Journal of Hydrology 625 (2023) 130014

Contents lists available at ScienceDirect

Journal of Hydrology
journal homepage: www.elsevier.com/locate/jhydrol

Research papers

HyetoClust method: Hyetograph design through cluster analysis

Carmen Patino a, Jose-Luis Molina b, *, Fernando Espejo b, Santiago Zazo b,
Abedin Mohammad-Hosseinpour b, Fernando Silla c
a
IGA Research Group. Department of Statistics, University of Salamanca, Campus Miguel de Unamuno, C/Alfonso X El Sabio s/n, 37007 Salamanca, Spain
b
IGA Research Group. Salamanca University, High Polytechnic School of Engineering Avila, Department of Hydraulic Engineering, Av. de los Hornos Caleros, 50, 05003
Ávila, Spain
c
Departamento de Biología Animal y Ecología, Facultad de Biología, Universidad de Salamanca, Campus Unamuno, 37071 Salamanca, Spain

A R T I C L E I N F O A B S T R A C T

Keywords: Given the increasing non-stationarity of hydrological processes largely due to Climate Change (CC), it is essential
Hyetoclust to create more accurate methods. The inherent spatiotemporal variability of hydrological processes reaches its
Hourly rainfall maximum when it comes to rainfall process. Rainfall is typically divided up into events in hydrological studies,
Inter-event Time Definition (IETD)
but this does not acquire the necessary analytical importance. The objective of the research is dual. First, to
Cluster analysis
Hyetograph
generate a robust method for the identification and characterization of homogeneous rainfall events, largely
based on the geometrical analysis of clustered hyetographs. Second, the generation of more accurate hyeto­
graphs. The study was built on the idea of Inter-Event Time Definition (IETD) and thresholds for rainfall depth,
duration, and intensity. The rainfalls were grouped due to the similarity of features as: duration, volume, in­
tensity, time maximum pick and interevent using hierarchical cluster analysis. The study area is placed in the
Muñogalindo station located in the province of Avila (Spain) and characterized by an altitude above 1,132 m.
Hierarchical cluster analysis showed to be particularly effective for selecting precipitation based on hyetograph
shape. Cluster method strongly revealed the identification of three rainfall eventś groups. HyetoClust is intended
to provide the scientific and technical community with a new tool for characterization of rainfall eventś and the
generation of more accurate hyetographs for supporting hydrological prediction.

1. Introduction mostly based on rainfall resources, But, because rain is dynamic in na­
ture, it is very difficult to anticipate with any degree of accuracy (Fahad
Rainfall is the climate variable that may cause the most widespread et al., 2023). Therefore, hybrid approaches should be used to solve the
social and economic impacts to society (Choo et al., 2020). Earth’­ shortcomings of rainfall forecasting because of the periodic nature and
s climate is increasingly changing and its impacts can be related to non-linear properties of rainfall. In order to forecast seasonal rainfall,
drought or floods (Lyra et al., 2014) which affects numerous productive some research has suggested using artificial neural networks (ANN),
activities. Additionally, during the past few decades, it has become fuzzy logic, neural networks, and autoregressive integrated moving
increasingly clear that old hydrological patterns are changing which is averages (ARIMA) (Thakur et al., 2021; Wong et al., 2003). It has also
causing climate variability and the hydrological processes that are become a relevant issue in recent decades, especially when considering
related to it to increase globally (Molina and Zazo, 2018). In order to the associated extreme events (Benabdelouahab et al., 2020).
evaluate the potential effects of climate change and variability, According to O’Gorman (2015) a warming climate will result in
numerous research on the trend of rainfall on various regional and more intense precipitation extremes. Extreme phenomena are of interest
temporal scales have been carried out (Abolverdi et al., 2016; Sa’adi in all disciplines. Some of them, such as strong storms, flooding or heat
et al., 2019). In developing regions, variations in rainfall have a negative waves cause more extensive damage and impacts on people, infra­
impact on crop productivity and impose harsh climatic conditions. structure, and ecosystems. Because they have the potential to do
Studies that are very focused on climatic variables, such as variability in considerable harm and have an impact on environmental, social, and
rainfall and temperature, are essential, especially in places that are technological systems, extreme events have gained the attention of

* Corresponding author.
E-mail addresses: [email protected] (C. Patino), [email protected] (J.-L. Molina), [email protected] (F. Silla).

https://doi.org/10.1016/j.jhydrol.2023.130014
Received 25 April 2023; Received in revised form 6 July 2023; Accepted 26 July 2023
Available online 9 August 2023
0022-1694/© 2023 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
C. Patino et al. Journal of Hydrology 625 (2023) 130014

