You are on page 1of 21

RESEARCH ARTICLE A New Adaptive Absolute Method for Homogenizing

10.1029/2021EA001716
GNSS-Derived Precipitable Water Vapor Time Series
Key Points:
Dantong Zhu1,2 , Kefei Zhang1,3, Zhen Shen1,2, Suqin Wu2 , Zhiping Liu2, and Laga Tong1,2
• A new absolute method is proposed
to homogenize a precipitable water 1
Jiangsu Key Laboratory of Resources and Environmental Information Engineering, China University of Mining and
vapor (PWV) time series derived
Technology, Xuzhou, China, 2School of Environment Science and Spatial Informatics, China University of Mining
from the Global Navigation Satellite
Systems (GNSS) and Technology, Xuzhou, China, 3Satellite Positioning for Atmosphere, Climate and Environment (SPACE) Research
• The success rate and false alarm Centre, RMIT University, Melbourne, VIC, Australia
rate were better than 74.4% and
12.2%; the accuracy of epochs and
magnitudes were 8.3 days and Abstract  The inhomogeneities, the so-called changepoints, inevitably occur in the precipitable water
0.144 mm
• PWV time series at 91 global GNSS vapor (PWV) time series derived from Global Navigation Satellite Systems (GNSS-PWV). Currently, the
stations were homogenized, and the two predominant types of homogenization methods, that is, absolute and relative methods, are limited in
impact of the homogenization was the poor performance and dependency on reference time series. In this study, a new absolute approach
also investigated
named adaptive absolute homogenization test (AAHT) by combining Seasonal-Trend decomposition
based on LOESS (STL) and Penalized Maximal F-test to account for the red noise (PMFred) is proposed.
Correspondence to: The performance of the new approach was validated using Monte Carlo simulations. Results showed
S. Wu,
the success rate and false alarm rate were better than 74.9% and 12.2%; the detection accuracy of the
sue.wu2018@gmail.com
changepoint epochs and the offset magnitudes were 8.0 days and 0.153 mm. AAHT was also applied to
homogenizing the real GNSS-PWV time series over 91 International GNSS Service stations, from which
Citation:
63 time series were identified as inhomogeneous containing 126 changepoints. Based on a comparison
Zhu, D., Zhang, K., Shen, Z., Wu,
S., Liu, Z., & Tong, L. (2021). A with the PWV time series from the fifth-generation European center for medium-range weather forecasts
new adaptive absolute method (ECMWF) reanalysis (ERA5), 43 changepoints were connected to climatic drivers, leaving 83 artificial
for homogenizing GNSS-derived
changepoints at 48 stations. The offset magnitudes at these artificial changepoints were estimated and
precipitable water vapor time
series. Earth and Space Science, corrected in the inhomogeneous time series. The long-term trends of those homogenized time series were
8, e2021EA001716. https://doi. compared with that of the PWV time series derived from ERA5 data sets (ERA-PWV) and the correlation
org/10.1029/2021EA001716
coefficient between the two sets of trends was 0.75, which was significantly larger than 0.23 before
Received 21 FEB 2021
homogenization.
Accepted 1 JUN 2021
Plain Language Summary  The precipitable water vapor time series derived from Global
Navigation Satellite Systems (GNSS-PWV) can be contaminated by inhomogeneities. This study aims to
propose a new reliable method to homogenize the global GNSS-PWV time series. This method, named
adaptive absolute homogenization test (AAHT), combines Seasonal-Trend decomposition based on LOESS
(STL) and Penalized Maximal F-test to account for the red noise (PMFred) with an overall success rate
and a false alarm rate better than 74.9% and 12.2%, respectively. An application on the GNSS-PWV over
91 GNSS stations showed that homogenization using AAHT significantly improved the trend correlation
between GNSS-PWV and PWV time series retrieved from ERA5 reanalysis.

1. Introduction
Precipitable water vapor (PWV) is an important greenhouse gas in the global energy and hydrologic cy-
cle. According to the Clausius-Clapeyron equation, a 1  K increase in surface temperature can lead to a
5%–7% increase in PWV under the assumption that the atmospheric relative humidity is constant (O'Gor-
man & Muller, 2010). An increase in PWV can cause a positive response to enhance the global warming
process, which is the source of a major climate feedback mechanism influencing climate change (Alshawaf
et al., 2017; Dai et al., 2002; Hodnebrog et al., 2019; Parracho et al., 2018; Sherwood et al., 2010; Trenberth
© 2021. The Authors. et al., 2005). Therefore, PWV is an important indicator for climatologists to analyze the atmospheric varia-
This is an open access article under
the terms of the Creative Commons bility. One of the main conventional atmosphere sensing techniques is radiosonde and its long-term sensing
Attribution-NonCommercial License, data can be used to retrieve PWV time series. However, due to the limitation of the poor spatiotemporal res-
which permits use, distribution and olutions on climatic applications. Currently, observations from Global Navigation Satellite Systems (GNSS)
reproduction in any medium, provided
the original work is properly cited and have been used to retrieve the PWV time series (GNSS-PWV) due to their high temporal resolution and
is not used for commercial purposes. low cost. Since Bevis (Bevis et al., 1992) proposed GNSS meteorology in the early 1990s, the GNSS-PWV

ZHU ET AL. 1 of 21
23335084, 2021, 7, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2021EA001716 by Nat Prov Indonesia, Wiley Online Library on [17/05/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Earth and Space Science 10.1029/2021EA001716

time series have been widely used in various climatic studies and applications (Klos et al., 2018; Nilsson &
Elgered, 2008; Ning & Elgered, 2012; X. Wang et al., 2018; Zhao et al., 2020). For all these studies or appli-
cations, homogeneous GNSS-PWV time series are desirable for trustworthy analyses. However, in practice,
climatic drivers and various artificial drivers can inevitably introduce some changepoints in most GNSS-
PWV time series. The changepoints caused by the two types of drivers are called the climatic changepoints
and artificial changepoints, the latter of which cause the inhomogeneities. Inhomogeneities are detrimental
to the climatic research and may lead to biased results, especially in the long-term climatic trend estimate.
Therefore, artificial changepoints in GNSS-PWV time series need to be identified and eliminated before the
time series is applied to climatic studies. The main focus of this study is on the detection of artificial change-
points and correction for the offsets in the GNSS-PWV time series, which is known as data homogenization.

The two main causes of changepoints in a GNSS-PWV time series are (1) different data processing
strategies used for different subsets in the time series (Steigenberger et al., 2007), and (2) station-related
changes including changes in the equipment and electromagnetic environment at the GNSS station
(Larson et al., 2010; Pierdicca et al., 2014). The cause (1) may involve the adoption of different ITRFs,
the use of different error correction models and/or different elevation cutting angels, etc. To address
this problem, unified processing strategies are commonly used (Ning & Elgered,  2012; Steigenberger
et al., 2007; Thomas et al., 2011; Vey et al., 2009). Regarding the cause (2), immediately after an equip-
ment is changed, an abrupt offset is introduced (Gradinarsky et al., 2002; Ning et al., 2011), which is
usually documented in the station metadata (Hoseini et al., 2020), but some changes might be missed
in some inaccurate and incomplete metadata (Lindau & Venema, 2013; Wang, Wen, & Wu, 2007; Wang,
Zhang, et  al.,  2007). A change in the electromagnetic environment around the GNSS station may be
related to the moisture condition and growing vegetation, which are rarely documented. Without the
aid of the station-related changepoints information, it is difficult to perform homogenization. In recent
years, several methods in detecting changepoints in climatic and geodetic time series were proposed,
and these methods can be classified into two categories: relative and absolute methods, which are elab-
orated below.

The relative methods use a differential time series formed by subtracting a reference time series from
the original (target) time series. The reference time series can be obtained from different sources, for
example, nearby GNSS stations or radiosonde models. This is because both the time series share similar
evolutions in the temporal domain, and the values at the same epoch in both the time series are regarded
as highly correlated in the spatial domain, that is, the difference between the two values is insignificant.
If the differential values at some epochs exhibit significant offsets, then changepoints are likely to occur
at these epochs. The significance of the offset is often evaluated by some statistical test methods, such
as the likelihood ratio method based on a modified two-phase regression model (Easterling & Peter-
son, 1995; Solow, 1987); a series of penalized test method based on the T-test and F-test (Wang, Wen,
& Wu,  2007; Wang, Zhang, et  al.,  2007; X. L. Wang,  2008a,  2008b); the Detection, Identification and
Adaptation method (Amiri-Simkooei et  al.,  2019; Teunissen,  2018); the segmental test method (Bock
et al., 2020). The reference data can be from a regional radiosonde station or a GNSS station near the tar-
get station (i.e., the GNSS station being investigated; Dai et al., 2011; Easterling & Peterson, 1995; Zhang
et al., 2017). However, it is not practical that such reference data are available for all GNSS stations, es-
pecially in the case that the GNSS stations from a regional network are sparsely deployed. To solve this
problem, other independent reference data sets were often introduced, for example, reanalysis data sets
from numerical weather prediction products, or DORIS and VLBI observations. Vey et al. (2009) used
the data from both neighbor GNSS stations and the National Centers for Environmental Prediction rea-
nalysis for the reference data of homogenizing the GNSS-PWV time series. Ning, Wickert, et al. (2016)
adopted several data sets for the references, and Bock and Parracho (2019) selected ERA-Interim data
as the reference for the same purpose under the assumption of heterogeneous white noise in the PWV
time series.

