Hindawi Publishing Corporation

Advances in Meteorology
Volume 2016, Article ID 6526341, 13 pages
http://dx.doi.org/10.1155/2016/6526341

Research Article
Comparison of Statistical Downscaling Methods for
Monthly Total Precipitation: Case Study for the Paute River
Basin in Southern Ecuador

L. Campozano,1,2,3 D. Tenelanda,1 E. Sanchez,1,2 E. Samaniego,1,2 and J. Feyen1

1
Departamento de Recursos Hı́dricos y Ciencias Ambientales, Universidad de Cuenca, 10150 Cuenca, Ecuador
2
Facultad de Ingenierı́a, Universidad de Cuenca, Avenida 12 de Abril s/n, 10150 Cuenca, Ecuador
3
Laboratory for Climatology and Remote Sensing (LCRS), Faculty of Geography, University of Marburg,
Deutschhausstrasse 10, 35032 Marburg, Germany

Correspondence should be addressed to L. Campozano; [email protected]

Received 27 October 2015; Accepted 27 December 2015

Academic Editor: Hung Soo Kim

Copyright © 2016 L. Campozano et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Downscaling improves considerably the results of General Circulation Models (GCMs). However, little information is available
on the performance of downscaling methods in the Andean mountain region. The paper presents the downscaling of monthly
precipitation estimates of the NCEP/NCAR reanalysis 1 applying the statistical downscaling model (SDSM), artificial neural
networks (ANNs), and the least squares support vector machines (LS-SVM) approach. Downscaled monthly precipitation estimates
after bias and variance correction were compared to the median and variance of the 30-year observations of 5 climate stations in the
Paute River basin in southern Ecuador, one of Ecuador’s main river basins. A preliminary comparison revealed that both artificial
intelligence methods, ANN and LS-SVM, performed equally. Results disclosed that ANN and LS-SVM methods depict, in general,
better skills in comparison to SDSM. However, in some months, SDSM estimates matched the median and variance of the observed
monthly precipitation depths better. Since synoptic variables do not always present local conditions, particularly in the period going
from September to December, it is recommended for future studies to refine estimates of downscaling, for example, by combining
dynamic and statistical methods, or to select sets of synoptic predictors for specific months or seasons.

1. Introduction GCM, which enables capturing the atmospheric phenomena

at a much higher resolution, in the order of tenths of
General Circulation Models (GCMs) are widely used to pre- kilometers. The SD techniques, on the other hand, are based
dict the impact of climate change on, for instance, the regional on the determination of statistical relations between large-
precipitation trend. The resolution of these models, typically scale synoptic predictors and local observations from ground
around 2∘ × 2∘ , is unsuitable for climate change impact stations, which are considered to be stationary, an assumption
estimations at basin scale [1, 2]. Additionally, these models that might not be true for future climate projections. The
do not capture well the subgrid processes, which can be very computational cost of these methods is low, they are relatively
complex in mountain regions, and fail to account properly easy to implement, and they present generally higher accu-
for the orographic features of those regions. Therefore, to racy than dynamical models. Particularly when the aim is not
obtain projections of the impact of climate change at basin to understand the change in weather processes provoked by
scale, particularly in a mountain region, downscaling is a climate change, the generation of future projections at basin
must. In general, the downscaling methods can be subdivided scale using SD methods might be convenient.
into two large groups: dynamical downscaling (DD) and Several SD techniques exist and among them the sta-
statistical downscaling (SD) methods. On the one hand, the tistical downscaling model (SDSM) is probably the most
DD methods integrate a regional climate model (RCM) in the widely used [3]. The SDSM approach facilitates the rapid
 1306, 2016, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2016/6526341 by Readcube (Labtiva Inc.), Wiley Online Library on [03/02/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
2 Advances in Meteorology

Table 1: Stations used for the present study.

Code Station Regime Latitude (∘ ) Longitude (∘ ) Elevation (m asl) Annual precipitation (mm)
M141 El Labrado TM −2.732 −79.073 3335 1286
M139 Gualaceo BM −2.882 −78.776 2230 820
M138 Paute TM −2.800 −78.763 2194 750
M045 Palmas UM −2.716 −78.630 2400 1341
M137 Biblián BM −2.709 −78.892 2640 1001

