Professional Documents
Culture Documents
ABSTRACT: Understanding the spatio-temporal variability of rainfall over mountainous regions such as the Andes is crucial
for the maintenance of water resources and ecosystems. This study provided a comprehensive analysis of the signal, statistical
significance and spatial pattern of rainfall trends in central and southern Peruvian Andes (CSPA) from 1965 to 2010. Rainfall
from 47 quality-controlled rain gauge stations was examined using the hydrologic calendar year. Total and extreme rainfall
indices proposed by the Expert Team on Climate Change Detection and Indices (ETCCDI) were calculated and trends were
examined with Mann–Kendall test and Sen’s slope estimator. Significant regional patterns of changes in rainfall extremes were
investigated and compared with previous studies. Four major regions of stations were identified based on principal component
analysis and clustering techniques: Amazon, central Pacific, southern Pacific and Titicaca basins. Statistically significant trend
patterns showed that the annual total rainfall has decreased in the Amazon basin, despite the increase in the number of rainy
days and some extreme rainfall indices. Decrease in 1-day and 5-day yearly maximum rainfall was observed in central Pacific,
along with an increase in the number of wet days. Positive trends in indices related to the intensity of very strong daily rainfall
were detected in southern Pacific. Titicaca basin showed an increase in the intensity of rainfall extremes. Rainfall variability
and trends were evaluated during contrasting El Niño-Southern Oscillation (ENSO) and Pacific Decadal Oscillation (PDO).
Most stations in the CSPA exhibited positive (negative) anomaly of total and extreme rainfall indices (consecutive dry days)
during La Niña (El Niño) years. Positive (negative) anomalies of annual total rainfall and consecutive dry days during positive
(negative) PDO occurred in the Amazon, southern Pacific and Titicaca basins, respectively. The ENSO–PDO conditions may
have influenced the complex and mostly non-statistically significant long-term trends in CSPA.
KEY WORDS climate change; rainfall trends; precipitation changes; extreme rainfall; extreme events; Andes; Altiplano; Peru
Received 3 September 2016; Revised 28 September 2017; Accepted 23 December 2017
distinctly modulates rainfall over central Peruvian Andes observation noise (Zhang et al., 2011). To evaluate impacts
(Garreaud et al., 2003) and southern Peruvian Andes of climate change on rainfall, the Expert Team on Cli-
(Vuille and Keimig, 2004; Perry et al., 2014). Lagos et al. mate Change Detection and Indices (ETCCDI) – a joint
(2008) found weakly positive, neutral and moderately neg- project between the World Meteorological Organization
ative correlation coefficients between sea surface temper- Commission for Climatology (CCI) and the World Climate
ature (SST) anomaly in the Niño3.4 region (5∘ N–5∘ S, Research Program Climate Variability and Predictability
120∘ –170∘ W) and rainfall in the northern, central and (CLIVAR) program – suggested the use of rainfall indices
southern Peru, respectively. Lavado and Espinoza (2014) (Haylock et al., 2006; Zhang et al., 2011). These indices
showed that rainfall increases during strong El Niño (EN) were developed to be non-region-specific and independent
events in the northern Pacific, decreases (increases) dur- (Folland et al., 1999; Nicholls and Murray, 1999).
ing strong EN (La Niña (LN)) in the southern Pacific Rainfall trends and extremes have been investigated in
and decreases during strong EN in the Titicaca basin. parts of the Peruvian Andes using gridded data and rainfall
Nonetheless, ENSO teleconnections and impacts on rain- indices. For instance, Alexander et al. (2006) used avail-
fall in the Peruvian Andes seem to vary over time. Bour- able daily rainfall from stations worldwide (3.75 × 2.5∘
rel et al. (2014) showed that with exception of the strong resolution) and showed a decrease in the number of heavy
1982/1983 and 1997/1998 ENSO events, the region expe- rainfall days (annual count of days when daily rainfall is
rienced increased (decreased) rainfall during EN years of greater or equal than 10 mm) in CSPA during 1951–2003.
the period 1964–1999 (2000–2011). Conversely, Donat et al. (2013) employed global gridded
The Pacific Decadal Oscillation (PDO) is the leading monthly and annual indices [the Hadley Centre Global
mode of monthly SST anomalies in the Pacific Ocean Climate Extremes Index 2 (HadEX2) at 3.75 × 2.5∘ reso-
north of 20∘ N and affects South American climate on lution] and found no trends in the number of heavy rainfall
decadal timescales (Mantua and Hare, 2002; Seiler days over the Andes during 1951–2010. However, Donat
et al., 2013; Newman et al., 2016). Negative (positive) et al. (2014) performed an inter-comparison of multiple
correlations between rainfall anomalies and PDO have global gridded observational and reanalysis data sets of
been observed in South America north (south) of ∼10∘ S rainfall extremes and showed a decrease in the HadEX2
(Garreaud et al., 2009). Total annual rainfall and number annual maximum consecutive 5-day rainfall (maximum
of extreme events in the Bolivian lowlands are higher 5-day rainfall) over central Peru during 1979–2008. Morin
during positive PDO phase (Seiler et al., 2013). That (2011) analysed the 0.5∘ monthly data set developed by
study also showed that monthly and annual standardized Beck et al. (2005), the Variability Analyses of Surface
rainfall anomalies increased from ∼1965 until 1984 and Climate Observations (VASClimO), based on the Global
decreased from ∼1985 until 2004, following the basic Precipitation Climatology Centre (GPCC), and found
structure of the PDO index. They also found that up to statistically significant negative trend in the total annual
15% of the 68 stations analysed in Bolivia have increased rainfall over CSPA during 1951–2000.
