Hydrological Sciences Journal

ISSN: 0262-6667 (Print) 2150-3435 (Online) Journal homepage: https://www.tandfonline.com/loi/thsj20

Construction of a high-resolution gridded rainfall

dataset for Peru from 1981 to the present day

Cesar Aybar, Carlos Fernández, Adrian Huerta, Waldo Lavado, Fiorella Vega
& Oscar Felipe-Obando

To cite this article: Cesar Aybar, Carlos Fernández, Adrian Huerta, Waldo Lavado,
Fiorella Vega & Oscar Felipe-Obando (2019): Construction of a high-resolution gridded
rainfall dataset for Peru from 1981 to the present day, Hydrological Sciences Journal, DOI:
10.1080/02626667.2019.1649411

To link to this article: https://doi.org/10.1080/02626667.2019.1649411

View supplementary material

Accepted author version posted online: 29

Jul 2019.
Published online: 13 Aug 2019.

Submit your article to this journal

Article views: 26

View Crossmark data

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=thsj20
HYDROLOGICAL SCIENCES JOURNAL
https://doi.org/10.1080/02626667.2019.1649411

SPECIAL ISSUE: HYDROLOGICAL DATA: OPPORTUNITIES AND BARRIERS

Construction of a high-resolution gridded rainfall dataset for Peru from 1981 to the
present day
Cesar Aybara, Carlos Fernándeza,b, Adrian Huertaa, Waldo Lavadoa, Fiorella Vegaa and Oscar Felipe-Obandoa
a
Servicio de Meteorología e Hidrología del Perú (SENAMHI), Dirección de Hidrología, Lima, Perú; bPotsdam Institute for Climate Impact Research,
Potsdam, Germany

ABSTRACT ARTICLE HISTORY

A new gridded rainfall dataset available for Peru is introduced, called PISCOp V2.1 (Peruvian Received 23 February 2019
Interpolated data of SENAMHI’s Climatological and Hydrological Observations). PISCOp has been Accepted 21 June 2019
developed for the period 1981 to the present, with an average latency of eight weeks at 0.1° spatial EDITOR
resolution. The merging algorithm is based on geostatistical and deterministic interpolation methods A. Castellarin
including three different rainfall sources: (i) the national quality-controlled and infilled raingauge
dataset, (ii) radar-gauge merged precipitation climatologies and (iii) the Climate Hazards Group GUEST EDITOR
Infrared Precipitation (CHIRP) estimates. The validation results suggest that precipitation estimates are H. Lins
acceptable showing the highest performance for the Pacific coast and the western flank of the Andes. KEYWORDS
Furthermore, a meticulous quality-control and gap-infilling procedure allowed us to reduce the forma- quality control; gap-infilling;
tion of inhomogeneities (non-climatic breaks). The dataset is publicly available at https://piscoprec. CHIRP; TRMM 2A25; quantile
github.io/ and is intended to support hydrological studies and water management practices. mapping; PISCOp V2.1

1 Introduction
Environmental Prediction – Climate Forecast System
Accurate spatio-temporal rainfall estimations are essential for Reanalysis (NCEP CFSR; Saha et al. 2010), tend to be of
the development of scientific and operational applications, coarser spatial resolution and show lower performance than
which allow to understand the water cycle and its impact on satellite products. However, GRD based on a blend of satel-
natural and human systems. Conventional observations from lite, reanalysis and gauge rainfall sources are available. The
raingauge stations are an ideal input for the aforementioned most recent blended GRDs, with global and near real-time
applications. Unfortunately, strong spatial variability coverage, are the Multi-Source Weighted-Ensemble
(Garreaud et al. 2009) and the heterogeneous and sparse Precipitation (MSWEP; Beck et al. 2019) and the Climate
distribution of raingauges combined with systematic data Hazards Group Infrared Precipitation with Stations
quality deficiencies (Hunziker et al. 2017b) precludes their (CHIRPS; Funk et al. 2015a).
widespread use within Peru. The MSWEP provides three-hour precipitation at a spatial
In the last decades, new algorithms based on the indirect resolution of ~10 km for the period 1979 to the present, while
estimations from advanced infrared and microwave satellites, CHIRPS covers daily precipitation at ~5 km for 1981–present.
have led to the construction of different gridded rainfall Only a few studies have been done to analyse the performance of
datasets (GRD) that are used as auxiliary data to overcome these new blended GRDs in adjacent regions of Peru. For
the lack of raingauge stations, increase the spatio-temporal instance, Zambrano-Bigiarini et al. (2017) show that both
resolution and reduce uncertainties in rainfall predictions CHIRPS and MSWEP perform well at high temporal scales,
(Baik et al. 2015, Verdin et al. 2015, Bi et al. 2017, Sun presenting problems of overestimation (underestimation) in
et al. 2017). GRD based on satellites have specific shortcom- events of light rain (extreme rain) in Chile, whereas
ings in Peru. For instance, the TRMM Multi-satellite Perdigón-Morales et al. (2017) and Javier et al. (2016) mention
Precipitation Analysis (TMPA) 3B42 V7 (Huffman et al. that CHIRPS is acceptable at reproducing climatological values
2007), the Climate Prediction Center MORPHing Technique of monthly accumulated precipitation in Mexico and Venezuela,
(CMORPH, Joyce et al. 2004) and the Precipitation respectively. Nonetheless, these studies re-used raingauges that
Estimation from Remotely Sensed Information using had been incorporated previously in the merging algorithm,
Artificial Neural Networks (PERSIANN, Sorooshian et al. which alters the reliability of blended GRD performance results.
2000) tend to have large sensor errors over the Peruvian In line with this, Beck et al. (2017b) reported that MSWEP
Andes (Derin et al. 2016), overestimate the magnitude of performs better than CHIRPS only in regions with extensive
precipitation on the Pacific Coast (Ochoa et al. 2014) and raingauge networks (e.g. in temperate regions). However, these
span a short time period (≤20 years). On the other hand, findings cannot be directly applied to regions with sparse and
reanalysis-based GRD, e.g. the National Centers for irregular monitoring networks, such as Peru.

CONTACT Cesar Aybar csaybar@gmail.com Servicio de Meteorología e Hidrología del Perú (SENAMHI), Dirección de Hidrología, Lima, Perú
Supplemental data for this article can be accessed here.
© 2019 IAHS
2 C. AYBAR ET AL.

Recently, several studies have indicated essential consid- detailed and transparent information about the construction
erations when generating daily and monthly blended GRD in of PISCOp V2.1 as well as to evaluate its performance and
data-scarce regions: (a) the efficiency of blended GRDs largely stress its limitations.
depends on the predictor, and the interpolation method used
must be able to adapt to scenarios with high spatial hetero-
geneity (Dinku et al. 2014); (b) geostatistical interpolation 2 Material and methods
methods outperformed deterministic methods at annual and
2.1 Study area
monthly time steps, whereas for the daily time step, geosta-
tistic and deterministic methods were proven to be compar- Peru is located in the central-western region of South
able (Ly et al. 2013); (c) there are difficulties in assuming the America (0°–18°S; 68°–82°W) (Fig. 1), covering climatically
space-stationary hypothesis and to establish an adequate the- extremely variable regions with diverse precipitation regimes
oretical semivariogram at the daily time step (Nerini et al. that result from the interaction between synoptic-scale atmo-
2015); (d) the optimized interpolation parameters in determi- spheric currents, the complex orography of the Andes, the
nistic methods significantly improve the final results (Chen cold Humboldt Current System (HCS) and the El Niño
and Liu 2012); (e) the gap-infilling in precipitation time series Southern Oscillation (ENSO; Garreaud et al. 2009, Lavado
is highly important to minimize inhomogeneities in the Casimiro et al. 2012).
gridded datasets for periods of missing data, especially in In the austral summer, easterly trade winds from the
heterogeneous regions (Peterson et al. 1998, Beguería et al. southerly position of the Intertropical Convergence Zone
2015, Yanto et al. 2017); and (f) the use of simple ratios based (ITCZ) transport humid air masses from the tropical
on very high rainfall climatologies can significantly decrease Atlantic towards the Amazon Basin (Carvalho et al. 2011,
the systematic bias (Strauch et al. 2016). Marengo et al. 2012, Manz et al. 2017) and to the south
On this basis, this study presents the development of along the Andes through the South American Low-Level Jet
PISCOp V2.1, a new local blended GRD, headed by the (SALLJ; Vera et al. 2006, Boers et al. 2013). This period
National Service of Meteorology and Hydrology of Peru determines a marked wet season in most of Peru (Marengo
(SENAMHI). PISCOp V.2.1 contains daily and monthly rain- et al. 2012). Conversely, when the ITCZ is located further
fall grids at 0.1° computed for 1981–2019 covering the whole north (austral winter), convection and, consequently, preci-
of Peru with an average latency of eight weeks. It is built pitation levels are significantly reduced.
using serially complete raingauge datasets, CHIRP V2.0 In the Peruvian Andes, the climate is complex and primarily
(without raingauge stations), radar-gauge merged precipita- controlled by orography that acts as a topographic barrier to
tion climatologies, geostatistics and deterministic interpola- moisture flow, causing the formation of strong precipitation
tion methods, and a simple monthly correction factor applied gradients on the eastern flanks of the Andes (Bookhagen and
to daily estimates. The objective of this paper is to provide Strecker 2008). The inter-Andean valleys (> ~500 mm/year)

