Environmental Modelling & Software 97 (2017) 243e258

Contents lists available at ScienceDirect

Environmental Modelling & Software

journal homepage: www.elsevier.com/locate/envsoft

A rainfall analysis and forecasting tool

Javier Diez-Sierra, Manuel del Jesus*

gico de Cantabria, 39011, Santander,
Environmental Hydraulics Institute, Universidad de Cantabria - Avda. Isabel Torres, 15, Parque Científico y Tecnolo
Spain

a r t i c l e i n f o a b s t r a c t

Article history: MENSEI-L is a stand-alone software tool for the automatic analysis of pluviometric networks, that also
Received 24 February 2017 provides three-day rainfall forecasts based on weather types. The software tool, implemented in Python
Received in revised form and R, is able to fill missing values in original daily data series and to generate synthetic pluviometers in
19 July 2017
ungauged locations, by means of kriging techniques. MENSEI-L also characterizes punctual and spatial,
Accepted 23 August 2017
average and extreme distributions of precipitation for the complete pluviometric network. Tenerife
(Canary Islands, Spain) is used as study site to evaluate the capabilities of MENSEI-L and the implicit
rainfall analysis methodology that it implements. MENSEI-L proves to be a useful tool to extract infor-
Keywords:
Rainfall modeling
mation from dense observation networks where manual analysis is not practical.
Weather types © 2017 Elsevier Ltd. All rights reserved.
Spatial analysis
Kriging
Rainfall extremes

1. Introduction
Rainfall is the main driver of the continental portion of the
water cycle below parallel 50+ . Rainfall characterization is
important for an accurate modeling of other hydrological fluxes
such as runoff and evapotranspiration, as rainfall controls their
statistical behavior, both average and extreme distributions. Even
more, in small arid and semiarid watersheds, where runoff and
discharge measurements may not be practical, an accurate spatio-
temporal rainfall characterization becomes an essential part of the
assessment of hydrological fluxes. These fluxes may be required
for infrastructure design or risk assessments, effectively turning
rainfall into an important part of the design and the decision-
making processes (Xu et al., 2010; Paixao et al., 2011; Arnbjerg-
Nielsen et al., 2013; Renschler et al., 1999; Sugiyama et al., 1995).
Furthermore, the incorporation of quantitative precipitation fore-
casting (QPF) may also be required for many hydrological appli-
cations especially for real-time flood forecasting, as it is a critical
component of public safety and quality of life. The potential risk
associated to flash floods is related to two main characteristics:
their rapid occurrence and to the spatial dispersion of the areas
affected by these floods. Both characteristics limit the ability to
issue timely flood warnings (Borga et al., 2011) turning flash floods
* Corresponding author.
E-mail address: manuel.deljesus@unican.es (M. del Jesus).
into one of the most dangerous natural hazards (Blo €schl et al.,
2008).
Rainfall is measured by pluviometers, which only capture rain-
fall data for a very small portion of the territory, and thus, tend to be
deployed in networks to cover larger areas. On top of collecting
rainfall data for a large area, pluviometric networks provide useful
information for the spatio-temporal characterization of rainfall and
are thus important to inform engineering design and decision-
making processes. In addition, the availability of dense pluvio-
metric networks offers the opportunity to build short-term sto-
chastic rainfall forecasting systems making predictions that cover
the entire area of interest. Rainfall data collection and treatment,
however, presents a series of complications. Measurements for
specific events may be missing, due to equipment malfunctions or
other reasons, creating discontinuities in the time series. Equip-
ment upgrades may lead to differences in time resolution for a
given gauge for different time periods. Newer gauges, with better
quality data, provide shorter time series, limiting their usability.
Data collection from pluviometric networks leads to generating
large databases from which manually extracting and interpreting
information is complicated and time consuming. Moreover, the
implementation of statistical short-term forecasting systems often
leads to deal with sophisticated techniques and large atmospheric
databases. For these reasons, statistical tools, machine learning
techniques and software applications become indispensable to
convert rainfall and atmospheric data into information, ready to
support professional practitioners and managers.