(decreased) total and extreme rainfall during 1965–1984 Using point-wise (gauge) data to investigate trends can
(1985–2009). Marengo (2004) associated the shift from improve the details about the spatial variability of local
negative to positive rainfall anomalies over the southern rainfall changes in regions with complex terrain such as
Amazon of Brazil in the 1970s with the switch from the the Peruvian Andes. For instance, Lavado et al. (2013)
cold to the warm PDO phase in 1976/1977. Carvalho et al. investigated monthly rainfall in 48 Andean and Amazon
(2011) showed that the amplitude of the SAMS increased rain gauge stations during 1965–2007 and found statisti-
after the 1976/1977 climate shift with impacts on rainfall. cally significant decrease (increase) of monthly rainfall in
Few studies have examined trends in rainfall in the Peru- only four (three) stations. Other studies also used gauged
vian Andes and they have used a limited number of gauge data at daily timescales to calculate ETCCDI indices and
stations. For example, using observed rainfall, paleocli- investigate rainfall trends over CSPA. For example, Vuille
mate data and global climate models Neukom et al. (2015) et al. (2003) found a statistically significant increase of
and Minvielle and Garreaud (2011) found a long-term total annual rainfall in some stations in central Andes
negative rainfall trend in the central and southern Peru- but non-significant decrease of total annual rainfall in the
vian Andes (CSPA), which seemed associated with the southern Peruvian Andes. Haylock et al. (2006) found a
strengthening of the upper-tropospheric westerly winds. significant increase of consecutive dry days in southern
Some other potential mechanisms have been proposed to Andes but a non-significant decrease of total and extreme
explain the negative rainfall trends in the Andean-Amazon rainfall in central and southern Andes. Marengo et al.
region such as the warming of the North Tropical Atlantic (2009) showed a significant increase of dry consecutive
Ocean (Yoon and Zeng, 2010; Espinoza et al., 2011; days in the southern Andes whereas Skansi et al. (2013)
Marengo et al., 2011) and changes in the characteristics of found significant drier conditions over the southern Andes
the SAMS in recent decades (Arias et al., 2015). All these and a decrease of extreme rainfall in central Andes.
mechanism are relevant for trends in extreme rainfall in the These previous studies suggested that the most con-
region. sistent trend in rainfall is likely the intensification of
The most appropriate way to detect changes in climate dry conditions over the southern Andes. However, these
extremes is by using a set of indices that are statisti- previous studies also revealed a large degree of spatial and
cally robust, cover a variety of climates and minimize temporal variability of rainfall in the Peruvian Andes and
anomalies in the Niño3.4 region based on centred 30-year edu/pdo/. Based on the PDO index, positive (negative)
base periods updated every 5 years. Events are defined as PDO phases occurred during 1977–1997 and 2002–2006
five consecutive overlapping 3-month periods at or above (1965–1976, 1998–2001 and 2007–2010).
the +0.5∘ ONI anomaly for warm (EN) events and at or
below the −0.5 ONI anomaly for cold (LN) events. Based 3. Methods
on this criterion, 17 EN and 14 LN years occurred during
1965–2010. 3.1. Regionalization of stations
PDO conditions were defined using the PDO index, Principal component analysis (PCA) was performed using
which is the leading principal component (PC) of North a correlation matrix obtained from the 47 gauged daily
Pacific monthly SST variability (Zhang et al., 1997), and rainfall time series (after removing the mean annual
it was obtained from http://research.jisao.-washington. cycle) to calculate the eigenvalues and eigenvectors
Figure 2. Mean (cross) and standard deviation (error bar) of gauged Figure 4. Cluster analysis mean Silhouette values after using different
1965–2010 monthly rainfall of the 47 central and southern Peruvian number of PC loadings and increasing number of clusters. [Colour figure
Andean stations analysed in this study. can be viewed at wileyonlinelibrary.com].
Table 2. Definition of the 12 total and extreme rainfall indices proposed by the ETCCDI.
of having any trend. The statistic S (Mann, 1945; Kendall, The MK test is affected by serial correlation. Positive
1975) is defined as: serial correlation in the time series leads to an underestima-
tion of the sampling variance. As a consequence, the statis-
∑ ∑
n−1 n
( ) tic Z (Equation (5)) increases in absolute value, yielding a
S= sgn xj − xi (2)
smaller p values and falsely rejecting the null hypothesis
i=1 j=i+1
(Wilks, 2011). There are various approaches to adjust the
where xi and xj are the sequential data, n is the total number test to avoid the first-type error caused by autocorrelation
of data in the time series and effects: (1) pre-whitening, (2) trend-free pre-whitening, (3)
variance correction and (4) block resampling techniques.