The absolute methods can directly detect changepoints in an original (i.e., zero-differential) time series
without the need for the reference data, which motivated the interest of some climatic and geodetic ana-
lysts. Rodionov (2004) presented a method of sequential t-test analysis of regime shifts based on local re-
gression, and the method was applied to the homogenization of the coordinate time series at GNSS stations.

ZHU ET AL. 2 of 21
23335084, 2021, 7, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2021EA001716 by Nat Prov Indonesia, Wiley Online Library on [17/05/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Earth and Space Science 10.1029/2021EA001716

Lindau  (2003) and Lindau and Venema  (2018) decomposed the total variance of an original time series
into an external variance and an internal variance, and the minimum internal variance was used to detect
changepoints. Wu et al. (2018) used an extensive L1 regularization method rather than the least squares
method to simultaneously estimate piecewise trends, periodic components and level offsets, which were
then used to detect changepoints and discontinuities. Hoseini et al. (2020) developed a zero-difference ap-
proach based on singular spectrum analysis (SSA) for the homogenization of PWV time series. All of these
absolute methods have acceptable performances (Van Malderen et  al.,  2020), but there is still room for
improvement.

The relative methods are more widely employed than the absolute methods, since the majority of the noise
and periodic evolutions in the original time series can be mitigated. However, there are some issues, espe-
cially in the case that the target time series has much higher temporal resolutions (e.g., daily, or even hour-
ly) than that of the reference data sets. The main issues are as follows:

1. Regarding the reference data set, if the reference data set is from a station far away from the target
station; or the reference data set has unsubstantiated inhomogeneity and uneven systematic represent-
ative discrepancies with the original data set in both spatial and temporal domains etc (Bock & Parra-
cho, 2019; Ferguson & Villarini, 2012; Venema et al., 2012);
2. The common part in changepoints due to a simultaneous change for upgrading the instruments or data
processing strategies at both the target and reference stations is canceled out or hidden by the differential
operator (Ning, Wickert, et al., 2016);
3. Identifying the origin of detected changepoints can be difficult when different data sets are used in the
homogenization process (Guo & Ding, 2011; Hoseini et al., 2020; Lindau & Venema, 2018).

In summary, the performance of the relative methods depends on the reference time series, and show lim-
ited suitability and university for real applications. The adaptation of the absolute methods can avoid these
problems, which is the main reason for the prevailing of the absolute methods for homogenizing the GNSS-
PWV time series. However, the performance of the absolute methods on the detection of changepoints still
needs to be improved.

Previous studies stated that the noise in an atmospheric time series, for example, the zenith wet delay
(ZWD) of GNSS, PWV, and temperature over a site, can be expressed by a combination of the white
noise and first-order autoregressive (AR(1)) noise (Alshawaf et al., 2017; Klos et al., 2018; Ning, Wickert,
et al., 2016). If the noise is large, it is difficult to correctly identify those true changepoints in the time
series, which may lead to an aliasing mechanism (the true changepoints with slight offsets may be hidden
by the large observation noise), if any, and also hide slight offsets and lead to more false alarms in the
detection results.

In this study, for improving the homogenization of the GNSS-PWV time series, a new absolute method,
named adaptive absolute homogenization test (AAHT), that combines a Seasonal-Trend decomposition
based on LOESS (STL) method (Cleveland et al., 1990) and a Penalized Maximal F-test to account for the
red noise (PMFred; X. L. Wang, 2008a) is proposed. Compared to other existing absolute methods, the
proposed method has the following characteristics: (1) It decomposes an original time series into a peri-
odic component, a trend component and a residual component, and almost all significant changepoints
are reflected in the trend component; thus, the use of the trend component leads to higher success rates
(SRs) in the detection; (2) The use of an adaptive criterion obtained from the residual component results
in better discrimination of false alarms from true changepoints, that is, leading to reduced numbers of
false alarms.

The outline of this study is as follows. Section 2 introduces the methodology for the new method proposed
in this study, including formulas, framework and procedures. Section 3 presents the performance assess-
ment of the new method using a simulated data set. In Section 4, the new method is used for homogenizing
the GNSS-PWV time series over 91 global International GNSS Service (IGS) stations, and the trend compo-
nents from some original and homogenized GNSS-PWV time series are compared and analyzed. Summary
and conclusions are given in Section 5.

ZHU ET AL. 3 of 21
23335084, 2021, 7, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2021EA001716 by Nat Prov Indonesia, Wiley Online Library on [17/05/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Earth and Space Science 10.1029/2021EA001716

2. Methodology
The AAHT method proposed in this study is a combination of the STL method and the PMFred test for
decomposing a time series and detecting changepoints in the time series, respectively. In this section, the
STL method and PMFred test are introduced first, and then the methodology and procedure for the AAHT
method are presented.

2.1.  STL and PMFred Test


2.1.1. STL

The STL method is a traditional non-parametric method for decomposing a time series into a trend compo-
nent, a periodic component, and a residual component, as expressed by

Y t   T t   S t   R t 
(1)

where Y  t  is the original target time series; T  t , S  t , and R  t  are the decomposed trend, periodic, and
residual components, respectively.

Unlike other decomposition methods, for example, SSA and empirical mode decomposition (Huang
et al., 1998), in which an extra step is needed for empirically identifying the trend and periodic components,
the STL can directly give the two components, which is the main reason for using the method in this study.
The STL is driven by a sequence of smoothing operators consisting of the locally estimated scatterplot
smoothing (LOESS) and moving average (MA) operators, and the LOESS operator is implemented multiple
times. The basic idea of LOESS is to smooth a time series using a local weighted polynomial fitting model
with an order of d. During the smoothing process, the smoothed data at an epoch, also named a target
epoch, can be determined through weighted fitting a subset time series. The subset time series is obtained
from the original entire time series at the q nearest neighbor epochs around the target epoch. The weight of
the data at each of the q epochs can be calculated as follows.

Suppose ti is the ith epoch, that is, the target epoch, in the original time series Y  t  to be smoothed; q  ti  is
the longest time interval between ti and all of the other neighbor epochs in the subset formed by the data at
the q nearest neighbor epochs of ti, then, the weight of the data at a neighbor epoch t j can be calculated by

 t j  ti 
Pj  ti   W 
(2) 
 q  ti  
 

where t j  ti is the time interval between ti and t j ; W   is the cubic weight function that is expressed as
follows:


 
3
 1  u3 , 0  u  1
W u  
(3)
 0 , u 1

The above weight function ensures that the largest and the least (zero) weight values are assigned to the
epochs nearest and furthest to the target epoch, respectively. Then, the smoothed value C  ti  at ti can be
obtained from a d-order weighted polynomial fitting model. For simplicity, the LOESS operator with the two
 
variables of d and q is denoted as LOE q, d .

Applying a LOESS operator onto the original time series, a smoothed series C  t , comprising all the periodic
component and part of the trend component in the original time series, can be generated. Then, applying
the LOESS and MA operators both T  t  and S  t  can be obtained (multiple times execution).

ZHU ET AL. 4 of 21
23335084, 2021, 7, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2021EA001716 by Nat Prov Indonesia, Wiley Online Library on [17/05/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Earth and Space Science 10.1029/2021EA001716

2.1.2.  PMFred Test

The PMFred test, proposed by Wang (X. L. Wang, 2008a), is a sequential method to detect multiple change-
points accounting for the lag-1 autocorrelation in the original time series by combining white noise and
AR(1) noise using the Penalized Maximal F-test (PMFT). The PMFT is an empirical statistic test embedded
in a two-phase linear model described below. The uneven U-shaped false alarm rate (FAR) of an F-test can
be effectively converted into a W-shaped curve by introducing an empirical penalty factor function into the
F-test. Therefore, the PMFred test can be used to detect potential changepoints accurately (Dai et al., 2011;
Venema et al., 2012).

Suppose LT  t  to be a linear time series without any periodicity, to test if an epoch tc is a changepoint, let
the null hypothesis H0 be: tc is not a changepoint; and the alternative hypothesis H a be: tc is a changepoint,
then the two hypotheses can be constructed as

H0 : LT  ti     ti   i , i 1, 2, , n
(4)  1   ti   i , i  1, 2, , c
H a : LT  ti   
   ti   i , i  c  1, c  2, , n
 2

where  , 1, and 2 are the constant intercepts;  is the long-term trend;  is the noise that can be expressed
by the combination of Gaussian white noise and AR(1) noise with a temporal correction of  ; and n is the
length of the time series.

To test H a against H0, the following statistic can be constructed

PF
(5)
c Pc  Fc

where PFc is the penalized F-test statistic value at tc; Pc is the penalty factor, which is generated empirically
based on a series of Monte Carlo simulation (X. L. Wang, 2008a); Fc is the F-test statistic which can be ex-
pressed by

 
SSE 0  SSE A n 2
, SSE0  LTi  ˆ  ˆ ti
Fc

SSE A / n  3 i 1
(6)
   
c 2 n 2
SSE
A  LTi  ˆ1  ˆ ti   LTi  ˆ 2  ˆ ti
i 1 i  c 1

where SSE 0 and SSE A are the sums of squared errors under H0 and H a, respectively; ̂ and ̂ are estimated
from the least-squares method under H0; ˆ1, ˆ 2, and ̂ are the estimates under H a.

 
To test the true hypothesis, a critical value CV  , N as a function of a selected confidence value  needs
to be pre-set. If the PFc value is above the critical value, tc is regarded as a changepoint and the magnitude
of the offset at the changepoint is Δ  2  1 ; otherwise H0 holds true, and    1 2. This test can
be used to find all possible changepoints with their penalized F-test statistic values in a time series, and the
point (or epoch) that results in the largest penalized F-test statistic value is identified as the possible change-
point. This changepoint is then used to divide the time series into two segments and the above procedure is
carried out for each of the segments to find the possible changepoint in the segment. This recursive process
is completed until no more possible changepoints can be found. More details on the PMFred can be found
in X. L. Wang (2008a).