http://dx.doi.org/10.1016/j.envsoft.2017.08.011
1364-8152/© 2017 Elsevier Ltd. All rights reserved.
244 J. Diez-Sierra, M. del Jesus / Environmental Modelling & Software 97 (2017) 243e258
Nowadays, different software applications exist to assist tech-
nicians in the process of rainfall analysis. STORMS 2010 (Sabourin
and Associates Inc., 2010) is a commercial application able to
analyze local rainfall information to fit average and extreme dis-
tributions. It computes Intensity-Duration-Frequency (IDF) curves,
performs intensity trend analysis and is able to create design
storms. CLIMATOL (Guijarro, 2016) is a time series homogenization
software distributed as an R (R Core Team, 2015) package under an
open source license. It was designed to deal with general climato-
logical time series, not only with rainfall. Similar software appli-
cations are MASH (Szentimrey, 2006), SPLIDHOM (Mestre et al.,
2011) and EXTRAQC (Aguilar and Prohom, 2011), tools designed
for homogenizing and checking the internal coherence of daily time
series, and AnCli (Stepanek, 2005), which also provides statistical
characterization tools.
Software libraries and packages have also been developed for
the treatment of rainfall data. One interesting example is reddPrec
(Serrano-Notivoli et al., 2017), an R package that contains a set of
functions for controlling the quality of rainfall time series, filling
missing values and creating virtual series (and gridded datasets) in
ungauged locations. Other softwares such as RainSim (Burton et al.,
2008) or MRS (Mehrotra et al., 2015a) are able to simulate sto-
chastic rainfall at multiple locations preserving the most important
statistics (Burton et al., 2013; Kretzschmar et al., 2014; Mehrotra
et al., 2015b).
To the best of the author's knowledge, there are not too many
tools specifically developed to perform short-term rainfall fore-
casting (Neiman et al., 2009; Wright et al., 2017). In most cases
rainfall forecasting is dealt with by means of numerical weather
prediction models or the extrapolation of remote sensing obser-
vations (radar data and satellite images) to the current weather
conditions. Unfortunately, while the first do not seem able to pro-
vide accurate rainfall forecasts at the temporal and spatial resolu-
tion required by many hydrologic applications, neither the outputs
from satellite or radar images allow a satisfactory assessment of
rainfall intensity yet (Toth et al., 2000; Krzysztofowicz, 1995; Liu
et al., 2013; Moser et al., 2015).
All the reviewed applications and techniques serve to charac-
terize, homogenize and interpolate punctual rainfall time series, to
generate stochastic rainfall series or to forecast rainfall. However,
no software package provides an integrated methodology to
automatically analyze the spatio-temporal characteristics of rainfall
for a pluviometric network and generate short-term rainfall fore-
casting. This paper presents such methodology, implemented into
an automated rainfall analysis software tool called MENSEI-L, that
allows to generate point and spatial characterizations of average
and extreme distributions of rainfall, as well as to make short-term
stochastic rainfall forecasts using an analogs method based on
weather typing. An application to Tenerife (Canary islands, Spain) is
used as a calibration and validation step for the methodology
developed in the present work.
In order to make the exposition clearer and easier to follow, the
theoretical basis of each section will be presented first and the
specific implementation for the Tenerife study case will follow. For
this reason, the paper begins with a description of the study area
and the databases used in the study. Then, a description of the base
methods used to carry out the rainfall analysis is presented. These
methods are grouped in three categories: the generation of
weather types, the characterization of point statistics of rainfall,
and the implementation of a spatio-temporal interpolation scheme
based on geostatistics. The paper continues presenting the work
flow followed to carry out the rainfall analysis, explaining how the
basic methods are used to construct robust rainfall analysis tools.
Afterwards, the scheme of the stochastic rainfall forecasting system
is described. The main body of the paper ends with a description of
the implementation details of MENSEI-L, to finally close with future
developments, limitations and the conclusions of the work.
2. Description of the study area and the information sources
2.1. Description of the study area
Tenerife is one of the seven islands that form the Spanish ar-
chipelago of the Canary Islands. Tenerife (see Fig. 1) is located in the
Atlantic Ocean, 300 Km west of the African coast, between parallels
28+ N and 29+ N, and between meridians 16+ W and 17+ W. Tenerife
covers a surface area of 2034:38 Km2 , with a population of 891; 111
inhabitants. It is the largest, and the most populated of the Canary
Islands and of the whole Macaronesia region.
Tenerife displays a strong topographic gradient from the coast to
the central part of the island. The central region of the island is a
mountain range, where El Teide (3:718 m) is the highest peak, that
separates the island in two well differentiated climatic areas: a
northern region, relatively wet due to the effect of the Trade winds,
and a southern region, drier due to the blocking induced by the
orography. Spatial heterogeneities of climate are also present
within each of the regions, configuring a widely varied range of
climates within a relatively small area.
The precipitation in Tenerife is very scarce, with an annual
average of 233 mm for the whole island. However, its temporal and
spatial distribution is very heterogeneous since a very high per-
centage of precipitation falls in short periods of time and with a
great spatial inhomogeneity. The singular orographic and climatic
characteristics of the island combined with the urbanization pro-
cesses carried out in the last years have transformed flash floods
into a serious problem. If fact, in the twentieth century, almost one
hundred floods were recorded in Tenerife causing important losses
both materials and humans.
2.2. Description of data sources
Information regarding synoptic patterns of atmospheric vari-
ables is obtained from two sources. On the one hand, historical
information is obtained from the global reanalysis NCEP Climate
Forecast System Reanalysis (CFSR) (Saha et al., 2010), together with
its continuation, the Climate Forecast System Version 2 (CFSV2),
developed by the National Oceanic and Atmospheric Administra-
tion (NOAA). In this paper, the term CFSR will be used to refer both
databases, CFSR and CFSV2. CFSR is a third-generation reanalysis
database, generated with a global high-resolution land-ocean-at-
mosphere model coupled with an ice sheet model. CFSR provides 6-
hourly data. The spatial resolution of the atmospheric model is
38 Km horizontally, with 64 vertical levels. The spatial resolution of
the oceanic model is approximately 0:25+ near the Equator, and
0:5+ beyond the Tropics, with 40 vertical levels. The land model
counts with 4 soil levels, and the ice model with 3 levels. CFSR and
CFSv2 together cover the period 1979e2015.
On the other hand, forecasting information is obtained from the
Global Forecast System (GFS), also developed by NOAA. GFS pro-
vides 16 days-into-the-future forecasts with a model similar to the
one used to generate CFSR.
Rainfall data is provided by Consejo Insular de Agua de Tenerife
(CIATF), the water planning and managing agency for Tenerife Is-
land. CIATF maintains a database of rainfall observations (Melia
n
et al., 2011) that includes the observation network of Agencia
Estatal de Meteorología (AEMET), the Spanish national meteorolog-
ical agency, and AgroCabildo, a local agency for agriculture devel-
opment. The CIATF database contains information for 125 gauges,
combining daily and sub-daily values, with an average coverage of
15 years, although the longest series start in 1890.
 J. Diez-Sierra, M. del Jesus / Environmental Modelling & Software 97 (2017) 243e258
Fig. 1. Tenerife island location.
CIATF also provided a digital elevation model (DEM) of 5 m
spatial resolution for the whole island.
3. Methodology
The rainfall analysis methodology and MENSEI-L's imple-
mentation are articulated around three major procedures. The first
one is the generation of weather types, that serves to classify the
time series of climatic conditions into a limited number of repre-
sentative synoptic patterns. These patterns capture most of the
observed variability and serve to cluster together similar climatic
situations (Huth et al., 2008). Weather types are computed for at-
mospheric variables and for rainfall.
The second procedure is the computation of point rainfall sta-
tistics, i.e. the analysis of isolated rainfall time series. Average and
extreme distributions are both considered. An automatic procedure
is developed to select the best fitting parametric distribution in
each case.
The last procedure is spatio-temporal interpolation, that makes
use of geostatistical methods. Rainfall measurements are taken at
specific locations, but some data may be missing, or may be
required at locations different from where the measuring devices
are deployed. Both situations require an interpolation procedure
able to make use of all the information in the database to provide
robust estimates.
It is impotant to notice that the methodologies applied in
MENSEI-L have been developed and validated at a daily temporal
resolution; due to the natural heterogeneity of precipitation at
small aggregation scales, it is likely that the higher the temporal
resolution, the less accurate the results will be.
The following sections describe in detail these three major
procedures that constitute the basis of the analysis.
3.1. Generating weather types
Atmospheric weather types. Atmospheric weather types, weather
types from now on, are representative synoptic patterns of the state
245
of the atmosphere that are selected in such a way that they
constitute an optimal basis to express the state of the atmosphere
and its natural variability through a reduced ensemble of situations.
For the purposes of this study, the state of the atmosphere at a
specific time is represented by the combined fields of sea level
pressure (SLP), elevation of the 500 hPa geopotential surface
(GH500) and elevation of the 850 hPa geopotential surface
(GH850). SLP, GH500 and GH850 fields, that cover the area between
45+ W and 5+ E, and 20+ N and 50+ N, are extracted from the CFSR
and CFSv2 databases (6-hourly information over a global 0:5+ x
0:5+ mesh) and aggregated to a daily temporal resolution. The se-
lection of variables, as well as their area coverage, were chosen
based on analysis found in literature (Herrera et al., 2001; Tullot,
1959).
Weather types are computed through a two-step process. In the
first place, a principal components analysis (PCA) (Abdi and
Williams, 2010) is applied to reduce the dimensionality of the
data. Afterwards, a clustering or a selection algorithm is applied to
the reduced-dimensional dataset to determine the optimal
weather types. Several algorithms can be used to determine the
optimal weather types: k means (KMA), maximum dissimilarity
algorithm (MDA) and self-organized maps (SOM). The different
algorithms may provide different weather types, all of them
optimal according to the criteria implicitly or explicitly established
by each specific algorithm. Further detail about clustering algo-
rithms and their application to climate characterization can be
found in Camus et al. (2011). In the present study, the KMA algo-
rithm is used (as implemented in the scikit-learn python library,
Pedregosa et al. (2011)), as it best serves the purpose of grouping all
the observed situations into homogeneous groups for statistical
purposes.
Fig. 2 shows the SLP component of the weather types deter-
mined for the rainfall analysis in Tenerife. The total number of