development of multiple, low cost, single-site scenarios of 79∘ 30󳰀 0󳰀󳰀 W 79∘ 0󳰀 0󳰀󳰀 W 78∘ 30󳰀 0󳰀󳰀 W
daily surface weather variables and is considered as a stochas- 2∘ 0󳰀 0󳰀󳰀 S 2∘ 0󳰀 0󳰀󳰀 S
tic weather generator on a daily scale. The limited use of
this technique in the Andean mountain region is perhaps
the consequence of its complex topography, location in the
tropical zone, and the influence of the warm and cold ENSO Ecuador
2∘ 30󳰀 0󳰀󳰀 S 2∘ 30󳰀 0󳰀󳰀 S
phases, El Niño and La Niña, respectively. All these influences
add complexity to the atmospheric processes, making the M141 M137 M045
# #
#
representation by downscaling techniques more difficult. # M138
# M139
Artificial neural networks (ANNs) are another SD technique
commonly used, however, to a limited extent, in the Andes 3∘ 0󳰀 0󳰀󳰀 S 3∘ 0󳰀 0󳰀󳰀 S
region. As stated by [4], the success of this technique is
primarily due to its ability to map highly nonlinear relations
between inputs and outputs of the model. Reference [5]
applied ANN for downscaling monthly precipitation and 79∘ 30󳰀 0󳰀󳰀 W 79∘ 0󳰀 0󳰀󳰀 W 78∘ 30󳰀 0󳰀󳰀 W
temperature in the Paute basin in the Andes of Ecuador. # Stations
The application of ANN based SD was conducted in order Paute basin
to evaluate the performance on seasonality representation
against DD using a regional climate model, Weather Research Figure 1: The study area Paute basin in the Andes of Ecuador.
and Forecasting, WRF, model [6]. With respect to rainfall
representation, they found that although both downscaling
approaches represent qualitatively well seasonality in this
2. Study Area and Data
highly complex terrain, ANN estimates of rainfall were
more accurate than WRF estimates. This fact highlights The Paute River basin, tributary of the Amazon basin and
the applicability of ANN when the understanding of the 6148 km2 in size, was selected for the comparative evaluation
processes involved is not required. Another technique with of the downscaling methods, a basin located between the
recent application for the downscaling of GCMs is SVM, eastern and western cordillera of the Andes in Ecuador. The
support vector machines [7]. In particular the least squares basin is characterized by a high spatial and temporal variation
support vector machines, LS-SVM [8], downscale GCM out- in precipitation that broadly can be classified into three
put even better, mainly due to the reduction of the optimiza- rainfall regimes, respectively, subregions with a uni-, bi-, and
tion problem to the resolution of a linear system reducing three-modal precipitation pattern [10, 11]. Data of 5 rainfall
considerably the computational requirements. Despite SD stations (see Table 1) were used of whom the geographical
methods presenting greater accuracy than DD methods, from distribution is given in Figure 1. The measured monthly
a statistical point of view, they possess two handicaps: bias rainfall for the 30-year period 01-1980 to 12-2009 was used
and low variance. Normally, the quantile mapping (QM) to quantify the differences in the performance of the selected
technique [9] is applied to improve the representation of the downscaling methods. The dataset was split in a first set for
distribution of derived applications. the calibration of the methods, encompassing 75% of the total
This paper presents a comparative evaluation of down- dataset, and the remaining 25% was used for the validation.
scaled GCM estimates of monthly precipitation at the scale The results presented herein belong all to the validation set.
of a large river basin situated in the Andean mountain region Quality control of the data was performed using double mass
in southern Ecuador, applying SDSM and two artificial intel- curves on the time series with gaps not exceeding 20% of the
ligence (AI) techniques: artificial neural networks (ANNs) observations. The infilling of the data was accomplished using
and the least squares support vector machines (LS-SVM) multiple linear regression with stations with higher Pearson
approach. The downscaled results were corrected for bias and correlation [12].
variance inflation applying the QM technique prior to the NCEP/NCAR reanalysis 1 data [13] with 2.5∘ × 2.5∘
comparative analysis. For the evaluation, historic data was resolution was used as input. The selected synoptic predictors
used. for the SDSM and AI models are presented in Table 2.
 1306, 2016, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2016/6526341 by Readcube (Labtiva Inc.), Wiley Online Library on [03/02/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Advances in Meteorology 3

Table 2: Synoptic predictors.

ANN/LS-SVM SDSM
# Synoptic predictors
(All stations) El Labrado Gualaceo Paute Palmas Biblián
1 Precipitation ∗ ∗ ∗ ∗ ∗ ∗
2 Pressure (surface) ∗ ∗ ∗ ∗
3 Relative humidity (surface) ∗ ∗ ∗ ∗ ∗
4 Specific humidity 700 hPa ∗
5 Sea level pressure ∗
6 Temperature 2 m ∗ ∗ ∗ ∗ ∗
7 Potential temperature 700 hPa ∗ ∗ ∗ ∗
8 Zonal wind (surface) ∗
9 Omega 500 hPa ∗ ∗ ∗
10 Geopotential height 200 hPa ∗ ∗
11 Geopotential height 500 hPa ∗
12 Geopotential height 850 hPa ∗ ∗

3. Methods pressure, relative humidity, and the temperature 2 m above

the surface. For the Palmas station, the only station with a
The methodology of the present study encompasses the
unimodal regime, the best results were obtained using three
following steps: (i) selection of the predictors, (ii) calibration
predictors, namely, precipitation, potential temperature at the
and validation of the SDSM and AI models, (iii) SDSM and
700 hPa level, and the geopotential height at the 850 hPa level.
AI ensemble generation, (iv) bias correction by quantile map-
Table 2 provides the list of the selected predictors as a function
ping, and (v) evaluation of results. SDSM can be conceived of the downscaling technique and station.
as a weather generator model, while ANN and LS-SVM are Using the selected predictors, daily precipitation was
both transfer functions in statistical downscaling models. downscaled. Given the low match with the observed time
Given the analogy between ANN and LS-SVM models, the series of daily data, monthly time series were generated. At
results of both formed one ensemble, which were compared the monthly scale the precipitation time series in the study
to the SDSM ensemble. Notwithstanding it is well known [14] basin are continuous; months with zero precipitation are not
that a selection of predictors for specific months or seasons present even during the dry months. Regarding the SDSM
might be a good option to improve the accuracy in climate model, 20 versions were applied [15]. Then, the median of the
projections; in the present study only one set of predictors for simulated results was calculated as the representative value
the whole year was considered given that the main interest for the ensemble of the SDSM model variants.
was the comparison of downscaling methods rather than the
analysis of climate projections. The same predictors were used 3.2. Downscaling Using ANN. An artificial neural network
for the AI ensemble of ANN and LS-SVM models, and one (ANN) is composed of several interconnected layers of
set of predictors in each station for the SDSM model. A more processing units (the neurons) that transform inputs into
detailed description of the tested downscaling models is given outputs. The inputs at the neurons are multiplied by weights
in the following paragraphs. and then inserted into an activation function. ANNs are
characterized by their topology, and probably the most widely
3.1. Downscaling Using SDSM. SDSM is a hybrid between known neural network is the multilayer perceptron (MLP).
a stochastic weather generator and a multilinear regression It consists of multiple layers of adaptive weights with full
method [14], forcing synoptic-scale weather variables to local connectivity between inputs and hidden units and between
meteorological variables using statistical relationships. In hidden units and outputs. MLP is feed-forward artificial
order to be better in agreement with the variance of the neural network mapping sets of input data onto a set of
observed time series, stochastic techniques are used to arti- appropriate outputs.
ficially inflate the variance of the downscaled weather time The neural network toolbox of Matlab [16] was used,
series. The reader is referred to [3] for a detailed description and optimization of the neural network was pursued using
of the SDSM technique, an approach widely used for the the Levenberg-Marquardt method, minimizing the mean
downscaling of large-scale meteorological, hydrological, and square error. The performance of a total of four ANNs was
environmental variables. tested, respectively, a model considering either one or two
Partial correlation at a significance level of 𝑝 = 0.05 was intermediate neural layers, and a linear or sigmoidal transfer
used to select the predictors that capture best for each of the function in the neurons (Table 3). For the input layer all
climate stations the effect of global climate. Precipitation data networks had 12 neurons (equal to the number of predictors;
from the reanalysis products was evaluated as the predictor see Table 2) and for the network with one hidden layer
that considerably helped improving the downscaling. For 8 neurons were used, which was determined by trial and
the climate stations Biblián, Paute, and Gualaceo, 5 predic- error. In a similar way for networks with two hidden layers,
tors were identified, of whom 4 are similar: precipitation, respectively, 9 and 5 neurons were used.
 1306, 2016, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2016/6526341 by Readcube (Labtiva Inc.), Wiley Online Library on [03/02/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
4 Advances in Meteorology