2.2. AAHT
2.2.1.  Combination of STL and PMFred

The PMFred, which is based on the linear regression method, has been widely applied to the homogeniza-
tion of the climatic time series due to its high detection power and good accuracy. Since it accounts for the
AR(1) noise (red noise) in time series, the PMFred test can be also applied to homogenize the GNSS-PWV
time series (Ning, Wickert, et al., 2016; Schröder et al., 2016). Previous studies have employed the PMFred

ZHU ET AL. 5 of 21
23335084, 2021, 7, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2021EA001716 by Nat Prov Indonesia, Wiley Online Library on [17/05/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Earth and Space Science 10.1029/2021EA001716

Figure 1.  Impact of inhomogeneity on the decomposed components of a simulated time series (in unit of mm): (a)
original inhomogeneous time series Y  t ; (b) smoothed residual Y  t   C  t ; (c) smoothed series C  t  and abnormal
trends; (d) periodic component S  t ; (e) trend component T  t ; (f) residual component R  t .

to detect changepoints in a differential time series with the aid of a selected reference data set, like the rela-
tive methods (Dai et al., 2011; Venema et al., 2012). This study treats the PMFred to be an absolute method
as previously mentioned. However, the PMFred is embedded in the linear regression model, which cannot
well fit the significant periodic evolution in GNSS-PWV time series. Hence, it is crucial to extract the linear
time series from the original GNSS-PWV time series for employing the PMFred.

The STL, as a flexible decomposition method, can retrieve the trend component, which can be regarded as
a linear time series, from a homogeneous time series. Then, the retrieved linear time series can be input to
the PMFred for detecting potential changepoints. However, for inhomogeneous time series, the inhomoge-
neity can contaminate the decomposition result, especially the smoothed series C  t  (the most important
intermediate component) and the trend component.

To illustrate the impact of inhomogeneity on decomposed components by the STL, Figure 1 shows a sim-
ulated inhomogeneous time series, and the contaminated decomposition results by the STL (the window
lengths qc, qt ,n p, and the polynomial order d are the same as the ones mentioned in section 3). In this simu-
lation, the annual and semiannual periods with the amplitudes of 4 and 0.5 mm, respectively, are included;
a 0.02 mm/year long-term linear trend is set; a red noise with a 0.15 temporal correlation is generated base
on a uniform white noise. To simulate the true inhomogeneity, an artificial changepoint with a 3-mm offset
3-mm offset is added to the time series after the 2,800th epoch. Subfigure (a) is the original inhomogeneous
time series; subfigure (b) is the smoothed series C  t ; subfigure (c) is the smoothed residual Y  t   C  t ; the
final decomposed components, that is, S  t , T  t  and R  t  obtained from the STL are shown in (d), (e) and
(f), respectively. Subfigure (b) shows obvious abnormal trends (0.12, 0.67 and 0.04 mm/year for different
segments) around the changepoint in C  t ; (c) shows an obvious jump and abnormal trends; (d) shows
slight abnormal trends (−0.04, −0.01, and 0.01 mm/year for different segments) in S  t ; while (e) and (f)
indicate that T  t  and R  t  are also contaminated by the artificial changepoint and both show the additional
constant jump besides the abnormal linear trend. These abnormalities are due to the over-smoothing of
the LOESS operator, which is beneficial for the avoidance of outliers in homogenous time series but harm-
ful for inhomogeneous time series. In other words, for significant enough changes, the final decomposed

ZHU ET AL. 6 of 21
23335084, 2021, 7, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2021EA001716 by Nat Prov Indonesia, Wiley Online Library on [17/05/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Earth and Space Science 10.1029/2021EA001716

components are affected by the inhomogeneity in the original time series through the over-smoothed C  t 
from the LOESS operator. Therefore, it is better to weaken the effect of the inhomogeneity on C  t  initially,
if it cannot be completely eliminated in decomposition. Fortunately, the PMFred can be used to weaken the
effect of inhomogeneity and this is the reason for the combination of the PMFred and STL is used in the
new AAHT method.

2.2.2.  Procedure of AAHT

The main steps involved in the AAHT method are as follows.

 
1. Weakening the impact of inhomogeneity. A regular LOESS operator LOE qc , d is employed to extract
the inhomogeneous smoothed series C  t  and V  t  Y  t   C  t  from Y  t ; then a PMFred with a
low confidence value of  c is employed to detect changepoints in V  t , since those significant periodic
evolutions are less likely to be in V  t . If any, the mc detected changepoints are recorded in a tempo-
 
rary list o j ; and Y  t  is divided into mc  1 segments Ys  t ( s  1,2, , mc  1) for extracting segmented
s s  c 
smoothed series Cˆ  t  from Y  t  using LOE q , d ; the magnitude of the offset at each changepoint is
estimated from the least squares method based on the following observation equation

 2
 1   ti   ak cos  k ti   k    i , i  1,2, , o1
 k 1
 2
 2   ti   ak cos k ti   k    i , i  o1  1, o1  2, , o2
Y  ti   
(7) k 1
 
 2

 m 1   ti   ak cos k ti   k    i , i omc  1, omc  2, , n
 k 1

where a,  ,  are the amplitude, angular frequency and phase, respectively; the subscripts k  1,2 are the
annual and semiannual periods.

2. Decomposing an original time series Y  t . For the decomposition of Y  t , the two additional time se-
ries are added to both ends of the above smoothed series Ĉ  t  for the avoidance of potential bias from
LOESS near both ends of Y  t . The length (n p) of the additional time series is the same as the length of
the strongest periodic cycle in Y  t , and the data are extrapolated from the d-order polynomial model
through fitting the data at the identical time of each period in Ĉ  t . Then, three MA operators (the
 
lengths of windows in MA operators are set to n p,n p, and 3) and a LOE n p , d operator are subsequently

employed to obtain the final periodic component S  t , which is homogeneous. Next, LOE qt , d is used 
to extract the final trend component T  t  (with almost all significant changepoints) from the inhomoge-
neous deseasonalized time series Y  t   S  t .
3. Detecting changepoints using the trend component. The PMFred test with a strict confidence value of  t
is used to detect multiple changepoints and the magnitudes of the corresponding offsets are estimated
   
based on Equation 7. Let c 0j and Δ 0j ( j  1,2,3, , m0 ) be the lists containing the detected change-
points and the magnitudes of the offsets, respectively. Replace Y  t  with Y  t   T  t  and repeat the above
process (i.e., go back to step [1]) until no more changepoint is detected to complete the recursive process.
4. Identifying false alarms and correcting final changepoints. All false alarms need to be identified for their
removal from the above detection result. Generally speaking, it is difficult to discriminate those false
alarms with a very small offset from true changepoints, since the offsets and strong observation noise
may cause aliasing. To reduce the aliasing effect, an adaptive criterion    R is used in this study, where
   0,1 is an empirical factor and  R is the standard deviation of the residual component R  t , to de-
termine the threshold for the offset. If the magnitude of the offset estimated for a changepoint is above
the    R value, the changepoint is classified as a significant changepoint; otherwise, it is regarded as
 
a false alarm. If the final list of significant changepoints is denoted as c j ( j  1,2,3, , m ), then the

ZHU ET AL. 7 of 21
23335084, 2021, 7, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2021EA001716 by Nat Prov Indonesia, Wiley Online Library on [17/05/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Earth and Space Science 10.1029/2021EA001716

Figure 2.  Procedure of the new proposed adaptive absolute homogenization test.

 
magnitudes of all the offsets in the list, denoted by Δ j , need to be re-estimated based on Equation 7.
After this, a correction process for the offsets at all significant changepoints in the original time series is
carried out, the homogenization process is completed.

The flowchart for the above procedure is illustrated in Figure 2.

3.  Performance Assessment


The performance of the AAHT method is assessed through a Monte Carlo simulation using a simulated
data set based on the characteristics of GNSS-PWV time series from global stations used in previous studies
(Jin et al., 2008; Klos et al., 2018; Ning & Elgered, 2012; Zhao et al., 2019). The reason for not using the real/
observed benchmark time series is to avoid the effect of potential unsubstantiated inhomogeneity in the
time series on the detection results.

3.1.  Simulated Data Set

Each daily PWV observation in a PWV time series consists of a signal and a noise, and the observation at the
ith epoch is generated or simulated using the following formula
2
Y  ti     ti   ak cos k ti   j   i
(8)  
k 1

where Y  ti  is the PWV observation;  i is the noise that is a combination of Gaussian white noise with an
amplitude (i.e., a standard deviation)  WN and an AR(1) noise with a temporal correlation coefficient  and
an amplitude  AR.

