardized watershed features. To assess and compare the performances of the methods, they were evaluated by ASW cluster validity index and the heterogeneity indices H1, H2, and H3. It should be noted that the words ‘cluster’ and ‘region’ may be used in the rest of article equivalently.
The ASW values for implementing regionalization by each method for two, three, four, and five regions are presented in Table 6. In all cases, ASW for all the CCA-WAKM implementation options and CCA-wWAKM, are higher than those of the single WAKM. This indicates a higher quality of the final clusters resulting from the application of the proposed method in comparison with the single WAKM. The reduction of the number of dimensions or regionalization features (from six watershed features to one, two, or three canonical variables) in the CCA-WAKM implementation options compared to the WAKM may be considered as an effective factor in increasing ASW and improving the clustering quality in terms of intra-cluster compactness and inter-cluster separation. However, for CCA-wWAKM the number of regionalization features (six weighted watershed features) is equal to that of WAKM (six watershed features) and so the increase in ASW can be interpreted as an increase in the quality of clustering.
The values of the cluster validity index ASW for clustering by using WAKM, CCA-wWAKM, and the four CCA-WAKM options
| Number of regions | 2 | 3 | 4 | 5 |
|---|---|---|---|---|
| WAKM | 0.418 | 0.492 | 0.420 | 0.415 |
| CCA-wWAKM | 0.557 | 0.545 | 0.484 | 0.438 |
| CCA-WAKM1 | 0.655 | 0.574 | 0.557 | 0.598 |
| CCA-WAKM1,2 | 0.451 | 0.516 | 0.550 | 0.430 |
| CCA-WAKM1,3 | 0459 | 0.468 | 0.525 | 0.564 |
| CCA-WAKM1,2,3 | 0.367 | 0.411 | 0.431 | 0.443 |
| Number of regions | 2 | 3 | 4 | 5 |
|---|---|---|---|---|
| WAKM | 0.418 | 0.492 | 0.420 | 0.415 |
| CCA-wWAKM | 0.557 | 0.545 | 0.484 | 0.438 |
| CCA-WAKM1 | 0.655 | 0.574 | 0.557 | 0.598 |
| CCA-WAKM1,2 | 0.451 | 0.516 | 0.550 | 0.430 |
| CCA-WAKM1,3 | 0459 | 0.468 | 0.525 | 0.564 |
| CCA-WAKM1,2,3 | 0.367 | 0.411 | 0.431 | 0.443 |
Figure 2 shows the values of the heterogeneity measures H1, H2, and H3 for two regions identified by WAKM, CCA-wWAKM, and the four CCA-WAKM options. All the options and methods result in identifying two homogeneous regions, and only by implementing CCA-WAKM1,2, one of the regions is relatively heterogeneous based on H1.
The values of the heterogeneity measures H1, H2, and H3 for two regions identified by WAKM, CCA-wWAKM, and CCA-WAKM.
The values of the heterogeneity measures H1, H2, and H3 for two regions identified by WAKM, CCA-wWAKM, and CCA-WAKM.
According to Figure 3, only CCA-wWAKM provides three homogeneous regions simultaneously. Both WAKM and the four CCA-WAKM options result in identifying a possibly heterogeneous region based on the values of one or two heterogeneity measures.
The values of the heterogeneity measures H1, H2, and H3 for three regions identified by WAKM, CCA-wWAKM, and CCA-WAKM.
The values of the heterogeneity measures H1, H2, and H3 for three regions identified by WAKM, CCA-wWAKM, and CCA-WAKM.
Figure 4 shows that the use of WAKM to identify four regions in the study area leads to identifying two homogeneous regions and two possibly heterogeneous regions. In addition, as seen in the four-region state, applying CCA-wWAKM results in satisfying the homogeneity conditions in all the regions. Also, while CCA-WAKM1, CCA-WAKM1,2, and CCA-WAKM1,2,3 provide four homogeneous regions, CCA-WAKM1,3 identifies a possibly heterogeneous region along with the three homogeneous regions.