⎧ 1, Δx > 0 In this study, we used the variance correction approach pro-
⎪
sgn (Δx) = ⎨0, Δx = 0 (3) posed by Hamed and Rao (1998). Therefore, the variance
⎪−1, Δx < 0 of this test statistic S was corrected by using an effective
⎩
sample size. The modified variance of the MK test statistic
A positive (negative) value of S indicates a positive is given by:
(negative) trend. For n ≥ 8, the statistic S is approximately Var∗ (S) = Var (S) ∗ CF (6)
normally distributed with a mean E(S) = 0 and variance
where Var(S) is the variance of the MK test statistic S for
(Wilks, 2011):
the original sample data and CF is a correction factor (Yue
∑
m ( ) and Wang, 2004) calculated as:
n (n − 1) (2n + 5) − ti ti − 1)(2ti + 5
Var (S) =
i=1
(4) ∑
n−1
rk (1 − k)
18 CF = 1 + 2 (7)
k=1
n
where n is the number of data points, m is the number of
tied groups and ti denotes the number of ties of extent i. A where rk is the lag-k serial correlation coefficient of data.
tied group is a set of sample data having the same value. In this study, we assumed that the time series was ade-
In cases where the sample size n > 10, the standard normal quately described by an autoregressive process of order
test statistic Z is (Wilks, 2011): one. Then, the corrected variance is:
( )
S−1 1 + r
⎧√ , S>0 Var∗ (S) = Var (S) (
1
) (8)
⎪ Var(S) 1 − r1
Z = ⎨0, S=0 (5)
⎪ √S+1 , S<0 where r1 is the autocorrelation of the detrended time series
⎩ Var(S)
(Wilks, 2011). The lag-1 serial correlation coefficient (r1 )
At a significance level ∝, the null hypothesis of no trend was computed as:
is rejected if the absolute value of Z is greater than the the- n−1 ( ( )) ( ( ))
1 ∑
oretical value Z1−∝L∕2 , which is obtained from the standard xi − E xi xi+1 − E xi
n−1
normal distribution table. In this study, a significance level r1 =
i=1
(9)
n ( ( ))2
𝛼 = 0.05 was used. Thus, the null hypothesis of no trend i ∑
xi − E xi
was rejected if |Z| > 1.96. n
i=1
where E(xi ) is the mean of sample data and n is the (14) independent EN (LN) events that occurred during
sample size. To test the significance of the autocorrelation, 1965–2010. Furthermore, to assess the strength and signal
a one-tailed test (alternative hypothesis is that true r1 is of the relationship between ENSO and rainfall indices over
greater than zero) was used and the critical value of r1 for time we also evaluated the anomalies of rainfall indices
a 5% significance level was computed as: during the periods 1965–1976 (P1), 1977–1999 (P2) and
√ 2000–2010 (P3) separately (excluding the strong EN of
′ −1 + 1.645 n − 2 1982–1983 and 1997–1998) and calculated the percent-
r1 = (10)
n−1 age of stations with different anomalies per each CSPA
Once a statistically significant autocorrelation of the region. P1, P2 and P3 were the same periods evaluated in
time series was found, Var*(S) was used instead of Var(S) Bourrel et al. (2014).
in Equation (5) to estimate Z. By performing that, we To evaluate the effect of PDO on rainfall indices in
removed the influence of serial correlation in the test’s CSPA, we calculated the anomalies of ETCCDI indices
statistics. during positive and negative PDO phases of the period
Additionally, the trend slope was calculated using the 1965–2010 for each CSPA region of stations. The signif-
method of Sen (1968). The magnitude of the slope of the icance of the anomalies were assessed based on a z-test
trend was estimated as (Gocic and Trajkovic, 2013): at 95% confidence level considering two (three) positive
xj − xi (negative) independent PDO events during 1965–2010.
Qo = for o = 1, … , N (11)
j−i
where xi and xj are data at time i and j (j > i), respectively. 4. Results
If there are n values in the time series, then N = n(n − 1)/2
4.1. Regional characterization of stations
slope estimates are possible. Then, the N values of Qo are
ranked from smallest to largest and the median of slope or The regionalization of stations described in Section 3.1
Sen’s slope estimator is: generated four major regions (denoted by the letters A,
B, C and D) and 14 subregions (denoted by the num-
⎧ bers 1–14) mapped in Figure 5. Fourteen subregions are
⎪Q[(N+1)∕2] , if N is odd
Qmed = ⎨ QN∕2 +Q[ N+2 ] (12) embedded into the four major regions: subregions 1, 2
⎪
2
, if N is even and 3 are embedded in region A; subregions 4, 5, 6 and
⎩ 2
7 in region B; subregions 8, 9, 10 and 11 in region C;
The Qmed sign reflects the data trend, while its value
indicates the steepness of the trend. To determine whether
the median slope is statistically different from zero, one
should obtain the confidence interval of Qmed at a specific
probability (Hollander and Wolfe, 1973; Gilbert, 1987) as:
√
C∝ = Z1−∝∕2 Var (S) (13)
where Var(S) is defined in Equation (4) or Var*(S) if the
time series is statistically significant autocorrelated, and
Z 1 − ∝ /2 is obtained from the standard normal distribu-
tion table. Then, M 1 = (N − C∝ )/2 and M 2 = (N + C∝ )/2
are computed. The lower and upper limits of confidence
interval, Qmin and Qmax , are the M 1 th largest and (M 2 + 1)th
largest of the N ordered slope estimates. The slope Qmed is
statistically different than zero if the two limits (Qmin and
Qmax ) have similar sign. To conclude that a time series has
a statistically significant trend, both the MK test and the
Qmed Sen’s slope have to be statistically significant; then
the sign and slope of that trend is given by Sen’s slope cal-
culations.