ZHU ET AL. 8 of 21
23335084, 2021, 7, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2021EA001716 by Nat Prov Indonesia, Wiley Online Library on [17/05/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Earth and Space Science 10.1029/2021EA001716

Table 1
Seven Parameters and Their Values Used in the Simulation (Forming 225 Schemes)
Parameter Value
Length/temporal resolution 15 years/1 day
Long-term trend  Random variables in the range from −0.5 to 0.5 mm/year
Annual amplitudes a1 4, 8, 12, 16, and 20 mm
Semiannual amplitude a2 1, 2, 3, 4, and 5 mm
Amplitude of white noise  WN 1, 2, and 3 mm
Temporal correlation coefficient of AR(1) noise  0, 0.2, and 0.4

Amplitude of AR(1) noise  AR 1 mm

Table 1 lists the seven parameters and the values selected for each of the parameters for the simulation,
which were based on the results of the GNSS-PWV time series and ZWD time series in various regions
(Alshawaf et al., 2017; Jin et al., 2008; Ning, Wang, et al., 2016; Van Malderen et al., 2020; Zhao et al., 2019).
Since the numbers used for the four parameters a1, a2,  WN , and  were 5, 5, 3, and 3, respectively, the num-
ber of combinations is then 225, that is, 225 sets of 7-parameter values or schemes. Each of the schemes was
performed 1,000 times to generate a 15-year homogeneous time series, where the trend value was a random
variable from the value range between −0.5 and 0.5  mm/year at each time. Based on the 225 schemes,
225,000 such time series were generated. Then, for simulating inhomogeneous data sets, three changepoints
with the offset magnitudes of 1, 2, and 3 mm (with randomly selected signs (negative or positive) and posi-
tions/epochs) were simultaneously inserted into each of the time series under the condition that the time
interval between two successive changepoints was at least one year. As a result, 225,000 new inhomogene-
ous time series were generated.

3.2.  Quality Indicator

The performance of the AAHT method can be measured by several quality indicators such as the com-
monly used SR, FAR, missing rate (MR), mean absolute error (MAE) of the epochs and the MAE of the
offset magnitudes at all detected changepoints. The SR, FAR, and MR are defined as the corresponding
numbers of success changepoints, false alarms, and missing changepoints divided by the total number of
the above three types of detections (Gazeaux et al., 2013; Van Malderen et al., 2020). In this study, three
success criteria indicating the successful changepoints within 30-day, 60-day, and 90-day time windows for
the following three scenarios were employed: (1) If the detected changepoint falls into the pre-selected time
window (i.e., one of the three time windows) that contains its corresponding true epoch, and is the closest
to the true epoch, then the detected changepoint is defined as a successful changepoint; (2) If the detected
changepoint is either beyond the above pre-selected window that contains its corresponding true epoch,
or within the pre-selected time window but is not the closest to the corresponding true epoch, then the
detected changepoint is defined as a false alarm; (3) If the true changepoint is not successfully detected, it
is called a missing changepoint.

The MAEs of an epoch and a magnitude are estimated using:

 1 ns
MAE  ˆ j  ˆ j
 ns j 1
(9)

MAE 1 ns

   Δˆ j  Δˆ j
 ns j 1

where  and Δ are the true epoch and magnitude of the offset at the successful changepoint, respectively; ˆ
and Δˆ are their corresponding estimates; ns is the number of all successful points in the time series.

ZHU ET AL. 9 of 21
23335084, 2021, 7, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2021EA001716 by Nat Prov Indonesia, Wiley Online Library on [17/05/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Earth and Space Science 10.1029/2021EA001716

3.3.  Detection Result


A set of values for the parameters (i.e., qc, qt , d , n p,  c, and  t ) need to be
set for a homogenization process. According to Cleveland et al. (1990), in
this study, 15 and 609 were set for qc and qt , respectively, for the LOESS
operator, 2 for the polynomial order d ; 365 for n p used in the MA opera-
tor; 0.95 and 0.99 for the confidence values  c and  t , respectively, used
in the PMFred. After a homogenized time series was obtained from the
AAHT procedure, the mean values of the aforementioned quality indi-
cators were obtained to evaluate the overall performance of the AAHT
method, and the results are shown in Figure 3 and Table 2. Note that the
reason for not showing the MR result in Figure 3 is that the SR and FAR
are more concerned.

In Figure 3, the results in the three columns of the SR indicate that, the
larger the size of the time window, the better the detection results. The
mean SR significantly decreases with the increase of the amplitude of
Figure 3.  Overall performance of the adaptive absolute homogenization
white noise, and the minimum SR value is 74.4%, which is resulted from
test method based on the simulated data set.  WN |  means the
combination of the white noise and temporal correlation coefficient. a 3 mm level of white noise. Table 2 indicates that our selection of noise
agrees well with existing studies (Klos et al., 2016), where a noise level of
global GNSS-PWV time series is under 3 mm. Therefore, the SR results
mean the AAHT method performs very well on a global scale.

An optimal homogenization method should not only maximize SR but also minimize FAR. The results from
the FAR columns also indicate the good performance of AAHT. The lowest FAR is resulting from the case of
the temporal correlation coefficient   0, that is, only white noise exists in the time series. As the increase
in  , more complexities are introduced into the noise component of each time series, and the FAR increases.
The highest value 12.3% occurs in the case that white noise is the minimum (1 mm) and  is the maximum
(0.4). In this case, the strong temporal correlation raises the aliasing effect between the noise and the true
offset. Fortunately, the adaptive criterion obtained from the residual component reduces the aliasing effect,
hence the FAR value is still small.

The results in columns of MAE and MAE  indicate the errors in the identification of the epoch and the
magnitude of the detected offsets. The mean values of MAE from the three time windows are 4.9 days,
7.0 days, and 8.3 days, respectively. The similarity in the variation tendency for both MAE and SR implies
that the increase in the size of the time window leads to more errors in the epoch ˆ and significantly in-
creases the MAE . Nevertheless, MAE  are all around 0.142 mm, meaning little variation with the window
size, which accounts for about 7.1% of the true magnitudes of the offsets. Theoretically, a large window size,
that is, a looser success criterion, means more changepoints are included in the list of final changepoints,
and some false changepoints are also inevitably identified as success detections. Those falsely identified
changepoints may introduce an error in the estimation of the offset magnitude. However, the above results
do not highlight the influence of such errors.

In the context of variable trend values, the effect of homogenization on the trend estimates was investigat-
ed by comparing the true trend used in the simulated time series with the trend estimate using the least
squares method before and after homogenization. Figure 4 shows the scatterplot of the true value (x-axis)
and estimate (y-axis) of the trend of the two time series (before and after homogenization). The correlation
coefficient between the true values and estimated trends of the original inhomogeneous time series (see
subfigure(a)) was 0.990, while the coefficient value after homogenization by the AAHT was improved to
0.999 (see (b)). The root mean squares (RMSs) of the two sets of fitting residuals were 0.3 and 0.05 mm
before and after homogenization. The difference value also reflects the effect of homogenization by the
proposed method on the trend estimates.

ZHU ET AL. 10 of 21
23335084, 2021, 7, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2021EA001716 by Nat Prov Indonesia, Wiley Online Library on [17/05/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Earth and Space Science 10.1029/2021EA001716

Table 2
Statistics of Quality Indicators for the AAHT Method

Parameter value SR (%) FAR (%) MAEτ (day) MAEµ (0.1mm)


σ WN SC
(mm) ϕ 30 60 90 30 60 90 30 60 90 30 60 90

0 93.4 97.9 97.9 4.0 1.3 1.9 2.8 3.6 4.1 0.79 0.81 0.82
1 0.2 88.9 94.2 96.6 6.7 4.7 3.1 3.2 4.3 4.9 0.91 0.93 0.94
0.4 81.9 89.4 91.1 12.3 8.9 7.2 4.0 5.8 6.6 1.19 1.22 1.24
0 81.1 90.1 93.2 9.3 5.1 3.1 4.7 6.7 8.3 1.25 1.27 1.28
2 0.2 79.1 89.4 92.7 9.6 5.4 3.9 4.9 7.6 9.3 1.37 1.39 1.41
0.4 77.2 87.1 92.1 11.2 5.9 4.1 5.3 8.0 9.7 1.54 1.54 1.56
0 77.3 85.8 89.2 9.2 3.4 2.6 6.3 8.8 10.0 1.68 1.69 1.70
3 0.2 75.9 84.8 87.8 9.6 4.6 3.1 6.6 8.9 10.8 1.82 1.83 1.84
0.4 74.4 84.1 87.1 10.4 5.0 3.5 6.6 9.2 11.2 2.14 2.14 2.14

Mean 81.0 89.2 92.0 9.1 4.9 3.6 4.9 7.0 8.3 1.42 1.42 1.44

Abbreviations: AAHT, adaptive absolute homogenization test; FAR, false alarm rate; MAE, mean absolute error; SC,
success criteria; SR, success rate. The bold values denote the maximum, minimum and mean values.

4.  Homogenization of Daily GNSS-PWV Time Series


In this section, real GNSS-PWV time series over 91 global IGS stations from 2000 to 2018 are first homog-
enized using the AAHT, where the same parameter values adopted in the above simulation are used since
these parameter values are based on the characteristics of the noise and variation of the global GNSS-PWV

Figure 4.  Scatterplots of the true trend values (x-axis) and the trend estimates (y-axis) calculated by least squares method: (a) the simulated inhomogeneous
time series; (b) homogenized time series by adaptive absolute homogenization test. The two lines are their fitting results (see the text in subfigures), where the
slopes of the lines indicate the correlation coefficients between the true values and estimates of the two sets of time series.

ZHU ET AL. 11 of 21
23335084, 2021, 7, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2021EA001716 by Nat Prov Indonesia, Wiley Online Library on [17/05/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Earth and Space Science 10.1029/2021EA001716

Figure 5.  Distribution of the selected 91 International Global Navigation Satellite Systems Service stations for testing
the proposed homogenization method.

time series. The homogenization process includes three steps. The first step is to detect all changepoints
and estimate the initial magnitudes of the offsets at the changepoints, and construct the initial changepoint
list. The second step is to identify those climatic and documented changepoints in the list by comparing
the GNSS-PWV time series against the PWV time series derived from the ERA-5 data set (ERA-PWV) and
station metadata. Then, the climatic changepoints are removed from the initial changepoints list. The last
step is to re-estimate the magnitudes of the offsets at all the rest changepoints (i.e., artificial changepoints)
in the changepoint list. The influence of the homogenization on the long-term trends of the GNSS-PWV
time series is also investigated.