weather types, 81 in this case, is selected by means of a sensitivity
analysis of the storm type statistics to the number of weather types.
The number of types is selected to minimize the total variance of
the rainfall clustering, as well as the intragroup variance.
246 J. Diez-Sierra, M. del Jesus / Environmental Modelling & Software 97 (2017) 243e258
Fig. 2. SLP Weather types. Image created with MENSEI-L.
Storm types. In the present work, storm types are the weather
types corresponding to representative rainfall patterns. Storm
types could be determined in the same way atmospheric weather
types were: by a combination of a dimensionality reduction and a
clustering algorithm. However, to be able to make stochastic rain-
fall forecasts, storm types are defined as the representative rainfall
patterns corresponding to every weather type. A storm type is the
average rainfall probability and the average rainfall depth pattern
induced by all the atmospheric situations contained within a
weather type. It is assumed that with a sufficiently high number of
weather types, every weather type induces a narrow collection of
rainfall patterns that can be properly captured and represented by
their average. When storm types are defined this way, their gen-
eration does not involve additional classifications, it derives from
the initial classification done for the atmospheric conditions.
Fig. 3 shows the spatial distribution of rainfall probability for
every storm type. Each colored point corresponds to the rainfall
probability (total depth larger than 0.1 mm) of a pluviometer. It can
be seen that there are storm types (6, 7, 8, 9, 41 and 66) with greater
rainfall probability. These storm types correspond to specific
weather type such as cold depressions or low-pressure systems
from temperate zones. The 81 weather types and their corre-
sponding storm types, allow to explore the whole spectrum of
synoptic patterns that affect Tenerife island. The figure also shows
that storm types represent different spatial patterns of rainfall,
indicating that the total number of types considered is able to
properly capture the overall variability.
3.2. Computing point statistics
Rainfall data captured by pluviometers are kept in time series; a
collection of paired values that register the time of occurrence
together with the depth or intensity of rainfall. Time series data
may be directly summarized through some statistics (average,
standard deviation, lag-one correlation, etc). However, a complete
statistical description of the data is normally useful and thus
parametric distributions are fitted for all the data.
Average distributions. Average distributions aim to characterize
the behavior of the complete population of a variable. The precip-
itation amount distribution tends to be strongly skewed at a daily
time scale, and is commonly assumed to be gamma distributed
(Ben Alaya et al., 2017; Stephenson et al., 1999; Yang et al., 2005).
Nevertheless, some authors suggest that the gamma distribution
can be unsuitable for modeling precipitation extremes since it is
very restrictive and cannot account for features like heavy tails.
Therefore, to treat this issue, average distributions are fitted to a
collection of parametric distributions that includes: the lognormal
distribution, the Pearson-III distribution, the Generalized Extreme
Value (GEV) distribution, the Rayleigh distribution, and the Gamma
distribution (Serinaldi and Kilsby, 2014).
It is important to remark that GEV is fitted to the complete
population as a general three parameter distribution, without any
specific treatment of extremes. Distributions are fitted by the
maximum likelihood method. The best-fit distribution is selected
with the Akaike Information Criterion (AIC) (Akaike, 1974; Aho
et al., 2014).
Extreme distributions. Extreme distributions aim to characterize
the largest values of the population of a variable. They are impor-
tant in infrastructure design and risk assessment because they
generate the most demanding loads or impacts on the systems.
Extreme distributions are assumed to follow a generalized extreme
value (GEV) distribution. The GEV encompasses, in a single func-
tional form, the Weibull, the Gumbel and the Fre
chet distributions
of extremes. GEV fitting is done for annual maxima by means of the
maximum likelihood method.
3.3. Interpolating rainfall
Spatio-temporal interpolation is an important part of charac-
terizing rainfall patterns over a territory. Observations in time se-
ries may be missing due to multiple causes: equipment
malfunction, transcription errors, information transfer errors,
 J. Diez-Sierra, M. del Jesus / Environmental Modelling & Software 97 (2017) 243e258
Fig. 3. Storm types of rainfall probability. Image created with MENSEI-L.
discontinued observation locations, etc. Data may not even exist at
a location where they are needed because no pluviometer was ever
deployed there. Spatio-temporal interpolation serves to complete
missing value in time series and to create virtual pluviometers, to
improve data coverage over the territory, allowing a better statis-
tical characterization of the spatial distribution of rainfall.
There are many statistical techniques in the literature that can
be used to interpolate precipitation such as: Kriging, Inverse Dis-
tance Weighting (IDW), Spline functions and Multiple regression
with residual correction, among others (Ninyerola et al., 2007; New
et al., 2000; Wagner et al., 2012). Interpolation techniques are
methods that are commonly used to predict precipitation at a given
point using the available adjacent information. One of the major
drawbacks of interpolation techniques is that the results are
smoothed and that they are not able to capture the spatial rainfall
heterogeneity associated to local processes. To deal with those is-
sues, some variants of kriging, such as Cokriging or Universal
kriging, multiple regression with residual correlations and other
techniques like PRISM (Parameter-elevation Relationships on In-
dependent Slopes Models) (Isotta et al., 2014; Daly et al., 2008) use
external information (elevation, coastal proximity, topographic
facet orientation, hydrometeorological variables, etc.) to improve
the results.
Cokriging is used to take advantage of the covariance between
two or more regionalized variables that are related to rainfall;
Universal kriging is a variant of the Ordinary Kriging with a local
trend; Multiple regression with residuals is a combination of sta-
tistical (multiple regression) and spatial interpolation (splines and
inverse distance weighting) tools; and PRISM calculates a regres-
sion between the climate variable and the elevation for each point,
assigning a weight to every station that is based primarily on the
physiographic similarity of the stations.
247
In this context, geostatistic kriging techniques were selected for
the present study because (i) several authors (Tabios and Salas,
1985; Phillips et al., 1992) have shown that kriging provides bet-
ter estimates of rainfall than conventional methods when there is a
high density of stations (as it is the case of Tenerife) and (ii) there
are software libraries specifically designed to be implemented in a
simple way.
Geostatistics, discipline that is based on the theory of region-
alized variables (Journel and Huijbregts, 1978; Goovaerts, 1997,
2000), is increasingly preferred to other statistical techniques
because it allows to account for the spatial correlation among
neighboring gauges to make predictions at ungauged locations.
Rainfall interpolation follows the two-step process presented in
Herrera et al. (2012): first, the interpolation predicts the rainfall
state of the location, wet or dry; then, in the case that a wet state
has been predicted, the total amount of rainfall is interpolated.
Different flavors of kriging were tested to be used as the spatio-
temporal interpolation method and the best one was selected
through cross-validation. Ordinary kriging, universal kriging using
elevations as the drift term, universal kriging using precipitation
averages as drift and universal kriging using both elevation and
precipitation as drifts, were tested. Cross-validation for each of the
methods used the leave-one-out approach: all the gauges but one
were used to reconstruct the information on the missing one. This
process was repeated for all the gauges and an average recon-
struction error was computed. Universal kriging using elevation
and average rainfall information was selected as it minimized the
average reconstruction error.
The error scores were computed for every station by comparing
the observed values with the corresponding gridded one in the
period 1950e2016. In particular we used bias, root-mean-square
error (RMSE) and Pearson correlation coefficient (R) as validation
248 J. Diez-Sierra, M. del Jesus / Environmental Modelling & Software 97 (2017) 243e258
measures; both RMSE and R scores were computed on the subset of
wet days (Precip > 0.1 mm) in order to reduce the effects of the
varying occurrence frequencies. It provided and average Pearson
correlation coefficient R of 0.80 for all the gauges, an average
standardized bias of 0.02 mm and a RMSE of 6.7 mm. Fig. 4 shows
the spatial distribution of the above scores for the kriging method
over Tenerife island. In general, a very good agreement between the
interpolated and recorded series, including the correlations, is
observed. Fig. 4 also shows that bigger errors are registered in re-
gions with complex topography or in areas where the density of
stations is lower.
(a) Bias/Std
(b) Root Mean Square Error (pp > 0.1 mm)
(c) Correlation (pp > 0.1 mm)
Kriging requires a specific parametric form of the variogram for
the interpolation process. The gstat (Pebesma, 2004) and automap
(Hiemstra, 2013) libraries used to implement the interpolation
procedure, select the optimal variogram for every interpolation
among 20 different ones. Most of the time the circular and expo-
nential variograms are selected as optimal.
The selected interpolation method serves both, as a temporal
and a spatial interpolation method. Indeed, temporal interpolation
is achieved through spatial interpolation. When a guage is missing
a value for any given time, all the other gauges with information for
Fig. 4. Cross validation results for the period 1950 to 2016. Scores are the standardized bias of the full series and the RMSE and Pearson correlation coefficient R of the wet days
(precip > 0.1 mm).
that particular time are used to interpolate, via kriging, the value at
the location of the gauge missing the observation. This kriging
procedure makes use of elevation and average rainfall information,
and therefore includes local information, about the specific rainfall
event, and also global information, the relative amount of precipi-
tation that the interpolation location receives. The creation of a
virtual pluviometer is equivalent to filling the missing values of a
gauge that contains no observations.
3.4. Forecasting rainfall
In recent years, time-series analysis, such as Generalized linear
models (GLM), Stochastic auto-regressive moving-average models,
Logistic regression or Artificial intelligence methods (LogitBoosting,
Random Forest, Support Vector Machine, Artificial neural networks,
K-nearest-neighbors or Clustering regression methods), have
become some of the most popular techniques for short-term
rainfall forecasting. Although these techniques present a limited
prediction skill owing to the low persistence in time that usually
characterizes rainfall, the moderate data availability and reduced
computer time required to use them, make their application very
attractive in the context of real-time flood forecasting (Toth et al.,
2000; Monira et al., 2010). In this context, the analogs method
based on weather types was selected for the forecasts due to the
reduced computational cost of application and the interpretability
of the results, which simplifies its application by non experts.
 J. Diez-Sierra, M. del Jesus / Environmental Modelling & Software 97 (2017) 243e258 249