Figure 7. As in Figure 2, but for each subregion of the central and southern Peruvian Andean stations located in regions: (a) A (Amazon basin), (b)
B (central Pacific basin), (c) C (southern Pacific basin) and (d) D (Titicaca basin). [Colour figure can be viewed at wileyonlinelibrary.com].
Figure 8. Mean (dashed) and standard deviation (error bar) of (a) latitude and longitude, (b) altitude, (c) slope and (d) aspect from all stations in
each major CSPA region. The major regions are A (Amazon Basin), B (central Pacific Basin), C (southern Pacific Basin) and D (Titicaca Basin).
Note: The aspect corresponds to a surface facing north (0–22.5∘ ), northeast (22.5∘ –67.5∘ ), east (67.5∘ –112.5∘ ), southeast (112.5∘ –157.5∘ ), south
(157.5∘ –202.5∘ ), southwest (202.5∘ –247.5∘ ), west (247.5∘ –292.5∘ ), northwest (292.5∘ –337.5∘ ) and north (337.5∘ –360∘ ). [Colour figure can be
viewed at wileyonlinelibrary.com].
Figure 9. As in Figure 8, but for the 14 central and southern Peruvian Andean subregions. [Colour figure can be viewed at wileyonlinelibrary.com].
Figure 10. Mean (dashed) and standard deviation (error bar) of (a) PRCPTOT, (b) Wdays, (c) CDD and (d) R95p of the 14 central and southern
Peruvian Andean subregions.
and altitude are the dominant factors for different rain- Table 3. Percentage of stations over CSPA with different trends
fall regimes in the Andes and that the method can cap- when using the hydrologic calendar rather than the Western
ture these differences. The 14 subregions have different calendar years for the calculation of rainfall indices during
mean latitude, longitude and altitude and smaller standard 1965–2009 hydrologic years.
deviation among stations (Figure 9) compared to the four Index Stations (%)
major regions (Figure 8). Each of the 14 subregions is
characterized by a unique combination of latitude and lon- PRCPTOT 15.1
SDII 20.8
gitude (Figure 9(a)) and altitude (Figure 9(b)) ranges. Alti-
CDD 41.5
tude varies greatly among subregions (Figure 9(b)). For CWD 20.8
instance, subregion 1 has the greatest altitude in region A; R10 20.8
subregion 5 in region B; and subregion 10 in region C. The R20 9.4
slope (Figure 9(c)) is more variable mainly in subregions R95p 32.1
located over the Amazon basin (1–3) and the Pacific basin R99p 34.0
(4–11), comparatively to the subregions located in the Titi- RX1day 24.5
RX5day 17.0
caca basin (12–14). The aspect (Figure 9(d)) varies greatly R95pTOT 26.4
among stations of the same subregion, except for the sub- R99pTOT 24.5
regions 3 and 13 (5 and 8) which have most stations facing Wdays 11.3
east (west).
Subregions also exhibit differences in total and extreme
rainfall indices. Figure 10 shows the mean and standard and the Amazon basins, and no evident changes in rainfall
deviation of some rainfall indices such as the PRCPTOT,
with altitude in the Pacific basin. Furthermore, stations
Wdays, CDD and R95p. The greatest amounts of PRCP-
facing east have greater amounts of rainfall as a result
TOT are evident in the subregions 1, 5, 10 and 13, which
of the enhanced transport of moisture east of the Andes
also show large Wdays and R95p but low CDD. Con-
(Garreaud, 2009). This is evident in subregion 13 whose
versely, the smallest amounts of PRCTOT are observed in
stations face east and consequently exhibit the greatest
subregions 6, 7, 8, 9 and 11 are also related to compara-
PRCPTOT among all other subregions embedded in
tively low Wdays and R95p, and great CDD. Interestingly,
region D (Titicaca basin).
subregions with the highest altitudes (such as subregions
1, 5, 10, 12, 13 and 14) show the greatest amounts of
PRCPTOT, Wdays and R95p. This relationship between 4.2. Trends in rainfall indices
altitude and rainfall over CSPA differs from the patterns To evaluate the effect of using a specific calendar on the
obtained in Lavado Casimiro et al., 2012, who described trend analysis of rainfall indices, we calculated the per-
a decrease of monthly rainfall with altitude in the Titicaca centage of stations that exhibited different trends (either
the signal, the statistical significance or both) when the Skansi et al. (2013) and our study showed non-significant
hydrologic year is replaced with the Western calendar trends for Progreso station located in the Titicaca basin.
(Table 3). The indices that are most sensitive to the use of With respect to extreme rainfall indices, there is a much
hydrologic year calendar are the CDD and the R99p as 41.5 larger discrepancy in trends among stations and studies.
and 34% of stations, respectively, exhibit different trends. For instance, Huayao and Marcapomacocha stations in the
The least sensitive index is the number of R20, the PRCP- Amazon basin exhibited statistically significant decrease
TOT and the annual percentage of wet days (Wdays), with of RX1day and RX5day in the current study in contrast to
only 9.4, 15.1 and 11.3% of stations, respectively, exhibit- Haylock et al. (2006) and Skansi et al. (2013) that found
ing a different trend after using hydrologic years. non-statistically significant trends. The possible reasons
Stations that showed statistically significant trends in for these discrepancies are different lengths of time series,
rainfall indices are displayed in Table 4. In region A (the the adoption of different calendars to calculate the rainfall
Amazon basin), approximately ∼20–40% out of a total of indices and the slightly different approaches used to assess
nine stations exhibited negative trend in PRCPTOT, SDII, trends.