scientists and management in recent years. The definition of an extreme An example of characterization of rainfall using probability distri­
event varies among studies and fields, where it is typically used inter­ butions functions and clustering analysis was given by Lima et al. (Lima
changeably with its repercussions (McPhillips et al., 2018). In Hydrol­ et al., 2021) who presented an empirical analysis of time series data of
ogy the term extreme event referred to those events that have values annual maximum daily rainfall which obtained from stations in the state
exceeding a predetermined threshold. To find the extreme levels, a of Rio de Janeiro. Their results showed five regions with homogeneous
threshold should be carefully determined (Hamitouche and Molina, pattern of extreme rainfall. On the other hand, to make rainfall design
2022a). To establish a threshold for analyzing and modeling severe easier, which typically necessitates having a solid understanding of the
occurrences, a variety of techniques have been explored (McPhillips temporal patterns of rainstorm events, it is required to analyze the
et al., 2018). Understanding the processes that cause hydrological and temporal rainfall pattern (Jun et al., 2021). There are several methods
hydraulic extremes is essential to reducing their negative effects on so­ for delimitating independent rainfall events, such as autocorrelation
ciety. One of the key tools in this quest for information is hydrological analysis (Grace and Eagleson, 1967), a minimum inter-event time dry
modeling, but it has limitations. For so, a thorough examination of how period (Dunkerley, 2008), the exponential method (Restrepo-Posada
the structure of hydrological models affects the frequency and timing of and Eagleson, 1982) or Inter-event Time Definition (IETD) (Howard,
extreme rainfall-runoff events is required (Molina et al., 2022). For this 1976). It is challenging to determine the beginning and conclusion of
reason, rainfall variability in space and time has been researched using events because of this intermittency, which also has an impact on event
different stochastic and statistical method, such as principal component characteristics. The delineation of rain events in the literature is
analysis, standard regression models and non-hierarchical classification different and is mostly dependent on the research topic and the data’s
(Chen et al., 2021; Sahu et al., 2022). According to Hamitouche and time resolution (Molina-Sanchis et al., 2016). The average intensity,
Molina’s (Hamitouche and Molina, 2022b) review about artificial in­ peak intensity, number peak, volume, duration and inter-event time
telligence methods for high-flow extremes prediction artificial neural (IET), velocity and their spatial and temporal variability are some of the
networks, support vector machine, wavelets, and Bayesian methods variables that used in the analysis and design of hydraulic systems (Hong
consistently perform the best for high-flow extremes prediction at all et al., 2018; Joo et al., 2013). Also, the identification of homogeneous
time scales. However, rainfall has several disadvantages regarding its rainfall events is essential when natural catastrophes that are brought on
treatment. In this sense, it requires accurate conditions for real-time by heavy rainfall can have a variety of negative effects on civilization
precipitation data, often difficult to obtain. Extreme rainfall for a spe­ (da Silva et al., 2020). Clustering analysis based on cumulative hierar­
cific return period is necessary to build or develop various hydraulic chical algorithms, has been used in various parts of the world, including
projects. For example, for the design of drainage systems in urban Brazil (Lima et al., 2021), Switzerland (Barton et al., 2016) or France
catchments, which are typically characterized by fast reaction, hourly (Knerr et al., 2020).
rainfall events or shorter are particularly required (Aronica and Freni, In general, rainfall depth-duration-frequency (DDF) or intensi­
2005). Also, it is necessary to precisely represent rainfall key parameters ty–duration–frequency (IDF) curves are used to estimate design rainfall
to be able to identify events. This largely depends on the criteria adopted magnitude in the context of Civil Engineering. Theses have the advan­
for the identification and on the availability of rainfall records (Iadanza tageous property of being analytical relationships capable of providing a
et al., 2016). According to Lima et al. (Lima et al., 2021) when taking design rainfall intensity or depth for an assigned duration and return
into account long-term series and statistical inferences of anomalies, period (Gioia et al., 2021). Also, synthetic hyetographs are used to
extreme rainfall occurrences are defined as records throughout a time determine the temporal patterns of design rainfall under various com­
period (daily, seasonal, annual) that are significantly greater than the binations of frequencies and durations. Hydrological analyses typically
average values for the region. It is important to specify the probability of presume that it is necessary to take into account the temporal variability
occurrence when taking into account extreme climate events, which is of point rainfall intensity, which is typically represented graphically by
typically expressed by their recurrence intervals (Chen and Singh, hyetographs (Chimene and Campos, 2020; Wartalska et al., 2020). The
2018). rainfall hydrograph is essential for watershed hydrology in general and
For hydrological and climatic studies, it is crucial to understand the managing rainfall water.
rain features. For this purpose, rainfall-runoff modelling can be done in Wartalska et al. (2020) examined the local precipitation hyetograph
two ways, continuous or event-driven (Tokar, 1996). Studies that cover shapes and confirmed the reference hyetograph shapes used for
real precipitation to hydrological or biological systems typically cate­ designing and simulating the functioning of drainage systems in Poland
gorize rainfall into events. Due to the very erratic rainfall distribution using Ward and k-means methods. And recently Mikołajewski et al.
across time, defining what constitutes a rain event is not an easy task. (2022) in a study also carried out in Poland, used the cluster analysis
For assessing the features of extreme events that set off the aforemen­ methodology at an urban precipitation field scale.
tioned physical processes, sub-daily rainfall data are required (Lewis The objectives of the present case study are to (1) determine the
et al., 2019). Many hydrological applications require sub-daily rainfall suitable value of IETD for rainfall events in a representative study-case
data, yet in areas with poor gauges around the world, ground-based data (2) summarize the characteristics of rainfall events (3) identify the ho­
availability is still a problem. In order to estimate sub-daily rainfall data mogeneous rainfall events and (4) investigate the characteristics of the
globally, satellite remote sensing missions like the Global Precipitation rainfall events found in each cluster (5) determine and verify the shape
Measurement (GPM) mission have proved crucial (Freitas et al., 2020). (geometry) of precipitation hyetographs based on the clusters design.
Hence, breaking down the rainfall into events is crucial for hydrological The research novelty is largely based on the fact that rainfall event
applications to have a better knowledge of these processes. For both represents the analytical core. The more precise the event is defined
surface and subsurface hydrology, rainfall partitioning at the soil- (pre-process), the more accurate will be the prediction (outcome).
atmosphere interface is crucial. It affects several significant hydrologic Consequently, this new method better reproduces and adapts to the
events, including runoff generation, aquifer recharge, and the move­ current increasing hydrological variability, which is even more evident
ment of contaminants in surface waters and the vadose zone (Morbidelli in rainfall processes.
et al., 2018). The structure of the paper is described as follows: first section covers
On the other side, when extreme events are affected by both the paper introduction where a rich state of the art is provided, mainly
dynamical as mesoscale and physiographic as land cover, lakes, terrain on rainfall analytical techniques. This section also outlines the main
factors, the identification of homogeneous zones of extreme occurrence research goals. Second section describes the study area, the data input,
is crucial (Knerr et al., 2020). According to Ragetteli (2017), semiarid which is largely based on precipitation data records, and provides the
and arid regions are characterized by intensity rainfall very high, and methods to assess homogeneous rainfall classes. Third section shows the
the rainfall duration is low. results obtained from this research and, finally, main conclusions and