The short-term rainfall forecasting system implemented in

MENSEI-L feeds on the atmospheric forecasts provided by NOAA's
GFS model. SLP, GH-500 and GH-850 forecasts are automatically
collected from NOAA for the three following days. The represen-
tative weather type corresponding to the forecast atmospheric
state is determined with the weather type classification developed
above. Forecasts for the rainfall probability and the average rainfall
depth for every period are computed by selecting the storm type
corresponding to the forecasted weather type.
The quality of the forecast is evaluated by a k-fold cross-
validation approach (Markatou et al., 2005), that is based on
dividing the observation period into k subperiods, and recon-
structing every subperiod using the information of the other ones
and evaluating the quality of the reconstruction. The k-fold cross-
validation approach avoids spurious effects of any particular
partition. The reconstruction performance is evaluated comparing
observed and CFSR-generated forecast maps of average rainfall
probability and average rainfall depth (see Fig. 2). It is important to Fig. 6. F-score index for the 17 gauges selected.
note that, as probabilities are involved the reconstructed time se-
ries are not unique. Therefore, reconstruction is repeated 100 times,
and the reconstruction performance is taken to be the average of The figure increases up to 50% for three-day accumulated and up to
the 100 realizations. As an example, Fig. 5 shows the reconstructed 70% when the observed rainfall intensity is greater than 30 mm.
series (red line), using the WT technique, and the observed series Fig. 6 shows the f-score index (Powers, 2011) for the 17 selected
for the year 2010 (blue line) of the “Santiago del Teide” station, gauges. The f-score index is very useful to compare the skill of
located in the northwest part of the island. The reconstructed series model predictions when working with imbalanced datasets
corresponds to one of the 100 simulations. The top panel shows the (datasets in which one class is much more frequent than the
daily series while the bottom panel corresponds to the three-day others), as in the case of Tenerife where rainfall occurrence is rare
moving average. It is possible to observe that the correlation be- (below 15%) and can be placed up to 25% in some gauges located in
tween the reconstructed series and the observed one is greater the north of the island and down to 5% in some gauges located in
once the moving mean is applied, which is expected since the the south. Fig. 6 shows that the skill of the forecast is substantially
reconstruction was performed with the mean values of each better in the north of the Island than in the south. This is probably
weather type. because weather types capture the frontal precipitation, which
In order to evaluate the skill of the forecasts, the 17 most usually occurs in the north of the island, better than the convective
complete gauges covering the period 1979e2014 along Tenerife processes prevailing in the south. A model that hits on average 30%
have been selected (see Fig. 6). Using the average conditions of the of the rainy days and with an average f-score of 0.28 may not seem
storm type to reconstruct rainfall makes bias (of number or rainy very skillful. If rainfall was forecast by a random model that only
days and average annual precipitation) almost zero, and minimizes preserved the probability of rainfall at each station, it would only
RMSE (Precip > 0.1 mm) to a value of 15.6 mm for all gauges. The correctly predict 13% of the days, with an f-score close to 0.13. That
daily rainfall forecasting procedure shows an average accuracy over is, it would perform only half as good as our model. If it was forecast
all pluviometers of 30%, combining rainfall occurrence and depth. by a Markov Chain preserving the transition probabilities, it would

Fig. 5. Reconstructed serie (red line), using WT technique, and the observed one (blue line) of the station know as “Santiago del Teide” located in the northwest part of the island for
the year 2010. The reconstructed series corresponds to one of the 100 simulations. The top panel shows the daily serie while the bottom panel corresponds to the three-day moving
average. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
250 J. Diez-Sierra, M. del Jesus / Environmental Modelling & Software 97 (2017) 243e258

correctly predict about 20% of the day, leaving an f-score of 0.23,
which is still one third worse than our model, and in this case the
next day forecast is made based on the observed rainfall the pre-
vious day, performance would decrease if computed for the three
day forecast.
On the positive side, most flooding events coincided with a
reduced set of weather types. This is the case of the flash flood
events registered on the dates of 4-12-1991, 13-12-1995 and 31-03-
2002 which fell in the weather types 37, 27 and 12 (see Fig. 7) and
with 300, 220 and 230 mm/day of rainfall, respectively. All of them
corresponding to atmospheric systems known as cut-off low
pressure system. This figure shows the 35 rainfall events with an
intensity higher than 200 mm/day in the period between 1979 and
2014. Blue and orange bars correspond to the number of events
with an intensity greater than 200 mm/day and to the associated
probability of occurrence of each weather type respectively. These
35 events fell in just 15 weather types whose probability in all the
cases, except two, is below the average 100=81 ¼ 1:23 (black hor-
izontal line). This means that special attention should be paid to the
weather types with a high number of extreme events and a low
probability of occurrence (e.g. 18 and 27), since they will have great
probability of generating torrential rains.
Therefore, the forecasting system, although not accurately pre-
dicting rainfall occurrence, is still useful for early warning of
flooding events, which in Tenerife present flash flood
characteristics.
4. Rainfall analysis work flow
The present section describes how the different tools presented
in the previous section are combined to generate a work flow for
rainfall analysis. Fig. 8 depicts a schematic representation of the
methodology. The description assumes the existence of a database
of precipitation observations. The main steps of the work flow are:
(1) quality control of the rainfall series, (2) selection of the refer-
ence series, (3) creation of a spatial map of average precipitation,
(4) reconstruction of time series, and (5) statistical analysis of
rainfall.
Controlling the quality of time series. The spatio-temporal rainfall
analysis begins with a quality control of all the rain gauges in the
database. Length of record, number of missing values and number
of outliers (defined as values more than five standard deviations
above the mean without concomitant occurrences on several
gauges) were considered to evaluate the quality of each series. Only