Figure 1. (a) Spatial extent of PISCOp V2.1. Points with × indicate stations with more than 95% of data within the 1981–2016 period. (b) Location and upstream
catchments of the selected stream gauges.
 HYDROLOGICAL SCIENCES JOURNAL 3

are principally dominated by convective processes (Garreaud as conservative because only gross errors are deleted if there is
1999, Campozano et al. 2016) channelling moisture intrusions strong evidence for implausibility. Hence, it is still expected to
of the Amazon (Chavez and Takahashi 2017). At the same find some remaining errors after QC, especially in areas with
time, the influence of cold and dry air masses originating a lower density of gauges. The following checks were applied:
from the HCS cause the driest conditions of our study area at
the Pacific coast and on the western flanks of the Andes (1) General problems (automatic): to delete obvious incon-
(< ~500 mm/year). However, during the occurrence of sistent values, such as negative and non-physical preci-
ENSO, the HCS weakens and the formation of severe convec- pitation, decimal point-related errors, repetitive dates,
tive storms can occur, especially over the northern Pacific coast repetitive consecutive values and unexpected changes
(Antico 2009). in latitude and longitude coordinates.
Based on the hydro-climatic heterogeneity described (2) Spatial extreme values (automatic): a threshold of
above and according to the classification of Manz et al. 200 years of return period of precipitation is used for
(2017), the study area was divided into five sub-regions the detection of extreme events, as in Keller et al.
(Fig. 1(a) and Table 1): (a) The Pacific coast (PC, average (2015). Then, if the extreme values occur in at least
annual precipitation of ~150 mm/year); (b) The western two neighbouring (<50 km) gauges for the same date,
Andean slopes (AW, ~400 mm/year), (c) the eastern they are preserved, otherwise they are deleted.
Andean slopes (AE, ~1100 mm/year), (d) the Andes- (3) Break and bad segments (manual): a visual control to
Amazon transition (AAT, ~3200 mm/year) and (e) the recognize segments with asymmetric rounding pat-
Amazon lowlands (AL, ~2250 mm/year). terns and obvious inhomogeneities.

2.2.2 Gap-infilling
2.2 Raingauge dataset Another source of uncertainty is found in the temporally
The raw gauge dataset comprised 945 daily observation data inconsistent gauge network which is liable to produce sys-
provided by SENAMHI. The raingauge data spans the period tematic bias during the merging phase (New et al. 2000). This
1981–2019 and is characterized by a high number of missing values is of primary importance in data-scarce regions, where several
and numerous quality issues (Hunziker et al. 2017b), being gen- raingauges come in and out of use (Hunziker et al. 2017b).
erally caused by the observer at the time of data collection (e.g. Similar to the QC approach, there is no established metho-
incorrect measurement of values ≤1 mm) and instrumentation dology for gap-infilling in Peru. Therefore, we propose a two-
malfunction. Even though metadata information would help to step approach to generate a serially complete raingauge
perform the data quality control, this was not used because of its dataset.
limited availability. Therefore, our analysis of raingauge data First, relying on neighbouring gauges, the relatively newer
mainly focused on gross error detection and gap-infilling methods. and effective spatio-temporal imputation method, CUTOFF
(Feng et al. 2014), was applied to infill the previously quality
controlled gauge datasets at daily and monthly time steps.
2.2.1 Quality control (QC) This method, unlike iterative imputation approaches (e.g.
Most methods for quality control (QC) in raingauge observa- missforest, Stekhoven and Bühlmann 2012) or matrix decom-
tions are designed for dense station networks (Vicente- position techniques (Lindstr et al. 2013), is principally
Serrano et al. 2010, Isotta et al. 2014, Notivoli et al. 2017), designed to handle missing values in raingauges by taking
which are difficult to assume in this study. As expected, into account the spatio-temporal rain distribution. Before
assuring the quality of a dataset is more problematic for data- CUTOFF is applied, each raingauge is grouped with other
scarce regions due to the reduced number of neighbouring raingauges if the following conditions are met: (a) distance <
stations (Hunziker et al. 2017b). Considering this fact and the 100 km, (b) sharing a minimum of 10 years of data, and (c)
lack of an established quality management system in Peru, we daily (monthly) linear relationship > 0.5 (> 0.8). Secondly, if
propose a three-step QC approach which can be considered the previous condition is neither fulfilled nor enough to
create serially complete time series, the quantile mapping
Table 1. Sub-regions defined for the analysis of PISCOp V.2.1. Adapted from bias correction (Qm), produced by matching the empirical
(Manz et al. 2016), N is the number of raingauges within each sub-region.
cumulative distribution of the collocated grid cell (CHIRPM,
Sub-region Elevation Climate driver Rainfall regime N
(m a.s.l.)
see Section 2.3) to the available gauge data (Gudmundsson
Peruvian Pacific 0–1500 ITCZ, HCS, ENOS Wet (Dec.–May) 97
et al. 2012), is used to infill the remaining gaps of each
Coast, PC Dry (Jun.–Nov.) raingauge.
Andes western >1500 Elevation, ITCZ Wet (Dec.–May) 151
slope, AW Dry (Jun.–Nov.)
Andes eastern >1500 Elevation, Weak seasonality, 128 2.3 Modification of CHIRP (CHIRPm)
slope, AE Orography, drier JJA
ITCZ CHIRP products at the monthly (CHIRPm) and daily
Andes-Amazon 500–1500 Orography, ITCZ, Weak seasonality, 26
transition, AAT SALLJ drier JJA
(CHIRPd) time step are initially calculated from the prelimin-
Amazon lowland, AL 0–500 ITCZ, trade winds Weak seasonality, 39 ary pentad time-step product (CHIRPpentad) using the follow-
drier JJA ing equations (Funk et al. 2015a):
4 C. AYBAR ET AL.


IRPpentad ¼ bo þ b1  TIR CCDpentad % (1) the delineation of the overpassing TPR pixel, proposed
by Manz et al. (2016).
IRPpentad (3) Application of a spatial bias thresholding filter to
CHIRPpentad ¼ CHPclim  (2) replace the pixels with large ratios (>5 median) by
IRPclim
the average of the surrounding 3 × 3 kernel.
where IRPpentad is the precipitation calculated from the linear (4) Smoothing of rain rate through cubic spline
model between thermal infrared cold cloud duration percen- interpolation.
tage (TIR CCD%) and TRMM 3B42 V7 product; CHPclim are (5) Merging with the long-term (1981–2010) monthly cli-
the monthly precipitation climatologies generated by Funk matologies raingauge dataset (Fig. 1(a)) using residual
et al. (2015b) and IRPclim represents the climatology of infra- ordinary kriging (ROK, Section 2.4.2).
red precipitation.
Even though the use of climatic or monthly correction
2.4 Merging phase of PISCOp V.2.1
factors can reduce the systematic bias (Funk et al. 2015a,
Keller et al. 2015, Parmentier et al. 2015, Strauch et al. The merging phase (Fig. 2) can be divided into four steps.
2016, Beck et al. 2017a, van Osnabrugge et al. 2017), it is Firstly, the provisional product P-PISCOpd is created by mer-
crucial that the predictor of larger temporary aggregations ging CHIRPMd and serially complete daily gauge datasets
(e.g. CHPclim) is well represented, otherwise a reverse process applying residual inverse distance weighting (RIDW). Secondly,
could occur. For Peru, we found that CHPclim extremely PISCOpm is estimated by merging CHIRPMm and completed
overestimates precipitation (>500%) at the Peruvian coast monthly gauge datasets using ROK. Thirdly, a monthly correc-
between 8°–18°S (see Supplementary material, Fig. S1). tion factor (Mcf) is derived from the comparison of PISCOpm
Furthermore, it does not adequately represent the orographic and P-PISCOpd aggregated at a monthly time step. Finally,
rainfall hotspots over the Andes-Amazon transition and con- PISCOpd is estimated by multiplying Mcf by P-PISCOpd.
siders raingauges with poor reliability. Based on this, CHPclim
was replaced by our own climatology PISCOpclim (see 2.4.1 Residual inverse distance weighting (RIDW)
Section 2.3.1), resulting in the modified form of CHIRP, Residual inverse distance weighting is used to generate P-
CHIRPM): PISCOpd. In this deterministic prediction method, the resi-
duals are defined in each gauged location si , as follows:
PISCOpclim þ ε
CHIRPMm ¼ CHIRPm  (3)
CHPclim þ ε ro ðsi Þ ¼ XB ðsi Þ  XO ðsi Þ (5)