4.1.  Data Source and Pre-Processing

The original GNSS-PWV time series for testing are converted from the latest zenith tropospheric delay
(ZTD) products from the IGS, referred to as IGS repro 2, which provides a continuous ZTD time series
from 1 January 2000 with a 5-min temporal resolution and typical errors (precision) of 1.5–5 mm (Byun &
Bar-Sever, 2009). The products are reprocessed by the GIPSY Precise Point Positioning software in a fully
“consistent” way using the latest models (e.g., identical antenna phase center models, mapping function,
ITRF, and IGS final products were used). Therefore, the inhomogeneity caused by the use of different pro-
cessing strategies can be excluded from the GNSS-PWV time series. More details about IGS repro 2 can be
found at http://acc.igs.org/reprocess2.html.

Stations with the available ZTD data from January 1, 2000 to December 31, 2018 were selected, and the
daily resolution, rather than the 5-min resolution provided in the original ZTD data sets, was adopted for
the GNSS-PWV times series used in the test. The following screening process for data pre-processing was
carried out for quality control. First, if the daily ZTD data contain less than 20-h data, then the daily data
were excluded. Second, all the 5-min ZTD observations on a day were averaged for the daily ZTD, so that the
data gaps, if any, can be examined. The stations that had either a continuous data gap spanning more than
100 days or the total number of data gaps more than 10% of the number of days were rejected. The purpose
to do so is to ensure that the final homogenization results would not be affected by the gap-filling process
(introduced later on). After all the above steps were carried out, 91 stations remained, and their geographi-
cal distribution is presented in Figure 5. Due to the different beginning epochs of the ZTD time series at the
91 stations, their lengths are different. The minimum, maximum, and mean length of the daily ZTD time
series (i.e., the number of days) at the 91 stations were 4,939, 6,458, and 6,040 days, respectively. The next
step was to identify outliers in each daily ZTD time series using the Inter-Quartile Range (IQR). If a ZTD
value is beyond 5 times IQR (far from the median ZTDs in the time series), this ZTD value is removed from

ZHU ET AL. 12 of 21
23335084, 2021, 7, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2021EA001716 by Nat Prov Indonesia, Wiley Online Library on [17/05/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Earth and Space Science 10.1029/2021EA001716

Figure 6.  Number of changepoints detected at each of the 63 inhomogeneous Global Navigation Satellite Systems
stations.

the time series. After this process was performed, a total of 30,134 days, accounting for 5.12% of the total
ZTD time series, were removed.

PWV was converted from the ZWD using

PWV Π  ZW
(10)

where ZW is the ZWD over the GNSS station, which can be obtained by subtracting the zenith hydrostatic
delay derived by Saastamoinen model (Saastamoinen, 1972) from the ZTD estimated from the GNSS data
processing; Π is the conversion factor, which is a function of weighted mean temperature (Wang, Wen, &
Wu, et al., 2007; Wang, Zhang, et al., 2007).

Atmospheric data (specific humidity, temperature, and pressure profiles) from the hourly ERA-5 pressure
levels data set were used in the conversion from the ZWD to PWV (Hersbach et al., 2020). Note that these
atmospheric data need to be interpolated or extrapolated for the position of the GNSS station in the spatial
domain and averaged to match the epoch of the ZTD time series in the temporal domain. The spatial do-
main included both vertical and horizontal domains, and the interpolation process started with the vertical
domain (note that in some cases an extrapolation, rather than interpolation, may be needed). The geopo-
tential heights of the profile data at the four grid points surrounding the GNSS station were transformed to
their geometric height first, and then the profile data at the four grid points were all vertically interpolated
or extrapolated to the geometric height of the GNSS station. These profile data, supposed to be in the same
horizontal plane, were then bi-linearly interpolated for the horizontal position of the GNSS station. The
profile data were then used to obtain the weighted mean temperature over the station and subsequently
the conversion factor (X. Wang et al., 2016). In the temporal domain, all the hourly conversion factor values
at the GNSS station on the day were averaged for the daily conversion factor, then Equation 10 was used
to obtain the daily GNSS-PWV. The PWV time series of the station was formed from all daily GNSS-PWV
values of the station. All data gaps in the time series were filled using the first-order autoregressive model
developed from the least squares method described by Alshawaf et al. (2017).

4.2.  Result of Changepoint Detection

After the AAHT method was used to detect the changepoints in the above GNSS-PWV time series at the 91
IGS stations, 126 changepoints at 63 stations were detected. The number of the changepoints at each station
is shown in Figure 6.

ZHU ET AL. 13 of 21
23335084, 2021, 7, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2021EA001716 by Nat Prov Indonesia, Wiley Online Library on [17/05/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Earth and Space Science 10.1029/2021EA001716

Table 3
Epoch, Magnitude, and Difference in the Magnitudes of the Offsets at Each of the 43 Climate Changepoints Identified
From GNSS-PWV and ERA-PWV Time Series by AAHT
GNSS-PWV ERA-PWV
Difference
Station Epoch Magnitude (mm) Epoch Magnitude (mm) (mm)
AMC2 15 Sep 2014 1.017 19 Aug 2014 0.985 0.032
BAIE 21 Feb 2013 −0.933 2 Feb 2013 −1.186 0.253
BOGT 05 Jan 2007 −0.917 25 Feb 2007 −0.411 −0.506
BOGT 24 Oct 2011 −0.446 31 Oct 2011 −0.776 0.330
BOR1 27 Sep 2002 −1.198 16 Aug 2002 −1.942 0.744
BREW 06 Jan 2007 −1.597 27 Feb 2007 −1.088 −0.510
CHUM 20 Aug 2015 2.054 5 Jun 2015 2.157 −0.104
DRAO 04 Dec 2016 −1.200 13 Nov 2016 −1.184 −0.016
FAIR 28 Aug 2009 1.125 24 Jun 2009 0.596 0.528
FLIN 08 Oct 2006 −1.355 7 Oct 2006 −1.332 −0.023
FLIN 29 Sep 2012 −0.622 8 Sep 2012 −1.187 0.565
FLIN 01 Jul 2017 −0.934 22 Apr 2017 −1.322 0.388
GODE 07 May 2018 2.462 24 Dec 2017 2.314 0.148
GRAZ 28 May 2018 0.802 29 May 2018 1.347 −0.544
HOLM 06 Jan 2007 −0.826 26 Feb 2007 −0.687 −0.139
IENG 02 May 2018 1.552 4 Jul 2018 1.418 0.134
INVK 17 Jul 2005 0.312 18 Sep 2005 0.726 −0.413
JOZE 24 Jun 2002 −1.280 1 Sep 2002 −1.664 0.384
JOZE 18 May 2006 1.072 21 Apr 2006 1.216 −0.145
KIR0 16 Jan 2008 −0.758 6 Jan 2008 −0.640 −0.117
KIRU 27 Dec 2007 −1.063 6 Jan 2008 −0.640 −0.422
KIRU 02 Mar 2011 0.654 23 Dec 2010 0.831 −0.177
MARS 05 Jun 2018 0.896 11 Jun 2018 2.660 −1.763
MAS1 29 Nov 2006 −1.675 24 Oct 2006 −1.217 −0.457
MAS1 05 May 2009 1.173 24 Feb 2009 1.472 −0.299
MAS1 21 Feb 2011 −1.181 27 Jan 2011 −1.608 0.427
NAIN 03 Apr 2013 −0.687 14 May 2013 −2.911 2.224
NRC1 16 Jan 2007 −1.515 7 Dec 2006 −1.099 −0.416
PETS 05 Jun 2008 0.588 27 Apr 2008 1.488 −0.900
PETS 24 Jul 2009 −1.246 2 May 2009 −1.084 −0.162
PETS 01 Jun 2010 1.024 16 Aug 2010 1.391 −0.366
PETS 06 Oct 2012 0.730 24 Aug 2012 0.892 −0.162
PTBB 02 Sep 2007 −1.112 1 Nov 2007 −0.980 −0.132
REYK 31 Dec 2014 −0.819 31 Dec 2014 −0.566 −0.253
REYK 07 Feb 2016 1.044 9 Mar 2016 1.232 −0.188
SFDM 07 Jul 2006 −1.732 15 Apr 2006 −1.544 −0.188
SPT0 08 Jul 2007 −1.122 8 Jul 2007 −1.653 0.531
TOW2 20 Mar 2010 5.407 30 Mar 2010 7.182 −1.775
TOW2 01 Apr 2011 −6.353 25 Feb 2011 −6.561 0.208
UCLU 08 Feb 2006 −1.619 20 Nov 2005 −1.011 −0.608
UCLU 25 Oct 2016 −1.793 27 Dec 2016 −0.681 −1.112

ZHU ET AL. 14 of 21
23335084, 2021, 7, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2021EA001716 by Nat Prov Indonesia, Wiley Online Library on [17/05/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Earth and Space Science 10.1029/2021EA001716

Table 3
Continued
GNSS-PWV ERA-PWV
Difference
Station Epoch Magnitude (mm) Epoch Magnitude (mm) (mm)
ZIMM 10 Aug 2001 −0.838 4 Jun 2001 −1.211 0.373
ZIMM 13 Jul 2007 −0.899 16 Sep 2007 −0.912 0.014
Abbreviations: AAHT, adaptive absolute homogenization test; GNSS-PWV, Global Navigation Satellite Systems-
precipitable water vapor.