Table 3: Artificial intelligence models. Finally the LS-SVM model for function estimation is
Model Type Transfer function 𝑁
1 Linear & 1 hidden layer 𝑦 (𝑥) = ∑ 𝛼𝑖 𝐾 (𝑥𝑖 , 𝑥) + 𝑏. (7)
2 Sigmoidal & 1 hidden layer 𝑖=1
Ann
3 Linear & 2 hidden layers
A Bayesian framework with three levels of inference was
4 Sigmoidal & 2 hidden layers
developed for the optimization of parameters [8]. The LS-
5 Linear SVMlab tool [17] developed in Matlab applying three Kernel
6 LS-SVM Polynomial tuning options (linear, polynomial, and RBF) was used within
7 Radial basis functions the ensemble of AI methods.

3.4. Bias Correction Using the Quantile Mapping Approach.

3.3. Downscaling Using LS-SVM. Support vector machines,
First the predictors were selected, followed by the application
SVM [7], solve nonlinear classifications and estimations of
of the multiple compositions of the SDSM model and the 7
functions and densities using quadratic programming. A
AI models, 4 ANNs, and 3 LS-SVM models (Table 3). The
least squares support vector machine, LS-SVM [8], is a
output distributions of the ensemble of SDSM models and
reformulation of a SVM replacing the solution of the convex
the 7 AI models were grouped into two distinct populations.
quadratic programming problem by the solution of a set of
Both these populations were corrected for bias and variance
linear equations. As explained by [8] it is possible to adopt
inflation applying the quantile mapping technique. The QM
into LS-SVM the robustness, sparseness, and weightings.
applied to SDSM population distributions is from now on
Due to the fact that LS-SVM has more recent application
called the SDSM QM and the QM applied to AI the AI QM.
than ANN and SDSM techniques for the downscaling of
The quantile mapping technique for bias correction and
GCMs, we present here a succinct description of LS-SVM
variance inflation can be regarded as a statistical transforma-
theory. For a more in deep description of LS-SVM see [8].
tion of the original distribution into a reference probability
Let us consider 𝑥 ∈ R𝑛 and 𝑦 ∈ R; the LS-SVM model,
distribution. It aims to find the function that maps the
mapping the 𝑥 into a feature space, is
distribution of the model variables into the distribution of the
observed variables. Basically three types of approaches exist
𝑦 = 𝑤𝑇 𝜑 (𝑥) + 𝑏. (1)
to find the mapping function; they are adjustment (i) based on
The optimization problem can be stated as theoretical distributions, (ii) based on parametric transfor-
mations, and (iii) based on nonparametric transformations.
After evaluation of several transformations the power
1 𝑇 1 𝑁 2
min Γ (𝑤, 𝑒) = 𝑤 𝑤 + 𝛾 ∑𝑒 , (2) parametric transformation was applied because of its good
𝑤,𝑒 2 2 𝑖=1 𝑖 results and the parsimony of the model [9]. The power
parametric transformation is defined as
where 𝑒 and 𝛾 are the error and the regularization parameter,
respectively. The minimization of the cost function Γ(𝑤, 𝑒) is 𝑎
𝑃0∗ = 𝑏 (𝑃𝑚 − 𝑥0 ) , (8)
subject to the constrains:
where 𝑃0∗ and 𝑃𝑚 are the corrected modeled and modeled
𝑦𝑖 = 𝑤𝑇 𝜑 (𝑥𝑖 ) + 𝑏 + 𝑒𝑖 . (3) variables, and 𝑎, 𝑏, and 𝑥0 are parameters to be determined.
To find the parameters 𝑎, 𝑏, and 𝑥0 , the tool 𝑅 QMAP
The Lagrangian of the optimization is
was used [18]. For SDSM QM and AI QM the parameters
𝑁 obtained after QM adjustment are presented in Table 4.
L (𝑤, 𝑏, 𝑒, 𝛼) = Γ (𝑤, 𝑒) − ∑ 𝛼𝑖 (𝑤𝑇 𝜑 (𝑥𝑖 ) + 𝑏 + 𝑒𝑖 − 𝑦𝑖 ) , (4)
𝑖=1 3.5. Criteria for Model Evaluation. For the qualitative assess-
where 𝛼𝑖 are the Lagrangian multipliers. ment of the used methods of downscaling the long-term
After considering the conditions for optimality: monthly mean precipitation of both ensembles, the SDSM
and AI, was compared with the long-term monthly mean
observed precipitation, and for the quantitative evaluation
𝜕L 𝜕L 𝜕L 𝜕L 𝑇
[ , , , ] = [0, 0, 0, 0]𝑇 . (5) statistical metrics of the distributions of the time series of
𝜕𝑤 𝜕𝑏 𝜕𝑒𝑖 𝜕𝛼𝑖 observed and modeled values for the validation period were
calculated. In addition to the statistical metrics, Box-Whisker
We obtain the matrix equation:
plots presenting both ensembles versus the observations were
0 1 𝑏 0 drawn. This type of graphical representation was selected
[ −1
][ ] = [ ], (6) because, in addition to the median, the Box-Whisker plot
1 Ω−𝛾 𝐼 𝛼 𝑦 depicts the extreme values, respectively, the minimum and
maximum (the caps at the end of each box), and the outliers
where Ω𝑖𝑗 = 𝜑(𝑥𝑖 )𝑇 𝜑(𝑥𝑗 ) is the Kernel function 𝐾(𝑥𝑖 , 𝑥𝑗 ) and falling more than 1 time of the interquartile range above the
𝐼 is the identity matrix. third or below the first quartile (the points in the graph).
 1306, 2016, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2016/6526341 by Readcube (Labtiva Inc.), Wiley Online Library on [03/02/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Advances in Meteorology 5