2
C. Patino et al. Journal of Hydrology 625 (2023) 130014

discussion are given in Section 4. condition for the statistical independence of storm arrivals. Therefore,
the definition of an independent rainfall event has not been agreed upon
2. Materials and methods by hydrologists. On the other hand, separate independent events should
be characterized. The prediction of hydrological extremes recognizes
2.1. Study area both qualitative and quantitative applications. One of the qualitative
predictions is risk assessment, and one of the quantitative forecasts is
The selected case study is the Muñogalindo station, located at Upper rainfall-runoff. There are several methods to characterize rainfall. For
Adaja river basin headwaters, which is part of the Cega-Eresma-Adaja example, Hamitouche and Molina (2022a) with the objective of char­
exploitation system of the Duero Hydrographic Demarcation (Fig. 1). acterization extreme high-flow events in the Upper Andarax watershed
The basin covers 763 km2 and a perimeter of 200 km (40◦ 36′09″ to in SE Spain adopted an event-based Bayesian Causal Modelling.
4◦ 53′42″), with the Muñogalindo station located 20 km upstream from
the city of Avila (Spain). Due to its geographic location, Ávila experi­ 2.2.2. Multivariate analysis for event type identification based on rainfall
ences a temperate Mediterranean climate (Köppen classification: Csb), characteristics
which is bordered by a cold semi-arid environment and features warm Multivariate statistical techniques including principle component
summers and cold winters with snowfall (BSk). The coldest month is analysis (PCA), factor analysis (FA), and numerous cluster methods have
January, with an average temperature of 3.0 ◦ C (37 ◦ F), while the hottest all been highlighted as multivariate techniques that are effective and
month is July, with an average temperature of 20.6 ◦ C (69 ◦ F). The efficient for classifying meteorological data, such as rainfall (Mrad et al.,
typical yearly rainfall is 416 mm (16.38 in). It appears to be in a rain 2020). Several studies have employed different variables, such as daily
shadow of the northern mountain ranges of the Central System (Sierra de precipitation or spatiotemporal data, to find hydrologically homoge­
Ávila) since the annual rainfall is lower than in the southern ranges of neous zones for precipitation. Wang et al. (2017) studied the depen­
the Central System. Due to its high altitude (1,132 m above sea level) dence structure of one-day and multi-day rainfall using copulas and
Ávila has the lowest winter temperatures among the provincial capital Nazeri Tahroudi et al. (2022) applied a method based on the conditional
cities of Spain. density of vine cases to drought the rainfall deficiency. Brito et al. (Brito
et al., 2017) analized spatial and temporal patterns of rainfall using a
linear regression model and cluster analysis in the state of Rio de Janeiro
2.2. Methods (Brazil). Santos et al. (2019) used hierarchical clustering to classify the
sites into different groups with similar trends in Paraíba state. Mrad
2.2.1. General approach et al. (2020) applied PCA and cluster analysis (CA) techniques to ana­
According to Restrepo-Posada and Eagleson (1982) rainfall tends to lized the variability of annual rainfall changes and identification of
arrive in clusters. On the one hand, some approximate criterions have homogeneous regions in the north-eastern region of Algeria. The annual
been proposed for the independent rainfall events and arbitrary sepa­ standardized precipitation index (SPI) based on multivariate analysis
ration. Grace and Eagleson (1967) and Sariahmed and Kisiel (1968) has been applied in the identification of rainfall in several recent
chose the storm depth for this purpose and other investigators used investigation studies in the Barbate River basin (S. Iberian Peninsula)
autocorrelation and rank correlation coefficients-based procedures. and in the State of Alagoas (Northeast Brazil) (Costa et al., 2021; Vélez-
Restrero-Posada and Eagleson (1982) proposed a criterion in parameter Nicolás et al., 2022).
estimation for Poisson models of annual rainfall frequency in an arid
climate and they used the exponentiality of the time between storms as a

Fig. 1. Case study situation and location of the Muñogalindo station at the Upper Adaja river basin.

3
C. Patino et al. Journal of Hydrology 625 (2023) 130014

2.3. Applied methodology not based on the statistical independence of rainfall events, but on
rainfall pulses. The autocorrelation coefficient ρt for a lag time t is
2.3.1. Overall methodological description computed as:
This research shows an innovative method for the characterization ∑N− t
and design of regional hyetographs with the following steps. First, due to (Y − Y)(Yi+t − Y)
ρt = i=1∑Ni 2
the large number of rainfall records, a grouping of rainfall events is i=1 (Yi − Y)

obtained by means of correlogram analysis. Then, cluster analysis is

applied to group the events into specific clusters using the silhouette where Yi is the sample of observations at time i, and Y is the sample
index. Based on the data resulting from cluster analysis (hierarchical), mean of N samples.
the hyetograph is built considering the maximum intensity found in each
of the analyzed clusters. And finally, a characterization of the clusters 2.3.3. Clustering analysis
found was carried out. Cluster analysis was used to identify groups of rainfall with similar
Statistical analysis of the 14-year hourly rainfall data recorded from physical characteristics using the following variables: volume, intensity,
2008 to 2022 was conducted. In this sense, 252,755 hourly 30-minute duration and time maximum peak.
data records were collected. The methodology followed in the study is The aim of cluster analysis is to find groups rainfall where the events
summarized in Fig. 2. in a cluster have very small differences in terms of their statistical
properties.
2.3.2. Identification of independent rainfall events It is an unsupervised analysis that divides data into similar, over­
The continuous historical rainfall series was separated into individ­ lapping, or non-overlapping groups. The advantage of clustering algo­
ual precipitation events using an Interevent Time Definition (IETD) rithms is that they generally do not assume any statistical distribution
estimation. The IETD defines a minimum dry period between indepen­ for the input data. A common method to identify groups in a dataset is
dent rainfall events as a criterion for grouping rainfall events. The pre­ Ward’s hierarchical method (Ward, 1963), considering the squared
cipitation record contains a series of precipitation pulses over time. The Euclidean distances to define events with similar rainfall. The distance,
minimum inter-event time between consecutive rain pulses (i.e. the de, is defined as:
definition of time to event, IETD) was used to extract a single storm [ ]0.5
∑ n
event from a record of continuous rain events (Adams and Papa, 2000). de = Pp.i − Pk.i
In case of the drought period between events is longer than the IETD, the i=1