To identify climatic changepoints in the detected changepoints, the ERA-PWV time series were examined
first by AAHT, and the detected changepoints were compared with the changepoints detected from the
GNSS-PWV time series. If the pair of changepoints detected in the ERA-PWV and GNSS-PWV time series
occur at almost the same epochs (i.e., with less than 90  days separation), and have similar magnitudes,
then the detected changepoint in the GNSS-PWV time series is defined as a climatic changepoint (Hoseini
et al., 2020). The results of the epoch and magnitude of the offset at each of the 43 climatic changepoints,
which accounted for 34.1% of the total changepoints detected initially, identified at 28 GNSS stations are
listed in Table 3. The magnitudes of the offsets from the GNSS-PWV vary from −6.35 to 5.41 mm, and from
the ERA-PWV vary from −6.56 to 7.18 mm, while the values of difference from both time series are in the
range from −1.77 to 2.22 mm. This difference is likely to be due to the fact that different meteorological data
assimilated into the integrated forecasting system have different feedbacks to climatic drivers and the ERA5
data set is based on various types of the meteorological data (Hoseini et al., 2020). In addition, each pair
of the 43 climate changepoints detected from both types of time series having the same sign of offsets and
a slight difference (under 0.5 mm) in the magnitudes for most offsets further verify that they are climatic
changepoints.

After the above 43 climatic changepoints were removed from the changepoint list, the rest 83 changepoints
are regraded as artificial changepoints. To analysis the origins of these artificial changepoints, station-relat-
ed changepoints in the time series were identified by comparing the epoch of the changepoints remaining
in the list with the station metadata (the criterion is 90 days). As a result, 26 station-related changepoints

Figure 7.  Decomposition results and three detected changepoints of the precipitable water vapor time series over
the DRAO station detected by adaptive absolute homogenization test (in unit of mm). The red lines in (d) denote the
changepoint epochs.

ZHU ET AL. 15 of 21
23335084, 2021, 7, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2021EA001716 by Nat Prov Indonesia, Wiley Online Library on [17/05/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Earth and Space Science 10.1029/2021EA001716

Table 4
Station-Related Changepoints Identified by Comparing With Station Metadata
Site Epoch Offset (mm) Documented epoch Documented type
CAS1 27 Feb 2011 0.56 1 Feb 2011 Receiver
CIT1 24 Jun 2006 −1.44 3 Feb 2006 Receiver
DRAO 9 Sep 2007 −1.69 28 Aug 2007 Sensors
FAIR 16 Jul 2002 0.94 8 Jul 2002 Receiver
GOPE 1 Aug 2002 −1.45 3 May 2002 Sensors
INVK 12 Oct 2017 −0.64 17 Nov 2017 Receiver
KIR0 15 Apr 2013 0.83 3 Apr 2013 Receiver
MAS1 9 Nov 2013 0.30 14 Aug 2013 Receiver
MATE 21 Jan 2002 1.37 25 Sep 2001 Receiver
MATE 24 Nov 2006 −0.77 3 Jan 2007 Sensors
MATE 18 Feb 2009 0.66 5 May 2009 Receiver
MATE 29 Nov 2016 −1.58 30 Sep 2016 Antenna
MATE 11 Jul 2018 1.87 25 Jul 2018 Receiver
ONSA 24 Feb 2008 −1.33 15 May 2008 Receiver
PDEL 8 Dec 2016 1.19 27 Dec 2016 Receiver
PETS 2 Jan 2006 −0.41 5 Oct 2005 Antenna
REYK 9 Sep 2013 0.67 9 Oct 2013 Receiver
SCH2 10 Sep 2009 1.17 9 Nov 2009 Receiver
SCH2 30 Apr 2012 0.40 14 Feb 2012 Sensors
THU3 31 Oct 2006 −0.42 29 Aug 2006 Receiver
VALD 21 Jul 2006 −1.18 27 Jun 2006 Sensors
VALD 4 Sep 2017 −0.75 6 Jun 2017 Antenna
VILL 7 Feb 2007 −1.43 18 Apr 2007 Antenna
VIS0 18 May 2008 −1.04 15 May 2008 Receiver
YEBE 11 Apr 2007 −0.96 2 Apr 2007 Sensors
YELL 15 Oct 2015 1.14 3 Dec 2015 Receiver

(∼19%) were identified, see Table 4, and these changepoints were further classified into three types: (1) 17
changepoints are caused by a change of GNSS receiver hardware, for example, the receiver at the SCH2 sta-
tion was changed on November 9, 2009, which was detected on September 10, 2009 with a 1.17 mm offset;
(2) Four changepoints are related to a change of antenna, such as a maintenance of the antenna monument
at the MATE station on September 30, 2016 leading to a 1.55 mm offset; (3) Five changepoints are related
to a changeover of sensors, for example, both the meteorological sensors and gravity sensors were changed.

Figure  7 shows the detected changepoints as well as the decomposition results of the PWV time series
over the DRAO station, as the example. Subfigure (a) is the original PWV time series; (b) and (c) are the
smoothed series Ĉ  t  and decomposed periodic component S  t , respectively; (d) shows the decomposed
trend components being divided into four segments by three changepoints detected. The epochs of the three
changepoints are July 29, 2006, September 19, 2007 and December 4, 2016 with the offsets of −1.02, −1.69,
and −1.2 mm, respectively. The second changepoint is supported by the record on August 28, 2007 in the
station metadata, and the third one is a climatic changepoint. The first changepoint is neither a climatic
changepoint nor a station-related changepoint.

From the analyses for the rest 58 unverified changepoints, the factors related to changes in the climatic
driver, station metadata and data processing strategy were all excluded; hence, these changepoints might be
related to inaccurate station metadata.

ZHU ET AL. 16 of 21
23335084, 2021, 7, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2021EA001716 by Nat Prov Indonesia, Wiley Online Library on [17/05/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Earth and Space Science 10.1029/2021EA001716

Figure 8.  Scatterplots of the trends between the Global Navigation Satellite Systems-precipitable water vapor (GNSS-
PWV; y-axis) and ERA-PWV (x-axis) time series at 48 stations before and after homogenization, respectively. The blue
and red dots are the trends of the original and homogenized GNSS-PWV time series, respectively, and the two lines are
their fitting results, where the slopes of the lines indicate the correlation coefficients of the two sets of the trends.

4.3.  Long-Term Trends

After the aforementioned 43 climatic changepoints (see Table 3) were removed from the list of the initially
detected 126 changepoints (at 48 GNSS stations), the 83 changepoints that remained in the changepoint
list were defined as final artificial changepoints since the low FAR of AAHT. Then Equation 7 was used to
re-estimate the magnitudes of the offsets for all the changepoints, and each of the estimated magnitudes
was added to the original PWV time series from the epoch of the changepoint till the next changepoint or
the last epoch of the time series. This process is the so-called homogenization process and the new time se-
ries are homogenized. The long-term trend estimates and their uncertainties (the standard errors estimated
from the least squares method) from both the original and homogenized GNSS-PWV time series obtained
from the least squares method based on the fitting model expressed by Equation 8, were compared with the
trends obtained from the ERA-PWV time series. It is noted that the ERA-PWV time series were uncorrected
since the aforementioned verification procedures cannot be applied.

Table 5 summarizes the estimation results of the above 48 GNSS-PWV time series. All the trends estimated
from the original time series are in the range from −0.100 to 0.173 mm/year; while after homogenization the
trend values become larger, in the range from −0.130 to 0.226 mm/year; the data in the Difference column
are in the range from −0.188 to 0.171 mm/year; the uncertainties shown in the first two columns at the
same station are almost same.

Figure 8 shows the scatterplot of the estimated trends of the GNSS-PWV time series (before and after ho-
mogenization) and the ERA-PWV time series at the 48 stations. Compared with the results of the original
GNSS-PWV, the range of the homogenized trends at the 48 stations become larger than that of original
trends, and its correlation with the ERA-PWV trend is much larger (0.23 vs. 0.75). The significant difference
in the two coefficients suggests the effectiveness of the new method.

5. Conclusions
Inhomogeneity is a well-known issue in PWV time series retrieved from the observations of GNSS-PWV,
which is caused by various factors and can adversely affect climatic analyses. The use of inhomogeneous
time series for climate applications is likely to result in biased or even wrong conclusions, especially in the
investigation of the variation trends of the long-term time series. Relative methods, which are applied to the
differential time series, can eliminate the common effects of climatic drivers and periodic evolution. How-
ever, the performance of these methods is affected by factors such as the spatial distribution of reference

ZHU ET AL. 17 of 21
23335084, 2021, 7, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2021EA001716 by Nat Prov Indonesia, Wiley Online Library on [17/05/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Earth and Space Science 10.1029/2021EA001716