PISCOpclim þ ε Si 2 S; i ¼ 1; . . . ; N (6)
CHIRPMd ¼ CHIRPd  (4)
CHPclim þ ε where N is the number of gauge observations, ro are the
where ε is a threshold defined as 0.5 in the denominator and residuals, XO is the daily raingauge value and XB is the
the numerator in order to deal with values of zero or near CHIRPMd value computed at each gauge location. The collo-
zero. This equation is applied to monthly and daily CHIRP cation of a raingauge to each CHIRPMd grid cell was per-
estimates, resulting in CHIRPMm and CHIRPMd, respec- formed using the smoothed merging (SM; Li and Shao 2010)
tively. CHIRP was previously resampled to a spatial resolu- approach. Unlike centre neighbour approximation, SM oper-
tion of 0.1 ° through cubic spline interpolation. ates over a 2 × 2 kernel considering the distances to the pixel
centroids as weights, resulting in a smoothed field that reduces
the boundary bias. To estimate the residual field ðμj ), ro is
2.3.1 PISCOp climatology (PISCOpclim )
interpolated by IDW at each grid point (j = 1, . . ., M), given by:
For Peru it has been found that the TRMM precipitation
8 PN
radar product 2A25 (TPR; Iguchi et al. 2000) is the most < P w ðjjSj Si jjÞro ðsi Þ  
i¼1 i
if Sj  Si Þ0
suitable rainfall data source for identifying spatial precipita- μj ¼ w ðjjSj Si jjÞ
N
  (7)
: i¼1 i
tion variability and seasonal patterns, even for the complex ui if Sj  Si  ¼ 0
orographic rainfall hotspots located in the eastern Andes
(Bookhagen and Strecker 2008, Nesbitt and Anders 2009,
  1
Manz et al. 2016). Based on this dataset, we constructed wi Sj  Si  ¼ (8)
jjSj  Si jjα
monthly climatologies at 0.1° spatial resolution. The TPR
data used corresponds to the 1998–2013 period, excluding where jj:jj is the euclidean distance, wi is the weight assigned
the year 2014 because during that year the satellite was to the gauge observation si and α is the power parameter. The
carrying out anomalous manoeuvres related to its dismantling α parameter controls the desired smoothness and the local
(Houze et al. 2015). The construction procedure of behaviour in the spatial prediction. High (low) values of α
PISCOpclim (Fig. 2) is summarized as follows: increase (decrease) the influence of the furthermost observa-
tions, generating low (high) variance in the residual field. For
(1) Extraction of suspicious pixels with a rain rate > additional details on IDW, see Babak and Deutsch (2009).
300 mm/h (Hamada and Takayabu 2014). Different studies have examined the effects of varying α for
(2) Aggregation of the entire TPR dataset to mean clima- the spatial prediction of rainfall (Chen and Liu 2012,
tology estimates for each calendar month considering Adhikary 2017). Accordingly, the optimal α was estimated
 HYDROLOGICAL SCIENCES JOURNAL 5

Figure 2. Schematic overview of the development of PISCOp V2.1.

by minimizing the root mean square error (RMSE) obtained compared with the monthly aggregation of P-PISCOpd.
from the 10-fold cross-validation. Finally, P-PISCOpd is Therefore, based on Keller et al. (2015), a monthly correction
defined as: factor (Mcf) was added after the creation of P-PISCOpd and
PISCOpm with two purposes: to provide higher spatial con-
P-PISCOpd ¼ CHIRPMd  μ (9)
sistency to daily predictions and to ensure that the monthly
aggregation of the daily product matches the monthly pro-
duct at each grid point. Thus, Mcf is calculated by:
2.4.2 Residual ordinary kriging (ROK)
Residual ordinary kriging is used for the generation of 8
>
> P PISCOpm if Mcf > 0
PISCOpm. Similar to RIDW, the residuals are estimated by < Ni¼1 PPISCOpd ðiÞ
Equations (5) and (6), with the main difference that XO Mcf ¼ P ... (10)
>
>
N
corresponds to monthly gauge estimates and XB represents : 1if P  PISCOpdðiÞ ¼ 0
i¼1
the CHIRPMm values computed at each gauge location.
However, unlike RIDW, the residual field ro is interpolated where N is the number of days of the corresponding month.
by ordinary kriging (Grimes et al. 1999) at each grid point Finally, PISCOpd is defined as:
and added back to CHIRPMm.
To ensure the non-stationarity assumption, the residuals PISCOpd ¼ P  PICSOpd Mcf (11)
are converted to a logarithmic scale and back-transformed
after the merging phase. In this study, the variogram adjust-
ment is automatically performed based on Hiemstra et al.
(2009). For more details on the implementation of ROK, 2.5 Evaluation of PISCOp V.2.1
refer to Goovaerts (2000).
The process for evaluating the performance of PISCOp V2.1
was performed on the period 1981–2016 in two steps: Firstly,
2.4.3 Monthly correction factor (Mcf) a pixel-to-point evaluation is carried out using an indepen-
Given that a higher spatial relationship is achieved at the dent rainfall network (ID) which consists of 100 raingauges
monthly rather than at the daily time step (Ly et al. 2013), (Fig. 1(a)) not previously used for the development of
PISCOpm is expected to present a higher performance PISCOp V2.1.
6 C. AYBAR ET AL.

We selected all raingauges with >12 months of data respectively. Unlike the pixel-to-point approach, the runoff
between January 1981 and December 2016 that are located ratios (RR and RRf) allowed us to assess the long-term capa-
at a minimum distance of 20 km from the PISCOp raingauge city of PISCOp V2.1 in more extensive areas. Due to the fact
network. Three continuous statistics are computed comparing that the annual time step is a sufficiently large time period, we
the time series of PISCOp V2.1 and ID (Table 2). The assume that the catchment storage is zero. Hence, Q and (P –
Pearson correlation coefficient (CC) is used to evaluate the ET) are expected to adopt similar values. The discharge gauge
capability of PISCOp V2.1 to capture rainfall variability; the data (Q) were obtained from the Environmental Research
RMSE measures the average magnitude of the error; and Observatory SO HYBAM (www.ore-hybam.org) and
the percent bias (PBIAS) indicates the degree to which each SENAMHI (Table 4 and Fig. 1(b)). Due to data scarcity, the
PISCOp V2.1 value is over- or under-estimated (Teng et al. ET was computed based on a modification of the Budyko
2014). hypothesis (Fu’s equation, Yao et al. 2016) and the gridded
Additionality, following the criteria defined by Zambrano- maximum and minimum temperature dataset generated for
Bigiarini et al. (2017), the three categorical statistics (Table 2) Peru by Vicente-Serrano et al. (2017). Four GRDs were used
probability of detection (POD), false alarm rate (FAR) and to estimate P: (a) PISCOp V2.1; (b) CHIRPM; (c) ORE-
threat score (TS) are used to determine PISCOp V2.1 rainfall HYBAM (HOP, Guimberteau et al. 2012), which is only
detection capabilities within five precipitation intensity classes available for the Amazon and generated at 1° spatial resolu-
(Table 3). The POD and FAR indicate which fraction of the tion using ordinary kriging; and (d) the previous PISCOp
observed events are correctly detected and which fraction of version (V1.0), which is based on the merging between
the events reported by the GRDs did not occur. The TS is CHIRPS and raingauges applying kriging with external drift
a general categorical statistic sensitive to hits and penalizes (Lavado et al. 2015). We used CHIRPM as a reference to
both missing and false alarms affected by the climatological explore possible improvements after the merging phase. We
frequency of the event. More information on the aforemen- compared PISCO V2.1 and PISCOp V1.0 to examine the
tioned indices can be found in Wilks (2006). repercussions of changing CHIRPS by CHIRPM. The com-
Secondly, a water balance evaluation using two simple parison to HOP serves to understand the role of the intro-
runoff ratios (RR and RRf) is carried out in 19 Peruvian duction of spatial predictors (CHIRPM). It is important to
catchments using the following equations: note that HOP, PISCOp V1.0 and PISCOp V2.1 present
almost the same availability of raingauges within selected
Q upstream catchments. Therefore, the influence of raingauge
RR ¼ (12)
P density can be handled as a common systematic variation for
all GRDs used.
Q
RRf ¼ (13)
ðP  ETÞ
3 Results and discussion
where Q, P and ET are the annual long-term average of the
discharge, precipitation and real evapotranspiration, 3.1 State and gap-infilling of the peruvian rainfall
dataset

Table 2. Continuous and categorical statistics. X: GRD estimate; Y: ID measure- According to the three-step QC approach (Section 2.2.1),
 GRD average; Y:
ment; X:  ID average; N: number of data pairs; A: number of hits; 3.51% of the total data (Table 5) had gross error and was
B: number of false alarms; C: number of misses; and D: number of correct deleted for the next steps. The most affected sub-region with
negatives.
data exclusion was the AW and the least affected was the PC.
Statistic Formula Perfect
score Hunziker et al. (2017b) indicated that a large fraction of gross
Continuous statistics errors are caused by observers during data recording, while
P