CWD, R10, RX1day, RX5day and Wdays; positive trends
were observed for R95pTOT, R99pTOT and Wdays. In 4.3. ENSO and PDO
region B (the central Pacific basin), about 30–40% out
of 14 stations indicated negative trends in SDII, CDD, While ENSO plays a significant role in the inter-annual
RX1day and RX5day whereas ∼20% of stations showed a variability of some of the indices calculated for the CSPA,
positive trend in Wdays. In region C (the southern Pacific the magnitude of this effect depends on the region and
basin), ∼20% of the 13 stations showed a negative trend on the considered period. Figure 12 shows PRCPTOT,
in CDD and a positive trend in RX1day. In region D Wdays, CDD and R95p anomalies during EN and LN
(the Titicaca basin), ∼30% of the 11 stations exhibited events of 1965–2009 water years (excluding the strong EN
a positive trend in RX1day. In general, we observe that of 1982–1983 and 1997–1998). During LN (EN) years,
statistically significant trends in the indices are consistent positive (negative) anomalies of PRCPTOT, Wdays and
among indices for most stations. R95p in all CSPA regions are evident, despite the large
The spatial variability of the signal along with the variability of R95p in regions A and B where the differ-
statistical significance of trends of 4 out of the 12 rainfall ence in mean anomalies are not statistically significant.
indices are shown in Figure 11. Trends exhibited a large Accordingly, positive (negative) anomalies of CDD occur
degree of variability over CSPA stations, and trends in during EN (LN) years in all CSPA regions. The rainfall
most stations were not statistically significant. Notice that indices anomalies calculated during 1965–2010 including
stations within the same major region or even within the the strong EN of 1982–1983 and 1997–1998 are slightly
same subregion may show different trends. Coherent spa- greater in magnitude due to the enhanced rainfall signal
tial trends were only observed in certain subregions and during strong ENSO events (not shown).
for some indices as, for instance, a decrease in PRCPTOT Table 6 shows the percentage of stations in each CSPA
over subregion 2, an increase of RX1day over subregion region with positive anomalies in PRCPTOT, Wdays,
4 and a decrease of RX5day over subregion 6. Possible CDD and R95p during EN and LN years separated in
reasons for these discrepancies could be the interactions the three periods 1965–1976 (P1), 1977–1999 (P2) and
of different atmospheric forcings with the complex orog- 2000–2010 (P3). Interestingly, the influence of ENSO
raphy of the study region, resulting in a large spatial on the calculated indices is not uniform throughout time.
variability of rainfall and rainfall trends. Nevertheless, the More (fewer) stations in regions A and D show positive
existence or not of statically significant trends should be anomalies in PRCPTOT, Wdays and R95p (CDD) in EN
interpreted with caution given the period of data availabil- events during the period P3 than during the periods P1
ity. For instance, Morin (2011) found that annual rainfall and P2. More stations in regions B and C exhibit positive
trends over the Tropics have a tendency to be undetectable anomalies in PRCPTOT, Wdays and R95p in LN events
because of high rainfall means and variability. In that study during periods P2 and P3 than during P1. Our results for
the author investigated rainfall trends in central Peruvian regions B and D corroborate to some extent with the results
Andes during 1951–2000 and showed that trends in the in Bourrel et al. (2014) and Vuille et al. (2000). Bourrel
eastern (western) side of the central Peruvian Andes were et al. (2014) described an increased rainfall pattern dur-
detected only where there was a change of about 20–40% ing EN of the period 1964–1999 compared to 2000–2011,
(40–75%) of the annual rainfall with respect to the mean and our results show that ∼60% (∼40%) of the stations
annual rainfall in the same period. These conditions are in region B have positive rainfall anomalies during EN of
hardly observed for the stations evaluated over CSPA. the period 1965–1976 (2000–2010). Vuille et al. (2000)
A comparison between trends of six representative rain- found below average rainfall during 1965–1990 in south-
fall indices obtained by Haylock et al. (2006), Skansi et al. ern Peruvian Andes during EN, while our results show
(2013) and the current study at common stations located that less than half of the stations in region D have positive
over CSPA is shown in Table 5. Regarding PRCPTOT, rainfall anomalies during 1965–1999 and that this num-
we found a statistically significant decrease in the Ama- ber of stations slightly increases in 2000–2010. Notice that
zon basin stations, while Haylock et al. (2006) and Skansi ENSO conditions are defined in this study using ONI and
et al. (2013) found a non-significant negative trend. Both does not account for different ENSO regimes (as defined
Index Region A: Amazon basin Region B: central Pacific basin Region C: southern Pacific basin Region D: Titicaca basin
Negative trend Positive trend Negative trend Positive trend Negative trend Positive trend Negative trend Positive trend
RX1day Huayao, Marcapomacocha Pisac Antioquia, Pariacancha, Huamantanga, Pirca Ichuna Chivay, Las – Chilligua,
Puquio, Santiago De Tuna Salinas, Talabaya Pampahuta,
Taraco
RX5day Huayao, Marcapomacocha, – Antioquia, Huangascar, Huamantanga, Parquin – Chivay, Las Huancane Chilligua
Pisac Pariacancha, Santiago De Salinas
Tuna
R95pTOT, Tanta Huayao, Parquin, Santiago De Tuna – – – – –
R99pTOT Marcapomacocha
Wdays Marcapomacocha, Pilchaca La Quinua, Tanta, – Pirca, San Lazaro De – – Progreso –
Urubamba,Yantac Escomarca, Santiago De
Tuna
Rainfall indices are denoted as PRCPTOT, SDII, CDD, CWD, R10, R20, R95p, R99p, RX1day, RX5day, R95pTOT, R99pTOT and Wdays.