Fig. 7. 35 rainfall events with an intensity higher than 200 mm/day in the period 1979e2014. Blue and orange bars correspond to the number of events with an intensity greater
than 200 mm/day and to the associated probability of occurrence of each weather type respectively. Black horizontal line correspond to the mean probability of occurrence for each
WT. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
 J. Diez-Sierra, M. del Jesus / Environmental Modelling & Software 97 (2017) 243e258
Fig. 8. Workflow of the methodology.
gauges with more than 10 years of information and less that 10
missing values per year were considered to compute the spatial
map of average precipitations.
Creating the spatial map of average rainfall. The spatial map of
average precipitation is constructed using universal kriging with
elevation because this method minimized the average reconstruc-
tion error. Only the reference gauges, those with the best quality
data, are used for this reconstruction.
Reconstructing time series. After having constructed the spatial
map of rainfall averages, the database is inspected to detect missing
values. Observation time series may begin in different dates. Newer
time series would present missing values from the starting date of
the oldest time series until their starting date. The kriging inter-
polation scheme requires a minimum number of gauges for accu-
rate interpolation. Therefore, the database reconstruction
procedure starts from the earliest date for which a minimum
number of observations, homogeneously distributed in space, are
consistently ensured. In the case of Tenerife island, the recon-
struction starts at 1950, the moment from which at least 25 ob-
servations are always present.
Database reconstruction is done through spatial interpolation as
described in section 3.3. For a given time, all the gauges with
measurements are used to reconstruct the spatial field of rainfall for
that particular day. The value from this rainfall field is introduced in
every time series with a missing value for the day. Following this
procedure, synthetic observations are provided at every location
251
where missing values appear and new virtual gauge can be created
wherever required. This process ensures complete rainfall time
series from the initial date of reconstruction of the database.
It is important to note here that information from all gauges is
used to create the spatial maps of rainfall; even those not used to
create the map of spatial rainfall averages. Therefore, the daily
kriging reconstruction of rainfall incorporates all the information
available.
Analyzing rainfall statistics. With a filled rainfall database, point
rainfall characterization is a straightforward procedure using the
tools described in section 3.2. Spatial statistics are created by
interpolation of the point analysis carried out for all the gauges in
the domain.
5. Software tool implementation
The rainfall analysis described in the previous sections has been
implemented into a software tool called MENSEI-L (see Fig. 9) that
simplifies and automatizes the application of the methodology. It is
implemented in python (Python Software Foundation, 2016) and R
(R Core Team, 2015), making use of the wxWidgets library (Smart
et al., 2005). The software is distributed in a way that Python
modules are frozen and embedded, so no python installation is
needed. The R packages required for MENSEI-L are installed directly
by the application once it is initialized. The software was developed
to be as easy to handle as possible.
252 J. Diez-Sierra, M. del Jesus / Environmental Modelling & Software 97 (2017) 243e258

Fig. 9. MENSEI-L main window.

Once the installation is complete, the main window of the
interface is automatically showed by the application (see Fig. 9). On
the left-hand panel of the main window the four main modules
explained in the methodology are found: weather types, point and
space precipitation distributions, interpolation and forecasting.
Each of the modules is divided into a series of submodules through
which all the analysis options are accessed. Fig. 11 shows a sche-
matic with the main options available in MENSEI-L; these options
will be described in-depth in the section number 6 Results.
MENSEI-L needs two input CSV files to work, stations.csv and

Fig. 10. Punctual rainfall regime submodule.

