Atmospheric Research 213 (2018) 437–449

Contents lists available at ScienceDirect

Atmospheric Research
journal homepage: www.elsevier.com/locate/atmosres

Validation of CHIRPS precipitation dataset along the Central Andes of T

Argentina
⁎
Juan A. Riveraa,b, , Georgina Marianettib, Sofía Hinrichsb
a
Instituto Argentino de Nivología, Glaciología y Ciencias Ambientales (CCT-Mendoza/CONICET), Av. Ruiz Leal s/n, Parque General San Martín, Mendoza 5500,
Argentina
b
Facultad de Ciencias Veterinarias y Ambientales, Universidad Juan Agustín Maza. Av. Acceso Este, Lateral Sur 2245, Guaymallén, Mendoza 5519, Argentina

A R T I C LE I N FO A B S T R A C T

Keywords: A validation covering a 30-year period (1987–2016) of the relatively new CHIRPS precipitation dataset was
CHIRPS performed over the Central Andes of Argentina (CAA), a semi-arid region with complex topography and sparse
Remote sensing ground observations. Precipitation data from 57 rain gauges and several well-known continuous and categorical
Precipitation validation statistics were assessed to evaluate the performance of CHIRPS estimations. The study area was di-
Spatio-temporal validation
vided into two zones based on the timing of the rainy season maximum, in order to determine regional differ-
Argentina
ences in the characterization of precipitation patterns. The results of this study indicate that CHIRPS data re-
Andes
produce adequately several characteristics of precipitation along the study area, as the seasonal and interannual
variability and the spatial patterns of precipitation. CHIRPS dataset is able to capture the rainy season char-
acteristics over the CAA, considering the Mediterranean climate features over the Andes ranges and the mon-
soonal regime in the lowlands. CHIRPS achieves better results for the stations located in the region with summer
precipitation maximum, mostly located over Cuyo region (correlation = 0.86, bias = 11%, mean absolute
error = 15.3 mm). Despite the strong correlation of 0.82 over Northern Patagonia region, CHIRPS showed a
significant overestimation of the seasonal precipitation totals during the cold semester (April to September,
bias = 65.8%, mean absolute error = 34.7 mm). These systematic errors can be attributed to the poor perfor-
mance of CHIRPS in reproducing the precipitation features over the zones above 1000 m.a.s.l. One of the reasons
behind the observed differences can be attributed to the limited number of anchor stations used in the CHIRPS
calculation procedure, which highlights this study as an independent validation given the amount of meteor-
ological stations used.

1. Introduction perspective based on operational monitoring and a future perspective

based on reliable precipitation forecasts.
The knowledge of precipitation variability, both in time and space, To achieve these goals, until some decades ago, the use of conven-
is crucial for water management strategies, environmental monitoring, tional rain gauge networks have provided the main source of relatively
agricultural practices and climate studies. Precipitation is one of the accurate point measurements of precipitation (Feidas, 2010; Katsanos
main input to hydrological models and to most of the indices used for et al., 2016a). However, estimations from rain gauges over mountai-
flood and drought monitoring. The impacts of precipitation variability, nous areas with complex topography are often subjected to large un-
associated to floods and droughts, often led to adverse economic effects, certainties due to the lack of accessibility to the rain gauges –which
particularly in regions that rely on agriculture or hydropower genera- habitually have an uneven and sparse distribution-, the limited sam-
tion. Changes in precipitation patterns can have profound societal pling area of the gauges, and the maintenance costs of the network
consequences, directly affecting ecological systems, food security, dis- (Farooq Iqbal and Athar, 2018; Feidas, 2010; Zambrano-Bigiarini et al.,
aster management and human lifes (Farooq Iqbal and Athar, 2018; 2017). In South America, the available rainfall network has significant
Maidment et al., 2015; Thiemig et al., 2012). Thus, it is of paramount limitations in its infrastructure, maintenance, density and frequency of
importance to provide accurate precipitation information, considering a observations (Hobouchian et al., 2017). Particularly in Argentina, to
climatological perspective based on past information, a current obtain accurate precipitation estimations based on rain gauges is a

⁎
Corresponding author at: Instituto Argentino de Nivología, Glaciología y Ciencias Ambientales (CCT-Mendoza/CONICET), Av. Ruiz Leal s/n, Parque General San Martín, Mendoza
5500, Argentina.
E-mail address: jrivera@mendoza-conicet.gob.ar (J.A. Rivera).

https://doi.org/10.1016/j.atmosres.2018.06.023
Received 29 January 2018; Received in revised form 2 June 2018; Accepted 27 June 2018
Available online 02 July 2018
0169-8095/ © 2018 Elsevier B.V. All rights reserved.
J.A. Rivera et al. Atmospheric Research 213 (2018) 437–449

challenging task due to the extension of the country, the sparse dis- Unit (CRU) gauge analysis. Le and Pricope (2017) used the CHIRPS
tribution of the stations and the complex topography that includes the dataset into a standard hydrologic model for the simulation of
Andes, the most important mountain range in the Southern Hemi- streamflow over Western Kenya. A meteorological drought assessment
sphere. In turn, 40% of the Argentinean meteorological network was over Southeast Asia was performed by Guo et al. (2017) using the
lost during the 1970s, particularly after 1976 (Boulanger et al., 2010; CHIRPS estimations to calculate the standardized precipitation index
Tencer et al., 2011), and most of the rain gauges located along the (SPI, McKee et al., 1993). Katsanos et al. (2016a) analyzed precipitation
adjacencies of the Andes were dismantled (Viale and Nuñez, 2011). extremes based on CHIRPS over Cyprus, after a validation study that
To overcome these limitations, satellite-based estimates are widely allowed to identify a trend in the difference between rain gauges and
used to measure precipitation given that provides quasi-global cov- CHIRPS over time (Katsanos et al., 2016b). The first quasi-global eva-
erage, high resolution, frequent sampling and easy access (Tang et al., luation of CHIRPS dataset was performed by Beck et al. (2017), finding
2015). Satellite derived precipitation data can support the study of that CHIRPS presents a viable choice for tropical regions.
precipitation patterns at different temporal and spatial scales, and are a This dataset was recently considered for validation studies over
crucial tool for hydrological applications, water management and de- South America. Particularly, Funk et al. (2015a) evaluated CHIRPS
cision-making (Hobouchian et al., 2017; Mantas et al., 2015). The re- during the rainy season in areas of complex terrain as Colombia and
markable capabilities of satellite estimations demanded constant ad- Peru. They obtained that CHIRPS estimations were comparable to the
vancement and development of new algorithms and methods for further variability of the Global Precipitation Climatology Centre (GPCC) da-
improving the quality of precipitation estimates (Ahmadalipour et al., taset. Paredes-Trejo et al. (2016) performed a validation of CHIRPS
2017). Moreover, temporal and spatial resolutions are continuously over Venezuela, finding an acceptable overall performance but with a
improving owing to steady advances in sensor technology and new low skill of rain detection, which prevents its use for agricultural pur-
methods for merging various data sources (Thiemig et al., 2012). poses. Over Northeastern Brazil, CHIRPS arise as a promising tool for
Nowadays, several high-resolution products based on satellite pre- drought monitoring, performing better outside the semi-arid region,
cipitation estimates are freely available at an operational stage, dif- although rain detection is deficient over the studied domain (Paredes-
fering in terms of design objective, data sources, spatial resolution, Trejo et al., 2017). This product has been recommended for (quasi-)
spatial coverage, published temporal resolution, temporal span and real-time monitoring and hydrological applications over Chile
latency (Beck et al., 2017). However, the use of these databases for (Zambrano-Bigiarini et al., 2017). Nevertheless, Zambrano et al. (2017)
climate studies is only possible if precipitation estimations resemble the showed that the product should be calibrated to adjust to rainfall
spatial and temporal variability based on rain gauges observations. In especially over northern Chile.
this sense, validations of satellite-based precipitation products are ne- The Central Andes of Argentina (CAA) is a region where the inter-
cessary to ascertain the accuracy of precipitation estimates on various play between the complex topography and the atmospheric circulation
spatial and temporal scales and to establish the direct usability of these determines a wide range of precipitation features, from intense winter
products (Feidas, 2010). The validation activity has two aims: 1) to help orographic precipitation (Viale and Norte, 2009), extreme summer
technique developers to improve their algorithms and 2) to provide the precipitation events leading to the occurrence of landslides along the
potential users with a reliable error structure of the precipitation pro- Andes (Santos et al., 2015) and hailstorms over the lowlands (Biles and
ducts (Porcù et al., 2014). Performance of different satellite products, Cobos, 2007), to multi-annual severe drought events (Penalba and
typically evaluated comparing them to ground measurements, is dif- Rivera, 2016; Rivera et al., 2017). In view of the lack of observations
ferent in different geographical regions, surface conditions and season over the region, it is crucial to advance in the understanding of pre-
(Hobouchian et al., 2017; Rahmawati and Lubczynski, 2017), which cipitation variability at diverse spatial and temporal scales by using
often lead to substantial biases and stochastic errors that need to be available satellite-derived precipitation databases, considering the large
reduced or corrected before transforming the satellite estimations into negative impacts associated to precipitation variations on agriculture
operational tools with potential hydrological application (Tang et al., and water resources sectors.
2015; Thiemig et al., 2012). The aim of this study is to assess the accuracy of monthly CHIRPS
Scientists at the United States Geological Survey (USGS), working satellite-based estimations along the CAA, in order to quantify its
closely with collaborators at the University of California, Santa Barbara suitability to represent the spatial patterns of precipitation, its season-
Climate Hazards Group, have developed a quasi-global (50°S–50°N, ality and interannual variability. To our best knowledge, given the long-
180°E–180°W), 0.05° resolution, 1981 to near-present gridded pre- record −1981 to near present- of CHIRPS estimations, this will be the
cipitation time series: the Climate Hazards Group InfraRed validation of a satellite-based precipitation product over the longer
Precipitation with Stations (CHIRPS) data archive (Funk et al., 2014, period of time ever performed in the region. Thus, the outcomes of this
2015a). The CHIRPS dataset is a new land-only climatic database of study will be a baseline for the study of climate variability and change,
precipitation, made available since early 2014; it encompasses three the assessment of precipitation at small-catchment scale and the de-
diverse types of information: global climatologies, satellite estimates velopment of meteorological, agricultural and hydrological drought
and in situ observations (Katsanos et al., 2016b). The CHIRPS pre- monitoring tools over the CAA. This paper is organized as follows:
cipitation data set blends in more station data than other products and Section 2 presents the study area, the rain gauge data used, the CHIRPS
uses a high-resolution background climatology, providing better esti- dataset and the validation statistics. Section 3 presents the results of the
mates of precipitation means and variations, resulting in a better hy- precipitation intercomparison. The discussion of the main findings is
drologic state (Shukla et al., 2014). presented in Section 4, while the conclusions are summarized in Section
The CHIRPS dataset has been subjected to diverse evaluations 5.
worldwide, although the number of studies is still limited. Shukla et al.
(2014) used CHIRPS estimations to generate seasonal soil moisture 2. Study area and data sources
forecasts for agricultural drought prediction over the equatorial East
Africa. Over the same region, Vigaud et al. (2017) analyzed precipita- In this study, the CAA was defined as the Argentinean sector of the
tion composites based on CHIRPS to identify teleconnections with tro- region located between 30°-40°S and 67°-71°W (Fig. 1). The study area
pical Pacific surface temperatures and regional atmospheric circulation. spans part of two geographical regions of the country: Cuyo region (30°-
Maidment et al. (2015) quantified the recent observed trends in pre- 36°S, comprising the provinces of San Juan and Mendoza, where the
cipitation over Africa with 8 gridded gauge-only and satellite-based Andes have a mean elevation of 3500 m) and northern Patagonia region
datasets, including CHIRPS estimations. The authors observed that (36°S-40°S, comprising Neuquén province, where the mean elevation of
CHIRPS show relatively small deviations from the Climate Research the Andes decrease to 1500 m). The Andes ranges strongly affect the