storm immediately preceding and the storm immediately following are

considered separate events. On the other hand, if the time without where de is the Euclidian distance, and Pp.i and Pk.i are the quantitative
precipitation is less than the reported IETD, multiple events were com­ variables i from elements p and k, respectively.
bined into a single one whose duration is equal to the sum of the du­ In addition, the within-group sum of squares of a group is defined as
rations of the two storms plus the drought period itself between the the sum of the distance between all objects in the group and their gravity
events. centre, which is expressed by equation:
To separate out independent rainfall events from continual precipi­ ∑
nc
tation, the IETD needs to be defined initially. The IETD was developed WtGSSc = de2 (xci − xc )
based on the correlation between observed data and temporal proximity i=1

using an autocorrelation analysis. The autocorrelation coefficient in­

where: nc and xc are the size and centroid of cluster c.
dicates the correlation of data at a given point in time with data at an
The silhouette index is utilized because the Ward method does not
earlier point in time. It uses the lag time as the IETD, which is sufficient
specify the ideal number of clusters that should be produced. The da­
to isolate independent rainfall events statistically within a specified level
ta were normalized because a variable with different units usually af­
of significance meaning that the coefficient of autocorrelation ap­
fects the clustering results.
proaches zero (Joo et al., 2013; Adams and Papa 2000). Therefore, it is

Fig. 2. Research methodology flow chart of this study.

4
C. Patino et al. Journal of Hydrology 625 (2023) 130014

2.3.3.1. Clusterś characterization. The application of a clustering determine the IETD which it is a standard for identifying an individual
methodology requires the definition of features that can be used to event from a continuous rainfall record and it is decided by whether or
characterize the rain events. Clusters were characterized for typical not there is a connection between nearby events. The autocorrelogram
characteristics (volume, duration, intensity, total depth, interevent time, represents a plot of autocorrelation coefficient ρt with respect to the lag
time maximum peak, maximum peak, relative peak and season). autocorrelogram indicated that the autocorrelation coefficient is close to
The quantitative variables are described as mean ± standard devia­ zero at approximately 23.5 hrs. (Fig. 3). However, the shape of the
tion. Kruskal Wallis H test was used to identify significant differences autocorrelation graph indicates that this process has an asymptotic
between cluster and Dunn’s post hoc tests (Dunn, 1964). Chi-squared behaviour to X Axe from approximately 2.5 h.
test was applied to analyse the association of independent qualitative
variable among subgroups. Threshold for significance was set at p ≤ 3.2. Statistics of precipitation events derived by application of IETD
0.05.
Based on the IETD technique, total 414 rain events were found
2.3.3.2. Defining features to characterize rain events. Several rainfall within a period of fourteen years. Descriptive statistics of rainfall
characteristics, over the time series were calculated in order to better characteristics are shown in Table 1. Average rainfall duration of 29.51
understand the rainfall phenomena that primarily define the clusters, hrs., intensity of 0.61 mm/hr, and total depth of 7.76 mm were obtained.
such as: Event volume is defined according to the IETD as the sum of the The interevent time, which is a time frame without rain, can be
precipitation per hour reported for a separate event, and precipitation compared to the typical dry time preceding a storm. The average
duration is the period from the beginning to the end of a precipitation interevent time of 273.78 hrs were found from these rainfall data.
event. Average event intensity can be calculated by dividing event The duration of precipitation ranged from 0.5 to 198.50 h. The
volume by duration. Total depth is the amount of water that falls per amount of precipitation also ranged from 0.60 to 77.29 mm during the
square meter. Interevent time is the rain-free interval is that the time entire period. Precipitation cases with a duration of one hour had the
between the end point of the previous rain event and the start point of highest frequency with 4.1% (17 times), and a duration of less than one
the next. Time maximum peak rainfall is the hours from the start of the hour was next with 2,7% (11 times). Overall, precipitation events with a
event until the highest peak. Maximum peak is the average maximum duration of<12 h occurred a total of 146 times, and this accounted for
intensity over the years 2008–2022 and relative peak is defined as the 35.3% of the total. Therefore, it should be noted that large duration
average maximum intensity in each event. precipitation events were much more prevalent in Muñogalindo. The
amount of precipitation<10 mm had a proportion of 73.9% (306 times)
2.3.4. Hyetographś design during the 14-year period.
This starts with an analysis of similarity features of rainfall hyeto­
graphś shapes for the designated clusters, considering the duration and 3.3. Clustering analysis
maximum intensity. Since time is a continuous variable, it is considered
to group its values in intervals of the same duration (d) (temporal range) For the work purpose, a detailed examination of the similarities
for each cluster. Consequently, the intervals are established as follows:
[t_min, t_min + d], (t_min + d, t_min + 2d],…,(t_min + nd, t_max),
Table 1
where n is the number of intervals, t_min and t_max are defined as the
Descriptive statistics of average rainfall data.
minimum/maximum value respectively of the time variable.
For each of the previous intervals, the value for the maximum rainfall Rainfall Total rainfall Rainfall lnterevent
duration (hr) depth (mm) intensity (mm/ time
intensity is taken (I_max). The previous data is represented graphically hr) (hr)
as a hyetograph in which the intervals are located on the abscissa axis
Minimum 0.50 0.60 0.2 24.00
and the maximum intensities for each of them on the ordinate axis.
Value
Maximum 198.50 77.29 11.60 2877.00
3. Results Value
Mean Value 29.51 7.76 0.61 273.78
Standar 29.70 8.29 1.12 372.07
3.1. IETD determination
Deviation

An analysis based on an autocorrelation method has been used to *414 events.