Table 4: Quantile mapping parameters for artificial intelligence and Table 5: Statistical metrics for ANN and LS-SVM ensembles.
SDSM ensembles.
Station Metric ANN LS-SVM
Parameters Pearson correlation 0.49 0.58
Model Stations
𝑎 𝑏 𝑥0 IRF 0.59 0.53
Biblián 0.03280 1.83727 0.87053 El Labrado Mean-bias 4.20 6.16
El Labrado 0.00820 1.98426 −8.57836 Cum bias 31.63 37.79
AI Palmas 0.00334 2.21628 6.04480 RMSE 40.73 37.63
Gualaceo 0.31943 1.40124 16.92140 Pearson correlation 0.70 0.71
Paute 0.02791 1.94766 0.03257 IRF 0.55 0.40
Biblián 0.00986 2.08840 3.21195 Gualaceo Mean-bias −2.29 −0.21
El Labrado 0.39313 1.19070 −45.84738 Cum bias 33.41 41.70
SDSM Palmas 0.00056 2.46774 −13.06023 RMSE 34.82 35.33
Gualaceo 0.00077 2.62716 −1.14051 Pearson correlation 0.55 0.54
Paute 0.04141 1.80403 3.91130 IRF 0.41 0.31
Paute Mean-bias −9.90 −8.34
Cum bias 44.16 48.21
As statistical metrics the following were used:
RMSE 31.61 31.04
(1) Pearson correlation (𝑅): to evaluate the linear corre- Pearson correlation 0.25 0.45
lation. IRF 0.45 0.53
(2) Mean bias (MB): to evaluate the mean (the 50th Palmas Mean-bias −1.85 7.87
percentile) difference between modeled and observed Cum bias 36.99 37.81
distributions. RMSE 52.44 47.51
(3) The root mean squared error (RMSE): to evaluate the Pearson correlation 0.67 0.65
error between modeled and observed time series. IRF 0.57 0.54
(4) The interquartile relative fraction (IRF): to evaluate Biblián Mean-bias −18.39 −7.05
the modeled variability representation relative to the Cum bias 52.34 35.49
observed: RMSE 44.75 40.97
𝑄3𝑚 − 𝑄1𝑚
IRF = , (9)
𝑄3𝑜 − 𝑄1𝑜 into one ensemble considering both models as transfer
function based models. Although previous studies [19, 20]
where IRF is the interquartile relative fraction. A demonstrated the relative superior performance of SVM
value of IRF > 1 represents overestimation of the based downscaling methods over other approaches including
variability, IRF = 1 is a perfect representation of the ANN based methods, we compared for the first time in the
variability, and IRF < 1 is an underestimation of region, to the best of our knowledge, an ANN ensemble
the variability; 𝑄3𝑚 and 𝑄3𝑜 and the 75th modeled and against a LS-SVM ensemble to evaluate their downscaling
observed percentile; 𝑄1𝑚 and 𝑄1𝑜 and the 25th modeled performance. The ensembles were not bias corrected in order
and observed percentile. to evaluate their actual performance. Table 5 presents the
(5) The absolute cumulative bias (ACB): to evaluate the values of 𝑅, RMSE, MB, IRF, and ACB comparing ANN
bias of the 25th, 50th, and 75th percentiles; or and LS-SVM ensembles. In El Labrado and Paute station
󵄨 󵄨 󵄨 󵄨 󵄨 󵄨 similar results of both ensembles are obtained. However,
ACB = 󵄨󵄨󵄨𝑄1𝑚 − 𝑄1𝑜 󵄨󵄨󵄨 + 󵄨󵄨󵄨𝑄2𝑚 − 𝑄2𝑜 󵄨󵄨󵄨 + 󵄨󵄨󵄨𝑄3𝑚 − 𝑄3𝑜 󵄨󵄨󵄨 , (10) both ensembles for Gualaceo station underrepresent the
variability. IRF is 0.55 and 0.40 for ANN and LS-SVM,
where ACB is the absolute cumulative bias. A value respectively. Therefore LS-SVM represents 15% less variability
of ACB = 0 is a perfect representation of the three than ANN ensemble. The MB for ANN is −2.29 mm whereas
percentiles (resp., the 25th, 50th, and 75th percentile) for LS-SVM it is −0.21 mm, which in monthly scale is very
of modeled and observed distributions, while under- low. The ACB metric for ANN is 33.41 mm whereas for LS-
or overestimation indicates a divergence of ACB from SVM it is 41.70 mm, meaning that although the MB is lower
zero to positive values. for LS-SVM, the bias in the 25th and 75th percentiles is higher
than for the ANN ensemble. In Palmas station 𝑅 is 0.25 and
4. Results and Discussion 0.45, respectively, for the ANN and LS-SVM ensembles. IRF
values of 0.45 and 0.53 for ANN and LS-SVM are obtained,
4.1. Evaluation of SDSM and AI Ensembles indicating a greater representation in variability by the latter
approach. However, MB values of −1.85 and 7.87 mean that
4.1.1. Relative Performance of ANN and LS-SVM Models. As LS-SVM presents more bias in the 50th percentile than ANN.
mentioned before, ANN and LS-SVM models were grouped For Biblián station, MB is −18.39 mm and −7.05 mm and ACB
 1306, 2016, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2016/6526341 by Readcube (Labtiva Inc.), Wiley Online Library on [03/02/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
6 Advances in Meteorology

Table 6: Statistical metrics for artificial intelligence and SDSM ensembles.