438
J.A. Rivera et al. Atmospheric Research 213 (2018) 437–449

Table 1
Geographical characteristics and percentage of missing data of the selected rain
gauges for the 1987–2016 period.
Name Province Lat (°S) Lon (°W) Elevation Source Missing
(masl) data (%)

San Juan San Juan 31.57 68.42 598 SMN 2.22

San Juan INTA San Juan 31.62 68.53 603 SMN 3.88
Jachal San Juan 30.23 68.75 1175 SMN 3.88
San Martín Mendoza 33.08 68.42 653 SMN 0.00
Mendoza Aero Mendoza 32.83 68.78 704 SMN 0.00
Mendoza Mendoza 32.88 68.85 827 SMN 0.00
Observato-
rio
Malargüe Mendoza 35.50 69.58 1425 SMN 0.00
San Rafael Mendoza 34.58 68.40 748 SMN 0.00
La Angostura Mendoza 35.09 68.87 1302 SSRH 0.00
La Jaula Mendoza 34.67 69.32 1457 SSRH 0.00
Rama Caída Mendoza 34.67 68.38 714 SSRH 1.67
El Nihuil Mendoza 35.03 68.67 1309 SSRH 0.00
Villa Atuel Mendoza 34.82 67.92 519 SSRH 0.00
Capitán Mendoza 34.58 68.45 820 SSRH 3.33
Montoya
Las Salinas Mendoza 34.93 68.81 1320 SSRH 5.56
Puesto Canales Mendoza 34.67 68.89 1574 SSRH 0.00
Puesto Mendoza 34.68 67.84 522 SSRH 0.00
Carmona
Arroyo Hondo Mendoza 34.48 69.28 1900 SSRH 0.00
Las Aucas Mendoza 34.70 69.54 1800 SSRH 1.67
Las Malvinas Mendoza 34.94 68.24 602 SSRH 0.00
Los Mayines Mendoza 35.66 70.20 1663 SSRH 2.50
Bardas Blancas Mendoza 35.87 69.81 1445 SSRH 1.67
Arroyo La Mendoza 35.92 69.99 1550 SSRH 1.67
Vaina
Fig. 1. Location of the study area with the main topographic features and the Puesto Las Mendoza 35.12 66.85 388 SSRH 0.00
rain gauges used for the validation of satellite-based estimates. Moras
Cacheuta Mendoza 33.01 69.12 1250 SSRH 6.39
Guido Mendoza 32.92 69.24 1408 SSRH 0.28
regional precipitation patterns over these regions, through interactions Valle de Uco Mendoza 33.78 69.27 1199 SSRH 1.67
with the continental atmospheric circulation and the incursion of moist Pincheira Mendoza 35.52 69.81 1775 SSRH 7.50
air masses from the Pacific Ocean (Rivera et al., 2017). Along the study San Rafael Mendoza 34.61 68.32 681 SSRH 0.28
Las Vertientes Mendoza 34.42 68.59 990 SSRH 1.11
area, the Andes act as a permanent barrier to the humid air masses from
Juncalito Mendoza 34.74 69.21 1593 SSRH 4.72
the mid-latitude South Pacific Ocean and the baroclinic precipitation Puesto Morales Mendoza 34.60 68.87 1455 SSRH 0.83
systems coming from the west. The climate at high elevations has a Polvaredas Mendoza 32.79 69.65 2250 SSRH 0.28
Mediterranean regime with higher precipitations during the cold season Potrerillos Mendoza 32.96 69.20 1427 SSRH 1.11
Puesto Mendoza 34.23 69.12 1529 SSRH 3.05
(April to September) and dry warm seasons (October to March), in re-
Papagayos
sponse to the seasonal displacement of the South Eastern Pacific High La Remonta Mendoza 33.71 69.29 1360 SSRH 0.28
(Falvey and Garreaud, 2007). Over this area precipitation is mainly San Alberto Mendoza 32.47 69.41 2180 SSRH 14.44
generated by the passage of cold fronts moving eastward from the Pa- Uspallata Mendoza 32.59 69.34 1890 SSRH 3.33
cific (Garreaud, 2009). Due to the strong rain shadow effect, climate Buta Ranquil Neuquén 37.07 69.75 816 SSRH 0.28
Paso de Indios Neuquén 38.53 69.41 498 SSRH 5.56
east of the Andes is arid to semi-arid, where convective warm season
El Cholar Neuquén 37.44 70.65 1230 SSRH 5.83
rainfalls favored by moist air masses from the Amazon and Atlantic Chos Malal Neuquén 37.37 70.27 856 SSRH 2.78
basins play a relevant role (Schwerdtfeger, 1976). There are very in- Andacollo Neuquén 37.18 70.68 1011 SSRH 12.22
tense storms east of the Andes during the warm season, capable to Junín de los Neuquén 40.05 71.10 750 SSRH 9.44
develop large hail over the study area (Biles and Cobos, 2007). Andes
Vilu Mallín Neuquén 37.46 70.76 1065 SSRH 1.67
The CAA is the major wine producer region in Argentina, and the Cajón Neuquén 36.96 70.39 1400 SSRH 4.17
agro-industrial activities in CAA depend largely on grape production, Curileuvú
an activity only possible through irrigation. Dams and reservoirs collect El Alamito Neuquén 37.26 70.42 1032 SSRH 5.28
the snowmelt contribution to the streamflows during the warm season. Las Ovejas Neuquén 36.98 70.75 1267 SSRH 2.78
Varvarco Neuquén 36.86 70.68 1180 SSRH 3.61
Along Cuyo region, 95% of its 2.5 million inhabitants are distributed
El Huecu Neuquén 37.65 70.58 1212 SSRH 1.67
just within the 4% of the territory, highlighting the vulnerability of the Tricao Malal Neuquén 37.04 70.32 1350 SSRH 0.00
region to periods of water shortages. Regarding Northern Patagonia, the Los Miches Neuquén 37.21 70.82 1219 SSRH 0.83
rivers which are born in the higher elevations of the Andes, fed by Chochoy Mallín Neuquén 37.36 70.79 1066 SSRH 1.94
snowmelt and rainfall, play an important role in the development of the Chorriaca Neuquén 37.94 70.10 1100 SSRH 7.22
Pichi Neuquén Neuquén 36.63 70.80 1350 SSRH 0.83
region, with several hydropower plants providing a significant part of Auquinco Neuquén 37.32 69.97 1520 SSRH 2.22
the electric power in Argentina (Seoane et al., 2005). Neuquén Neuquén 38.95 68.13 271 SMN 0.00