Fig. 3. Autocorrelogram for determining IETD.

5
C. Patino et al. Journal of Hydrology 625 (2023) 130014

between rainfall hyetograph shapes in groups of their duration is done, Table 2

considering the maximum rainfall intensity. To separate the events in Characteristics of events according to characteristics rainfalls by clusters.
the population under study into homogeneous related subsets, a cluster Cluster 1 Cluster 2 Cluster 3 p-value
analysis was first applied. The determination of the optimum number of n = 176 n = 136 n = 102
clusters k = 3 was followed by the hierarchical clustering process. Mean ± SD Mean ± SD Mean ± SD
Cluster analysis was done using the Ward method with the following Intensitya,b 0.91 ± 1.58 0.37 ± 0.44 0.43 ± 0.56 <0.001
rainfall parameters: duration, volume, intensity, hours from the start of Durationa,b,cbase 13.03 ± 58.52 ± 19.26 ± <0.001
the event until the highest peak and interevent time. Ward method is a 13.51 32.67 12.21
Volumena,c 4.39 ± 3.93 14.64 ± 4.41 ± 3.14
hierarchical clustering that uses the increase in the total sum of squares
<0.001
10.59
within the cluster as a result of joining clusters. The average value of the lnterevent timea,b 398.11 ± 157.73 ± 215.18 ± <0.001
silhouette index, S, was calculated for 2–10 clusters. Based on the 497.38 163.52 233.23
silhouette index, the optimum number of clusters for storm events is Time maximum 3.02 ± 4.87 7.38 ± 7.72 21.53 ± <0.001
picka,b,c 7.52
three. Maximum picka,c 1.58 ± 1.52 2.34 ± 2.48 1.46 ± 1.23 <0.001
Number of relative 3.5 ± 2.66 10.71 ± 4.12 ± 2.49 <0.001
3.3.1. Clusterś characterisation peaksa,c 5.64
To characterize the groups of independent events it has been Season 53 43 30 0.260
Spring (42.1%)34 (34.1%)16 (23.8%)10
considered as illustrated in Fig. 4, that if the time increment between
Summer (56.7%)38 (26.7%)32 (16.7%)31
two events is greater than the IETD they are two separate events. If the Winter (37.6%)51 (31.7%)45 (30.7%)31
time between events is less than or equal to the IETD, it is the same Autumn (40.2%) (35.4%) (24.4%)
event.
P values are for comparison of clusters by Kruskal-Wallis H test with Dunn’s post
Table 2 shows a general description of the clusters according to the hoc comparisons for quantitative variables and test chi-squares for categorical
selected variables. variables.
Cluster 1 consists of 176 events and includes the largest number of Abbreviations: SD, standard deviation.
events. In this cluster, there was a high average rainfall intensity with a a
p < 0.05 between cluster 1 and cluster 2 (p < 0.05).
value of 0.91 mm/hr and SD of 1.58. Also, it has the highest average b
p < 0.05 between cluster 1 and cluster 3 (p < 0.05).
c
value of the interevent time. The average and the standard deviation of p<:0.05 between cluster 2 and cluster 3 (p < 0.05).
interevent time is of 398.11hr and 497.38 respectively.
Cluster 2 consists of 136 events (42.5% of the events). This group had 3.4. Hyetographs
the highest duration overall at 58.52 hrs (SD = 32.67), volume (mean =
58.52 mm, SD = 32.67), time maximum peak with a mean value of the In cluster 1, the values for the eventś duration varies from 0.5 to 81 h.
7.38 hrs and SD equal to 7.72 and, it presents top average event intensity On the other hand, the amplitude of the interval is defined from 0 to 3.
(mean = 2.34 mm/h, SD = 2.48). The average and the standard devia­ The rainfall intensity values in this cluster ranged from 0.5 to 11.60 mm/
tion of interevent time is of 157.73 hr (SD = 163.52). h. I_max value is, therefore, 11.60 mm/h. In cluster 2 the range of the
Cluster 3 represents 24.64% (n = 102) of the events. This clusters variable time varies from 8.5 to 198.5 h, and the amplitude of the
reflects the highest number of relative peaks (mean = 21.53, SD = 7.52). considered values oscillates from 0 to 9. In cluster 2, the I_max reaches
The average of interevent time is 215.18 hrs. 13.36 mm/h. And finally, in cluster 3 the duration varies from 1 to 48 h
Kruskal-Wallis ANOVA confirmed distinct differences between the and the amplitude of the considered interval varies from 0 to 3. The
clusters based on characteristics rainfalls. The results of the Chi-Square minimum (I_min) and maximum rainfall intensity (I_max) values are
statistical test of independence show a calculated p-value > 0.05. This 0.19 and 3.4 respectively.
reveals the nonexistence of a statistically significant association and The most numerous cluster No. 1 covers 42% of the analyzed pre­
independence between clusters and season. cipitation events. Maximum height rises in this cluster of precipitation
occur during the first third of their duration. This cluster’s precipitation

Fig. 4. Separation of independent rainfall events (IETD).