Correlation coefficient, CC ðXXÞ YY 1 Hunziker et al. (2017a) mentioned that due to these errors
CC ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P
ðXXÞ ðYYÞ
2 2
and a large amount of missing data, 40% of available rain-
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P
Root mean squared error,
RMSE ¼ N1 ðX  Y Þ2
0 gauges are inappropriate for climate analyses. Following our
RMSE hP i
Percentage bias, PBIAS ðXY Þ PX 0
approach, and considering that within the 1981–2016 period
PBIAS ¼ 100  =
at least 10 years of continuous information must be available
Categorical statistics
Probability of detection, POD POD ¼ A=ðA þ CÞ 1
after applying QC, the number of stations had reduced from
False alarm ratio, FAR FAR ¼ B=ðA þ BÞ 0 945 to 441 (Fig. 1(a)). These raingauges (henceforth called
Threat score, TS TS ¼ A=ðA þ B þ CÞ 1 PISCOp rainfall network) form the basis and constitute the
most valuable source of information for the construction of
the gridded dataset.
Table 3. Classification of rainfall events based on quantiles. The density of the PISCOp rainfall network lies at around
Quantile Daily rain (mm/d) Daily rainfall event 282 per 106 km2) for Peru (Table 5), with maximum density
[0–0.1>] * [0–1.5>] No rain in AW and minimum in AL. These results suggest
[0.1–0.5>] [1.5–5.3>] Light rain a remarkably heterogeneous distribution and very sparse con-
[0.5–0.9>] [5.3–19.5>] Moderate rain
[0.9–0.975 >] [19.5–38.4>] Heavy rain ditions across the whole of Peru. Despite the data scarcity, the
0.975 > 38.4> Violent rain Peruvian rainfall network (Table 5) lies within the minimum
*This rainfall class is considered as no rain. requirements for hydrological analyses defined by the World
 HYDROLOGICAL SCIENCES JOURNAL 7

Table 4. Summary of in situ discharge gauge characteristics. ETP: potential evapotranspiration; ET: real evapotranspiration.
No. Name Code Catchment area Raingauge density Mean discharge PISCOp V2.1 Precipitation ETP ET
(km2 × 10−3) (per 106 km2) (m3/s) (mm/year)
1 Borja BRJ 94.21 371.5 49.43 1730 973 799
2 Chazuta CZT 69.57 445.6 30.76 2333 1043 936
3 Pucallpa PCP 267.38 254.3 101.78 2047 996 881
4 Requena RQN 359.26 203.2 123.34 2236 1351 1116
5 Ardilla ARD 11.89 1093.2 1.49 916 1051 554
6 Puchaca PCH 0.74 0 0.06 717 834 448
7 Condorcerro CON 10.54 664.3 1.39 743 653 380
8 Yanapampa YNP 4.27 703.3 0.41 470 678 199
9 Santo Domingo SDG 1.89 1587.2 0.16 463 676 206
10 La Capilla LCP 2.19 1370.8 0.19 541 695 242
11 S&T Imperial SYT 5.96 1341.3 0.55 477 658 210
12 Conta CNT 3.12 960.0 0.11 374 727 268
13 Letrayoc LTY 3.57 1121.4s 0.26 502 669 304
14 Huatiapa HTP 13.04 997.2 0.76 513 551 328
15 Chucarapi CCP 13.51 592.2 0.32 300 604 235
16 La Tranca LTC 2.01 993.1 0.02 123 665 93
17 Bella Union BUN 4.30 465.5 0.12 204 698 122
18 Puente Ilave ILV 8.12 369.2 0.32 304 523 173
19 Puente Ramis RMS 15.09 596.2 0.72 597 549 422

Table 5. Overview of the state and gap-infilling of the PISCOp rainfall network: BCC: regional runoff simulation in adjacent regions of Peru
percentage of gaps completed by bias-corrected CHIRPM (%); Dn: spatial average of
KS statistic; and MKbef–af: number of spurious trends after the gap-infilling procedure.
(Guimberteau et al. 2012, Zulkafli et al. 2013, Getirana et al.
Sub-region (total rain gauges: 441)
2014). Most of these raingauges are located in mountain
recharge zones, which is an important condition for rain-
PC AW AE AAT AL Total
fall–runoff simulation in Pacific and Andean drainage catch-
Monthly
Density (/106 km2) 480 754 381 170 59 282 ments (Fig. 1). Nonetheless, the density of the PISCOp
Gross error (%) 0.62 7.5 2.19 3.60 1.2 3.51 rainfall network is far below conventional climatological data-
No data (%) 34.8 30.94 41.44 45.14 41.72 37.41 sets worldwide (Hofstra et al. 2009, Yatagai et al. 2012,
CUTOFF (%) 14.16 11.20 12.49 35.28 40.76 16.40
BCC (%) 20.64 19.74 28.95 10.05 0.96 21.01 Newman et al. 2015, Lussana et al. 2018). Therefore, it is
Dn 0.07 0.05 0.04 0.03 0.02 0.05 expected that regions with lower density introduce biases in
Daily the mean and variance of the gridded dataset.
No data (%) 32.98 29.39 39.82 43.09 39.40 35.65
CUTOFF (%) 12.22 13.88 27.04 34.07 33.72 20.56 The gaps in the raingauge series were identified as the
BCC (%) 20.76 15.51 12.78 9.02 5.68 15.09 main problem for the construction of a serially complete
Dn 0.03 0.03 0.05 0.07 0.03 0.03 gauge dataset in Peru. In general, most raingauges belonging
MKbef–af 4 1 7 3 3 18
to the PISCOp rainfall network were installed after the year
2000, which explains the high percentage of missing data
Meteorological Organization (WMO 1994) and it is similar to (34.74%) before this date (Fig. 3). For this reason, the amount
the raingauge density presented in previous works related to of data available for the period 2001–2016 exceeds by 117%

Figure 3. Missing data for each sub-region. Note: the raingauges are ordered from lowest missing values (upper line) to highest missing values (lower line).
8 C. AYBAR ET AL.

that for the period 1981–2000, with the most significant rainfall characterization in most of the sub-regions in com-
difference (204%) in the AL. The monthly gap-infilling parison to CHIRPm. This improvement is explained by the
approach based on neighbouring stations allowed us to infill use of PISCOclim instead of CHPclim in the CHIRPMm con-
44% of total missing data, while the daily approach allowed struction. However, it is still unable to detect the convective
58% to be infilled (Table 5). storms caused by the ENSO in the northern PC.
Considering the fact that the neighbouring information In contrast to CHIRPm and CHIRPMm, the PISCOpm
was not enough to create serially complete time series for all product (Fig. 4(d)) underlies different rainfall intensities
raingauges, bias-corrected CHIRPM was added. Similar to and spatial structure of the entire study area. These changes
other studies (Chaney et al. 2014, Teegavarapu and Nayak are the result of the interaction of: (a) spatial autocorrela-
2017), we used the Mann-Kendall (MK) and Kolmogorov- tion among residuals (measured through the semivario-
Smirnov (KS) statistical tests to determine whether the per- gram), (b) the distribution of the PISCOp rainfall
formance of the gap-infilling procedure is appropriate. network, and (c) the magnitude and sign of the residuals.
Firstly, the KS nonparametric test, that does not use any For the analysed month, negative residuals and
distributional assumptions, was employed as a metric to esti- a considerable number of raingauges led to a rainfall
mate if the distribution before and after the gap-infilling increase in the AE and AW, as well as to a better repre-
procedure was identical. Considering a 5% significance level, sentation of the spatial pattern caused by ENSO in the
54 raingauges at the daily time step and 41 at the monthly northern PC. Likewise, the Peruvian Amazon areas (AAT
time step were flagged for showing a poor match between the and AL) also show negative residuals, although the reduced
cumulative distribution functions (Dn > 0.1, see number of raingauges led to a high negative local average.
Supplementary material, Fig. S2). In addition, Table 5 indi- Therefore, a scenario with high rainfall amounts is
cates a spatial average of Dn below 0.07 for all the sub-regions. expected, especially when the de-correlation distance is
Next, to identify the possible formation of spurious trends, overcome.
the MK test was performed on annual time series from the Even though the use of geostatistical interpolation meth-
period 1981–2016. At a 5% significance level, 18 raingauges ods allows the spatial coherence to be maximized, it must be
(see Supplementary material, Fig. S2) were identified with taken into account that the predicted values in PISCOpm may
positive trends (τ > 0.1). differ from the raingauge values particularly related to the
Both results suggest that the gap-infilling procedure formation of large residuals. This behaviour has already been
adopted in this study is acceptable, showing an unclear spatial extensively described in Tozer et al. (2012) and further stu-
pattern in the results (Supplementary material, Fig. S2). dies (Ensor and Robeson 2008, Hofstra et al. 2009, Erdin et al.
Nonetheless, note that ~14% of the raingauges were flagged 2012). Therefore, special care must be taken when using
by at least one of the tests. Therefore, it is expected that PISCOp V2.1 for analysing extreme events, such as those
inhomogeneities are present in the serially complete gauge related to ENSO.
dataset. For an individual inspection of the gap-infilling pro- For the blending process at a daily time step, several
cedure, see http://piscoprec.github.io/gauge. products (CHIRPMd, P-PISCOpd and Mcf), were created
before generating the final dataset PISCOpd. In previous
studies, it has been demonstrated that in data-scarce regions
3.2 Spatial description of PISCOp
the areal rainfall estimates are better represented in blended
The rainfall pattern for January 1998 (Fig. 4) and day 25 of than in the only gauge-based GRDs (Buytaert et al. 2006,
this month (Fig. 5) were used to perform a visual check of Schuurmans et al. 2007, Nerini et al. 2015). Hence,
PISCOp products (PISCOpm and PISCOpd, Fig. 2). This date P-PISCOp (Fig. 5(c)) was produced by merging the serially
was selected as an atypical and violent rainfall event (Table 3), complete PISCOpd rainfall network (Fig. 5(a)) with
caused by the ENSO phenomenon, was experienced on CHIRPMd (Fig. 5(b)). Unlike CHIRPMd, P-PISCOp allowed
this day in the north of Peru. an admissible characterization of convective storms that
The output analysis for January 1998 was performed con- occurred at the northern PC and improved the rainfall sce-
sidering as a reference the rainfall measured in this month by nario in the AW. Nonetheless, a marked underestimation
the PISCOp rainfall network (Fig. 4(a)). Thus, CHIRPm concerning CHIRPMd, was found for the rainy pixels located
(Fig. 4(b)) only shows an acceptable representation of the at the centre of the AAT and AL. The explanation of this
spatial structure of the rainfall field within the AW, while change is similar to the previous month analysed, with the
for the AE, AAT and AL, a remarkable underestimation is difference that at daily time step the predictor (CHIRPMd)
evident with increasing rainfall rates. Underestimation of represents the spatial variance being worse, causing higher
TIR-based GRDs, such as CHIRPm, for high rainfall values instability in the residuals. Due to the absence of raingauges,
is a well-known condition (Kidd and Huffman 2011). In the residuals of the western sub-regions (PC, AW and AE) are
contrast, the PC shows unrealistically high rain values for continuously and omnidirectionally transferred to the eastern
the entire study period (1981–2016). This pattern is sub-regions (AAT and AL) by IDW. The efficiency of this
a consequence of the CHPclim used as a spatial predictor; process directly depends on the residuals variance and inter-
CHPclim reasonably depicts the spatial structure in the PC. mittent rainfall regime (Chappell et al. 2012). A scenario with
Nonetheless, the lack of raingauges for calibration produces low variance and intermittency should improve the systema-
severe overestimation of rainfall amounts (Fig. S1). Regarding tic bias,; otherwise, this transfer would result in inaccurate
the CHIRPMm product (Fig. 4(c)), in general, it improved the precipitation values (Fig. 5(c)). Finally, PISCOpd (Fig. 5(d)),
 HYDROLOGICAL SCIENCES JOURNAL 9