Table 5
Long-Term Trend and its Uncertainty of the 43 GNSS-PWV Time Series (at 43 Stations) Before and After
Homogenization (the ERA-PWV Trend was the Reference)
GNSS-PWV trend (mm/year)
ERA-PWV trend
Station Before After Difference (mm/year)
ALBH 0.009 ± 0.011 0.187 ± 0.011 −0.178 0.058
AMC2 −0.010 ± 0.014 0.085 ± 0.014 −0.094 0.036
BAIE 0.066 ± 0.011 0.085 ± 0.011 −0.019 0.031
BOGT 0.041 ± 0.012 0.130 ± 0.012 −0.089 0.050
BOR1 −0.017 ± 0.004 −0.021 ± 0.004 0.004 −0.023
BREW −0.043 ± 0.009 −0.049 ± 0.009 0.006 −0.026
CAS1 0.084 ± 0.012 0.185 ± 0.012 −0.102 0.020
CHUM −0.028 ± 0.011 −0.103 ± 0.011 0.075 −0.008
CIT1 0.021 ± 0.009 0.122 ± 0.009 −0.100 0.022
DRAO 0.048 ± 0.010 0.120 ± 0.009 −0.072 −0.005
FAIR −0.037 ± 0.007 −0.130 ± 0.006 0.093 0.014
FLIN −0.011 ± 0.012 0.001 ± 0.012 −0.011 0.062
GODE 0.023 ± 0.009 0.045 ± 0.009 −0.023 0.041
GOLD −0.058 ± 0.011 −0.014 ± 0.011 −0.043 0.040
GOPE 0.005 ± 0.015 0.108 ± 0.015 −0.103 0.084
GRAZ −0.013 ± 0.009 −0.107 ± 0.009 0.094 0.038
HERT 0.009 ± 0.004 0.018 ± 0.004 −0.009 0.035
HOLM 0.049 ± 0.006 0.114 ± 0.006 −0.065 0.058
IENG −0.015 ± 0.011 0.119 ± 0.011 −0.134 0.024
INVK −0.053 ± 0.007 −0.101 ± 0.007 0.047 0.017
JOZE −0.044 ± 0.011 −0.123 ± 0.011 0.079 0.050
JPLM 0.001 ± 0.016 0.124 ± 0.016 −0.124 0.086
KIR0 −0.047 ± 0.010 −0.093 ± 0.010 0.046 0.002
KIRU −0.035 ± 0.011 −0.102 ± 0.011 0.067 0.070
MARS 0.042 ± 0.007 0.002 ± 0.007 0.040 0.043
MAS1 −0.046 ± 0.009 −0.017 ± 0.009 −0.029 −0.030
MATE −0.012 ± 0.011 0.160 ± 0.011 −0.172 0.050
NAIN 0.000 ± 0.013 0.013 ± 0.013 −0.013 0.050
NRC1 −0.041 ± 0.011 0.065 ± 0.011 −0.105 0.049
ONSA −0.090 ± 0.014 −0.130 ± 0.014 0.040 0.018
PDEL 0.090 ± 0.009 0.169 ± 0.009 −0.078 0.008
PETS 0.040 ± 0.008 0.122 ± 0.008 −0.082 0.042
PTBB −0.021 ± 0.005 −0.001 ± 0.005 −0.020 −0.007
REYK −0.044 ± 0.008 −0.100 ± 0.008 0.056 0.015
SASK 0.002 ± 0.010 0.063 ± 0.010 −0.062 0.015
SCH2 −0.005 ± 0.007 −0.052 ± 0.007 0.048 −0.002
SFDM 0.053 ± 0.011 0.152 ± 0.011 −0.099 0.037
SPT0 −0.004 ± 0.009 −0.014 ± 0.009 0.010 0.049
THU3 −0.014 ± 0.003 0.016 ± 0.003 −0.030 −0.012
TOW2 0.002 ± 0.013 0.097 ± 0.013 −0.096 0.038

ZHU ET AL. 18 of 21
23335084, 2021, 7, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2021EA001716 by Nat Prov Indonesia, Wiley Online Library on [17/05/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Earth and Space Science 10.1029/2021EA001716

Table 5
Continued
GNSS-PWV trend (mm/year)
ERA-PWV trend
Station Before After Difference (mm/year)
UCLU −0.031 ± 0.011 0.076 ± 0.011 −0.107 0.054
VALD −0.033 ± 0.010 0.051 ± 0.010 −0.083 0.033
VILL 0.093 ± 0.012 0.126 ± 0.012 −0.033 0.042
VIS0 0.049 ± 0.010 0.162 ± 0.009 −0.113 0.055
YEBE −0.016 ± 0.012 −0.091 ± 0.012 0.075 0.042
YELL −0.010 ± 0.012 −0.126 ± 0.012 0.115 0.097
ZIMM −0.008 ± 0.010 0.065 ± 0.010 −0.073 0.052
Abbreviation: GNSS-PWV, Global Navigation Satellite Systems-precipitable water vapor.

stations, potential inherent inhomogeneities in the reference time series, and the systematic discrepancy
between the reference and original time series. To address these issues, a new absolute method combining
the STL and PMFred, named the AAHT, is proposed in this study. The performance of the proposed method
was evaluated using Monte Carlo simulations with a data set of 225,000 time series based on the characteris-
tics of the GNSS-PWV time series on a global scale. Statistical results illustrated that the overall success rate
and false alarm rate were better than 74.9% and less than 12.2%, respectively. The precision of the estimates
of the epochs and the offset magnitudes at detected changepoints were 8.0 days and 0.153 mm, respectively.

The AAHT method was also tested using the GNSS-PWV time series over 91 global GNSS stations and
results showed 126 changepoints at 63 stations were detected. The source/type of these changepoints
was validated by comparing them against the changepoints detected in the corresponding PWV time
series retrieved by the ERA5 reanalysis data set (ERA-PWV) and GNSS station metadata. It was found
that 43 changepoints are climatic changepoints, other 83 detections are the artificial changepoints at 48
inhomogeneous GNSS-PWV time series, 26 out of which are related to the station instrument changes.
Among these validated changepoints, both the directions and magnitudes of all climatic changepoints
in both the GNSS-PWV and ERA-PWV agreed well. The offsets of all the 83 artificial changepoints were
re-estimated and used to homogenize these 48 inhomogeneous time series. Then the trends of the 48
corrected time series were estimated by the least squares method. Results showed that the trend ranges
of these time series became larger in comparison with that before homogenization, while the uncer-
tainties of the trends varied slightly. Moreover, the correlation between the long-term homogenized
GNSS-PWV and ERA-PWV time series was 0.75, in comparison with the value of 0.23 from the original
non-homogenized time series, which is a significant improvement. Our future work will be primarily
focused on further performance improvement of the AAHT method on the detection of changepoints
and the estimation of the magnitude of the offsets by lowering the aliasing mechanism caused by the
slight offsets and the strong noise.

Data Availability Statement


Acknowledgments The homogenized GNSS-PWV time series over 91 IGS stations in this study are available on the zenodo
This research study was supported by
the National Natural Science Foun- at https://zenodo.org/record/4567709#.YDnLSdgzaUn. ERA5 reanalysis data set can be freely available
dations of China (No. 41730109), the through the ECMWF Copernicus Climate Data Store at https://www.ecmwf.int/en/forecasts/datasets/rea-
China Postdoctoral Science Foundation nalysis-datasets/era5. ZTD products from IGS repro 2 can be available from IGS Analysis Centers.
(No. 2020M671645), the Natural Sci-
ence Foundation of Jiangsu Province,
China (No. BK20200646). The authors
thank the European Centre for Medi-
um-Range Weather Forecasts (ECMWF,
References
https://www.ecmwf.int/) for providing Alshawaf, F., Balidakis, K., Dick, G., Heise, S., & Wickert, J. (2017). Estimating trends in atmospheric water vapor and temperature time
ERA5 reanalysis data set and Interna- series over Germany. Atmospheric Measurement Techniques, 10(9), 3117–3132. https://doi.org/10.5194/amt-10-3117-2017
tional GNSS Service (IGS, http://www. Amiri-Simkooei, A. R., Hosseini-Asl, M., Asgari, J., & Zangeneh-Nejad, F. (2019). Offset detection in GPS position time series using multi-
igs.org) for providing ZTD products. variate analysis. GPS Solutions, 23(1). https://doi.org/10.1007/s10291-018-0805-z

ZHU ET AL. 19 of 21
23335084, 2021, 7, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2021EA001716 by Nat Prov Indonesia, Wiley Online Library on [17/05/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Earth and Space Science 10.1029/2021EA001716