2.1. Rain gauge observations

(Subsecretaría de Recursos Hídricos, SSRH; http://bdhi.
hidricosargentina.gob.ar/) databases. A filtering method was used to
Initially, monthly precipitation records from 103 rain gauges were
select the time series with < 15% of missing data, with a data length
collected from the National Weather Service (Servicio Meteorológico
requirement of at least 30 years, comprising ideally the most recent
Nacional, SMN) and the Water Resources Agency of Argentina

439
J.A. Rivera et al. Atmospheric Research 213 (2018) 437–449

period of records. After these procedures, the final database consisted during the last part of 1980s, 6 stations during the 1990s and 8 between
on monthly precipitation observations from 57 stations (see Fig. 1) with 2000 and 2016 (Fig. 2). Starting with the national datasets, the list of
measurements over the 1987–2016 period. In addition, the names and anchor stations to create CHIRPS were increased adding regional and
relevant particulars of the selected rain gauges are summarized in global sources (Funk et al., 2014). Nevertheless, the incorporation of a
Table 1. As can be observed in Fig. 1, the region has scarcely available valuable national source as the SSRH was discarded, perhaps given the
meteorological data for a long-term climatic assessment, particularly hydrological focus of the rain gauges information of this dataset or the
considering the stations from the SMN (Table 1). Precipitation time lack of knowledge about its existence and availability. As can be ob-
series were subjected to quality control procedures, considering the served in Table 1, just 9 of the 57 rain gauges used for this validation
techniques applied by González (2013) to the time series over the study where provided by the SMN, which indicates that the incorporation of
area. Moreover, we analyzed the spells of months without precipitation, data from the SSRH could result in an improvement of CHIRPS esti-
removing from the assessment the periods larger than 5 months, a mations over the study area and, particularly, over Argentina. More-
threshold identified by Llano and Penalba (2011) over the study area over, just one of the anchor stations was located in Neuquén province,
based on high-quality precipitation time series. Reasons behind the with two of the anchor stations having a precipitation maximum during
occurrence of these long spells can be attributed due to the in- the cold semester (Neuquén and Malargüe stations). Considering this
accessibility to the stations mainly during winter months. Extreme limitation, the validation of CHIRPS precipitation dataset over the CAA
precipitation totals were also analyzed considering the spatial dis- it can be considered as an independent validation.
tribution of precipitation in the nearest rain gauges, in the cases were
monthly totals exceeded four times the standard deviation above the
3. Methodology
mean. This criteria has been previously used by Penalba et al. (2014)
considering daily precipitation totals and is in line with the threshold
3.1. Continuous validation statistics
level used by Paredes-Trejo et al. (2016, 2017) for monthly data. Ad-
ditionally, homogeneity control using the Standard Normal Homo-
A point-to-pixel analysis was performed to compare the time series
geneity Test (Alexandersson, 1986) for a confidence level of 95% al-
of rain gauges observations to the corresponding CHIRPS pixel. This
lowed to identify inhomogeneities in just 6 of the 57 precipitation time
comparison allows capturing the small scale variability of monthly
series. These inhomogeneities were not corrected given that occurred
precipitation totals (Thiemig et al., 2012), largely subjected to its sea-
during the first or the last year of record. Missing data were replaced by
sonality and the topography of the study area. Therefore, we selected
applying linear regressions with neighboring stations, only for reference
the CHIRPS estimations for the 57 grid points conrresponding to the
stations that explain > 80% (R2 > 0.8) of the temporal behavior of
location of the rain gauges. An interpolation of the observed pre-
precipitation. Even when the final set of precipitation time series still
cipitation would involve large uncertainties given the lack of a high-
have some missing months –either because of removal of suspicious
density rain gauges database to reproduce adequately the precipitation
records or to the lack of close reference stations to fill missing gaps- this
gradients along the complex terrain, although this alternative was
aspect is not expected to affect the comparability of the results or the
previously carried out over the study area (Hobouchian et al., 2017)
calculation of the metrics for CHIRPS comparison.
and other regions of the world (Dinku et al., 2008; Feidas, 2010).
To evaluate the performance of CHIRPS in estimating the amount of
2.2. CHIRPS precipitation estimates
the precipitation we used the following comparison statistics: the
Pearson correlation coefficient (PCC), the mean absolute error (MAE),
The CHIRPS monthly precipitation product was obtained through
the Nash-Sutcliffe efficiency (NSE) and the percent bias (PB) (Table 2).
the Climate Hazards Group (CHG) of the University of California at
The PCC measures the linear relationship strength between the satellite
Santa Barbara (UCSB) webpage (http://chg.geog.ucsb.edu/data/
estimations and the rain gauges observations, bounded by −1 and 1
chirps/index.html) for the years 1987–2016, which overlaps the
with an optimal value of 1. The MAE provides information on the
period of ground-based precipitation data. This dataset is updated at
average magnitude of error estimations, considering both systematic
near-real time, has quasi-global coverage (land only, 50°S-50°N) with a
and random errors. Several studies are replacing the widely used root
spatial resolution of 0.05° (approximately 5 Km) and several temporal
mean squared error (RMSE) by the MAE because the errors are unlikely
scales (monthly, decadal, pentadal or daily time steps). The CHIRPS
to be unbiased or to follow a normal distribution (Beck et al., 2017;
dataset was produced blending precipitation estimates based on in-
Willmott et al., 2017). The perfect score for this statistic is 0. The NSE
frared cold cloud duration (CCD) observations calibrated using Tropical
(Nash and Sutcliffe, 1970) determines the relative magnitude of the
Rainfall Measuring Mission (TRMM) Multi-Satellite Precipitation
variance of the residuals compared to the variance of the observed
Analysis (MSPA) with in-situ station data from a variety of sources in-
values of precipitation. This statistic vary from minus infinity to 1,
cluding national and regional meteorological services (see Funk et al.,
being negative in cases of poor precipitation estimation and an optimal
2014, 2015a, 2015b for details). The incorporation of station data also
value equal to 1, which would indicate that the estimated values mat-
helps to correct for estimates that often underestimate the intensity of
ched observed precipitation exactly. The PB measures the average
precipitation events (Le and Pricope, 2017). This relatively new product
tendency of the estimated precipitation to be larger or smaller than the
was designed for drought monitoring in places with complex topo-
observed precipitation, with an optimal value of 0. Positive values in-
graphy, changing observation networks and deep convective pre-
dicate overestimation bias, whereas negative values indicate under-
cipitation systems (Funk et al., 2015b), features that are present in the
estimation bias.
study area. Moreover, the spatial resolution of CHIRPS is higher than
other satellite-based global precipitation datasets, making it favorable
to analyze precipitation variations at small basin scales. 3.2. Systematic and random error components
The temporal evolution of the anchor stations used in the creation of
CHIRPS over the study area is presented in Fig. 2 for the 1987–2016 Characterizing satellite precipitation errors and their random and
period. Between 2 and 12 anchor stations were used, all of them pro- systematic components is essential to develop bias reduction techni-
vided by the SMN and obtained through either the Global Historical ques, and thus to improve precipitation retrieval algorithms, and for
Climate Network (GHCN) or the Global Summary of the Day (GSOD), as many operational applications (Maggioni et al., 2016). According to
can be observed in ftp://ftp.chg.ucsb.edu/pub/org/chg/products/ Willmott (1981), the error in the numerical weather prediction models
CHIRPS-2.0/diagnostics/monthly_station_data/. This temporal evolu- can be separated into systematic and random error components of the
tion shows a large decadal variability, with an average of 11 stations mean squared error (MSE):

440
J.A. Rivera et al. Atmospheric Research 213 (2018) 437–449

Fig. 2. Temporal evolution of the number of anchor stations used for the estimation of CHIRPS dataset within the study area for the period 1987–2016.