Figure 4. Spatial distribution of rainfall for January 1998: (a) rainfall network, (b) CHIRPm, (c) CHIRPMm and (d) PISCOpm.

unlike P-PISCOpd, decreases the bull’s-eye effect formation the systematic PBIAS is close to 0. For the AE and AAT, the
around the gauge observations relying on the spatial structure performance of PISCOp V2.1 continues to indicate
of PISCOpm. Besides, a clear precipitation increase can be a substantial increase (reduction) of the CC (RMSE) score
observed for the AL and AAT due to the formation of Mcf with respect to CHIRPM, although these values are worse in
values > 1 in these sub-regions. comparison to the PC and AW. This can be explained by
a lower density of raingauges and that TIR-based retrieval
algorithms imply a poorer performance under high influence
3.3 Performance of PISCOp V.2.1
of topographic complexity (Thouret et al. 2013, Mantas et al.
As may be seen in Fig. 6, the sub-regions PC and AW show 2015, Derin et al. 2016). At the monthly time step, PISCOpm
the most significant improvements for PISCOp V2.1 com- provides higher accuracy in capturing the influence of the
pared to CHIRPM, although there is a slight increment in the ITCZ migration through the tropical Andes, despite
spread of their scores. In these sub-regions, P-PISCOpd and a remarkable underestimation of the precipitation gradients
PISCOpd (PISCOpm) increase the accuracy of the CC to 213% for the eastern Andean slopes. In contrast, at the daily time
and 210% (14%) compared to CHIRPMd (CHIRPMm). The step, P-PISCOpd and PISCOpd showed a poor performance
RMSE values show consistent reduction in random error and that did not lead to any improvement compared to
10 C. AYBAR ET AL.

Figure 5. Spatial distribution of rainfall for 25 January 1998: (a) the rainfall network, (b) CHIRPMd, (c) P-PISCOpd and (d) PISCOpd. Only values >1 mm are plotted.

CHIRPMd. Finally, the AL presented the largest RMSE score respectively. The TS showed that the daily products,
and lowest CC of our study area. Even at the monthly time CHIRPMd, PISCOpd and P-PISCOpd, were not able to cor-
step, PISCOpm presented a lower performance than rectly capture the fraction of rainfall events for all sub-
CHIRPMm. Similar results have recently been reported for regions. These results show that these daily PISCOp products
the Ecuadorian Amazon by Ulloa et al. (2017), who state that are most likely not sufficiently accurate for capturing heavy
the reduced number of raingauges used for the merging phase rainfall events. Hence, the use of PISCOp products to
generates spatial inconsistencies for the entire analysis period. describe the intensity of extreme precipitation events is not
According to the categorical indices applied in this study, recommended if no high-density rainfall network exists
both P-PISCOpd and PISCOpd presented similar scores that nearby.
were higher than CHIRPMd in the entire study area (Fig. 7). Figure 8 illustrates the water balance evaluation of PISCOp
In general, the detection capacity of the three products weak- V2.1, CHIRPM and two other GRDs (PISCOp V1.0 and
ens as the precipitation intensity category increases (Table 3), HOP) using runoff ratios (RR and RRf). The widest spread
regardless of the geographical position. Comparing the cate- in RR and RRf scores was observed within the Amazon basin,
gories “no rain” and “violent rain”, the POD score decreased indicating that PISCOp V.2.1 reaches a similar score to HOP,
and FAR increased drastically by 452% and 245%, while it has an underestimation of 15% for CHIRPM and an
 HYDROLOGICAL SCIENCES JOURNAL 11

Figure 6. Boxplots of continuous statistics (CC, RMSE and PBIAS) between the daily and monthly products of PISCOp V.2.1 and ID. The cross represents the spatial average.

overestimation of 28% for PISCOp V1.0. For the Amazon, RR differences in their RR and RRf scores, with values mostly
estimates are typically below 0.8 (Costa and Foley 1997, below 1. Unlike the Amazon, in the Andes-Pacific the change
Rudorff et al. 2014, Gusev et al. 2017), while the average RR from CHPclim to PISCOpclim did not affect the areal rainfall
values for PISCOp V.2.1, CHIRPM, PISCOp V1.0 and HOP estimations, possibly due to a better distribution and higher
are around 0.89, 0.75, 1.05 and 0.88, respectively. Similar density of the rainfall network. The RRf values of CHIRPM,
results were found for RRf, with PISCOp V2.1, CHIRPM PISCOp V2.1 and PISCOp V1.0 exhibited a systematic over-
and HOP having values close to 1 (–0.2 < RRf < 0.2). estimation mainly for catchments >10 000 km2 (Fig. 8). This
Although these high runoff ratios could be explained by uncertainty could be related to streamflow alteration caused
excessive groundwater contribution (Zubieta et al. 2015), by anthropogenic factors or low ET estimates. However, it is
a rainfall underestimation scenario is a more likely explana- difficult to predict and beyond the scope of this investigation.
tion, especially when considering the independent validation
results which indicate the PBIAS trend to be negative (Fig. 6).
The RRf within the Amazon shows that CHIRPM achieves the 3.4 Impacts and detection of inhomogeneities
best agreement with discharge values. Additionally, the cli- All inputs used for the development of PISCOp V2.1 are not
mate correction based on PISCOpclim (used in CHIRPM and thoroughly homogenized. The Peruvian long-term gauge
PISCOp V2.1) led to a better performance than CHPclim dataset is affected by a plethora of non-climatic factors, such
(used in PISCOp V1.0) in order to eliminate the underesti- as changes of instruments or bad observer practices (Peterson
mation of rainfall, particularly on the eastern slopes of the et al. 1998, New et al. 2000, Brönnimann 2015). Also, incon-
Andes. For catchments covering the Andes or Pacific, sistencies are also present in the CHIRP algorithm. For
CHIRPM, PISCOp V1.0 and PISCOp V2.1 present very slight instance, they arise when infilling missing CHIRP values
12 C. AYBAR ET AL.