Figure 11. Trends in total and extreme rainfall indices over CSPA during 1965–2009 water years: (a) PRCPTOT, (b) CDD, (c) R95p and (d) Wdays.
[Colour figure can be viewed at wileyonlinelibrary.com].
in Takahashi et al., 2011), which could eventually result in respectively, had positive (negative) trends in PRCPTOT.
regional differences. This investigation is beyond the scope During 1977–1997, 22, 36, 85 and 9% (78, 64, 15 and
of the present study. 91%) of stations in regions A, B, C and D, respectively,
PDO contrasting phases are associated with statistically exhibit positive (negative) trend in PRCPTOT. It is notice-
significant differences in mean anomalies of PRCPTOT in able that the region that showed greater dependence on
region D and CDD in regions A, C and D (Figure 13). PDO phase is region D (Titicaca basin). Even though
Positive (negative) anomalies of PRCPTOT in regions D this region showed positive (negative) PRCPTOT anoma-
are evident during positive (negative) PDO. Positive (neg- lies during positive (negative) PDO (Figure 13), most sta-
ative) anomalies in CDD are observed in regions A, C and tions show negative (positive) trends during the period
D during positive (negative) PDO. The positive anomalies 1977–1997 (1965–1976) which corresponds to a positive
of total annual rainfall in region D (Titicaca basin or Alti- (negative) PDO phase.
plano) during positive PDO agree with previous studies Figure 14 shows the 1965–2009 time series of the PDO
(Garreaud et al., 2009; Seiler et al., 2013). Furthermore, index and standardized anomalies of PRCPTOT and R95p
CSPA stations show non-statistically significant rainfall in each CSPA region. In general, these rainfall indices are
trends during 1965–1976 (negative PDO) and 1977–1997 poorly correlated with the PDO index, except for the weak
(positive PDO). During 1965–1976, 44, 50, 62 and 73% anti-correlation in the first two decades of the analysed
(44, 50, 31 and 27%) of stations in regions A, B, C and D, period. Positive standardized anomalies of PRCPTOT in
Table 5. Comparison between sign and statistical significance of trends obtained for some rainfall indices at common stations
analysed in Haylock et al. (2006), Skansi et al. (2013) and the present study over CSPA.
Figure 12. Mean (cross) and standard deviation (error bar) of (a) PRCPTOT, (b) Wdays, (c) CDD and (d) R95p anomalies of the four central and
southern Peruvian Andean regions (A, B, C and D) during EN and LN events of 1965–2010. Regions (letters) accompanied by an asterisk (*)
represent statistically significant difference in the means of rainfall indices during EN compared to LN years, based on a z-test with 95% level of
confidence and independent number of events. [Colour figure can be viewed at wileyonlinelibrary.com].
regions A and C (region D) are predominant during nega- diversity of ENSO (e.g. strong, moderate, central Pacific
tive (positive) PDO phase that occurred from 1965 to 1976 or eastern Pacific El Niño) on regional rainfall.
(1977–1997). Similar patterns are observed using Wdays
(not shown). The inter-annual variability, largely domi-
nated by ENSO events, may have obscured the real effect 5. Conclusions
of the PDO on rainfall, as in Seiler et al. (2013). Further Total and extreme yearly rainfall indices characterizing
studies are necessary to examine mechanisms connecting the intensity and frequency of rainfall were calculated
coupled modes of inter-annual-to-decadal timescales in the using daily rainfall from 47 quality-controlled stations
Pacific and Atlantic to variations in rainfall in the CSPA. located in CSPA. These indices were calculated using the
In addition to ENSO and PDO, other coupled climatic hydrologic calendar instead of the Western calendar. The
modes such as the North Atlantic Oscillation (NAO), the present study demonstrated that the use of the hydrological
Atlantic Dipole (AD) and the Atlantic Multidecadal Oscil- calendar, which adequately characterizes the wet season,
lation (AMO), among others, may have played important is important to evaluate trends in all CSPA stations. To
role in rainfall patterns over the region. The independent analyse trends in the indices, we applied a MK test and
influence of ENSO, PDO, NAO, AD and AMO should be calculated Sen’s slope with a correction for time series
explored in future research, as well as the impacts of the autocorrelation for over 30 hydrologic years (1965–2009).