6
C. Patino et al. Journal of Hydrology 625 (2023) 130014

is characterized by generally even downward height fluctuations. more detail, the results obtained in this study identify three well-
Cluster No. 2 (33%) of events is the second most common, in which differentiated clusters to obtain a multivariate characterisation of the
maximum height increases occur in the first half of the duration. Cluster rainfall: the first group consisted of strong intensity, moderately long-
No. 3 has a proportion of 25% and the rainfall belonging to the cluster lasting events with the peak intensity centred in the first quarter, just
has a similar to cluster 2. There are no obvious associations between the before the event started. The second group was characterized by oc­
three clusters that were examined, which simply serves to highlight how currences of long duration, moderate intensity, presumably related with
random the events under study are. The outcomes of the hierarchical a major precipitation front sustained throughout time, while the third
cluster analysis used to group the precipitation enabled for the verifi­ group was characterized by events of extended duration till the
cation of the maxim and relative peak. Fig. 5 shows the calculated maximum peak is reached with moderate intensity and duration. Hye­
averaged dimensionless cumulative storm rainfall hyetographs. toclust becomes a general tool that can be applied to any meteorolog­
ical/hydrological context. This new tool can assist on the prediction of
4. Discussion and conclusions events (ordinary and extremal) and the effective integration to Early
Warning Systems. Through Hyetoclust application it is expected to
Despite the need for improved quantitative rainfall analysis has reduce the amount of losses and damage associated to extreme hydro­
grown in recent years, it is not possible yet to robustly determine or logical phenomena.
predict the duration and intensity of associated extremal events such as
droughts or floods and their months of onset and termination. This re­ CRediT authorship contribution statement
quires modernisation of methods dealing with recorded precipitation
series processing. Necessary analytical elements of such modernization Carmen Patino: Conceptualization, Methodology, Formal analysis,
is modelled or synthetic hyetographs. The usage of data mining tech­ Visualization, Investigation. Jose-Luis Molina: Conceptualization,
nologies, particularly the cluster analysis, is justified and required due to Methodology, Formal analysis, Visualization, Investigation. Fernando
the ever-growing sets of rainfall records. The classification of storm Espejo: Methodology, Software, Formal analysis, Investigation. San­
rainfall hyetographs has already used the application of cluster analysis tiago Zazo: Conceptualization, Methodology. Abedin Mohammad-
in the examination of the temporal distributions of rainfall. However, Hosseinpour: Supervision, Validation, Formal analysis. Fernando
there are not studies using the silhouette index to make clustering al­ Silla: Supervision, Validation, Formal analysis.
gorithm, i.e. an internal clustering validation index. In addition, this
study is pioneer in constructing the hyetograph considering the Declaration of Competing Interest
maximum rainfall intensity in each group found by cluster analysis and
then, characterizing the clusters according to the physical characteris­ The authors declare that they have no known competing financial
tics of the rainfall. Therefore, this analysis becomes an easy interpretable interests or personal relationships that could have appeared to influence
graphical representation of the maximum intensity of the rain over time. the work reported in this paper.
In conclusion, the way of identifying clusters based on the previous
methodological approach has allowed to robustly differentiate the three
main of hyetographs that represents the three main types of rainfall. In

Fig. 5. Model rainfall hyetographs developed by clustering method. The horizontal axis shows the maximum rainfall intensity in cluster 1 every three hours, cluster 2
every 9 h and cluster 3 every 3 h and the vertical axis shows the vertical axis shows the time intervals.