Figure 7. Categorical validation statistics of PISCOpd products in five quantile classes of rainfall intensity for the five sub-regions (see Table 3).

with the Coupled Forecast System version 2 (Saha et al. 2014), Based on the breakpoints detected for each cell, a sensitive
or in the overlap between the TIR archives: Globally Gridded area (Fig. 9(a)) was defined to analyse the plausibility of the
Satellite (GriSat) and NOAA Climate Prediction Center time series in more detail. As shown in Fig. 8(b), the inten-
(CPC; Funk et al. 2015a). In order to detect a spatial pattern sities, breakpoint year at 5% significance level (red dotted
in these inhomogeneities, we applied the Pettitt test, with line) and seasonality of the three GRDs at the monthly time
a significance level of 5%, for annual time series of each step vary considerably despite using similar inputs. The inho-
PISCOp V2.1 grid cell, and for the water balance estimation, mogeneities in the GRDs imply severe impacts for the analy-
we used CHIRPM, PISCOp V1.0, and HOP as a reference. sis of the seasonal (not shown here) and annual trend
Figure 9(a) shows the years when a breakpoint was detected. (Fig. 9(c)). For the assessment of these impacts, Sen’s slope
In general, the three GRDs indicate wide variability in their estimator at the 95% confidence level was used. For the
breakpoint years and spatial extent. CHIRPM obtained the 1981–2016 period, PISCOp V2.1, PISCOp V1.0 and HOP
smaller inhomogeneity area within the Amazon basin (8.1%), revealed a significant positive trend that exceeded 55%, 81%
followed by PISCOp V.2.1 (8.2%), HOP (34%) and PISCOp and 72%, respectively, in the slope of CHIRPM. Artificial
V1.0 (89%). The breakpoint year observed for CHIRPM is trends (Hofstra et al. 2010, Nicolas and Bromwich 2011,
associated with the transition from GriSat to CPC, whereas Tozer et al. 2012, Kingdom 2014) are principally spread
the breakpoint for PISCOp V1.0 notably coincides with the across the entire Peruvian Amazon, especially where data
changes in the density of the PISCOp rainfall network scarcity prevails and there is a high amount of gaps.
(Fig. 3). Regarding PISCOp V2.1 and HOP, the inhomogene-
ity area drastically decreased due to the data gap infilling
4 Summary and conclusions
performed at each station and a further balancing-out during
the geostatistical interpolation. Although breakpoints in rain- In this paper, we presented the development of PISCOp V2.1,
fall time series can naturally occur, no evidence was found for a new daily and monthly long-term GRD for the period 1981
any of the 56 raingauges with more than 95% of complete to the present. This gridded product was generated based on
time series (Fig. 1(a)). the integration of serially complete gauge datasets, CHIRP
 HYDROLOGICAL SCIENCES JOURNAL 13

Figure 8. Runoff ratios (RR and RRf) between different GRDs (PISCOp V2.1, CHIRPM, PISCOp V1.0 and HOP) and discharge observations for catchments draining
within the Andes-Pacific (left) and Amazon (right).

Figure 9. (a) Spatial distribution of break year (calculated by Pettit test) in annual time series of PISCOp V2.1, CHIRPM, PISCOp V1.0 and HOP. Only values at 95%
significance level (p < 0.05) are plotted. (b) Areal monthly precipitation for the S1 region. The breaks at 95% significance level are plotted on the vertical dotted (red)
line. (c) Evolution of the average annual rainfall in the S1 region. Only trend lines with a significant level of 95% are plotted.
14 C. AYBAR ET AL.

data, radar-based climatologies and geostatistical and deter- References

ministic interpolation methods. The quality of PISCOp V2.1
Adhikary, S.K., 2017. Cokriging for enhanced spatial interpolation of
was assessed within six hydro-meteorological sub-regions rainfall in two Australian catchments. Hydrological Processes, 31
with an independent raingauge network. Additionality, the (March), 2143–2161. doi:10.1002/hyp.11163
runoff ratio estimation within 19 catchments allowed us to Antico, P.L., 2009. Relationships between autumn precipitation anoma-
evaluate the performance of PISCOp V2.1 in more extensive lies in Southeastern South America and El Niño event classification.
areas. International Journal of Climatology, 727 (July 2008), 719–727.
doi:10.1002/joc
For the first time, PISCOp V2.1 comprehensively presents Babak, O. and Deutsch, C.V., 2009. Statistical approach to inverse dis-
the state of the Peruvian raingauge dataset for the period tance interpolation. Stochastic Environmental Research and Risk
1981–2016. It has been identified that the gaps in rainfall Assessment, 23 (5), 543–553. doi:10.1007/s00477-008-0226-6
time series represent the most determining problem for the Baik, J., et al., 2015. Geospatial blending to improve spatial mapping of
construction of a temporally consistent GRDs. This study precipitation with high spatial resolution by merging satellite- and
ground based data. Hydrological Processes, 30 (16), 2789–2803.
shows that the combinations between both CUTOFF and doi:10.1002/hpy.10786
Qm are a conservative and efficient method for successful Beck, H.E., et al., 2017a. MSWEP: 3-hourly 0.25° global gridded pre-
data gap-infilling, especially for large data gaps that prevail cipitation (1979–2015) by merging gauge, satellite, and reanalysis
at the beginning of the PISCOp V2.1 period. However, this data. Hydrology and Earth System Sciences, 21 (1), 589–615.
method strongly depends on the predictor and the proximity doi:10.5194/hess-21-589-2017
Beck, H.E., et al., 2017b. Global-scale evaluation of 23 precipitation
of raingauges. Hence, it is expected that application of our datasets using gauge observations and hydrological modeling.
approach to areas with low station density and poor perfor- Hydrology and Earth System Sciences Discussions, 21 (August), 1–23.
mance of CHIRPM might lead to unsatisfactory results. doi:10.5194/hess-2017-508
The independent and water balance evaluations confirm that Beck, H.E., et al., 2019. MSWEP V2 global 3-hourly 0.1° precipitation.
PISCOp V2.1 is the most suitable product for representing areal Bulletin of the American Meteorological Society. 100 (March),
473–500. doi:10.1175/BAMS-D-17-0138.1
rainfall estimates, except for the Amazon lowland where Beguería, S., et al., 2015. Bias in the variance of gridded data sets leads to
CHIRPM had improved outcomes. Additionally, we note that misleading conclusions about changes in climate variability bias in the
the climatological correction based on PISCOpclim significantly variance of gridded data sets leads to misleading. International
improved the results compared to CHPclim. At the daily time Journal of Climatology, 36 (December), 3413–3422. doi:10.1002/
step, PISCOpd did not capture the convective storm intensity, joc.4561
Bi, S., et al., 2017. A double-smoothing algorithm for integrating satellite
regardless of the geographical position. Although it seems precipitation products in areas with sparsely distributed in situ
highly attractive to use gridded data with full spatial and tem- networks. Isprs International Journal of Geo-Information, 6 (1), 28.
poral coverage, such as PISCOp V2.1, all inhomogeneities doi:10.3390/ijgi6010028
inherent to this merged product and presented in this study Boers, N., et al., 2013. Complex networks identify spatial patterns of
must be entirely taken into account. Therefore, as for other extreme rainfall events of the South American monsoon system.
Geophysical Research Letters, 40, 1–7. doi:10.1002/grl.50681
blended GRD products (Yanto et al. 2017, Beck et al. 2017a), we Bookhagen, B. and Strecker, M.R., 2008. Orographic barriers,
recommend users to take special care when using PISCOp V.2.1 high-resolution TRMM rainfall, and relief variations along the
for the analysis of trends, extreme events or other applications Eastern Andes. Geophysical Research Letters, 35 (6), 1–6.
related to e.g. climate change. doi:10.1029/2007GL032011
New versions of PISCOp have been planned and it is Brönnimann, S., 2015. Climatic changes since 1700. Cham, Switzerland:
Springer International Publishing. doi:10.1007/978-3-319-19042-6_4
expected to make use of new existing information sources, Buytaert, W., et al., 2006. Spatial and temporal rainfall variability in moun-
such as IMERG (Huffman et al. 2015), cross-border raingauges, tainous areas: a case study from the South Ecuadorian Andes. Journal of
as well as to improve the data gap-infilling of rainfall time series. Hydrology, 329 (3–4), 413–421. doi:10.1016/j.jhydrol.2006.02.031
Campozano, L., et al., 2016. Rainfall and cloud dynamics in the Andes:
a Southern Ecuador case study. Advances in Meteorology, 2016, 1–15.
5 Data access doi:10.1155/2016/3192765
Carvalho, L.M.V., et al., 2011. Moisture transport and intraseasonal
The PISCOp V2.1 product, source code and additional infor- variability in the South America monsoon system. Climate dynamics,
mation are freely available to users in NetCDF (1981–2016) 36 (9–10), 1865–1880. doi:10.1007/s00382-010-0806-2
and GeoTIFF (1981–present) format at the following website: Chaney, N.W., et al., 2014. Development of a high-resolution gridded
https://piscoprec.github.io/. daily meteorological dataset over Sub-Saharan Africa: spatial analysis
of trends in climate extremes. Journal of Climate, 27 (15), 5815–5835.
doi:10.1175/JCLI-D-13-00423.1
Acknowledgements Chappell, A., Renzullo, L., and Haylock, M., 2012. Spatial uncertainty to
determine reliable daily precipitation maps. Journal of Geophysical
We thank SENAMHI and CLIMANDES for supporting this research Research: Atmospheres, 117 (September), 1–14. doi:10.1029/
and acknowledge hundreds of hydrometeorological observers around 2012JD017718
Peru for their work, which has made this study possible. The authors Chavez, S.P. and Takahashi, K., 2017. Orographic rainfall hot spots in
would like to thank Stephanie Gleixner, Fred F. Hattermann, Fabian the Andes-Amazon transition according to the TRMM precipitation
Drenkhan and the anonymous reviewers for their valuable comments. radar and in situ data. Journal of Geophysical Research: Atmospheres.
doi:10.1002/2016JD026282
Chen, F.W. and Liu, C.W., 2012. Estimation of the spatial rainfall
Disclosure statement distribution using inverse distance weighting (IDW) in the middle
of Taiwan. Paddy and Water Environment, 10 (3), 209–222.
No potential conflict of interest was reported by the authors. doi:10.1007/s10333-012-0319-1
 HYDROLOGICAL SCIENCES JOURNAL 15