Table 2 Table 4
Statistical measures of performance used for analysis based on continuous Statistical measures of performance used for analysis based on categorical
metrics, where: G = rain gauge observations, G = average rain gauge ob- metrics, where: A = number of hits, B = number of false alarms, C = number of
servations, S = CHIRPS estimations, S = average CHIRPS estimations and misses, D = number of correct negatives, N = number of data pairs and Ar
N = number of data pairs. stands for hits that could occur by chance.
Statistic Ecuation Statistic Equation

Pearson correlation coefficient (PCC) ∑ (G − G )(S − S ) Probability of detection (POD) POD =

A
r= A+C
(G − G )2 (S − S )2
False alarm rate (FAR) FAR =
B
Mean absolute error (MAE) MAE =
1
∑ (|S − G|) A+B
N
Equitable threat score (ETS) ETS =
A − Ar
where Ar =
(A + B )(A + C )
Nash-sutcliffe efficiency (NSE) ∑ (S − G )2 A + B + C − Ar N
NSE = 1 −
∑ (G − G )2 Hanssen-Kuipers discriminant A B
HK = −
Percent bias (PB) ∑ (S − G ) (HK) A+C B+D
PB = 100
∑G
Heidke skill ccore (HSS) HSS =
2(AD − BC )
(A + C )(C + D) + (A + B )(B + D)
Frequency bias index (FBI) FBI =
A+B

∑ (S − G )2
A+C
∑ (S − G )2
MSE = MSESyst + MSERand and =
N N
∑ (S − S )2 The validation statistics include the probability of detection (POD),
+
N the false alarm rate (FAR), the equitable threat score (ETS), the
Hanssen-Kuipers discriminant (HK), the Heidke skill score (HSS) and
where Ŝ is defined as the frequency bias index (FBI) (Table 4). The POD gives the fraction of
S = aS + b precipitation occurrences that were correctly detected; it ranges from 0
to a perfect score of 1. The FAR gives the fraction of events for which
where a and b are parameters (slope and intercept, respectively) to be CHIRPS detected precipitation but was not observed; ranging from 0 to
calibrated. The systematic error is defined as the part of error to which 1 with a perfect score of 0. The ETS measures the fraction of observed
a linear function can be fitted (Habib et al., 2009). Following and/or estimated events that are correctly predicted, adjusted by the
AghaKouchak et al. (2012), the systematic and random components of frequency of hits that would be expected to occur simply by random
error were expressed as MSESyst/MSE × 100 and MSERand/MSE × 100, chance. The range of ETS is −1/3 to 1, with a perfect score of 1, 0 for
respectively. no skill and negative values indicate that chance estimation of the event
should be preferred (Zambrano-Bigiarini et al., 2017). The HK shows
3.3. Categorical validation statistics how well the satellite estimates discriminate between precipitation and
no-precipitation events; ranging from −1 to 1, with a perfect score of 1
Complementary, several categorical validation statistics were used and no skill for 0. The HSS, which ranges from minus infinity to 1,
to assess CHIRPS rain-detection capabilities. The statistics were derived measures the accuracy of the estimates accounting for matches due to
from a contingency table involving four event combinations: hits (A) random chance. An HSS < 0 indicates that random chance is better
–months when both rain gauge and CHIRPS detect precipitation; false than the satellite product, an HSS of 0 means the product has no skill,
alarms (B) –months when CHIRPS detects precipitation and rain gauge and an HSS of 1 indicates a perfect estimation of precipitation by the
not; misses (C) –months when rain gauge recorded precipitation and product (Diem et al., 2014). Finally, the FBI reveals systematic differ-
CHIRPS not; and correct negatives (D) –both rain gauge and CHIRPS ences between precipitation events frequency in raingauge observations
detect no precipitation (see Table 3). Following the assessments of Toté and CHIRPS-based precipitation estimates. It can indicate whether
et al. (2015) and Paredes-Trejo et al. (2017), we used a precipitation there is a tendency to underestimate (FBI < 1) or overestimate
threshold of 5 mm. (FBI > 1) precipitation events, ranging from 0 to infinity with a perfect
score of 1.
Table 3
Contingency table for comparing rain gauge observations and satellite-based
4. Results
precipitation estimates (precipitation threshold used is 5 mm).
Gauge ≥ threshold Gauge < threshold 4.1. Annual precipitation climatology over the Central Andes of Argentina
Satellite ≥ threshold A B
Satellite < threshold C D As a first step, the CHIRPS precipitation estimates were compared
with the rain gauges observations considering the spatial distribution of

441
J.A. Rivera et al. Atmospheric Research 213 (2018) 437–449

Fig. 3. Mean annual precipitation over the CAA: a) based on rain gauges observations and b) based on CHIRPS estimations.

the annual average precipitation totals over the study area. This as- according to the seasonal precipitation features, an assessment shown
sessment is shown in Fig. 3, considering the climatology for the in Fig. 4 a. The spatial distribution of the rain gauges with higher
1987–2016 period. Regarding Northern Patagonia, observed pre- precipitation during the warm season is restricted to the region east of
cipitation ranges from over 700 mm at the high elevations of the con- 69.5° W approximately. Conversely, rain gauges west of this longitude
tinental divide, between 37° and 40° S, to < 300 mm east of the Andes. show higher precipitation values during the cold season (Fig. 4 a).
At these latitudes, there is a strong precipitation gradient from west to CHIRPS estimations reproduce this spatial pattern adequately, therefore
east due to the rain-shadow effect of the Andes. This spatial pattern of highlighting its ability to discriminate the seasonal climatic features of
annual precipitation is properly reproduced by the CHIRPS estimations, precipitation over the CAA. Just in 6 of the 57 locations the CHIRPS
although with an overestimation of the mean annual totals close to the estimates of the annual cycle of precipitation differed from the ob-
Andes ranges and an underestimation east of the Andes, between 68° to servations, with 5 pixels showing higher precipitations during the cold
70° S. This feature indicates that the precipitation gradient associated to season (corresponding to the locations of Guido, La Jaula, Arroyo
the rain-shadow effect is magnified considering the satellite estima- Hondo, Juncalito and Puesto Papagayos) and one location showing
tions. In contrast to the observed 500–800 mm close to the higher higher precipitations during the warm season (Paso de Indios), with
elevations of the Andes, CHIRPS shows annual precipitation totals be- observations showing an opposite behavior (Fig. 4 a). These stations are
tween 900 and 1300 mm (Fig. 3), while precipitation tends to be un- located in the geographical transition between the influence of the
derestimated by 50 mm approximately in the lowlands of Neuquén frontal systems during the cold season and the convective systems
province. Considering Cuyo region, the observations show annual pre- during the warm season. Hence, in this transitional region it is expected
cipitation totals below 150 mm over San Juan province (north of 32° S), a reduction in the seasonality of precipitation.
while higher values are distributed in the central-eastern portion of To exemplify the seasonal precipitation variability over the region,
Mendoza province (approximately 34° - 35° S). The mean annual Fig. 4 b shows the boxplots of monthly precipitation for six selected
CHIRPS precipitation estimates over this region show a similar spatial locations. There is a good agreement between rain gauges observations
pattern than the observations, although some regional features are and CHIRPS estimations considering Mendoza Aero, Rama Caída and
poorly captured by CHIRPS, as the relative maximum over Uco Valley Chochoy Mallín time series. The seasonality of precipitation is correctly
(around 34° S; 69.5° W) or the annual totals over the region east of 68° reproduced by satellite estimations, although precipitation values
W (Fig. 3). Moreover, CHIRPS estimations tend to overestimate pre- during the warm semester are slightly underestimated at Mendoza Aero
cipitation totals over the southwestern portion of Mendoza, close to the –especially during January, February and March- and overestimated at
Andes (35.5° S; 70° W), increasing the precipitation gradient as in the Rama Caída –particularly during November, December, January and
case of Northern Patagonia. March- (Fig. 4 b). For these 3 sites it can be observed that the variability
of precipitation during the rainy season –represented by the inter-
quartile range- is larger than for the dry season, considering both ob-
4.2. Annual cycle of precipitation
servations and CHIRPS estimations, although CHIRPS precipitation es-
timates exhibit less variability than the observations particularly during
As previously described in Section 2, the annual cycle of pre-
the dry season. Additionally, we analyzed the performance of CHIRPS
cipitation shows two distinct regional features related to the atmo-
for 3 locations over the transitional region (Guido, La Jaula and Au-
spheric circulation over the region and its interaction with the Andes
quinco, Fig. 4 b). As shown in Fig. 4 a, CHIRPS precipitation estimations
range: a Mediterranean regime close to the higher elevations of the
for Guido and La Jaula show that the larger precipitation totals are in
Andes, where the average precipitation is higher in the cold season than
the cold season, while observations indicate the opposite, although the
in the warm season, and a monsoonal regime over the low lands
annual cycle in precipitation is not clearly monsoonal. CHIRPS esti-
dominated by convective warm season rainfalls and a relatively dry
mations show a double maximum over Guido station, with a summer
cold season. Fig. 4 shows some of the main features of the annual cycle
peak in February and a winter peak in May and August, while
of precipitation over the CAA. Firstly we classified the rain gauges