7
C. Patino et al. Journal of Hydrology 625 (2023) 130014

Data availability Hong, M., Kim, J., Jeong, S., 2018. Rainfall intensity-duration thresholds for landslide
prediction in South Korea by considering the effects of antecedent rainfall.
Landslides 15, 523–534. https://doi.org/10.1007/s10346-017-0892-x.
Data will be made available on request. Howard, C.D.D., 1976. Theory of storage and treatment-plant overflows. J. Environ. Eng.
Div. 102, 709–722. https://doi.org/10.1061/JEEGAV.0000508.
Acknowledgments Iadanza, C., Trigila, A., Napolitano, F., 2016. Identification and characterization of
rainfall events responsible for triggering of debris flows and shallow landslides.
J. Hydrol. 541, 230–245. https://doi.org/10.1016/j.jhydrol.2016.01.018.
The development of this research was partially funded by Call Joo, J., Lee, J., Kim, J., Jun, H., Jo, D., 2013. Inter-event time definition setting
TCUE_2022. PC_TCUE21-23_002 project. Funding: This research was procedure for urban drainage systems. Water 6, 45–58. https://doi.org/10.3390/
w6010045.
also funded by Ministry of Science and Innovation, TED2021-129478B- Jun, C., Qin, X., Chen, M., Seo, H., 2021. Investigating event-based temporal patterns of
I00, through Research Project “Intelligent and Digital System for the design rainfall in a tropical region. Hydrol. Sci. J. 66, 1986–1996. https://doi.org/
ecological Restoration of DEgraded reservoirS (SID_REDES)”, granted to 10.1080/02626667.2021.1967958.
Knerr, I., Trachte, K., Garel, E., Huneau, F., Santoni, S., Bendix, J., 2020. Partitioning of
the Research Group “INGENIERÍA Y GESTIÓN DEL AGUA (IGA)” of large-scale and local-scale precipitation events by means of spatio-temporal
Salamanca University, Spain. precipitation regimes on corsica. Atmosphere (Basel). 11, 417. https://doi.org/
10.3390/atmos11040417.
Lewis, E., Fowler, H., Alexander, L., Dunn, R., McClean, F., Barbero, R., Guerreiro, S.,
References Li, X.-F., Blenkinsop, S., 2019. GSDR: a global sub-daily rainfall dataset. J. Clim. 32,
4715–4729. https://doi.org/10.1175/JCLI-D-18-0143.1.
Abolverdi, J., Ferdosifar, G., Khalili, D., Kamgar-Haghighi, A.A., 2016. Spatial and Lima, A.O., Lyra, G.B., Abreu, M.C., Oliveira-Júnior, J.F., Zeri, M., Cunha-Zeri, G., 2021.
temporal changes of precipitation concentration in Fars province, southwestern Iran. Extreme rainfall events over Rio de Janeiro State, Brazil: characterization using
Meteorol. Atmos. Phys. 128, 181–196. https://doi.org/10.1007/s00703-015-0414- probability distribution functions and clustering analysis. Atmos. Res. 247, 105221
0. https://doi.org/10.1016/j.atmosres.2020.105221.
Aronica, G.T., Freni, G., 2005. Estimation of sub-hourly DDF curves using scaling Lyra, G.B., Oliveira-Júnior, J.F., Zeri, M., 2014. Cluster analysis applied to the spatial and
properties of hourly and sub-hourly data at partially gauged site. Atmos. Res. 77, temporal variability of monthly rainfall in Alagoas state, Northeast of Brazil. Int. J.
114–123. https://doi.org/10.1016/j.atmosres.2004.10.025. Climatol. 34, 3546–3558. https://doi.org/10.1002/joc.3926.
Barton, Y., Giannakaki, P., von Waldow, H., Chevalier, C., Pfahl, S., Martius, O., 2016. McPhillips, L.E., Chang, H., Chester, M.V., Depietri, Y., Friedman, E., Grimm, N.B.,
Clustering of Regional-Scale Extreme Precipitation Events in Southern Switzerland. Kominoski, J.S., McPhearson, T., Méndez-Lázaro, P., Rosi, E.J., Shafiei Shiva, J.,
Mon. Weather Rev. 144, 347–369. https://doi.org/10.1175/MWR-D-15-0205.1. 2018. Defining extreme events: a cross-disciplinary review. Earth’s Futur. 6,
Benabdelouahab, T., Gadouali, F., Boudhar, A., Lebrini, Y., Hadria, R., Salhi, A., 2020. 441–455. https://doi.org/10.1002/2017EF000686.
Analysis and trends of rainfall amounts and extreme events in the Western Mikołajewski, K., Ruman, M., Kosek, K., Glixelli, M., Dzimińska, P., Ziętara, P.,
Mediterranean region. Theor. Appl. Climatol. 141, 309–320. https://doi.org/ Licznar, P., 2022. Development of cluster analysis methodology for identification of
10.1007/s00704-020-03205-4. model rainfall hyetographs and its application at an urban precipitation field scale.
Brito, T.T., Oliveira-Júnior, J.F., Lyra, G.B., Gois, G., Zeri, M., 2017. Multivariate analysis Sci. Total Environ. 829, 154588 https://doi.org/10.1016/j.scitotenv.2022.154588.
applied to monthly rainfall over Rio de Janeiro state. Brazil. Meteorol. Atmos. Phys. Molina, J.-L., Espejo, F., Zazo, S., Molina, M.-C., Hamitouche, M., García-Aróstegui, J.-L.,
129, 469–478. https://doi.org/10.1007/s00703-016-0481-x. 2022. HydroPredicT_Extreme: a probabilistic method for the prediction of extremal
Chen, Y., Paschalis, A., Wang, L.P., Onof, C., 2021. Can we estimate flood frequency with high-flow hydrological events. J. Hydrol. 610, 127929 https://doi.org/10.1016/j.
point-process spatial-temporal rainfall models? J. Hydrol. 600, 126667 https://doi. jhydrol.2022.127929.
org/10.1016/j.jhydrol.2021.126667. Molina, J.-L., Zazo, S., 2018. Assessment of temporally conditioned runoff fractions in
Chen, L., Singh, V.P., 2018. Entropy-based derivation of generalized distributions for unregulated rivers. J. Hydrol. Eng. 23 https://doi.org/10.1061/(ASCE)HE.1943-
hydrometeorological frequency analysis. J. Hydrol. 557, 699–712. https://doi.org/ 5584.0001645.
10.1016/j.jhydrol.2017.12.066. Molina-Sanchis, I., Lázaro, R., Arnau-Rosalén, E., Calvo-Cases, A., 2016. Rainfall timing
Chimene, C.A., Campos, J.N.B., 2020. The design flood under two approaches: synthetic and runoff: the influence of the criterion for rain event separation. J. Hydrol.
storm hyetograph and observed storm hyetograph. J. Appl. Water Eng. Res. 8, Hydromechanics 64, 226–236. https://doi.org/10.1515/johh-2016-0024.
171–182. https://doi.org/10.1080/23249676.2020.1787242. Morbidelli, R., Saltalippi, C., Flammini, A., Govindaraju, R.S., 2018. Role of slope on