 J. Diez-Sierra, M. del Jesus / Environmental Modelling & Software 97 (2017) 243e258
Fig. 11. MENSEI-L structure. Modules and options available from MENSEI-L.
series.csv. The first one contains the identifier of each station (ID), its
name, its location (longitude and latitude) and its elevation; while
the second corresponds to a CSV file in which each column contain
the time series of each station. CSV files can be introduced by hand,
however due to the technical specifications of the project, MENSEI-
L is designed to read the information directly from a SQL database.
Once the software is connected to the database, the two CSV files
are created locally by MENSEI-L. This way, it is possible to work
disconnected from the database, reducing the bandwidth required
by the application.
The first time MENSEI-L is used, it only allows the user to choose
the Complete interpolation option (see Fig. 11). This option accesses
the database, resamples the rainfall series to a daily temporal res-
olution and performs the interpolation day by day; filling the gaps
and generating new series in those new stations created by the user
in the stations.csv file. The rest of the modules use automatically the
interpolated information for their analysis. The Complete Interpo-
lation option may take a couple of hours so once it has been
completed the user will be able to use the Incremental interpolation
(see Fig. 11) option, which directly identifies the new pluviometric
records added to the database, performs the interpolation for those
new records, generates the CSV files and saves the results in the
database. In addition to the CSV files, the application contains the
results of the weather type classification, the DEM of the study area,
and a connection file for the database.
One of strengths of MENSEI-L, is that the software interface
adapts to the required analysis. For example, in the Punctual rainfall
regimen submodule (Punctual in Fig. 11), a canvas is shown where
gauges can be selected graphically for analysis (see Fig. 10). The left-
hand panel allows to access different analysis, such as time series
plotting, average and extreme distribution fitting, and to select the
gauge by name. All analysis can be saved to disk for further pro-
cessing outside MENSEI-L. The upper panel offers different options
for visualization, for instance showing shape files to improve visual
aspects, and for image storage in different formats, PNG, CSV and
GIS raster.
6. Results
This section presents the main results and options that can be
253
used in MENSEI-L. The results section is organized in such a way
that every figure summarizes the functionality of each of the main
modules: weather types, rainfall distributions module (punctual
and spatial), interpolation module and forecasting module.
Weather types module. This module (see Fig. 11) makes use of the
weather types classification to generate all the analysis. The clas-
sification results are stored in the application reducing in this way,
both the space installation requirements and the computational
cost that would take perform this process every time. However, the
81 storm types are calculated every time since the user might add
new additional information or generate virtual stations.
The weather types module generates a visualization of the
weather type classification for the atmospheric variables SLP,
GH500 and GH850 (see Fig. 2). Furthermore, it generates a visual-
ization of the storm types of average rainfall depth and average
rainfall probability associated with each weather type (see Fig. 3).
In addition, the user has access to the continuous daily time series
of weather types which is very useful, for instance, when
comparing each pattern with the historical records of flooding,
droughts or infrastructures damages. The spatial fields of the 81
weather types can be exported to PNG and GIS raster formats with
the same spatial resolution of the CFSR database. Moreover the
storm types of average rainfall depth and average rainfall proba-
bility can be exported to CSV format.
(a) Time series visualization for the gauge “Orotava Palmeras” in
the north of the island. The blue line corresponds to the
reconstructed time series while the dashed red line corre-
sponds to the observed one.
(b) Medium value analysis of the “Orotava Palmeras” rain gauge.
The left-hand panel shows the empirical cumulative distri-
bution function (CDF), as well as the fitted parametric one
(Gumbel). On the right-hand panel, a Q-Q plot shows the
empirical data (red dots), as well as the parametric fit (black
line).
(c) Extreme value analysis of the “Orotava Palmeras” rain gauge.
The left-hand panel shows the empirical cumulative distri-
bution function (CDF), as well as the fitted parametric one
(GEV). On the right-hand panel, a return-level plot shows the
empirical data (red dots), as well as the parametric fit
254 J. Diez-Sierra, M. del Jesus / Environmental Modelling & Software 97 (2017) 243e258
(black line). It also shows the values of the fitted distribution
(GEV).
Rainfall regime module. This module is divided in two sub-
modules: punctual and spatial (see Fig. 11). The first one grants
access to the interface showed in Fig. 10, where any pluviometer
can be selected to plot the complete daily time series, compute the
average distribution or carry an extreme value analysis, all of them
for the complete daily series. Fig. 12 shows the resulting plot of
these principal functionalities of the punctual rainfall submodule.
Panel (a), located at the top of the figure, shows the time series of
the “Orotava Palmeras” gauge, located in the north of the island. It
can be seen that the original data of the gauge (dashed red line) just
counts with 5 years of records. The missing value estimation
implemented in the interpolation module allows to generate a
continuous time series with more than 60 years of data (blue line),
significantly improving the statistical characterization of the series.
The average distribution characterization is shown in panel (b). The
panel shows the average distribution for the “Orotava Palmeras”
rain gauge. On the left-hand side of the panel, the empirical cu-
mulative distribution function (CDF), as well as the fitted para-
metric one (Gumbel) are shown. On the right-hand panel, a Q-Q
plot shows the empirical data (red dots) and the parametric fit
(black line). Some representative percentiles are shown as blue dots
in the image, as well as the fitted parameters for the distribution.
Panel (c), at the bottom of the figure, shows an extreme value
analysis for the “Orotava Palmeras” rain gauge. On the left-hand
side of the panel, the empirical cumulative distribution function
(CDF), as well as the fitted parametric one (GEV) are shown. On the
right-hand side of the panel, a return-level plot shows the empirical
data (red dots), as well as the parametric fit (black line). It also
shows the values of the fitted distribution (GEV). The results can be
exported to PNG and CSV formats. On one side, the selected time
series, the fixed distributional parameters of the average and
extreme analysis and the empirical and parametric quantiles and
return periods can be exported to CSV; on the other side the three
panels showed in Fig. 12 for the selected serie can be saved to PNG
format.
The spatial rainfall submodule is designed to generate spatial
precipitation fields using universal kriging with elevation and
mean precipitation as drifts. MENSEI-L contains a digital elevation
model, in this case of Tenerife island, which is directly stored in the
application. This submodule involves two different analysis:
computing the spatial field of average, cumulative or maximum
rainfall for a specific period of time since 1950, and computing the
spatial rainfall field for a return period between 2 and 500 years. In
this last option, the application uses the entire time series to
perform the extreme value analysis. These maps make use of the
information of all gauges, and not just the reference gauges used to
generate the drift term for the kriging interpolation. Panel (a) of
Fig. 13 shows the cumulative rainfall field for the period 1986-01-01
to 1987-01-01. The map correctly captures the strong gradient of
precipitation that exists between the north and the south of the
island. It also correctly captures the correlation that exists between
precipitation and elevation; larger elevations receive larger
amounts of precipitation. Panel (b) shows the spatial rainfall map
for the 100 years return period. The spatial rainfall map for a spe-
cific return period is generated by means of calculating the return
period for every station individually and using universal kriging
with elevation and mean precipitation as drifts. Results of this
submodule can be exported to PNG and GIS raster formats with the
same spatial resolution of the DEM (5 m in the case of Tenerife).
(a) Accumulative rainfall map for the period 1986-01-01 to
1987-01-01.
(b) Return period rainfall map (T ¼ 100 years).
Interpolation module. As mention before, the Interpolation
module has two options: Complete interpolation and Incremental
interpolation (see Fig. 11). Complete interpolation has no specific
display since it just interpolates missing data from time series from
1950 to the current date. The first time the application is opened,
the user can only select this option, since the rest of modules feed
on the results generated by the complete interpolation process. To
generate a synthetic gauge in a location where there is no real one,
its location and name information must be included in the database
(stations.csv file); then the software automatically takes it into ac-
count in the interpolation process. In addition, in the case of new
records being incorporated to the database, the Incremental
interpolation option can be carried out from the previous ending
date in order save computational time. As discussed above the
Complete Interpolation submodule could take a couple of hours, but
it mainly depends on the number of stations and the interpolation
period.
Forecasting module. In the Forecasting module (see Fig. 11) the
user can visualize the spatial fields of SLP, GH500 and GH850,
automatically downloaded from the GFS model for the next three
days. From this information the most similar weather type to the
GFS data is found and their respective storm types of average
rainfall probability and average rainfall depth are shown. Fig. 14
shows three-day GH-500 forecasts from the GFS model for the
dates 07-06-2017, 08-06-2017 and 09-06-2017 and their respective
storm types of average rainfall probability and average rainfall
depth. When the Storm types submodule is selected two options are
available: (i) showing the rainfall probability and depth in the
gauges like in Fig. 3, or (ii) constructing the spatial maps by using
universal kriging with elevation and average rainfall information as
drifts. Results of this module can be exported as follows: export the
images in PNG format, export the spatial precipitation forecast
fields in raster format and export the rainfall forecast for all the
stations in CSV format.
(a) GH-500 (m)
(b) Average rainfall probability (%)
(c) Average rainfall depth (mm/day)
7. Future developments, limitations and conclusions
Universal kriging using elevation and average rainfall as drifts
has proven to be a powerful technique to fill and reconstruct
rainfall in areas with a high density of gauges. Pluviometer filling
with information from neighboring gauges allows to extend the
length of time series, improving the stability of extreme value
estimation. Regular and extreme analysis, both punctual and
spatial, are implemented in MENSEI-L in order to be applied to
hydrological projects.
One of the main limitations is its applicability to areas where
there is a low number of gauges, or where they are far from each
other, since the quality of the reconstruction is strongly affected by
this lack of original information (Beguería et al., 2016). In such
areas, interpolation techniques that integrate relationships be-
tween the climatic variables and the topography must be applied in
order to capture the underlying heterogeneity in rainfall (Isotta
et al., 2014; Daly et al., 2008).
Rainfall forecasting requires more sophisticated models than
average predictions based on weather types. A 50% forecasting
accuracy on three-day periods is acceptable as a pre-warning sys-
tem, but is still not good enough for measure implementation and
mitigation means deployment. Different regression techniques
 J. Diez-Sierra, M. del Jesus / Environmental Modelling & Software 97 (2017) 243e258 255

Fig. 12. Functionality of the Punctual distribution submodule. Images created with MENSEI-L.
256 J. Diez-Sierra, M. del Jesus / Environmental Modelling & Software 97 (2017) 243e258

Fig. 13. Functionality of the Spatial distribution submodule. Images created with MENSEI-L.

combined with weather type decomposition are already under potential as a flood warning system, since it identifies those
analysis to improve the forecasting procedure. It has been shown, weather types that have led to extreme rainfall and flooding events
however, that the rainfall forecasting system shows a great on Tenerife island in the past.