The values of the heterogeneity measures H1, H2, and H3 for four regions identified by WAKM, CCA-wWAKM, and CCA-WAKM.
The values of the heterogeneity measures H1, H2, and H3 for four regions identified by WAKM, CCA-wWAKM, and CCA-WAKM.
As seen in Figure 5, using WAKM to identify five regions results in identifying two possibly heterogeneous regions, while the other three regions satisfy the homogeneity conditions. Among CCA-WAKM implementation options, CCA-WAKM1, CCA-WAKM1,2, and CCA-WAKM1,3, provide four homogeneous regions and one possibly heterogeneous region, whereas the option CCA-WAKM1,2,3 identifies five homogeneous regions. In this case, the use of CCA-wWAKM yields identifying five homogenous regions.
The values of the heterogeneity measures H1, H2, and H3 for five regions identified by WAKM, CCA-wWAKM, and CCA-WAKM.
The values of the heterogeneity measures H1, H2, and H3 for five regions identified by WAKM, CCA-wWAKM, and CCA-WAKM.
is the percentage of homogeneous regions, 
represents the number of homogeneous regions and 
denotes the total number of regions in two, three, four, and five-region states 
.The percentage of homogeneous regions (%) identified in regionalization for two, three, four, and five regions according to the heterogeneity measure H
| Regionalization method | Heterogeneity measure | |||
|---|---|---|---|---|
| H1 | H2 | H3 | H1, H2, H3 | |
| WAKM | 92.9 | 78.6 | 78.6 | 64.3 |
| CCA-wWAKM | 100 | 100 | 100 | 100 |
| CCA-WAKM1 | 92.9 | 85.7 | 100 | 85.7 |
| CCA-WAKM1,2 | 85.7 | 92.9 | 100 | 78.6 |
| CCA-WAKM1,3 | 100 | 78.6 | 78.6 | 78.6 |
| CCA-WAKM1,2,3 | 92.9 | 100 | 100 | 92.9 |
| Regionalization method | Heterogeneity measure | |||
|---|---|---|---|---|
| H1 | H2 | H3 | H1, H2, H3 | |
| WAKM | 92.9 | 78.6 | 78.6 | 64.3 |
| CCA-wWAKM | 100 | 100 | 100 | 100 |
| CCA-WAKM1 | 92.9 | 85.7 | 100 | 85.7 |
| CCA-WAKM1,2 | 85.7 | 92.9 | 100 | 78.6 |
| CCA-WAKM1,3 | 100 | 78.6 | 78.6 | 78.6 |
| CCA-WAKM1,2,3 | 92.9 | 100 | 100 | 92.9 |
The results indicate that among the methods and their implementation options, the best performance in providing the homogeneous regions is related to CCA-wWAKM. All 14 regions identified by CCA-wWAKM in two, three, four, and five-region states are identified as homogeneous according to all the three heterogeneity measures. CCA-wWAKM shows perfect efficiency (100%) in identifying homogeneous regions in the study area.
A probable reason for the superiority of CCA-wWAKM over the other options is to use information related to all the watershed features. While in the implementation options of CCA-WAKM the multiples of watershed features are used in a linear combination to provide values of a regionalization feature, in CCA-wWAKM each watershed feature is included in the regionalization feature vector separately. In addition, the weight of each feature is determined only based on the absolute magnitude of its coefficient in the linear combination of the canonical variable V1.