Choo, K.-S., Kang, D.-H., Kim, B.-S., 2020. Impact assessment of urban flood on traffic infiltration: a review. J. Hydrol. 557, 878–886. https://doi.org/10.1016/j.
disruption using rainfall–depth–vehicle speed relationship. Water 12, 926. https:// jhydrol.2018.01.019.
doi.org/10.3390/w12040926. Mrad, D., Dairi, S., Boukhari, S., Djebbar, Y., 2020. Applied multivariate analysis on
Costa, M.d.S., Oliveira-Júnior, J.F.d., Santos, P.J.D., Correia Filho, W.L.F., Gois, G.d., annual rainfall in the northeast of Algeria. J. Water Clim. Chang. 11, 1165–1176.
Blanco, C.J.C., Teodoro, P.E., Silva Junior, C.A.d., Santiago, D.d.B., Souza, E.d.O., https://doi.org/10.2166/wcc.2019.272.
Jardim, A.M.d.R.F., 2021. Rainfall extremes and drought in Northeast Brazil and its Nazeri Tahroudi, M., Ramezani, Y., Michele, C., Mirabbasi, R., 2022. Multivariate
relationship with El Niño-Southern Oscillation. Int. J. Climatol. 41 (S1) https://doi. analysis of rainfall and its deficiency signatures using vine copulas. Int. J. Climatol.
org/10.1002/joc.6835. 42, 2005–2018. https://doi.org/10.1002/joc.7349.
da Silva, F.P., Rotunno Filho, O.C., Justi da Silva, M.G.A., Sampaio, R.J., Pires, G.D., de O’Gorman, P.A., 2015. Precipitation extremes under climate change. Curr. Clim. Chang.
Araújo, A.A.M., 2020. Identification of rainfall and atmospheric patterns associated Reports 1, 49–59. https://doi.org/10.1007/s40641-015-0009-3.
with Quitandinha River flooding events in Petropolis, Rio de Janeiro (Brazil). Nat. Restrepo-Posada, P.J., Eagleson, P.S., 1982. Identification of independent rainstorms.
Hazards 103, 3745–3764. https://doi.org/10.1007/s11069-020-04153-y. J. Hidrol. 55 (1-4), 303–319.
Dunkerley, D., 2008. Identifying individual rain events from pluviograph records: a Sa’adi, Z., Shahid, S., Ismail, T., Chung, E.-S., Wang, X.-J., 2019. Trends analysis of
review with analysis of data from an Australian dryland site. Hydrol. Process. 22, rainfall and rainfall extremes in Sarawak, Malaysia using modified Mann-Kendall
5024–5036. https://doi.org/10.1002/hyp.7122. test. Meteorol. Atmos. Phys. 131 (3), 263–277.
Dunn, O.J., 1964. Multiple comparisons using rank sums. Technometrics 6 (3), 241–252. Sahu, R., Verma, M., Ahmad, I., 2022. Segmental variability of precipitation in the
Fahad, S., Su, F., Khan, S.U., Naeem, M.R., Wei, K., 2023. Implementing a novel deep Mahanadi River basin from 1901 to 2017. Geocarto Int. 1–22.
learning technique for rainfall forecasting via climatic variables: an approach via Santos, C., Brasil Neto, R., da Silva, R., Costa, S., 2019. Cluster analysis applied to
hierarchical clustering analysis. Sci. Total Environ. 854, 158760 https://doi.org/ spatiotemporal variability of monthly precipitation over paraíba state using tropical
10.1016/j.scitotenv.2022.158760. rainfall measuring mission (TRMM) data. Remote Sens. 11, 637. https://doi.org/
Freitas, E. da S., Coelho, V.H.R., Xuan, Y., Melo, D. de C.D., Gadelha, A.N., Santos, E.A., 10.3390/rs11060637.
Galvão, C. de O., Ramos Filho, G.M., Barbosa, L.R., Huffman, G.J., Petersen, W.A., Sariahmed, A., Kisiel, C., 1968. Synthesis of sequences of summer thunderstorm volumes
Almeida, C. das N., 2020. The performance of the IMERG satellite-based product in for the Atterbury watershed in the Tucson area, in: Proc. Int. Assoc. Hydrol. Sci.
identifying sub-daily rainfall events and their properties. J. Hydrol. 589, 125128. Syrup. on Use of Analog and Digital Computers in Hydrology. pp. 439–447.
https://doi.org/10.1016/j.jhydrol.2020.125128. Thakur, N., Karmakar, S., Soni, S., 2021. Rainfall Forecasting Using Various Artificial
Gioia, A., Lioi, B., Totaro, V., Molfetta, M.G., Apollonio, C., Bisantino, T., Iacobellis, V., Neural Network Techniques - A Review. Int. J. Sci. Res. Comput. Sci. Eng. Inf.
2021. Estimation of peak discharges under different rainfall Technol. 506–526. https://doi.org/10.32628/CSEIT2173159.
depth–duration–frequency formulations. Hydrology 8, 150. https://doi.org/ Tokar, A.S., 1996. Rainfall-runoff modeling in an uncertain environment. University of
10.3390/hydrology8040150. Maryland, College Park.
Grace, R., Eagleson, P., 1967. A model for generating synthetic sequences of short-time- Vélez-Nicolás, M., García-López, S., Ruiz-Ortiz, V., Zazo, S., Molina, J.L., 2022.
interval rainfall depths, in: Proc. Int. Hydrol. Syrup., Fort Collins, Colo.,. pp. Precipitation variability and drought assessment using the SPI: application to long-
268–276. term series in the strait of Gibraltar area. Water 14, 884. https://doi.org/10.3390/
Hamitouche, M., Molina, J.-L., 2022a. Event-based Bayesian causal modelling for flood w14060884.
hydrograph prediction, Upper Andarax intermittent stream. Spain. J. Hydrol. Reg. Wang, Y., Li, C., Liu, J., Yu, F., Qiu, Q., Tian, J., Zhang, M., 2017. Multivariate analysis of
Stud. 44, 101260 https://doi.org/10.1016/j.ejrh.2022.101260. joint probability of different rainfall frequencies based on copulas. Water 9, 198.
Hamitouche, M., Molina, J.-L., 2022b. A review of AI methods for the prediction of high- https://doi.org/10.3390/w9030198.
flow extremal hydrology. Water Resour. Manag. 36, 3859–3876. https://doi.org/
10.1007/s11269-022-03240-y.

8
C. Patino et al. Journal of Hydrology 625 (2023) 130014

Ward, J.H., 1963. Hierarchical grouping to optimize an objective function. J. Am. Stat. Wong, K.W., Wong, P.M., Gedeon, T.D., Fung, C.C., 2003. Rainfall prediction model
Assoc. 58 (301), 236–244. using soft computing technique. Soft comput. - a fusion found. Methodol. Appl. 7,
Wartalska, K., Kaźmierczak, B., Nowakowska, M., Kotowski, A., 2020. Analysis of 434–438. https://doi.org/10.1007/s00500-002-0232-4.
hyetographs for drainage system modeling. Water 12, 149. https://doi.org/10.3390/
w12010149.