Table 6. Percentage of stations in CSPA regions (A, B, C and of stations. The regionalization of stations was based on
D) with positive anomalies of PRCPTOT, Wdays, CDD and PCA and cluster analysis of the gauged time series that
R95p during EN and LN years of the periods: 1965–1976 (P1), efficiently separated the stations into four major regions
1977–1999 (P2) and 2000–2010 (P3). and 14 subregions. The following major regions of stations
Index ENSO Period A B C D were identified: Amazon, central Pacific, southern Pacific
and Titicaca basins. These regions showed consistent dif-
PRCPTOT EN P1 44 57 69 9 ferences in latitude, longitude, altitude and, in some cases,
P2 22 0 0 45 slope. These geographical features also affected the exis-
P3 67 36 0 55
LN P1 89 43 77 36 tence of common features in the total and extreme rainfall
P2 56 100 100 64 indices among stations of the same subregion. Neverthe-
P3 33 93 69 55 less, trends in rainfall indices did not show a clear spa-
Wdays EN P1 33 36 46 27 tial pattern over these subregions, except for a few subre-
P2 0 0 0 18 gions and specific indices. Therefore, we summarized the
P3 78 50 8 73 main conclusions based on statistically significant trends
LN P1 78 57 77 73
P2 78 100 100 45
observed in about 30% of stations in each one of the four
P3 67 100 77 64 major regions in CSPA.
CDD EN P1 67 50 92 64 Despite the increase in the frequency of wet days,
P2 78 86 92 100 approximately 30–44% of the stations located over the
P3 56 71 62 64 Amazon basin (region A) exhibited negative trends in the
LN P1 33 50 38 55 yearly rainfall intensity, the number of heavy rainfall days
P2 44 21 31 18
and the maximum 5-day rainfall during 1965–2009. This
P3 22 50 38 36
R95p EN P1 56 43 54 9 means that although the number of rainy days increased,
P2 44 21 23 36 this decrease has not contributed to the increase in the
P3 67 21 0 45 total and mean rainfall indices for this region. Similar neg-
LN P1 67 50 38 36 ative trends were also detected in the stations located in
P2 44 50 54 64 central Pacific basin (region B); about 30–43% of the sta-
P3 33 36 69 36 tions in this region showed a decrease in the yearly rainfall
intensity, consecutive dry days, the maximum 1-day rain-
fall and the maximum 5-day rainfall. Despite the indica-
Given the influences of the complex terrain on the oro- tion of consistent negative trends in some indices related
graphic rainfall, we proposed a new method to regionalize to extreme rainfall in this region, about 20% of these sta-
stations according to their rainfall characteristics, such that tions also showed an increase in the frequency of wet days.
we could identify trend patterns over homogeneous groups In southern Pacific basin (region C), only about 20% of
Figure 13. As in Figure 12, but for positive (PDO+) and negative (PDO−) PDO phases. [Colour figure can be viewed at wileyonlinelibrary.com].
Figure 14. Standardized anomalies of PRCPTOT and R95p of each region of the central and southern Peruvian Andean stations: (a) A (Amazon
basin), (b) B (central Pacific basin), (c) C (southern Pacific basin) and (d) D (Titicaca basin). The Pacific Decal Oscillation (PDO) index is shown in
grey.
the stations showed a decrease in consecutive dry days In summary, this study demonstrated that generaliza-
and an increase in maximum 1-day rainfall, indicating that tions regarding trends in ETCCDI indices are not pos-
the drying signal of southern Peruvian Andes is actually sible in regions with complex terrain and, therefore, a
more complex than pointed out in previous studies. Finally, regionalization that accounts for the intrinsic behaviour
one-third of the stations located over the Titicaca basin of daily precipitation is necessary to investigate trends in
(region D) showed an increase in the intensity of rainfall the Peruvian Andes (and other regions with similar land-
extremes (maximum 1-day rainfall), but these trends did scape). Therefore, a clear pattern of geographic trends
not affect the total annual rainfall in the region. based on proximity did not emerge when using a large net-
During 1965–2009 hydrologic years, the four CSPA work of stations, as suggested in previous studies that used
regions showed consistent positive (negative) anomalies of less stations or other regionalization approaches (e.g. Hay-
total and extreme rainfall indices (consecutive dry days) lock et al., 2006; Skansi et al., 2013). Decadal variations in
during LN (EN) years. Positive (negative) anomalies of ENSO may have obscured other climatic signals, therefore
total annual rainfall were evident in region D during posi- contributing to a large degree to the observed long-term
tive (negative) PDO. Positive (negative) anomalies of con- variability of extreme precipitation indices. The proposed
secutive dry days were observed in regions A, C and D. methodology proved to be adequate to investigate changes
Nevertheless, rainfall indices showed variable behaviour in rainfall in complex terrain using rain gauges and could
during different ENSO–PDO conditions depending on the be adapted to other high mountain regions.
evaluated decade.
The predominant absence of long-term trends of
rainfall indices over CSPA could have resulted from
Acknowledgements
a decade-dependent variability of the ENSO–rainfall
relationship. Nevertheless, numerous factors may have This research was funded by the Consultative Group for
influenced the low-frequency variability in CSPA’s rain- International Agricultural Research (CGIAR) Research
fall which were not possible to examine with the present Program on Climate Change, Agriculture and Food Secu-
methodology and record length. Modelling studies could rity (CCAFS). H. Heidinger acknowledges CONCYTEC
help to further investigate mechanisms causing rainfall (Consejo Nacional de Ciencia, Tecnología e Innovación
variations and changes in CSPA and proxies could help to Tecnológica) for partially funding her doctoral studies at
improve our understanding about low-frequency climatic UCSB. The authors would like to acknowledge the anony-
variability in the region. mous reviewers for their constructive comments.