Costa, M.H. and Foley, J.A., 1997. Water balance of the Amazon Basin: Houze, R.A., et al., 2015. The variable nature of convection in the tropics
dependence on vegetation cover and canopy conductance. Journal of and subtropics: a legacy of 16 years of the tropical rainfall measuring
Geophysical Research, 102 (D20), 23973. doi:10.1029/97JD01865 mission satellite. Reviews of Geophysics, 53, 994–1021. doi:10.1002/
Derin, Y., et al., 2016. Multiregional satellite precipitation products 2015RG000488.Received
evaluation over complex terrain. Journal of Hydrometeorology, 17 Huffman, G.J., et al., 2007. The TRMM multisatellite precipitation ana-
(6), 1817–1836. doi:10.1175/JHM-D-15-0197.1 lysis (TMPA): quasi-global, multiyear, combined-sensor precipitation
Dinku, T., et al., 2014. Combined use of satellite estimates and rain estimates at fine scales. Journal of Hydrometeorology, 8 (1), 38–55.
gauge observations to generate high-quality historical rainfall time doi:10.1175/JHM560.1
series over ethiopia. International Journal of Climatology, 34 (7), Huffman, G.J., Bolvin, D.T., and Nelkin, E.J., 2015. Integrated
2489–2504. doi:10.1002/joc.3855 multi-satellite retrievals for GPM (IMERG) technical
Ensor, L.A. and Robeson, S.M., 2008. Statistical characteristics of daily documentation. NASA/GSFC Code, 612 (February), 47. doi:10.1136/
precipitation: comparisons of gridded and point datasets. Journal of openhrt-2016-000469
Applied Meteorology and Climatology, 47 (9), 2468–2476. doi:10.1175/ Hunziker, S., et al., 2017a. Effects of undetected data quality issues on
2008JAMC1757.1 climatological analyses. Climate of the Past Discussions. 14 (May),
Erdin, R., Frei, C., and Künsch, H.R., 2012. Data transformation and 1–31. doi:10.5194/cp-2017-64
uncertainty in geostatistical combination of radar and rain gauges. Hunziker, S., et al., 2017b. Identifying, attributing, and overcoming
Journal of Hydrometeorology, 13 (4), 1332–1346. doi:10.1175/JHM- common data quality issues of manned station observations.
D-11-096.1 International Journal of Climatology, 37, 4131–4145. doi:10.1002/
Feng, L., et al., 2014. CUTOFF: a spatio-temporal imputation method. joc.5037
Journal of Hydrology, 519, 3591–3605. doi:10.1016/j.jhydrol.2014.11.012 Iguchi, T., et al., 2000. Rain-profiling algorithm for the TRMM precipi-
Funk, C., et al., 2015a. The climate hazards infrared precipitation with tation radar. Journal of Applied Meteorology, 39 (12), 2038–2052.
stations — a new environmental record for monitoring extremes. doi:10.1175/1520-0450(2001)040<2038:RPAFTT>2.0.CO;2
Scientific data, 2, 1–21. doi:10.1038/sdata.2015.66 Isotta, F.A., et al., 2014. The climate of daily precipitation in the Alps:
Funk, C., et al., 2015b. A global satellite-assisted precipitation development and analysis of a high-resolution grid dataset from
climatology. Earth System Science Data, 7 (2), 275–287. doi:10.5194/ pan-alpine rain-gauge data. International Journal of Climatology, 34
essd-7-275-2015 (5), 1657–1675. doi:10.1002/joc.3794
Garreaud, R., 1999. Multiscale analysis of the summertime precipitation Javier, F., et al., 2016. Intercomparison of improved satellite rainfall
over the Central Andes. Monthly Weather Review, 127 (5), 901–921. estimation with CHIRPS gridded product and rain gauge data over
doi:10.1175/1520-0493(1999)127<0901:MAOTSP>2.0.CO;2 venezuela. Atmósfera, 29 (4), 323–342. doi:10.20937/ATM.2016.
Garreaud, R.D., et al., 2009. Present-day South American climate. 29.04.04
Palaeogeography, Palaeoclimatology, Palaeoecology, 281 (3–4), Joyce, R.J., et al., 2004. CMORPH: a method that produces global
180–195. doi:10.1016/j.palaeo.2007.10.032 precipitation estimates from passive microwave and infrared data at
Getirana, A.C.V., et al., 2014. Water balance in the Amazon Basin from high spatial and temporal resolution. Journal of Hydrometeorology, 5,
a land surface model ensemble. Journal of Hydrometeorology, 15 (6), 487–503. doi:10.1175/1525-7541(2004)005<0487:CAMTPG>2.0.CO;2
2586–2614. doi:10.1175/JHM-D-14-0068.1 Keller, V.D.J., et al., 2015. CEH-GEAR: 1 Km resolution daily and
Goovaerts, P., 2000. Geostatistical approaches for incorporating eleva- monthly areal rainfall estimates for the UK for hydrological and
tion into the Spatial interpolation of rainfall. Journal of Hydrology, other applications. Earth System Science Data, 7 (1), 143–155.
228 (1–2), 113–129. doi:10.1016/S0022-1694(00)00144-X doi:10.5194/essd-7-143-2015
Grimes, D.I.F., Pardo-Igúzquiza, E., and Bonifacio, R., 1999. Kidd, C. and Huffman, G., 2011. Review global precipitation measurement.
Optimal areal rainfall estimation using raingauges and satellite Meteorological Applications, 353, 334–353. doi:10.1002/met.284
data. Journal of Hydrology, 222 (1–4), 93–108. doi:10.1016/S0022- Kingdom, U., 2014. Advances in the stochastic modeling of satellite-derived
1694(99)00092-X rainfall estimates using a sparse calibration dataset. Journal of
Gudmundsson, L., et al., 2012. Technical note: downscaling RCM pre- Hydrometeorology, 15 (5), 1810–1831. doi:10.1175/JHM-D-13-0145.1
cipitation to the station scale using statistical transformations. Lavado Casimiro, W.S., et al., 2012. Basin-scale analysis of rainfall and
A comparison of methods. Hydrology and Earth System Sciences, 16 runoff in Peru (1969–2004): Pacific, Titicaca and Amazonas
(9), 3383–3390. doi:10.5194/hess-16-3383-2012 Drainages. Hydrological Sciences Journal, 57 (4), 625–642.
Guimberteau, M., et al., 2012. Discharge simulation in the Sub-Basins of doi:10.1080/02626667.2012.672985
the Amazon using ORCHIDEE forced by new datasets. Hydrology and Lavado, W., et al., 2015. PISCO: peruvian Interpolated Data of the SENAMHI’s
Earth System Sciences, 16 (3), 911–935. doi:10.5194/hess-16-911-2012 Climatological and Hydrological Observations. Precipítación v1.0.
Gusev, Y.M., et al., 2017. Modelling river runoff and estimating its Li, M. and Shao, Q., 2010. An improved statistical approach to merge
weather-related uncertainty for 11 large-scale rivers located in differ- satellite rainfall estimates and raingauge data. Journal of Hydrology,
ent regions of the globe. Hydrology Research, 2013, nh2017015. 385 (1–4), 51–64. doi:10.1016/j.jhydrol.2010.01.023
doi:10.2166/nh.2017.015 Lindstr, J., et al., 2013. Spatiotemporal: An R package for spatio-temporal
Hamada, A., and Takayabu, Y.N., 2014. A removal filter for suspicious modelling of air-pollution. R package, 33.
extreme rainfall profiles in TRMM PR 2A25 version-7 data. Journal of Lussana, C., et al., 2018. SeNorge2 daily precipitation, an observational
Applied Meteorology and Climatology, 53, 1252–1271. doi:10.1175/ gridded dataset over Norway from 1957 to the present day. Earth
JAMC-D-13-099.1 System Science Data, 10 (1), 235–249. doi:10.5194/essd-10-235-2018
Hiemstra, P.H., Pebesma, E.J., and Twenho, C.J.W., 2009. Computers and Ly, S., Charles, C., and Degré, A., 2013. Different methods for spatial
geosciences real-time automatic interpolation of ambient gamma dose interpolation of rainfall data for operational hydrology and hydro-
rates from the Dutch radioactivity monitoring network. Computers & logical modeling at watershed scale: a review. Base, 17 (2), 392–406.
Geosciences, 35, 1711–1721. doi:10.1016/j.cageo.2008.10.011 Available from: http://popups.ulg.ac.be/1780-4507/index.php?id=
Hofstra, N., et al., 2009. Testing E-OBS European high-resolution 10003%5Cnhttp://www.scopus.com/record/display.uri?eid=2-s2.
gridded data set of daily precipitation and surface temperature. 0-84878998500&origin=inward&txGid=0
Journal of Geophysical Research: Atmospheres, 114 (November). Mantas, V.M., et al., 2015. Validation of TRMM multi-satellite precipitation
doi:10.1029/2009JD011799 analysis (TMPA) products in the Peruvian Andes. Atmospheric Research,
Hofstra, N., New, M., and McSweeney, C., 2010. The influence of inter- 163, 132–145. Elsevier B.V. doi:10.1016/j.atmosres.2014.11.012
polation and station network density on the distributions and trends Manz, B., et al., 2016. High-resolution satellite-gauge merged precipitation
of climate variables in gridded daily data. Climate Dynamics, 35 (5), climatologies of the Tropical Andes. Journal of Geophysical Research:
841–858. doi:10.1007/s00382-009-0698-1 Atmospheres, 121, 1190–1207. doi:10.1002/2015JD023788.Received
16 C. AYBAR ET AL.