Fig. 14. GH-500 (m) three days forecast from GFS for the dates 2017-06-07, 2017-06-08 and 2017-06-09 ðaÞ average rainfall WT probability ðbÞ and average rainfall WT depth
selected for these days ðcÞ. The spatial maps are constructed using universal kriging with elevation and average rainfall information as drifts. The number placed in the lower left
corner corresponds to the storm type selected for each day. Images created with MENSEI-L.
 J. Diez-Sierra, M. del Jesus / Environmental Modelling & Software 97 (2017) 243e258
Weather types constitute a good representation of optimal basis
for hydroclimatic variables. Although storm types derived from
weather types show distinct rainfall patterns, further investigation
is needed to establish a strong relation between synoptic patterns
of atmospheric conditions and rainfall patterns. In this study, a
blind classification (without information about rainfall patterns)
was carried out for the atmospheric conditions. However, for
optimal results, the classification should consider the final objec-
tive: to group atmospheric situations into homogeneous rainfall
patterns.
MENSEI-L constitutes a powerful tool for the analysis of a plu-
viometric network. It simplifies the application of advanced
mathematical techniques (kriging interpolation) and automatizes
the process, allowing non-specialized technical staff to have access
to powerful rainfall analysis techniques. This advantage can be
transferred into risk assessments and infrastructure design pro-
cesses, improving the overall value obtained by society.
Software and data availability
Software name: MENSEI-L.
Developers: Environmental Hydraulics Institute, Universidad de
Cantabria.
Contact information: manuel.deljesus@unican.es.
Hardware required: 1 GB HD and 8 GB RAM.
Software required: 64-bit Windows.
Program languages: python 2.7 and R.
Availability: MENSEI-L rainfall analysis tools and graphical user
interface (GUI) could be made available for research purposes upon
request.
Acknowledgments
The authors would like to thank Omar Castellanos for his work
in software implementation and Pedro Delgado Melia
n for his
comments to improve MENSEI-L. They also would like to thank
“Consejo Insular de Aguas de Tenerife (CIATF)” for granting
permission to use their rainfall database for this manuscript and for
funding the implementation of the software in Tenerife.
References
Abdi, H., Williams, L.J., 2010. Principal component analysis. Wiley Interdiscip. Rev.
Comput. Stat. 2, 433e459. URL. http://dx.doi.org/10.1002/wics.101.
Aguilar, E., Prohom, M., 2011. Extraqc. Centre for Climate Change and Servei
Meteorolgic de Catalunya, Spain. URL. www.c3.urv.cat/data.html.
Aho, K., Derryberry, D., Peterson, T., 2014. Model selection for ecologists: the
worldviews of aic and bic. Ecology 95, 631e636. URL. http://dx.doi.org/10.1890/
13-1452.1.
Akaike, H., 1974. A new look at the statistical model identification. IEEE Trans.
automatic control 19, 716e723.
Arnbjerg-Nielsen, K., Willems, P., Olsson, J., Beecham, S., Pathirana, A.,
Gregersen, I.B., Madsen, H., Nguyen, V.T.V., 2013. Impacts of climate change on
rainfall extremes and urban drainage systems: a review. Water Sci. Technol. 68,
16e28.
Beguería, S., Vicente-Serrano, S., Tom
as-Burguera, M., Maneta, M., 2016. Bias in the
variance of gridded data sets leads to misleading conclusions about changes in
climate variability. Int. J. Climatol. 36, 3413e3422. http://dx.doi.org/10.1002/
joc.4561.
Ben Alaya, M., Ouarda, T., Chebana, F., 2017. Non-gaussian spatiotemporal simula-
tion of multisite daily precipitation: downscaling framework. Clim. Dyn. 1e15.
http://dx.doi.org/10.1007/s00382-017-3578-0 (in press).
€ schl, G., Reszler, C., Komma, J., 2008. A spatially distributed flash flood fore-
Blo
casting model. Environ. Model. Softw. 23, 464e478. http://dx.doi.org/10.1016/
j.envsoft.2007.06.010.
Borga, M., Anagnostou, E., Blo € schl, G., Creutin, J.D., 2011. Flash flood forecasting,
warning and risk management: the hydrate project. Environ. Sci. Policy 14,
834e844. http://dx.doi.org/10.1016/j.envsci.2011.05.017.
Burton, A., Kilsby, C., Fowler, H., Cowpertwait, P., O'Connell, P., 2008. Rainsim: a
spatial-temporal stochastic rainfall modelling system. Environ. Model. Softw. 23,
1356e1369. http://dx.doi.org/10.1016/j.envsoft.2008.04.003.
Burton, A., Glenis, V., Jones, M., Kilsby, C., 2013. Models of daily rainfall cross-
257
correlation for the United Kingdom. Environ. Model. Softw. 49, 22e33. http://
dx.doi.org/10.1016/j.envsoft.2013.06.001.
~ o, A.S., 2011. Analysis of clustering and
Camus, P., Mendez, F.J., Medina, R., Cofin
selection algorithms for the study of multivariate wave climate. Coast. Eng. 58,
453e462.
Daly, C., Halbleib, M., Smith, J., Gibson, W., Doggett, M., Taylor, G., Curtis, J.,
Pasteris, P., 2008. Physiographically sensitive mapping of climatological tem-
perature and precipitation across the conterminous United States. Int. J. Cli-
matol. 28, 2031e2064. http://dx.doi.org/10.1002/joc.1688.
Goovaerts, P., 1997. Geostatistics for Natural Resources Evaluation. Oxford Univer-
sity Press on Demand.
Goovaerts, P., 2000. Geostatistical approaches for incorporating elevation into the
spatial interpolation of rainfall. J. hydrology 228, 113e129.
Guijarro, J.A., 2016. Climatol. URL. http://www.climatol.eu/.
Herrera, R.G., Puyol, D.G., Martín, E.H., Presa, L.G., Rodríguez, P.R., 2001. Influence of
the north atlantic oscillation on the canary islands precipitation. J. Clim. 14,
3889e3903.
Herrera, S., Gutie
rrez, J.M., Ancell, R., Pons, M., Frías, M., Fern
andez, J., 2012.
Development and analysis of a 50-year high-resolution daily gridded precipi-
tation dataset over Spain (Spain02). Int. J. Climatol. 32, 74e85.
Hiemstra, P., 2013. Automap. URL. https://cran.r-project.org/web/packages/
automap.
Huth, R., Beck, C., Philipp, A., Demuzere, M., Ustrnul, Z., Cahynova
, M., Kyselỳ, J.,
Tveito, O.E., 2008. Classifications of atmospheric circulation patterns. Ann. N. Y.
Acad. Sci. 1146, 105e152.
Isotta, F., Frei, C., Weilguni, V., Percec Tadi
gues, P., Rudolf, B., Pavan, V.,
c, M., Lasse
Cacciamani, C., Antolini, G., Ratto, S., Munari, M., Micheletti, S., Bonati, V.,
Lussana, C., Ronchi, C., Panettieri, E., Marigo, G., Verta cnik, G., 2014. The climate
of daily precipitation in the alps: development and analysis of a high-resolution
grid dataset from pan-alpine rain-gauge data. Int. J. Climatol. 34, 1657e1675.
http://dx.doi.org/10.1002/joc. 3794.
Journel, A.G., Huijbregts, C.J., 1978. Mining Geostatistics. Academic press.
Kretzschmar, A., Tych, W., Chappell, N., 2014. Reversing hydrology: estimation of
sub-hourly rainfall time-series from streamflow. Environ. Model. Softw. 60,
290e301. http://dx.doi.org/10.1016/j.envsoft.2014.06.017.
Krzysztofowicz, R., 1995. Recent advances associated with flood forecast and