442
J.A. Rivera et al. Atmospheric Research 213 (2018) 437–449

Fig. 4. a) Spatial distribution of the season of higher precipitation amounts for the 57 rain gauges and its correspondent CHIRPS pixel. b) Boxplot of the mean
monthly precipitation for 6 selected stations along the CAA. Each boxplot shows the median and first and third quartiles, while the whiskers extend to the data values
that are 1.5 times the interquartile range above or below the quartiles. Notice that the scale of the monthly precipitation varies among the selected stations.

observations show higher precipitations during the warm semester. (PCC = 0.86), and similar results can be found considering the stations
Regarding La Jaula station, the seasonal variability of observed pre- with higher precipitation during the warm season (PCC = 0.85) or the
cipitation is not clearly defined. Monthly precipitation totals based on stations with higher precipitation during the cold season (PCC = 0.89).
rain gauges exhibit larger variability than the estimations based on Nevertheless, CHIRPS estimations tends to show an overestimation of
CHIRPS, with the exception of the month of May. A double maximum precipitation totals mostly over North Patagonia and southwestern
during the cold season is estimated by CHIRPS during May and Sep- Cuyo –the stations with rainy season during the cold semester-, while
tember, a feature that is also evident at Auquinco station (Fig. 4 b), with there is a slight underestimation considering the locations with the
an overestimation of precipitation totals especially during the cold rainy season during the warm semester –most of Cuyo region- (Fig. 5 a).
season. Regarding the dry season estimations, the linear regression considering
To further illustrate the differences in the representation of the the 57 locations indicates a poor agreement between observations and
annual cycle of precipitation, Fig. 5 shows the scatterplots of the ob- CHIRPS estimations (PCC = 0.27), which indicates that CHIRPS data
served precipitation and the CHIRPS estimations averaged over the misrepresent the precipitation totals during this season. This is parti-
rainy and dry season of each location. Considering the 57 locations, the cularly evident considering the stations with dry season during the
rainy season estimations show a good agreement with the observations warm semester (PCC = 0.15), while most of the stations located over

Fig. 5. a) Rainy season and b) dry season precipitation scatter plots comparing the 57 rain gauges data with the corresponding grids of CHIRPS estimations over the
1987–2016 period, with linear regression fits (black lines) for the stations showing higher precipitation totals during the cold season (squares) and the stations
showing higher precipitation totals during the warm season (circles). The grey dashed line represents a 1:1 relation.

443
J.A. Rivera et al. Atmospheric Research 213 (2018) 437–449

Fig. 6. Spatial distribution of the continuous statistics used to measure the agreement between observations and CHIRPS estimations: a) Pearson correlation
coefficient, b) mean absolute error, c) Nash-Sutcliffe efficiency and d) percent bias.

Fig. 7. Systematic and random error components (%) over the CAA for CHIRPS monthly precipitation estimations.

Cuyo region show an acceptable skill (PCC = 0.66) (Fig. 5 b). Once in the lowlands of Cuyo region, with values higher than r = 0.6 (Fig. 6
again, CHIRPS estimations show an overestimation of precipitation a). Low correlations are observed over the mountainous region of
over North Patagonia, with a slight underestimation of dry season Mendoza, particularly in the central portion of the province where the
precipitation over most of Cuyo region. annual cycle of precipitation was poorly captured by CHIRPS. Re-
garding the spatial behavior of the MAE, Fig. 6 b) indicates that higher
errors are located over the northwest of North Patagonia and the
4.3. Spatial distribution of continuous validation statistics
southwestern tip of Cuyo region, with values ranging from 30 to over
50 mm. The NSE is above 0.4 in a large portion of Cuyo region, with 16
The spatial variations of the continuous statistics used for the vali-
locations with values higher than 0.65, a threshold that indicates that
dation of CHIRPS estimations across the CAA are shown in Fig. 6. The
CHIRPS estimations are in agreement with observations (Förster et al.,
values of PCC show a better agreement between observations and sa-
2016). The NSE values decrease over North Patagonia, even with
tellite estimations close to the Andean portion of North Patagonia and

444
J.A. Rivera et al. Atmospheric Research 213 (2018) 437–449

negative values of efficiency (Fig. 6 c), in coincidence with the region the FBI, it can be observed an overestimation of precipitation totals in
where larger MAE are observed. The spatial distribution of the PB in- all the thresholds, considering all the months and the cold season
dicates that CHIRPS estimations over most of North Patagonia region (Fig. 8f). For the thresholds between 5 and 20 mm, the warm season
tend to overestimate precipitation values, with bias larger than 20% shows the largest overestimations, while the cold season shows the
(Fig. 6 d). Over Cuyo region the spatial pattern is heterogeneous but lower overestimations. For thresholds of 50 and 100 mm there is an
with a large number of locations showing an underestimation of pre- opposite seasonal behavior in the FBI, with large overestimation during
cipitation values over the center of Mendoza province, a result pre- the cold season and underestimation of precipitation totals during the
viously shown in Fig. 5. The wet bias over the Andean portion of warm season.
Neuquén province is in line with the poor performances considering the
MAE and NSE, even when the PCC between observations and CHIRPS 4.6. Precipitation performance at different elevation
estimations is above 0.7. Moreover, this result is supported by Fig. 5
considering both the rainy and dry seasons. The MAE, the NSE and the Given the complex topography of the CAA, this study assessed the
PB show that the CHIRPS estimations provide a good approximation to dependence of CHIRPS's performance on elevation. In order to achieve
observed precipitation over northern Cuyo region (north of 34° S), even this, the rain gauges were classified in five groups considering their
when the PCC have low values closer to the Andes. Conversely, the elevation: 0–700 m (10 stations), 700–1000 m (9 stations),
Andean region of North Patagonia has large PCC but the overestimation 1000–1300 m (13 stations), 1300–1500 m (13 stations), and > 1500 m
of precipitation is translated to high MAE and low NSE. (12 stations) (see Table 1). The results for the continuous validation
statistics are shown in Table 5. For the PCC, NS and PB, stations located
4.4. Systematic and random errors between 700 and 1000 m.a.s.l. showed the best performance, while the
MAE shows lower values for the stations located at elevations lower
Fig. 7 shows the spatial distribution of the systematic and random than 700 m. As the elevation increase, the statistics indicate that the
components of error over the study area. Considering the systematic performance of CHIRPS tends to decrease. For example, the lower PCC
component of CHIRPS estimations, the higher values are observed over and NS are observed for the stations located above 1500 m (Table 5).
the region that showed large positive PB, high MAE and low NSE (see Nevertheless, the higher values in MAE and PB are observed for the
Fig. 6), located south of 35° S and west of 69° W (Fig. 7). As observed in stations located between 1000 and 1300 m. In order to complement
Fig. 4, the overestimation of precipitation over this region, especially in these results, we plotted the categorical POD, FAR, HSS and FBI sta-
the winter, and the misrepresentation of the dry season totals can be tistics for the five selected elevation ranges and different thresholds to
factors that contribute to the systematic error. The areas with low identify precipitation (Fig. 9). The best performance considering the
systematic error can be associated with the blending procedure using POD is observed in the stations located between 1000 and 1300 m, with
rain gauges measurements, given that the spatial pattern resembles the values between 0.8 and 1 (Fig. 9a). This behavior resembles the result
distribution of the anchor stations used for CHIRPS estimations (not obtained for the POD of the winter season (Fig. 8), which is typically
shown). Over this region, the systematic component is less than the the rainy season over the stations located at this altitude. Higher dif-
random error component (Fig. 7) and corresponds to the areas with low ferences in POD for each altitude range are observed for thresholds
MAE, PB and high NSE (see Fig. 6). higher than 50 mm, with the worst performance observed for pre-
cipitation higher than 100 mm in the stations located between 0 and
4.5. Performance based on categorical validation statistics 700 m. Nevertheless, the FAR for this group of stations shows the best
performance for all the selected thresholds (Fig. 9b). Higher FAR values
To measure algorithm performance for different rain rates it is are observed for the stations located at higher altitudes. The HSS shows
useful to plot the categorical scores as a function of an increasing that the best performance is observed for the stations located between 0
precipitation threshold (Ebert, 2007). For this purpose, multiple and 1000 m, particularly for thresholds of 10 and 20 mm (Fig. 9c). The
monthly precipitation thresholds have been considered to calculate the group of stations located between 1000 and 1300 m have the best
statistics: 1, 5, 10, 20, 50 and 100 mm. Fig. 8 shows the results of the performance for precipitation higher than 100 mm. Finally, the FBI
performance based on the categorical statistics. The POD (Fig. 8a) shows precipitation overestimation for the stations located over 1000 m
ranges between 0.8 and 1 for low precipitation values (1, 5 and 10 mm for all the selected thresholds (Fig. 9d). Underestimation of precipita-
thresholds) considering all the months and both warm and cold seasons. tion totals is observed for the stations located below 1000 m, con-
This good performance is observed also for large precipitation values sidering a threshold higher than 50 mm. This result is in line with the
during the cold season, with stabilization around POD = 0.75; never- findings observed in Fig. 8f, given that the rainy season of the stations
theless, during the warm season the performance for large precipitation located at lower elevations is recorded during summer. Additionally,
totals shows POD values lower than 0.1 for the 100 mm threshold large positive FBI is observed for the stations located between 1000 and
(Fig. 8a). The behavior of the FAR shows an increase as the precipita- 1300 m, particularly for high precipitation thresholds, a result that is
tion threshold increases, with larger values for the warm season in mainly associated to winter precipitation totals.
comparison to the cold season, particularly for the months with pre-
cipitation larger than 100 mm (Fig. 8b). In this sense, 80% of the pre- 4.7. Temporal evolution of regional precipitation
cipitation detections above 100 mm during the warm season were in-
correct, with most of these events occurring over Cuyo region. The ETS In order to illustrate the temporal behavior of observed and esti-
shows that the maximum detection skill during the warm season was mated precipitation, the time series of the 57 locations were averaged
achieved for a threshold of 20 mm, with a value of 0.23 (Fig. 8c), with considering its rainy season (see Fig. 4a for the spatial distribution).
almost no skill for precipitation exceeding 100 mm. Considering all the This allowed to obtain two regionally-averaged time series comprising
months, precipitation thresholds exceeding 20 mm shows the best de- the stations that have higher precipitation amounts during the warm
tection skill, with a value closer to 0.3. The best performance is ob- and cold seasons. Even when this clustering seems arbitrary, it is phy-
served during the cold season for monthly precipitation higher than sically meaningful given that separates one of the main features of
20 mm (ETS ~ 0.4). A similar behavior is observed for the HSS, al- precipitation over the CAA. Fig. 10 show the comparison of the re-
though with larger values for this categorical metric (Fig. 8d). Perfor- gionally-averaged time series of observed and estimated precipitation,
mance evaluated through the HK score indicates a low skill for all the together with the difference between the time series for each region.
thresholds, with negative values for higher monthly precipitation totals, Precipitation variability at the stations with warm season maximum
i.e. performance inferior to a random estimation (Fig. 8e). Regarding shows that the peaks in precipitation are not appropriately represented