Station Metric AI SDSM AI QM SDSM QM

Pearson correlation 0.58 0.37 0.58 0.38
IRF 0.49 1.00 0.87 1.14
El Labrado Mean-bias 4.37 −41.49 −3.24 −0.54
Cum bias 38.23 126.89 12.19 9.96
RMSE 37.70 65.41 41.67 51.44
Pearson correlation 0.74 0.53 0.72 0.50
IRF 0.52 0.46 1.04 1.03
Gualaceo Mean-bias −1.01 10.28 4.33 1.85
Cum bias 34.02 47.50 7.35 9.01
RMSE 33.92 38.26 32.87 44.86
Pearson correlation 0.59 0.47 0.57 0.47
IRF 0.36 0.47 0.80 0.89
Paute Mean-bias −10.46 1.14 −4.26 1.34
Cum bias 47.61 31.73 15.59 7.77
RMSE 30.60 30.26 31.13 35.03
Pearson correlation 0.44 0.16 0.44 0.14
IRF 0.52 0.54 1.12 1.02
Palmas Mean-bias 7.39 16.31 0.97 −3.66
Cum bias 38.14 56.93 8.77 12.28
RMSE 47.45 56.09 55.73 66.62
Pearson correlation 0.66 0.46 0.67 0.45
IRF 0.61 0.56 1.29 1.25
Biblián Mean-bias −11.12 −2.87 −1.21 1.66
Cum bias 35.01 29.99 18.91 17.30
RMSE 41.73 44.89 39.87 51.81

52.34 mm and 35.49 mm for ANN and LS-SVM ensembles, 350

respectively, meaning a strong bias for the ANN ensemble 300
with respect to LS-SVM ensemble.
250
Precipitation (mm)

For a qualitative evaluation of ANN and LS-SVM ensem-

bles the Box-Whisker plots for the results during the valida- 200
tion period are presented in Figure 2. For El Labrado station
both ensembles similarly represent the median, although 150
the low variance is clearly showed, as measured by IRF in 100
Table 5. For Gualaceo and Paute stations both ensembles
represent less variance, with LS-SVM presenting lower val- 50
ues. The percentiles above 75th are strongly underestimated 0
in both stations making the necessity of correction on the M141 M139 M138 M045 M137
distribution of the ensembles evident. In Palmas station El Labrado Gualaceo Paute Palmas Biblián
both ensembles underrepresent the variance, and the median
Observations
is rightly represented by ANN but overestimated by LS-
ANN
SVM. Finally for Biblián station the variance is strongly LS-SVM
underestimated as well as the higher percentiles. The median
is better represented by LS-SVM and underestimated by Figure 2: Box plots for ANN and LS-SVM ensembles evaluated in
ANN ensemble. Both methods were able to perform similarly station El Labrado (M141), Gualaceo (M139), Paute (M138), Palmas
well for the downscaling of monthly precipitation in the (M045), and Biblián (M137).
selected stations. In addition, comparison of the quantitative
analysis based on the statistical metrics and the qualitative
analysis based on Box-Whisker plots shed light on the relative the derived SDSM and AI ensembles were compared after
performance of ANN and LS-SVM methods. QM correction. Table 6 depicts for the 5 climate stations in the
Paute River basin, of which data were used, the comparison
4.1.2. Comparison of SDSM and AI Ensembles. Once the between the SDSM and AI versus SDSM QM and AI QM.
ANN and LS-SVM ensembles were evaluated, in a next step The evaluation between both sets is based on the statistical
 1306, 2016, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2016/6526341 by Readcube (Labtiva Inc.), Wiley Online Library on [03/02/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Advances in Meteorology 7

350 evaluation of seasonality representation might not help to

300
quantify, for example, flooding events, it is very important
for issues related to water availability for hydroelectricity
250
Precipitation (mm)

generation, drinking water availability, and agriculture. Fur-

200
ther, the evaluation of seasonality representation is of special
importance in the study region due to the low resolution
150 of GCMs, unable to depict the precipitation regime due to
100
mesoscale influences [5].