After CCA-wWAKM, CCA-WAKM1,2,3 displays the best performance by identifying 13 homogeneous regions based on all three heterogeneity measures H. In fact, according to Figures 2–5, all the regions identified by this option are homogeneous according to H2 and H3, and only in the three-region state, a region is identified as possibly heterogeneous by H1. As a result, the efficiency of this option in identifying homogeneous regions in the study area can be estimated at 93%. For the option CCA-WAKM1, the efficiency is equal to 86% approximately. Also, for the options CCA-WAKM1,2 and CCA-WAKM1,3, among 14 identified regions, 11 regions are shown to be homogeneous on the basis of all three measures H. Thus, the efficiency of these options in identifying the homogeneous regions is approximately 79%. According to Hosking & Wallis (1997), the heterogeneity measure H1 is more sensitive to heterogeneity of regions in comparison with the measures H2 and H3. However, as seen in Table 6, this is not observed for CCA-WAKM1,3, because in the regionalization states with four and five regions, in each regionalization, CCA-WAKM1,3 identified one region including a group of sites with a relatively high value of standard deviation of the statistics L-skewness and L-kurtosis. L-skewness and L-kurtosis play key roles in definitions of the measures H2 and H3 (Hosking & Wallis 1997). By applying WAKM, 9 homogeneous regions were identified among 14 regions, which yields the lowest efficiency among the methods used for regionalization in this study (about 64%). This means that all options of the proposed method are more efficient in providing homogeneous regions than WAKM. Moreover, the CCA-wWAKM is superior to all four CCA-WAKM options by providing 100% efficiency.
The better performance of CCA-WAKM1,2,3 in comparison with other options of CCA-WAKM is because of adding the second and third canonical variables V2 and V3 to regionalization feature vectors. V2 and V3 show higher correlations with L-moment ratios L-skewness and L-kurtosis which play important roles in calculation of the heterogeneity measures H2 and H3. According to Table 7, by adding V2 and V3 to the regionalization feature vectors, homogeneity of the regions is improved to some extent based on the heterogeneity measures H2 and H3. Of course, as seen in Table 6, the effect of V2 is more considerable than V3 on the homogeneity improvement.
In general, the results of calculating the heterogeneity indices H for the regions show that CCA-wWAKM and all the four CCA-WAKM options outperform WAKM in effectively identifying homogeneous regions for the study area. Therefore, all of the implementation options of the proposed method can be used as effective alternatives to common regionalization methods in order to improve the homogeneity of the identified regions. Among the CCA-WAKM implementation options, the CCA-WAKM1,2,3 outperforms the other options in terms of the percentage of homogeneous regions. In addition, CCA-wWAKM is the optimum options, because of its excellent performance in identifying the regions satisfying homogeneity condition completely. This method even outperforms CCA-WAKM1,2,3. Of course, the difference between the results of these two options is only related to the measure H1 in the second region that exceeds the threshold 
in the three-region state for CCA-WAKM1,2,3.
After the homogeneity assessment, the size of regions, i.e., the total number of flood data contained by each region, was evaluated. Also, the assignment of the sites to the regions was studied. For this evaluation, CCA-WAKM1,2,3 is selected as the best option for representing CCA-WAKM, due to its better performance in identifying homogeneous regions than other CCA-WAKM options. For both WAKM and CCA-wWAKM, it is not needed to choose an optimal option, because there is only one implementation option for each of them.
Figures 6–9 show the assignment of the sites to the regions identified by WAKM, CCA-WAKM1,2,3, and CCA-wWAKM, respectively. According to the figures, the geographical contiguity in the regions identified by the CCA-wWAKM is more considerable than that in the regions provided by the single WAKM and CCA-WAKM1,2,3. In other words, delineating the crisp geographical boundaries for the regions identified by CCA-wWAKM in the study area is more feasible compared to the regions provided by the other options. In fact, for this study area and the selected features, implementation of regionalization using the CCA-wWAKM method results in the assignment of greater weights to the features related to the geographical location of the sites, especially the longitude. It should be noted that all the positive and negative coefficients of the watershed features in the first canonical variable of the watershed features used as clustering feature weights are squared in the Euclidean distance. Thus, only the absolute magnitude of these coefficients or weights affects the clustering.
Geographical dispersion of the sites over two regions.
Geographical dispersion of the sites over three regions.
Geographical dispersion of the sites over four regions.
Geographical dispersion of the sites over five regions.
In addition, according to the number of sites assigned to the regions, the dispersion of the sites across the regions provided by the single WAKM and CCA-wWAKM is more balanced than that across the regions identified by CCA-WAKM1,2,3. Figure 10 displays the sizes of the regions identified by the regionalization methods in terms of the number of flood data recorded in each region (station-years) for two, three, four, and five regions.