445
J.A. Rivera et al. Atmospheric Research 213 (2018) 437–449

Fig. 8. Categorical verification of monthly CHIRPS precipitation estimates for the complete period and during the warm and cold seasons over the CAA as a function
of the threshold value chosen to separate the precipitation/no-precipitation events: a) probability of detection, b) false alarm rate, c) equitable threat score, d)
Hanssen-Kuipers discriminant, e) Heidke skill score and f) frequency bias index.

Table 5 precipitation peak during the cold season, Fig. 10b shows that the an-
Performance of the continuous validation statistics as a function of different nual cycle is accurately captured by CHIRPS estimations (PCC = 0.82),
elevation. even in the years where a double maximum in precipitation was ob-
Elevation categories (m.a.s.l.) served (see for example the years 1999 and 2001; Fig. 10b). However,
the difference between observations and estimations shows a wet bias
Continuous metrics 0–700 700–1000 1000–1300 1300–1500 > 1500 that is mostly evident during the cold season, with values that can be
PCC 0.708 0.744 0.715 0.606 0.470
larger than 100 mm especially during the months of May of 1992, 1993,
MAE 14.464 16.008 33.653 26.077 24.353
NS 0.466 0.520 0.133 −0.149 −0.377 1999, 2001, 2005, 2012 and 2016. The large values of precipitation
PB −19.560 6.678 64.477 30.158 34.640 estimated by CHIRPS during the month of May are in line with the
extremes of Fig. 4b in La Jaula and Auquinco stations during this par-
ticular month (see the whisker above the quartiles). Just in few months
by CHIRPS, with an underestimation of the precipitation amounts the CHIRPS estimations showed a dry bias, a result that is in line with
mostly during the warm season (Fig. 10a). It was observed that in the PB statistic for this region considering the complete period of
several years the magnitude of this dry bias was higher than 50 mm analysis (PB = 65.8%).
(1994, 1997, 1998, 2007 and 2014), representing almost half of the Considering Fig. 10, we focused mainly on the assessment of pre-
precipitation accumulated during the summer season over the region. cipitation bias during the rainy season, which is usually translated to an
Nevertheless, there is a good agreement between both regional ob- annual precipitation bias as observed in Figs. 3 and 5. Nevertheless,
servations and CHIRPS estimations (PCC = 0.86), with a PB of −11% during the dry season CHIRPS estimations show precipitation occur-
and a MAE of 15.3 mm. Considering the time series of the stations with rences in most of the months, with just few cases of zero precipitation

446
J.A. Rivera et al. Atmospheric Research 213 (2018) 437–449

Fig. 9. Categorical metrics performance at different elevations as a function of the threshold value chosen to separate the precipitation/no-precipitation events: a)
probability of detection, b) false alarm rate, c) Heidke skill score and d) frequency bias index.

during the 30 years period. This characteristic arise from a screening estimated precipitation can be associated to the number of anchor
procedure developed to remove “false zeros” in the CHIRPS estima- stations used in the creation of CHIRPS, which was shown in section 2
tions, given that national and global precipitation datasets have a as a deficient aspect in the generation of the product and highlights this
substantial number of values that have been incorrectly reported as assessment as an independent validation. Another source of uncertainty
zeros (Funk et al., 2014). Similar results have been previously reported was the duplication of data from different sources. We observed that
over Cyprus (Katsanos et al., 2016b) and Chile (Zambrano et al., 2017). several anchor stations were included from global datasets like the
Global Summary of the Day (GSOD) and the Global Historical Clima-
5. Discussion tology Network (GHCN) with the same name but with different loca-
tion. As an example, during March 2010 we found that data from San
The point-to-pixel comparison of monthly precipitation data from Juan and Neuquén stations was duplicated from these two global da-
57 rain gauges over the CAA with the relatively new CHIRPS tasets. Considering San Juan station, coordinates from GHCN are 31.5°
dataset allowed to identify its accuracy for estimating the amount, S, 68.699° W, elevation of 630 m; while GSOD data shows that location
timing and spatial distribution of regional precipitation. This dataset is is 31.4° S, 68.417° W, with an elevation of 597 m. In the case of Neu-
unique regarding spatial resolution (0.05°) and allows the study of quén station, GHCN data has a location in 39° S, 68.27° W with an
precipitation variability for the past 37 years in high resolution. Given elevation of 270 m; while the coordinates from GSOD data are 38.95° S,
the lack of rain gauges and the complexity of the Andes range, accurate 68.13° W, elevation of 273 m. The information for this example can be
estimations of precipitation gradients by interpolation methods can be accessed at ftp://ftp.chg.ucsb.edu/pub/org/chg/products/CHIRPS-2.0/
affected by large uncertainties. The CHIRPS precipitation dataset blends diagnostics/monthly_station_data/2010.03.csv, and the following link
in more station data than other products and uses a high-resolution has the information about the monthly station data for the complete
background climatology, providing better estimates of precipitation CHIRPS period: ftp://ftp.chg.ucsb.edu/pub/org/chg/products/
means and variations, resulting in a better hydrologic state (Shukla CHIRPS-2.0/diagnostics/monthly_station_data/. For CHIRPS calcula-
et al., 2014). Our results showed that the validation statistics indicated tion, stations were only added to the anchor list if they were outside of a
a relatively good agreement between the CHIRPS precipitation esti- 10-km radius from any station already in the list, assuming that stations
mates and ground precipitation measurements over the CAA. However, within the 10-km radius to be duplicates (Funk et al., 2014). In this
a large wet bias (PB = 65.8%) was observed in monthly precipitation sense, a difference in the location of the rain gauges depending on the
totals especially over the region where the seasonal cycle of precipita- data source could led to a deficient blending, resulting in a mis-
tion has a maximum during the cold season (North Patagonia, western representation of precipitation based on CHIRPS estimates.
and southern Cuyo). The annual cycle and the interannual variability of Finally, another cause that can led to the overestimation of CHIRPS
precipitation over North Patagonia was accurately represented by over North Patagonia region can be related to the precipitation esti-
CHIRPS estimations (PCC = 0.82), but an overestimation of precipita- mates based on infrared cold cloud duration observations. In a valida-
tion during the wet season (cold semester) resulted in large errors tion of daily satellite precipitation estimates over South America, Salio
(MAE = 34.7 mm) particularly for high precipitation totals. et al. (2015) found that the greatest difficulties for these products to
One of the causes of this discrepancy between observed and reproduce the main features of precipitation were shown to arise in