50 4.2.1. The Added Value of Quantile Mapping. The QM cor-

0
rection parameters for the power parametric transformation
M141 M139 M138 M045 M137 applied to AI and SDSM ensembles are presented in Table 4.
El Labrado Gualaceo Paute Palmas Biblián The comparison of the multiyear monthly mean (mymm)
precipitation of the SDSM with the SDSM QM ensemble
Observations
is presented in Figures 4(a)–4(e). As shown in Figure 4(a),
AI_QM
SDSM_QM
SDSM applied to the El Labrado station fails to capture
the observed seasonality. However, the performance, bias,
Figure 3: Monthly precipitation box plots for artificial intelligence and variance improved considerably after applying QM.
and SDSM ensembles bias corrected. Results evaluated in station El Seasonality applying SDSM to the Gualaceo (Figure 4(b))
Labrado (M141), Gualaceo (M139), Paute (M138), Palmas (M045), station is less correctly presented and fails to capture the
and Biblián (M137). maximum in April and overestimates precipitation during the
dry season in August. Application of QM only corrects the
representation in August, but not in April. The ensemble of
metrics R, RMSE, MB, ACB, and IRF. For El Labrado station SDSM compositions represents well seasonality but under-
SDSM QM presents lower correlation than AI QM with estimates significantly the November precipitation depth.
0.38 and 0.58 values. Although SDSM QM presents a lower Application of QM improves the representation of seasonality
bias than AI QM, the RMSE of AI QM is a bit lower. For but does not improve the November estimate. The SDSM in
Gualaceo station 𝑅 for AI QM and SDSM QM are 0.72 and Paute station represents seasonality well, but underestimates
0.5. RMSE as in El Labrado station is lower for AI QM than significantly the maximum in November. QM applied to
for SDSM QM with 32.87 and 44.86, respectively. For Paute SDSM in Paute improves the performance of seasonality,
station also 𝑅 is higher for AI QM with 0.57 and 0.47 for yet fails to improve the representation of the November
SDSM QM, although the ACB is higher for AI QM with precipitation (Figure 4(c)). The SDSM approach calibrated
15.59 and 7.77 for SDSM QM. For Palmas there is a marked to the observations of the Palmas station (Figure 4(d)), a
difference in 𝑅 values with 0.44 for AI QM and 0.14 for station with unimodal regime (UM), depicts fairly correct
SDSM QM, depicting for the former a lower RMSE than seasonality notwithstanding the limited spatial extent of the
the latter. Similar for the Biblián station is the 𝑅 value for UM regime and the poor representation of the mesoscale
AI QM higher than for SDSM QM, respectively, 0.67 versus influences in the synoptic predictors. Application of QM
0.45. Analogous to the Palmas station, a lower RMSE value negatively affects the SDSM representation during the first 6
for AI QM equal to 39.87 was obtained compared to the months of the year but improves slightly the representation
calculated RMSE of 51.81 for SDSM QM. during the remaining period of the year. The seasonality of
All other metrics in Table 6 present similar values. The the Biblián station (Figure 4(e)) is properly represented by
stronger differences arise generally in the RMSE and 𝑅 the SDSM ensemble and as well as the SDSM QM. Only the
statistical metrics, which might be related to the fact that November peak in precipitation is captured by neither SDSM
QM corrects only the characteristic of the distribution, as nor SDSM QM.
can be seen in Figure 3. This figure presents the monthly The comparison of mymm of the AI with the AI QM
precipitation Box-Whisker plots for AI and SDSM ensem- ensembles is presented in Figures 4(f)–4(j). The AI emsemble
bles bias corrected for the 5 climate stations. As can be underestimates precipitation in April, overestimate it in July
observed the distributions are fairly alike. From the analysis and August, and underestimate it in November, although
for all stations AI QM presented higher values of 𝑅 than at some extent it represents seasonality. QM improves the
SDSM QM. Similarly AI QM presents better agreement with intra-annual variability, but still the November maximum
the observed data with the exception of the Paute station. is not correctly estimated. The AI ensemble in Gualaceo
This fact might point to a slightly better representation of (Figure 4(g)) station captures well both peaks in April
the observed monthly precipitation distribution by AI QM and November but underestimates precipitation in August.
ensemble, for this specific region. The estimation in August is considerably improved after
correction of the bias and the inflation of the variance. In the
4.2. Evaluation of Intra-Annual Precipitation Seasonality Rep- Paute station (Figure 4(h)) are both the peaks, respectively, in
resentation. Whereas in previous section the entire distribu- April and November, and the minimum in August were well
tion of downscaled estimates was evaluated, in the following captured by the AI ensemble, while QM further improves the
the representation of seasonality is evaluated. Although the distribution of the median of the monthly precipitation. The
 8

Precipitation (mm) Precipitation (mm) Precipitation (mm)

0
20
40
60
80
100
120
140
160
180
200
0
20
40
60
80
100
120
140
160
180
200

0
20
40
60
80
100
120
140
160
180
200
January January January
February February February

SDSM

SDSM

SDSM
March March March

Obs._Paute
SDSM_QM

SDSM_QM

SDSM_QM
April April April

Obs._Biblian
Obs._El Labrado
May May May

(e)
(c)
(a)
June June June
July July July
August August August
September September September
October October October
November November November
December December December

Precipitation (mm) Precipitation (mm) Precipitation (mm)

0
20
40
60
80
100
120
140
160
180
200

0
20
40
60
80
100
120
140
160
180
200

0
20
40
60
80
100
120
140
160
180
200

Figure 4: Continued.
January January January

AI
February February February

SDSM
SDSM

AI_QM
March March March

SDSM_QM
SDSM_QM

April April

Obs._Palmas
April
Obs._Gualaceo

Obs._El Labrado
May May May
(b)

(f)
(d)
June June June
July July July
August August August
September September September
October October October
November November November
December December December
Advances in Meteorology

1306, 2016, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2016/6526341 by Readcube (Labtiva Inc.), Wiley Online Library on [03/02/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
 1306, 2016, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2016/6526341 by Readcube (Labtiva Inc.), Wiley Online Library on [03/02/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Advances in Meteorology 9

200 200
180 180
160 160

Precipitation (mm)
140 140
Precipitation (mm)

120 120
100 100
80 80
60 60
40 40
20 20
0 0

October
March
April

June

August
September

November
December
January
February

May

July
October
March
April

June

August
September

November
December
January
February

May

July

Obs._Gualaceo Obs._Paute
AI AI
AI_QM AI_QM
(g) (h)
200 200
180 180
160 160
Precipitation (mm)

140 140
Precipitation (mm)

120 120
100 100
80 80
60 60
40 40
20 20
0 0
October
March
April

June

August
September

November
December
January
February

May

July

October
March
April

June

August
September

November
December
January
February

May

July

Obs._Palmas ́
Obs._Biblian
AI AI
AI_QM AI_QM
(i) (j)

Figure 4: Multiyear monthly mean precipitation for SDSM ensemble and SDSM ensemble bias corrected (a–e) and multiyear monthly mean
precipitation for AI and AI bias corrected (f–j).

AI ensemble represents poorly the distribution of the Palmas Table 7: Pearson correlations between observations and SDSM QM
station (Figure 4(i), UM regime), even after QM application. and AI QM models.
For the Biblián station (Figure 4(j)) the AI QM captures the
Station AI QM SDSM QM
April peak one month earlier but fails to correctly depict the
El Labrado 0.741 0.516
magnitude of the November peak.
Results clearly reveal that the application of QM to Gualaceo 0.946 0.672
the output of both modeling approaches, SDSM and AI, Paute 0.730 0.629
overall improves the representation of seasonality, as well Palmas 0.788 0.593
as the representation of rainy and dry periods. However, Biblián 0.691 0.532
both approaches underestimate the median value of the
precipitation depth in November. This fact could indicate that
the set of synoptic predictors do not include a variable that 4.2.2. Representation of Monthly Variability by Downscaled
is related to an enhancement of precipitation in this period. Results. To compare the representativeness of SDSM QM
Further studies are needed to determine the variables and and AI QM, the correlation of the mymm time series
related phenomena. with the observed values is shown in Table 7. For all
 10