Manz, B., et al., 2017. Comparative ground validation of IMERG and TMPA Sun, Q., et al., 2017. A review of global precipitation datasets: data
2 at variable spatio-temporal scales in the Tropical Andes. Journal of sources, estimation, and intercomparisons. Reviews of Geophysics,
Hydrometeorology, 18, 2469–2489. doi:10.1175/JHM-D-16-0277.1 79–107. doi:10.1002/2017RG000574
Marengo, J.A., et al., 2012. Recent developments on the South American Teegavarapu, R.S.V. and Nayak, A., 2017. Evaluation of long-term trends
monsoon system. International Journal of Climatology, 32 (1), 1–21. in extreme precipitation: implications of in-filled historical data use
doi:10.1002/joc.2254 for analysis. Journal of Hydrology, 550, 616–634. doi:10.1016/j.
Nerini, D., et al., 2015. A comparative analysis of TRMM-rain gauge data jhydrol.2017.05.030
merging techniques at the daily time scale for distributed Teng, H., et al., 2014. Estimating spatially downscaled rainfall by regres-
rainfall-runoff modelling applications. Journal of Hydrometeorology, sion kriging using TRMM precipitation and elevation in Zhejiang
150904104740009. doi:10.1175/JHM-D-14-0197.1 Province, Southeast China. International Journal of Remote Sensing,
Nesbitt, S.W. and Anders, A.M., 2009. Very high resolution precipitation 35 (22), 7775–7794. doi:10.1080/01431161.2014.976888
climatologies from the tropical rainfall measuring mission precipita- Thouret, J.C., et al., 2013. Combining criteria for delineating lahar-and
tion radar. Geophysical Research Letters, 36 (15), 1–5. doi:10.1029/ flash-flood-prone hazard and risk zones for the City of Arequipa,
2009GL038026 Peru. Natural Hazards and Earth System Science, 13 (2), 339–360.
New, M., Hulme, M., and Jones, P., 2000. Representing twentieth- doi:10.5194/nhess-13-339-2013
century space – time climate variability. Part II: development of Tozer, C.R., Kiem, A.S., and Verdon-Kidd, D.C., 2012. On the uncer-
1901 – 96 monthly grids of terrestrial surface climate. Journal of tainties associated with using gridded rainfall data as a proxy for
Climate, 13 (13), 2217–2238. doi:10.1175/1520-0442(2000)013<2217: observed. Hydrology and Earth System Sciences, 16 (5), 1481–1499.
RTCSTC>2.0.CO;2 doi:10.5194/hess-16-1481-2012
Newman, A.J., et al., 2015. Gridded ensemble precipitation and tem- Ulloa, J., et al., 2017. Two-step downscaling of Trmm 3b43 V7 precipita-
perature estimates for the contiguous United States. Journal of tion in contrasting climatic regions with sparse monitoring: the case
Hydrometeorology, 16 (6), 2481–2500. doi:10.1175/JHM-D-15-0026.1 of ecuador in Tropical South America. Remote Sensing, 9 (7), 758.
Nicolas, J.P. and Bromwich, D.H., 2011. Precipitation changes in high doi:10.3390/rs9070758
southern latitudes from global reanalyses: a cautionary tale. Surveys in van Osnabrugge, B., Weerts, A.H., and Uijlenhoet, R., 2017. GenRE:
Geophysics, 32 (4–5), 475–494. doi:10.1007/s10712-011-9114-6 a method to extend gridded precipitation climatology data sets in
Notivoli, R.S., et al., 2017. Reconstrucción e incertidumbre en series de near real-time for hydrological forecasting purposes. Water Resources
precipitación diaria instrumental. X Congreso Internacional AEC: Research, 53 (11), 9284–9303. doi:10.1002/2017WR021201
Clima, Sociedad, Riesgos y Ordenación Del Territorio, 397–406. Vera, C., et al., 2006. The South American low-level jet experiment.
Ochoa, A., et al., 2014. Evaluation of TRMM 3B42 precipitation esti- Bulletin of the American Meteorological Society, 87 (January), 63–77.
mates and WRF retrospective precipitation simulation over the doi:10.1175/BAMS-87-1-63
Pacific-Andean Region of Ecuador and Peru. Hydrology and Earth Verdin, A., et al., 2015. A bayesian kriging approach for blending
System Sciences, 18 (8), 3179–3193. doi:10.5194/hess-18-3179-2014 satellite and ground precipitation observations. Water Resources
Parmentier, B., et al., 2015. Using multi-timescale methods and Research, 1359–1371. doi:10.1002/2014WR015432.Received
satellite-derived land surface temperature for the interpolation of Vicente-Serrano, S.M., et al., 2010. A complete daily precipitation data-
daily maximum air temperature in Oregon. International Journal of base for Northeast Spain: reconstruction, quality control, and
Climatology, 35 (13), 3862–3878. doi:10.1002/joc.4251 homogeneity. International Journal of Climatology, 30 (8),
Perdigón-Morales, J., et al., 2017. The midsummer drought in Mexico: 1146–1163. doi:10.1002/joc.1850
perspectives on duration and intensity from the CHIRPS precipitation Vicente-Serrano, S.M., et al., 2017. Recent changes in monthly surface
database. International Journal of Climatology. doi:10.1002/joc.5322 air temperature over Peru, 1964–2014. International Journal of
Peterson, T.C., et al., 1998. Homogeneity adjustments of in of in situ Climatology. doi:10.1002/joc.5176
atmospheric climate data: a review. International Journal of Wilks, D.S., 2006. Statistical methods in the atmospheric sciences. 2nd ed.
Climatology, 1517, 1493–1517. WMO, 1994. Guide to hydrological practices. 5th ed. Geneva,
Rudorff, C.M., Melack, J.M., and Bates, P.D., 2014. Flooding dynamics Switzerland: World Meteorological Organization.
on the lower amazon floodplain: 2. seasonal and interannual hydro- Yanto, B.L., et al., 2017. Development of a gridded meteorological
logical variability. Water Resources Research, 50 (1), 635–649. dataset over Java Island, Indonesia 1985–2014. Scientific Data, 4,
doi:10.1002/2013WR014714 170072. doi:10.1038/sdata.2017.72
Saha, S., et al., 2014. The NCEP climate forecast system version 2. Yao, J., et al., 2016. Annual actual evapotranspiration in Inland River
Journal of Climate, 27 (6), 2185–2208. doi:10.1175/JCLI-D-12-00823.1 catchments of China based on the Budyko framework. Stochastic
Saha, S., Moorthi, S., and Pan, H.-L., 2010. The NCEP climate forecast Environmental Research and Risk Assessment. doi:10.1007/s00477-
system reanalysis. Bulletin of the American Meteorological Society, 91 016-1271-1
(August), 1015–1058. doi:10.1175/2010BAMS3001.1 Yatagai, A., et al., 2012. Aphrodite constructing a long-term daily
Schuurmans, J.M., et al., 2007. Automatic prediction of high-resolution gridded precipitation dataset for Asia based on a dense network of
daily rainfall fields for multiple extents: the potential of operational rain gauges. Bulletin of the American Meteorological Society, 93 (9),
radar. Journal of Hydrometeorology, 8 (6), 1204–1224. doi:10.1175/ 1401–1415. doi:10.1175/BAMS-D-11-00122.1
2007JHM792.1 Zambrano-Bigiarini, M., et al., 2017. Temporal and spatial evaluation of
Sorooshian, S., et al., 2000. Evaluation of PERSIANN system satellite-based rainfall estimates across the complex topographical and
satellite-based estimates of tropical rainfall. Bulletin of the American climatic gradients of Chile. Hydrology and Earth System Sciences, 21
Meteorological Society, 81 (9), 2035–2046. doi:10.1175/1520-0477- (2), 1295–1320. doi:10.5194/hess-21-1295-2017
(2000)081<2035:EOPSSE>2.3.CO;2 Zubieta, R., et al., 2015. Impacts of satellite-based precipitation datasets
Stekhoven, D.J. and Bühlmann, P., 2012. MissForest — non-parametric on rainfall-runoff modeling of the Western Amazon Basin of Peru
missing value imputation for mixed-type data. Bioinformatics, 28 (1), and Ecuador. Journal of Hydrology, 528, 599–612. Elsevier B.V.
112–118. doi:10.1093/bioinformatics/btr597 doi:10.1016/j.jhydrol.2015.06.064
Strauch, M., et al., 2016. Adjustment of global precipitation data for enhanced Zulkali, Z., et al., 2013. A critical assessment of the JULES land surface
hydrologic modeling of Tropical Andean watersheds. Climatic Change, model hydrology for humid tropical environments. Hydrology and Earth
1–14. doi:10.1007/s10584-016-1706-1 System Sciences, 17 (3), 1113–1132. doi:10.5194/hess-17-1113-2013