Rau P, Bourrel L, Labat D, Melo P, Dewitte B, Frappart F. 2016. Clim. Change 59(1–2): 75–99. https://doi.org/10.1023/a:102440642
Regionalization of rainfall over the Peruvian Pacific slope and coast. 7519.
Int. J. Climatol. 37: 143–158. https://doi.org/10.1002/joc.4693. Wang XL. 2008a. Penalized maximal F test for detecting undocumented
Rousseeuw PJ. 1987. Silhouettes: a graphical aid to the interpretation mean shift without trend change. J. Atmos. Oceanic Technol. 25(3):
and validation of cluster analysis. J. Comput. Appl. Math. 20: 53–65. 368–384. https://doi.org/10.1175/2007JTECHA982.1.
https://doi.org/10.1016/0377-0427(87)90125-7. Wang XL. 2008b. Accounting for autocorrelation in detecting mean
Seiler C, Hutjes RWA, Kabat P. 2013. Climate variability and trends in shifts in climate data series using the penalized maximal t or F test. J.
Bolivia. J. Appl. Meteorol. Climatol. 52(1): 130–146. https://doi.org/ Appl. Meteorol. Climatol. 47(9): 2423–2444. https://doi.org/10.1175/
10.1175/JAMC-D-12-0105.1. 2008JAMC1741.1.
Sen PK. 1968. Estimates of the regression coefficient based on Kendall’s Wang XL, Wen QH, Wu Y. 2007. Penalized maximal t test for detecting
tau. J. Am. Stat. Assoc. 63(324): 1379–1389. https://doi.org/10.2307/ undocumented mean change in climate data series. J. Appl. Meteorol.
2285891. Climatol. 46(6): 916–931. https://doi.org/10.1175/JAM2504.1.
Skansi MDLM, Brunet M, Sigró J, Aguilar E, Arevalo Groening JA, Weston KJ, Roy MG. 1994. The directional-dependence of the enhance-
Bentancur OJ, Castellón Geier YR, Correa Amaya RL, Jácome H, ment of rainfall over complex orography. Meteorol. Appl. 1(3):
Malheiros Ramos A, Oria Rojas C, Pasten AM, Sallons Mitro S, 267–275.
Villaroel Jiménez C, Martínez R, Alexander LV, Jones PD. 2013. Wilks DS. 2011. Statistical Methods in the Atmospheric Sciences. Aca-
Warming and wetting signals emerging from analysis of changes in demic Press: Oxford.
climate extreme indices over South America. Global Planet. Change Yoon JH, Zeng N. 2010. An Atlantic influence on Amazon rainfall. Clim.
100: 295–307. https://doi.org/10.1016/j.gloplacha.2012.11.004. Dyn. 34(2): 249–264. https://doi.org/10.1007/s00382-009-0551-6.
Takahashi K, Montecinos A, Goubanova K, Dewitte B. 2011. ENSO Yue S, Wang C. 2004. The Mann–Kendall test modified by effective
regimes: reinterpreting the canonical and Modoki El Niño. Geophys. sample size to detect trend in serially correlated hydrological series.
Res. Lett. 38(10): L10704. https://doi.org/10.1029/2011GL047364. Water Resour. Manage. 18(3): 201–218. https://doi.org/10.1023/B:
Vuille M. 1999. Atmospheric circulation over the Bolivian Altiplano WARM.0000043140.61082.60.
during dry and wet periods and extreme phases of the Southern Zhang Y, Wallace JM, Battisti DS. 1997. ENSO-like inter-
Oscillation. Int. J. Climatol. 19(14): 1579–1600. decadal variability: 1900–93. J. Clim. 10(5): 1004–1020.
Vuille M, Keimig F. 2004. Interannual variability of summertime con- https://doi.org/10.1175/1520-0442(1997)010<1004:ELIV>2.0.CO;2.
vective cloudiness and precipitation in the central Andes derived from Zhang X, Alexander L, Hegerl GC, Jones P, Tank AK, Peterson TC,
ISCCP-B3 data. J. Clim. 17(17): 3334–3348. Trewin B, Zwiers FW. 2011. Indices for monitoring changes in
Vuille M, Bradley R, Keimig F. 2000. Interannual climate variability extremes based on daily temperature and precipitation data. WIREs
in the central Andes and its relation to tropical Pacific and Atlantic Clim. Change 2(6): 851–870. https://doi.org/10.1002/wcc.147.
forcing. J. Geophys. Res. 105(12): 447–460. https://doi.org/10.1029/ Zilli MT, Carvalho LMV, Liebmann B, Silva Dias MA. 2016. A compre-
2000JD900134. hensive analysis of trends in extreme precipitation over southeastern
Vuille M, Bradley R, Werner M, Keimig F. 2003. 20th century cli- coast of Brazil. Int. J. Climatol. 37: 2269–2279. https://doi.org/10
mate change in the tropical Andes: observations and model results. .1002/joc.4840.