warning systems. Rev. Geophys. 33, 1139e1147. http://dx.doi.org/10.1029/
95RG00873.
Liu, J., Bray, M., Han, D., 2013. A study on wrf radar data assimilation for hydrological
rainfall prediction. Hydrology Earth Syst. Sci. 17, 3095e3110. http://dx.doi.org/
10.5194/hess-17-3095-2013.
Markatou, M., Tian, H., Biswas, S., Hripcsak, G., 2005. Analysis of variance of cross-
validation estimators of the generalization error. J. Mach. Learn. Res. 6.
Mehrotra, R., Li, J., Westra, S., Sharma, A., 2015a. A programming tool to generate
multi-site daily rainfall using a two-stage semi parametric model. Environ.
Model. Softw. 63, 230e239. http://dx.doi.org/10.1016/j.envsoft.2014.10.016.
Mehrotra, R., Li, J., Westra, S., Sharma, A., 2015b. A programming tool to generate
multi-site daily rainfall using a two-stage semi parametric model. Environ.
Model. Softw. 63, 230e239. http://dx.doi.org/10.1016/j.envsoft.2014.10.016.
Melia
n, P.D., Ruiz, J.J.B., Díez, A.M., Salete, E.G., 2011. “Mo
dulo de gestio
n de tor-
mentas” en la modelizacio
n hidrolo
gica de superficie de tenerife. In: II Jornadas
de Ingeniería del Agua. Fundacio
n para el Fomento de la Ingeniería del Agua,
Barcelona, Spain.
Mestre, O., Gruber, C., Prieur, C., Caussinus, H., Jourdain, S., 2011. Splidhom: a
method for homogenization of daily temperature observations. J. Appl. Mete-
orology Climatol. 50, 2343e2358. http://dx.doi.org/10.1175/2011JAMC2641.1.
Monira, S., Faisal, Z., Hirose, H., 2010. Comparison of Artificially Intelligent Methods
in Short Term Rainfall Forecast, pp. 39e44. http://dx.doi.org/10.1109/
ICCITECHN.2010.5723826.
Moser, B., Gallus, W.A.J., Mantilla, R., 2015. An initial assessment of radar data
assimilation on warm season rainfall forecasts for use in hydrologic models.
Weather Forecast. 30, 1491e1520. http://dx.doi.org/10.1175/WAF-D-14-00125.1.
Neiman, P., White, A., Ralph, F., Gottas, D., Gutman, S., 2009. A water vapour flux tool
for precipitation forecasting. Proceedings of the Institution of Civil Engineers.
Water Manag. 162, 83e94. http://dx.doi.org/10.1680/wama.2009.162.2.83.
New, M., Hulme, M., Jones, P., 2000. Representing twentieth-century space-time
climate variability. part ii: development of 1901-96 monthly grids of terrestrial
surface climate. J. Clim. 13, 2217e2238.
Ninyerola, M., Pons, X., Roure, J., 2007. Monthly precipitation mapping of the iberian
peninsula using spatial interpolation tools implemented in a geographic in-
formation system. Theor. Appl. Climatol. 89, 195e209. http://dx.doi.org/
10.1007/s00704-006-0264-2.
Paixao, E., Auld, H., Mirza, M.M.Q., Klaassen, J., Shephard, M.W., 2011. Regionaliza-
tion of heavy rainfall to improve climatic design values for infrastructure: case
study in southern ontario, Canada. Hydrological Sci. J. 56, 1067e1089.
Pebesma, E.J., 2004. Multivariable geostatistics in s: the gstat package. Comput.
Geosciences 30, 683e691.
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O.,
Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A.,
Cournapeau, D., Brucher, M., Perrot, M., Duchesnay, E., 2011. Scikit-learn: ma-
chine learning in Python. J. Mach. Learn. Res. 12, 2825e2830.
Phillips, D.L., Dolph, J., Marks, D., 1992. A comparison of geostatistical procedures for
spatial analysis of precipitation in mountainous terrain. Agric. For. Meteorology
58, 119e141.
Powers, David M.W., 2011. Evaluation: from precision, recall and f-measure to roc,
258 J. Diez-Sierra, M. del Jesus / Environmental Modelling & Software 97 (2017) 243e258
informedness, markedness and correlation. J. Mach. Learn. Technol. 2, 37e63.
Python Software Foundation, 2016. Python language Reference. URL:, version 2.7.
http://www.python.org.
R Core Team, 2015. R: a language and environment for statistical computing. URL.
https://www.R-project.org.
Renschler, C., Mannaerts, C., Diekkrüger, B., 1999. Evaluating spatial and temporal
variability in soil erosion riskrainfall erosivity and soil loss ratios in andalusia,
Spain. Catena 34, 209e225.
Sabourin, J.F., Associates Inc, 2010. Storms 2010. URL. http://www.jfsa.com/rainfall-
analysis-storms.php.
Saha, S., Moorthi, S., Pan, H.L., Wu, X., Wang, J., Nadiga, S., Tripp, P., Kistler, R.,
Woollen, J., Behringer, D., et al., 2010. The ncep climate forecast system rean-
alysis. Bull. Am. Meteorological Soc. 91, 1015e1057.
Serinaldi, F., Kilsby, C., 2014. Simulating daily rainfall fields over large areas for
collective risk estimation. J. Hydrology 512, 285e302. http://dx.doi.org/10.1016/
j.jhydrol.2014.02.043.
Serrano-Notivoli, R., de Luis, M., Beguería, S., 2017. An r package for daily precipi-
tation climate series reconstruction. Environ. Model. Softw. 89, 190e195. http://
dx.doi.org/10.1016/j.envsoft.2016.11.005.
Smart, J., Csomor, S., et al., 2005. Cross-platform GUI programming with
wxWidgets. Prentice Hall Professional.
Stepanek, P., 2005. Anclim-software for Time Series Analysis. Dept. of Geography,
Fac. of Sciences, MU, Brno.
Stephenson, D., Rupa Kumar, K., Doblas-Reyes, F., Royer, J.F., Chauvin, F., Pezzulli, S.,
1999. Extreme daily rainfall events and their impact on ensemble forecasts of
the indian monsoon. Mon. Weather Rev. 127, 1954e1966.
Sugiyama, T., Okada, K., Muraishi, H., Noguchi, T., Samizo, M., 1995. Statistical
rainfall risk estimating method for a deep collapse of a cut slope. Soils Found.
35, 37e48.
Szentimrey, T., 2006. Development of mash homogenization procedure for daily
data. In: Proceedings of the Fifth Seminar for Homogenization and Quality
Control in Climatological Databases, Budapest, Hungary, pp. 123e130.
Tabios, G.Q., Salas, J.D., 1985. A comparative analysis of techniques for spatial
interpolation of precipitation. JAWRA J. Am. Water Resour. Assoc. 21, 365e380.
Toth, E., Brath, A., Montanari, A., 2000. Comparison of short-term rainfall prediction
models for real-time flood forecasting. J. Hydrology 239, 132e147. http://
dx.doi.org/10.1016/S0022-1694(00)00344-9.

nticos 1, 57e103.
Tullot, I.F., 1959. El clima de las islas canarias. Anu. Estud. atla
Wagner, P., Fiener, P., Wilken, F., Kumar, S., Schneider, K., 2012. Comparison and
evaluation of spatial interpolation schemes for daily rainfall in data scarce re-
gions. J. Hydrology 464e465, 388e400. http://dx.doi.org/10.1016/
j.jhydrol.2012.07.026.
Wright, D., Mantilla, R., Peters-Lidard, C., 2017. A remote sensing-based tool for
assessing rainfall-driven hazards. Environ. Model. Softw. 90, 34e54. http://
dx.doi.org/10.1016/j.envsoft.2016.12.006.
Xu, Y.P., Booij, M., Tong, Y.B., 2010. Uncertainty analysis in statistical modeling of
extreme hydrological events. Stoch. Environ. Res. Risk Assess. 24, 567e578.
http://dx.doi.org/10.1007/s00477-009-0337-8.
Yang, C., Chandler, R., Isham, V., Wheater, H., 2005. Spatial-temporal rainfall
simulation using generalized linear models. Water Resour. Res. 41, 1e13. http://
dx.doi.org/10.1029/2004WR003739.