447
J.A. Rivera et al. Atmospheric Research 213 (2018) 437–449

Fig. 10. Monthly regional time series of observed and estimated precipitation for the period 1987–2016. The difference between observations and CHIRPS is also
presented. a) average of stations with rainy season during the warm season, b) average of stations with rainy season during the cold season.

mountainous areas and in non-convective precipitation events, features to an average mean absolute error larger than 34 mm. Over the rainiest
present over North Patagonia region during the rainy season. Moreover, locations of North Patagonia, the overestimation of the higher values of
the overestimation generated by evaporation of precipitation below the precipitation can be larger than 100 mm per month, showing that the
cloud base before reaching the surface, a condition present in arid to performance of CHIRPS is dependent on the representation of the
semi-arid regions like the study area (Hobouchian et al., 2017), could dominant precipitation systems. Moreover, the complexity of the to-
be playing a major role in the performance of CHIRPS estimations. pography can introduce uncertainties in precipitation estimations that
Further studies are needed to quantify in detail these possible sources of the blending with stations could not reduce, given the limited number
error, for which observations at higher altitudes will be necessary to of anchor stations used in the construction of CHIRPS. In this sense,
provide a proper characterization of precipitation over the study area. stations located above 1000 m.a.s.l. showed the lowest performance in
most of the validation statistics, although precipitation detection is
6. Conclusions accurately captured. The use of the freely available data from the SSRH
in the blending procedure could improve the final CHIRPS product,
We performed a validation of CHIRPS precipitation dataset along although the number of rain gauges is still scarce for the study area.
the Central Andes of Argentina with monthly data that covers a recent In general, the CHIRPS dataset, conceived primarily to support
period of 30 years (1987–2016). This kind of validation is un- agricultural drought monitoring, is a promising tool to assess pre-
precedented in the study area given the number of rain gauges used and cipitation variability over the CAA and particularly over Cuyo region.
the available period of records for a satellite product. This study used This semi-arid region is vulnerable to climate variability, ranging from
observations from at least 45 rain gauges not included as anchor sta- drought conditions that generate wildfires to extreme precipitation
tions for the calculation of CHIRPS estimations, therefore it can be events that produce landslides and debris flows over the Andes. In this
considered an independent validation. sense, a continuous observing system updated regularly is essential to
CHIRPS estimations reproduce adequately several characteristics of monitor precipitation variations over Cuyo. Drought monitoring over
precipitation along the CAA as the seasonal and interannual variability this region was traditionally based on the SPI (Penalba and Rivera,
and the spatial patterns of precipitation. In general, the ability of 2016). The use of the SPI with the CHIRPS estimations –a combination
CHIRPS data to capture the seasonal precipitation totals is higher in the recently applied over Africa (López-Carr et al., 2014); Chile (Zambrano
rainy season rather than in the dry season. This dataset shows a slight et al., 2017) and Southeast Asia (Guo et al., 2017)- can contribute to
underestimation (PB = −11%) of precipitation along Cuyo region, improve flood and drought monitoring and understanding with an
particularly associated to the areas where the rainy season takes place adequate spatial detail and over a variety of time scales, likely leading
during the warm semester. However, there is a wet bias over North to enhanced climate services for water managers and agricultural
Patagonia region (PB = 65.8%) as a consequence of an overestimation planners.
of precipitation mainly during the wet season –cold semester-, leading