Precipitation (mm) Precipitation (mm) Precipitation (mm)

0
20
40
60
80
100
120
140
160
180
200

200
200

0
20
40
60
80
100
120
140
160
180
0
20
40
60
80
100
120
140
160
180
January January January
February February February

AI_QM
AI_QM

AI_QM
March March March

Obs._Paute

SDSM_QM
SDSM_QM

SDSM_QM
April April April

Obs._Gualaceo
Obs._El Labrado
May May May

(a)

(e)
(c)
June June June
July July July
August August August
September September September
October October October
November November November
December December December

Precipitation (mm) Precipitation (mm) Precipitation (mm)

0
50
100
150
200
250

200
250
200
250

0
50
100
150
0
50
100
150
January January January

Figure 5: Continued.
February February February
AI_QM

AI_QM

AI_QM
March March March

SDSM_QM
SDSM_QM
SDSM_QM

Observations
Observations
Observations

April April April

May May May

(f)
(b)

(d)
June June June
July July July
August August August
September September September
October October October
November November November
December December December
Advances in Meteorology

1306, 2016, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2016/6526341 by Readcube (Labtiva Inc.), Wiley Online Library on [03/02/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
 1306, 2016, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2016/6526341 by Readcube (Labtiva Inc.), Wiley Online Library on [03/02/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Advances in Meteorology 11

200 350
180
300
160
Precipitation (mm)

140 250

Precipitation (mm)
120
200
100
80 150
60 100
40
20 50
0 0

October

October
March
April

June

August
September

November
December

March

April

June

August

September

November

December
January
February

May

July

February

May

July
January
Obs._Palmas Observations
AI_QM SDSM_QM
SDSM_QM AI_QM
(g) (h)
200 250
180
160 200
Precipitation (mm)

140
Precipitation (mm)

120 150
100
80 100
60
40 50
20
0 0
October
March
April

June

August
September

November
December
February

May

July
January

October
March

April

June

August

September

November

December
February

May

July
January

Obs._Biblián Observations
AI_QM SDSM_QM
SDSM_QM AI_QM
(i) (j)

Figure 5: Comparison of SDSM and AI ensembles bias corrected (a, c, e, g, i) and Box-Whisker plots for SDSM and AI ensembles bias
corrected, from January through December (b, d, f, h, j).

stations AI QM presents greater Pearson correlation coeffi- for SD should be based on several models. The median
cient than SDSM QM. The multiyear median observed and and interquartile range are relatively well captured for the
estimated monthly precipitation depth, using, respectively, Gualaceo station (Figures 5(c) and 5(d)); the variability of
the SDSM QM and AI QM model ensembles are presented the months from January to September is similar for the
in Figures 5(a), 5(c), 5(e), 5(g), and 5(i). For the graphical SDSM-QM and AI QM estimates. October and November
presentation the Box-Whisker plot type was selected, as to variability is different for the two models, but the median
show in addition to the median the variation in estimates (see is well represented. The variability and the median are
Figures 5(b), 5(d), 5(f), 5(h), and 5(j)). better represented by AI QM and slightly overestimated by
For the El Labrado station (Figures 5(a) and 5(b)) the SDSM QM in the period of June to August. Figures 5(e)
observed interquartile range is higher for the period of and 5(f) depict the results for the Paute station, illustrating
January to April, with lower values in August and September. that both models relatively well represent the median and
It is worthwhile noticing that although AI QM captures interquartile ranges in the period of January to September but
seasonality, the intra-annual variability is not captured. Even fail to do so for the period of October to December. This fact
for some months SDSM QM captures the variability better, highlights the need to further explore the relation between
as is the case for March. This fact suggests that an assessment the synoptic conditions and rainfall. Neither the SDSM QM
 1306, 2016, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2016/6526341 by Readcube (Labtiva Inc.), Wiley Online Library on [03/02/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
12 Advances in Meteorology