Sizes of the identified regions (station-years) by WAKM, CCA-wWAKM, and the selected CCA-WAKM option.
Sizes of the identified regions (station-years) by WAKM, CCA-wWAKM, and the selected CCA-WAKM option.
As seen in Figure 10, the number of regions with a size lower than 100 station-years identified by CCA-WAKM1,2,3 is greater than those identified by the other methods. Considering the fact that the average flood data record length in the selected sites is about 23 years, a region with size greater than 100 station-years can include at least four watersheds with the average flood data record length. It should be noted that in RFFA, the main goal is to increase the reliability of flood estimates by increasing the number of flood data pooled from several sites in the homogeneous regions. The regionalization that yields the identification of small regions may not be so useful to achieve this goal and so, it cannot be the optimal option for RFFA. Indeed, RFFA is characterized by a trade-off between the size of the region (i.e., number of flood data in station-years) and its homogeneity: usually, the higher the size of the identified pooling group of sites, the higher the expected heterogeneity. Moreover, the target-size of the region depends on the return period associated with the target flood quantile (see, e.g., Cunnane 1988; Jakob et al. 1999). Therefore, the use of the CCA-wWAKM method in three, four, and five-region states seems to provide better results than the CCA-WAKM1,2,3.
Considering the excellent performance of CCA-wWAKM in the identification of homogeneous regions with appropriate spatial proximity and more balanced assignment of the sites to the regions, it seems that this method can be selected as the optimal option for regionalization of watersheds in Sefidrud basin for RFFA.
CONCLUSIONS
In this study, a hybrid regionalization method was proposed by combining CCA and WAKM clustering algorithms in order to increase the homogeneity of the identified regions for RFFA. Performances of the methods in the Sefidrud basin in northern Iran were evaluated based on ASW as a cluster validity and the measures H1, H2, and H3 as the heterogeneity measures.
According to the values of the ASW cluster validity index, the quality of clustering performed by all the options of the proposed method was higher than that of the clustering done by WAKM.
Also, the homogeneity assessment of the regions based on the values of the heterogeneity measures indicated that CCA-wWAKM and all four implementation options of CCA-WAKM were more efficient in identifying homogeneous regions than WAKM. Among the CCA-WAKM options, CCA-WAKM1,2,3 with an efficiency of 93% in the identification of homogeneous regions showed the best performance and so, it was identified as the optimal CCA-WAKM option. However, the best performance among all the options discussed was related to CCA-wWAKM. All the identified regions by CCA-WWAKM in two- to five-region states satisfied the homogeneity conditions completely. Thus, this option resulted in a 100% efficiency in providing the homogeneous regions. Therefore, CCA-wWAKM can be regarded as the optimal option for identifying the most homogeneous regions, among all the options discussed in this study.
The evaluation of the assignment of the sites to the regions identified by the regionalization methods showed that the geographical proximity of the sites in the regions identified by the CCA-wWAKM is clearer than those of the other options and methods. This may be because of the high weight of the geographical features, especially the longitude in comparison with the other features, in regionalization by CCA-wWAKM.
In addition, it was observed that the distribution of the sites across the regions identified by CCA-wWAKM is more balanced in terms of the number of flood data contained by the regions compared to that of CCA-WAKM. In fact, the use of CCA-WAKM1,2,3 in some cases led to the identification of large regions (in terms of station-years) along with small regions. Identifying small regions is not so desirable for RFFA because it is not possible to provide reliable flood estimates for the sites of these regions. Thus, while both CCA-WAKM and CCA-wWAKM seem efficient in identifying homogeneous regions in comparison with WAKM, CCA-wWAKM can be the more appropriate option for regionalization of watersheds in Sefidrud basin.
As the final remark, it should be noted that examining the effectiveness of the proposed method in case studies with the larger total area makes it possible to apply the regionalization methods for a higher number of regions. Of course, it depends on the target-size of the region, which is related to the return period considered for flood quantile estimation. Also, access to a higher number of watershed features can lead to a more accurate judgment about the advantages and disadvantages of the proposed method.