448
J.A. Rivera et al. Atmospheric Research 213 (2018) 437–449

Acknowledgements Llano, M.P., Penalba, O.C., 2011. A climatic analysis of dry sequences in Argentina. Int. J.
Climatol. 31, 504–513.
López-Carr, D., Pricope, N.G., Aukema, J.E., Jankowska, M.M., Funk, C., Husak, G.,
This work was supported by the National Agency for Scientific and Michaelsen, J., 2014. A spatial analysis of population dynamics and climate change in
Technological Promotion (ANPCyT) [grant number PICT-2016-0431]. Africa: potential vulnerability hot spots emerge where precipitation declines and
We thank Subsecretaría de Recursos Hídricos de Argentina and Servicio demographic pressures coincide. Popul. Environ. 35, 323–339.
Maggioni, V., Sapiano, M.R.P., Adler, R.F., 2016. Estimating uncertainties in high-re-
Meteorológico Nacional for providing the precipitation records used in solution satellite precipitation products: systematic or random error? J.
the study. Hydrometeorol. 17, 1119–1129.
Maidment, R.I., Allan, R.P., Black, E., 2015. Recent observed and simulated changes in
precipitation over Africa. Geophys. Res. Lett. 42, 8155–8164.
References Mantas, V.M., Liu, Z., Caro, C., Pereira, A.J.S.C., 2015. Validation of TRMM multi-satellite
precipitation analysis (TMPA) products in the Peruvian Andes. Atmos. Res. 163,
Aghakouchak, A., Mehran, A., Norouzi, H., Behrangi, A., 2012. Systematic and random 132–145.
error components in satellite precipitation data sets. Geophys. Res. Lett. 39, L09406. McKee, T.B., Doesken, N.J., Kleist, J., 1993. The relationship of drought frequency and
http://dx.doi.org/10.1029/2012GL051592. duration to time scales. In: Proceedings of the Eight Conference on Applied
Ahmadalipour, A., Moradkhani, H., Yan, H., Zarekarizi, M., 2017. Remote sensing of Climatology. American Meteorological Society, Anaheim, pp. 179–184.
drought: vegetation, soil moisture, and data assimilation. In: Lakshmi, V. (Ed.), Nash, J.E., Sutcliffe, J.V., 1970. River flow forecasting through conceptual models part
Remote Sensing of Hydrological Extremes. Springer, Switzerland, pp. 121–149. I—a discussion of principles. J. Hydrol. 10, 282–290.
Alexandersson, H., 1986. A homogeneity test applied to precipitation data. J. Climatol. 6, Paredes-Trejo, F.J., Barbosa, H.A., Peñaloza-Murillo, M.A., Moreno, M.A., Farías, A.,
661–675. 2016. Intercomparison of improved satellite rainfall estimation with CHIRPS gridded
Beck, H.E., Vergopolan, N., Pan, M., Levizzani, V., van Dijk, A.I.J.M., Weedon, G., Brocca, product and rain gauge data over Venezuela. Atmósfera 29, 323–342.
L., Pappenberger, F., Huffman, G.J., Wood, E.F., 2017. Global-scale evaluation of 23 Paredes-Trejo, F.J., Barbosa, H.A., Lakshmi Kumar, T.V., 2017. Validating CHIRPS-based
precipitation datasets using gauge observations and hydrological modeling. Hydrol. satellite precipitation estimates in Northeast Brazil. J. Arid Environ. 139, 26–40.
Earth Syst. Sci. 21, 6201–6217. Penalba, O.C., Rivera, J.A., 2016. Precipitation response to El Niño/La Niña events in
Biles, J.J., Cobos, D., 2007. Natural disasters and their impact in Latin America. In: southern South America - emphasis in regional drought occurrences. Adv. Geosci. 42,
Stoltman, J.P., Lidstone, J., DeChano, L.M. (Eds.), International Perspectives on 1–14.
Natural Disasters: Occurrence, Mitigation, and Consequences. Springer, Dordrecht, Penalba, O.C., Rivera, J.A., Pántano, V.C., 2014. The CLARIS LPB database: constructing a
The Netherlands, pp. 281–302. long-term daily hydro-meteorological dataset for La Plata Basin, southern South
Boulanger, J.P., Aizpuru, J., Leggieri, L., Marino, M., 2010. A procedure for automated America. Geosci. Data J. 1, 20–29.
quality control and homogenization of historical daily temperature and precipitation Porcù, F., Milani, L., Petracca, M., 2014. On the uncertainties in validating satellite in-
data (APACH). Part 1: quality control and application to the argentine weather ser- stantaneous rainfall estimates with raingauge operational network. Atmos. Res. 144,
vice stations. Clim. Chang. 98, 471–491. 73–81.
Diem, J.E., Hartter, J., Ryan, S.J., Palace, M.W., 2014. Validation of satellite rainfall Rahmawati, N., Lubczynski, M.W., 2017. Validation of satellite daily rainfall estimates in
products for western Uganda. J. Hydrometeorol. 15, 2030–2038. complex terrain of Bali Island, Indonesia. Theor. Appl. Climatol. http://dx.doi.org/
Dinku, T., Chidzambwa, S., Ceccato, P., Connor, S.J., Ropelewski, C.F., 2008. Validation 10.1007/s00704-017-2290-7.
of high-resolution satellite rainfall products over complex terrain. Int. J. Remote Rivera, J.A., Penalba, O.C., Villalba, R., Araneo, D.C., 2017. Spatio-temporal patterns of
Sens. 29, 4097–4110. the 2010-2015 extreme hydrological drought across the Central Andes, Argentina.
Ebert, E., 2007. Methods for verifying satellite precipitation estimates. In: Levizzani, V., Water 9, 652. http://dx.doi.org/10.3390/w9090652.
Bauer, P., Turkeds, F.J. (Eds.), Measuring Precipitation from Space – EURAINSAT and Salio, P., Hobouchian, M.P., García Skabar, Y., Vila, D., 2015. Evaluation of high-re-
the Future. Springer, pp. 345–356. solution satellite precipitation estimates over southern South America using a dense
Falvey, M., Garreaud, R., 2007. Wintertime precipitation episodes in Central Chile: as- rain gauge network. Atmos. Res. 163, 146–161.
sociated meteorological conditions and orographic influences. J. Hydrometeorol. 8, Santos, J.R., Norte, F., Moreiras, S., Araneo, D., Simonelli, S., 2015. Predicción de epi-
171–193. sodios de precipitación que ocasionan aludes en el área montañosa del noroeste de la
Farooq Iqbal, M., Athar, H., 2018. Validation of satellite based precipitation over diverse provincia de Mendoza. Argentina. Geoacta 40, 65–75.
topography of Pakistan. Atmos. Res. 201, 247–260. Schwerdtfeger, W., 1976. The atmospheric circulation over Central and South America.
Feidas, H., 2010. Validation of satellite rainfall products over Greece. Theor. Appl. In: Schwerdtfeger, W. (Ed.), Climates of Central and South America. Elsevier, New
Climatol. 99, 193–216. York, USA. Vol. 2. pp. 2–12.
Förster, K., Oesterle, F., Hanzer, F., Schöber, J., Huttenlau, M., Strasser, U., 2016. A snow Seoane, R.S., Valdés, J.B., Mata, L.J., 2005. Climate variability and climate change in
and ice melt seasonal prediction modeling system for alpine reservoirs. Proc. IAHS Patagonian rivers. In: Wagener, T. (Ed.), Regional Hydrological Impacts of Climatic
374, 143–150. Change – Impact Assessment and Decision Making. Vol. 295. IAHS Publication,
Funk, C.C., Peterson, P.J., Landsfeld, M.F., Pedreros, D.H., Verdin, J.P., Rowland, J.D., Wallingford, UK, pp. 26–36.
Romero, B.E., Husak, G.J., Michaelsen, J.C., Verdin, A.P., 2014. A quasi-global pre- Shukla, S., McNally, A., Husak, G., Funk, C., 2014. A seasonal agricultural drought
cipitation time series for drought monitoring. US Geol. Survey Data Series 832. forecast system for food-insecure regions of East Africa. Hydrol. Earth Syst. Sci. 18,
http://dx.doi.org/10.3133/ds832. 3907–3921.
Funk, C., Peterson, P., Landsfeld, M., Pedreros, D., Verdin, J., Shukla, S., Husak, G., Tang, L., Tian, Y., Yan, F., Habib, E., 2015. An improved procedure for the validation of
Rowland, J., Harrison, L., Hoell, A., Michaelsen, J., 2015a. The climate hazards in- satellite-based precipitation estimates. Atmos. Res. 163, 61–73.
frared precipitation with stations - a new environmental record for monitoring ex- Tencer, B., Rusticucci, M., Jones, P., Lister, D., 2011. A southeastern South America daily
tremes. Sci. Data 2, 150066. http://dx.doi.org/10.1038/sdata.2015.66. gridded dataset of observed surface minimum and maximum temperature for 1961-
Funk, C., Verdin, A., Michaelsen, J., Peterson, P., Pedreros, D., Husak, G., 2015b. A global 2000. Bull. Am. Meteorol. Soc. 92, 1339–1346.
satellite-assisted precipitation climatology. Earth Syst. Sci. Data 7, 275–287. Thiemig, V., Rojas, R., Zambrano-Bigiarini, M., Levizzani, V., De Roo, A., 2012.
Garreaud, R.D., 2009. The Andes climate and weather. Adv. Geosci. 22, 3–11. Validation of satellite-based precipitation products over sparsely gauged African
González, M.H., 2013. Some indicators of interannual rainfall variability in patagonia River basins. J. Hydrometeorol. 13, 1760–1783.
(Argentina). In: Tarhule, A. (Ed.), Climate Variability – Regional and Thematic Toté, C., Patricio, D., Boogaard, H., Van Der Wijngaart, R., Tarnavsky, E., Funk, C., 2015.
Patterns. InTech, Rijeka, Croatia, pp. 133–161. Evaluation of satellite rainfall estimates for drought and flood monitoring in
Guo, H., Bao, A., Liu, T., Ndayisaba, F., He, D., Kurban, A., De Maeyer, P., 2017. Mozambique. Remote Sens. 7, 1758–1776.
Meteorological drought analysis in the lower Mekong Basin using satellite-based Viale, M., Norte, F., 2009. Strong Cross-Barrier Flow under Stable Conditions Producing
long-term CHIRPS product. Sustainability 9, 901. http://dx.doi.org/10.3390/ Intense Winter Orographic Precipitation: A Case Study over the Subtropical Central
su9060901. Andes, Wea. Forecast. 24, 1009–1031.
Habib, E., Henschke, A., Adler, R., 2009. Evaluation of TMPA satellite-based research and Viale, M., Nuñez, M.N., 2011. Climatology of winter orographic precipitation over the
real-time rainfall estimates during six tropical-related heavy rainfall events over subtropical Central Andes and associated synoptic and regional characteristics. J.
Louisiana, USA. Atmos. Res. 94, 373–388. Hydrometeorol. 12, 481–507.
Hobouchian, M.P., Salio, P., García Skabar, Y., Vila, D., Garreaud, R., 2017. Assessment of Vigaud, N., Lyon, B., Giannini, A., 2017. Sub-seasonal teleconnections between convec-
satellite precipitation estimates over the slopes of the subtropical Andes. Atmos. Res. tion over the Indian Ocean, the east African long rains and tropical Pacific surface
190, 43–54. temperatures. Int. J. Climatol. 37, 1167–1180.
Katsanos, D., Retalis, A., Tymvios, F., Michaelides, S., 2016a. Analysis of precipitation Willmott, C.J., 1981. On the validation of model. Phys. Geogr. 2, 184–194.
extremes based on satellite (CHIRPS) and in situ dataset over Cyprus. Nat. Hazards Willmott, C.J., Robeson, S.M., Matsuura, K., 2017. Climate and other models may be more
83, 53–63. accurate than reported. Eos 98. http://dx.doi.org/10.1029/2017EO074939.
Katsanos, D., Retalis, A., Michaelides, S., 2016b. Validation of a high-resolution pre- Zambrano, F., Wardlow, B., Tadesse, T., Lillo-Saavedra, M., Lagos, O., 2017. Evaluating
cipitation database (CHIRPS) over Cyprus for a 30-year period. Atmos. Res. 169, satellite-derived long-term historical precipitation datasets for drought monitoring in
459–464. Chile. Atmos. Res. 186, 26–42.
Le, A.M., Pricope, N.G., 2017. Increasing the accuracy of runoff and Streamflow simu- Zambrano-Bigiarini, M., Nauditt, A., Birkel, C., Verbist, K., Ribbe, L., 2017. Temporal and
lation in the Nzoia Basin, western Kenya, through the incorporation of satellite-de- spatial evaluation of satellite-based rainfall estimates across the complex topo-
rived CHIRPS data. Water 9, 114. http://dx.doi.org/10.3390/w9020114. graphical and climatic gradients of Chile. Hydrol. Earth Syst. Sci. 21, 1295–1320.

449