nor the AI QM model estimates correctly the median of Acknowledgments

the Palmas station (Figures 5(g) and 5(h)), the only station
with a unimodal regime. Because both model approaches This work was funded by the Dirección de Investigación de la
indistinctly overestimate or underestimate the median in Universidad de Cuenca (DIUC) through the project “Análisis
some months, it might be worthwhile to examine more in de los Efectos del Cambio Climático en los Caudales en las
detail the representation of an ensemble of both models. Cuencas Andinas del Sur del Ecuador (Paute), Debido a los
The interquartile range of each month is relatively well Cambios en los Patrones de Iluvia y Temperatura” and by
represented except in a distinct number of months, such as the Secretaria de Educación Superior, Ciencia, Tecnologı́a e
January, April, June, July, September, and October. This could Innovación (SENESCYT) through a Ph.D. grant for the first
mean that in those months the influences of the mesoscale author. The authors would also like to thank INAMHI for
factors are not properly represented in the synoptic variables. providing the meteorological data.
An option for its remediation could be a methodology in
which the influences of mesoscale factors are considered, for
example, dynamic downscaling, followed by the application References
of statistical downscaling to regional predictors. However,
[1] W. Buytaert, M. Vuille, A. Dewulf, R. Urrutia, A. Karmalkar,
further studies are necessary to support the applicability of and R. Célleri, “Uncertainties in climate change projections and
such an approach in mountain regions. For the Biblián station regional downscaling in the tropical Andes: implications for
(Figures 5(i) and 5(j)), the two approaches overestimate water resources management,” Hydrology and Earth System
precipitation in the period of January to March. AI QM Sciences, vol. 14, no. 7, pp. 1247–1258, 2010.
captures adequately the median from April to December, [2] D. E. Mora, L. Campozano, F. Cisneros, G. Wyseure, and P.
with exception of November, as was the case for the other Willems, “Climate changes of hydrometeorological and hydro-
stations. Overall, the AI QM depicts fairly well the variability, logical extremes in the Paute basin, Ecuadorean Andes,” Hydrol-
except for October and November, whereas SDSM QM ogy and Earth System Sciences, vol. 18, no. 2, pp. 631–648, 2014.
underestimates the variability throughout the year. [3] R. L. Wilby, C. W. Dawson, and E. M. Barrow, “SDSM—a deci-
sion support tool for the assessment of regional climate change
5. Conclusions impacts,” Environmental Modelling & Software, vol. 17, no. 2,
The evaluation of downscaling methods in mountain regions pp. 145–157, 2002.
is of major importance due to the misrepresentation of cli- [4] P. Coulibaly, Y. B. Dibike, and F. Anctil, “Downscaling precipi-
mate by GCMs. The low resolution of GCMs limits the accu- tation and temperature with temporal neural networks,” Journal
of Hydrometeorology, vol. 6, no. 4, pp. 483–496, 2005.
rate prediction of the probable impacts of climate change at
basin scale. In the present work, the applicability of monthly [5] A. Ochoa, L. Campozano, E. Sánchez, R. Gualán, and E.
precipitation downscaling of global climate models by SDSM, Samaniego, “Evaluation of downscaled estimates of monthly
and methods of artificial intelligence, as neural networks and temperature and precipitation for a Southern Ecuador case
study,” International Journal of Climatology, 2015.
least squares support vector machines, was studied. Also a
comparative analysis of the applied downscaling methods [6] W. C. Skamarock, J. B. Klemp, J. Dudhia et al., A Description
was conducted. Comparative analysis revealed that with of the Advanced Research WRF Version 3, National Center for
respect to the downscaling of monthly precipitation neural Atmospheric Research, Boulder, Colo, USA, 2008.
networks and least squares support vector machine models [7] V. N. Vapnik, The Nature of Statistical Learning Theory, Springer,
perform equally. Considering the statistical metrics, such as New York, NY, USA, 1995.
Pearson correlation, root mean square error, and percentiles [8] J. A. K. Suykens, T. Van Gestel, J. De Brabanter, B. De Moor, and
biases, overall the artificial intelligence methods showed J. Vandewalle, Least Squares Support Vector Machines, World
better skills in relation to SDSM, although, in some stations Scientific, Singapore, 2002.
and some months, the importance of considering both model [9] L. Gudmundsson, J. B. Bremnes, J. E. Haugen, and T. Engen-
approaches was necessary in order to derive robust conclu- Skaugen, “Technical note: downscaling RCM precipitation to
sions. In general, although the representation of precipitation the station scale using statistical transformations—a compari-
from January to August is adequate, especially in November, son of methods,” Hydrology and Earth System Sciences, vol. 16,
both approaches failed to represent precipitation in some no. 9, pp. 3383–3390, 2012.
stations. Further analysis of the synoptic conditions for [10] L. Campozano, R. Célleri, K. Trachte, J. Bendix, and E. Sama-
this period is therefore recommended and a methodology niego, “Rainfall and cloud dynamics in the Andes: a southern
considering downscaling with specific predictors by month Ecuador case study,” Advances in Meteorology, In press.
or season might be advisable. From the analysis on Palmas [11] R. Célleri, P. Willems, W. Buytaert, and J. Feyen, “Space-
station, a station with important mesoscale influences, we time rainfall variability in the Paute basin, Ecuadorian Andes,”
could derive that further evaluation of a methodology of Hydrological Processes, vol. 21, no. 24, pp. 3316–3327, 2007.
downscaling using dynamic and statistical methods in cas- [12] L. Campozano, E. Sánchez, A. Avilés, and E. Samaniego, “Eval-
cade could help capture features that GCMs are not able to uation of infilling methods for time series of daily precipitation
represent. and temperature: the case of the Ecuadorian Andes,” Maskana,
vol. 5, no. 1, pp. 99–115, 2014, http://www.ucuenca.edu.ec/ojs/
index.php/maskana/article/view/431.
Conflict of Interests
[13] E. Kalnay, M. Kanamitsu, R. Kistler et al., “The NCEP/NCAR
The authors declare that there is no conflict of interests 40-year reanalysis project,” Bulletin of the American Meteorolog-
regarding the publication of this paper. ical Society, vol. 77, no. 3, pp. 437–471, 1996.
 1306, 2016, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2016/6526341 by Readcube (Labtiva Inc.), Wiley Online Library on [03/02/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Advances in Meteorology 13

[14] M. S. Pervez and G. M. Henebry, “Projections of the Ganges-

Brahmaputra precipitation-Downscaled from GCM predic-
tors,” Journal of Hydrology, vol. 517, pp. 120–134, 2014.
[15] R. Mahmood and M. S. Babel, “Future changes in extreme
temperature events using the statistical downscaling model
(SDSM) in the trans-boundary region of the Jhelum river basin,”
Weather and Climate Extremes, vol. 5-6, pp. 56–66, 2014.
[16] M. H. Beale, M. T. Hagan, and H. B. Demuth, “Neural Network
Toolbox™ User’s Guide R,” MathWorks, Natick, Mass, USA, 2014.
[17] K. De Brabanter, P. Karsmakers, F. Ojeda et al., LS-SVMlab
Toolbox User’s Guide, Version 1.8, 2011, http://www.esat
.kuleuven.be/sista/lssvmlab/downloads/tutorialv1 8.pdf.
[18] A. L. Gudmundsson, J. B. Bremnes, J. E. Haugen1, and T. Engen-
Skaugen, “Technical note: downscaling RCM precipitation to
the station scale using statistical transformations—a compari-
son of methods,” Hydrology and Earth System Sciences, vol. 16,
no. 9, pp. 3383–3390, 2012.
[19] S.-T. Chen, P.-S. Yu, and Y.-H. Tang, “Statistical downscaling of
daily precipitation using support vector machines and multi-
variate analysis,” Journal of Hydrology, vol. 385, no. 1–4, pp. 13–
22, 2010.
[20] S. Tripathi, V. V. Srinivas, and R. S. Nanjundiah, “Downscaling
of precipitation for climate change scenarios: a support vector
machine approach,” Journal of Hydrology, vol. 330, no. 3-4, pp.
621–640, 2006.