Bevis, M., Businger, S., Herring, T. A., Rocken, C., Anthes, R. A., & Ware, R. H. (1992). GPS meteorology: Remote sensing of atmospheric
water vapor using the global positioning system. Journal of Geophysical Research, 97(D14), 15787. https://doi.org/10.1029/92jd01517
Bock, O., Collilieux, X., Guillamon, F., Lebarbier, E., & Pascal, C. (2020). A breakpoint detection in the mean model with heterogeneous
variance on fixed time intervals. Statistics and Computing, 30(1), 195–207. https://doi.org/10.1007/s11222-019-09853-5
Bock, O., & Parracho, A. C. (2019). Consistency and representativeness of integrated water vapor from ground-based GPS observations and
ERA-Interim reanalysis. Atmospheric Chemistry and Physics, 19(14), 9453–9468. https://doi.org/10.5194/acp-19-9453-2019
Byun, S. H., & Bar-Sever, Y. E. (2009). A new type of troposphere zenith path delay product of the international GNSS service. Journal of
Geodesy, 83(3–4), 367–7. https://doi.org/10.1007/s00190-008-0288-8
Cleveland, R. B., Cleveland, W. S., McRae, J. E., & Terpenning, I. (1990). STL: A Seasonal-Trend decomposition procedure based on Loess
(with discussion). Journal of Official Statistics, 6, 3–73.
Dai, A., Wang, J., Thorne, P. W., Parker, D. E., Haimberger, L., & Wang, X. L. (2011). A new approach to homogenize daily radiosonde
humidity data. Journal of Climate, 24(4), 965–991. https://doi.org/10.1175/2010JCLI3816.1
Dai, A., Wang, J., Ware, R. H., & Van Hove, T. (2002). Diurnal variation in water vapor over North America and its implications for sampling
errors in radiosonde humidity. Journal of Geophysical Research: Atmospheres, 107(D10), 11–1. https://doi.org/10.1029/2001JD000642
Easterling, D. R., & Peterson, T. C. (1995). A new method for detecting undocumented discontinuities in climatological time series. Inter-
national Journal of Climatology, 15(4), 369–377. https://doi.org/10.1002/joc.3370150403
Ferguson, C. R., & Villarini, G. (2012). Detecting inhomogeneities in the Twentieth Century Reanalysis over the central United States.
Journal of Geophysical Research: Atmospheres, 117(D5). https://doi.org/10.1029/2011JD016988
Gazeaux, J., Williams, S., King, M., Bos, M., Dach, R., Deo, M., et al. (2013). Detecting offsets in GPS time series: First results from the de-
tection of offsets in GPS experiment. Journal of Geophysical Research: Solid Earth, 118(5), 2397–2407. https://doi.org/10.1002/jgrb.50152
Gradinarsky, L. P., Johansson, J. M., Bouma, H. R., Scherneck, H. G., & Elgered, G. (2002). Climate monitoring using GPS. Physics and
Chemistry of the Earth, Parts A/B/C, 27(4–5), 335–340. https://doi.org/10.1016/S1474-7065(02)00009-8
Guo, Y., & Ding, Y. (2011). Impacts of reference time series on the homogenization of radiosonde temperature. Advances in Atmospheric
Sciences, 28(5), 1011–1022. https://doi.org/10.1007/s00376-010-9211-3
Hersbach, H., Bell, B., Berrisford, P., Hirahara, S., Horányi, A., Muñoz-Sabater, J., et al. (2020). The ERA5 global reanalysis. Quarterly
Journal of the Royal Meteorological Society, 146(730), 1999–2049. https://doi.org/10.1002/qj.3803
Hodnebrog, Ø., Myhre, G., Samset, B. H., Alterskjær, K., Andrews, T., Boucher, O., et al. (2019). Water vapor adjustments and responses
differ between climate drivers. Atmospheric Chemistry and Physics, 19(20), 12887–12899. https://doi.org/10.5194/acp-19-12887-2019
Hoseini, M., Alshawaf, F., Nahavandchi, H., Dick, G., & Wickert, J. (2020). Toward a zero-difference approach for homogenizing GNSS
tropospheric products. GPS Solutions, 24(1). https://doi.org/10.1007/s10291-019-0915-2
Huang, N. E., Shen, Z., Long, S. R., Wu, M. C., Shih, H. H., Zheng, Q., et al. (1998). The empirical mode decomposition and the Hilbert
spectrum for nonlinear and non-stationary time series analysis. Proceedings of the Royal Society of London. Series A: Mathematical,
Physical and Engineering Sciences, 454, 903–995. https://doi.org/10.1098/rspa.1998.0193
Jin, S., Li, Z., & Cho, J. (2008). Integrated water vapor field and multiscale variations over China from GPS measurements. Journal of
Applied Meteorology and Climatology, 47(11), 3008–3015. https://doi.org/10.1175/2008JAMC1920.1
Klos, A., Hunegnaw, A., Teferle, F. N., Abraha, K. E., Ahmed, F., & Bogusz, J. (2016). Noise characteristics in Zenith Total Delay from homo-
geneously reprocessed GPS time series. Atmospheric Measurement Techniques Discussions, 1–28. https://doi.org/10.5194/amt-2016-385
Klos, A., Hunegnaw, A., Teferle, F. N., Abraha, K. E., Ahmed, F., & Bogusz, J. (2018). Statistical significance of trends in Zenith Wet Delay
from re-processed GPS solutions. GPS Solutions, 22(2). https://doi.org/10.1007/s10291-018-0717-y
Larson, K. M., Braun, J. J., Small, E. E., Zavorotny, V. U., Gutmann, E. D., & Bilich, A. L. (2010). GPS multipath and its relation to near-sur-
face soil moisture content. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 3(1), 91–99. https://doi.
org/10.1109/JSTARS.2009.2033612
Lindau, R. (2003). Errors of Atlantic air-sea fluxes derived from ship observations. Journal of Climate, 16(4), 783–788. https://doi.
org/10.1175/1520-0442(2003)016<0783:EOAASF>2.0.CO;2
Lindau, R., & Venema, V. (2013). On the multiple breakpoint problem and the number of significant breaks in homogenization of climate
records. Idojaras, 117(1), 1–34.
Lindau, R., & Venema, V. (2018). The joint influence of break and noise variance on the break detection capability in time series homog-
enization. Advances in Statistical Climatology, Meteorology and Oceanography, 4(1/2), 1–18. https://doi.org/10.5194/ascmo-4-1-2018
Nilsson, T., & Elgered, G. (2008). Long-term trends in the atmospheric water vapor content estimated from ground-based GPS data. Journal
of Geophysical Research Atmospheres, 113(19). https://doi.org/10.1029/2008JD010110
Ning, T., & Elgered, G. (2012). Trends in the atmospheric water vapor content from ground-based GPS: The impact of the elevation cut-
off angle. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 5(3), 744–751. https://doi.org/10.1109/
JSTARS.2012.2191392
Ning, T., Elgered, G., & Johansson, J. M. (2011). The impact of microwave absorber and radome geometries on GNSS measurements of
station coordinates and atmospheric water vapor. Advances in Space Research, 47(2), 186–196. https://doi.org/10.1016/j.asr.2010.06.023
Ning, T., Wang, J., Elgered, G., Dick, G., Wickert, J., Bradke, M., et al. (2016). The uncertainty of the atmospheric integrated water vapor
estimated from GNSS observations. Atmospheric Measurement Techniques, 9(1), 79–92. https://doi.org/10.5194/amt-9-79-2016
Ning, T., Wickert, J., Deng, Z., Heise, S., Dick, G., Vey, S., & Schöne, T. (2016). Homogenized time series of the atmospheric water vapor
content obtained from the GNSS reprocessed data. Journal of Climate, 29(7), 2443–2456. https://doi.org/10.1175/JCLI-D-15-0158.1
O'Gorman, P. A., & Muller, C. J. (2010). How closely do changes in surface and column water vapor follow Clausius–Clapeyron scaling in
climate change simulations? Environmental Research Letters, 5(2), 025207. https://doi.org/10.1088/1748-9326/5/2/025207
Parracho, A. C., Bock, O., & Bastin, S. (2018). Global IWV trends and variability in atmospheric reanalyzes and GPS observations. Atmos-
pheric Chemistry and Physics, 18(22), 16213–16237. https://doi.org/10.5194/acp-18-16213-2018
Pierdicca, N., Guerriero, L., Giusto, R., Brogioni, M., & Egido, A. (2014). SAVERS: A simulator of GNSS reflections from bare and vegetated
soils. IEEE Transactions on Geoscience and Remote Sensing, 52(10), 6542–6554. https://doi.org/10.1109/TGRS.2013.2297572
Rodionov, S. N. (2004). A sequential algorithm for testing climate regime shifts. Geophysical Research Letters, 31(9). https://doi.
org/10.1029/2004GL019448
Saastamoinen, J. (1972). Atmospheric correction for the troposphere and stratosphere in radio ranging satellites. The Use of Artificial
Satellites for Geodesy, 15, 247–251.
Schröder, M., Lockhoff, M., Forsythe, J. M., Cronk, H. Q., Haar, T. H. V., & Bennartz, R. (2016). The GEWEX water vapor assessment:
Results from intercomparison, trend, and homogeneity analysis of total column water vapor. Journal of Applied Meteorology and Clima-
tology, 55(7), 1633–1649. https://doi.org/10.1175/jamc-d-15-0304.1

ZHU ET AL. 20 of 21
23335084, 2021, 7, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2021EA001716 by Nat Prov Indonesia, Wiley Online Library on [17/05/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Earth and Space Science 10.1029/2021EA001716

Sherwood, S. C., Roca, R., Weckwerth, T. M., & Andronova, N. G. (2010). Tropospheric water vapor, convection, and climate. Reviews of
Geophysics, 48(2). https://doi.org/10.1029/2009rg000301
Solow, A. R. (1987). Testing for Climate Change: An Application of the Two-Phase Regression Model. Journal of Climate and Applied
Meteorology, 26(10). https://doi.org/10.1175/1520-0450(1987)026<1401:TFCCAA>2.0.CO;2
Steigenberger, P., Tesmer, V., Krügel, M., Thaller, D., Schmid, R., Vey, S., & Rothacher, M. (2007). Comparisons of homogeneously repro-
cessed GPS and VLBI long time-series of troposphere zenith delays and gradients. Journal of Geodesy, 81(6–8), 503–514. https://doi.
org/10.1007/s00190-006-0124-y
Teunissen, P. J. G. (2018). Distributional theory for the DIA method. Journal of Geodesy, 92(1), 59–80. https://doi.org/10.1007/
s00190-017-1045-7
Thomas, I. D., King, M. A., Clarke, P. J., & Penna, N. T. (2011). Precipitable water vapor estimates from homogeneously reprocessed GPS
data: An intertechnique comparison in Antarctica. Journal of Geophysical Research, 116(D4). https://doi.org/10.1029/2010JD013889
Trenberth, K. E., Fasullo, J., & Smith, L. (2005). Trends and variability in column-integrated atmospheric water vapor. Climate Dynamics,
24(7–8), 741–758. https://doi.org/10.1007/s00382-005-0017-4
VanMalderen, R., Pottiaux, E., Klos, A., Domonkos, P., Elias, M., Ning, T., et al. (2020). Homogenizing GPS integrated water vapor time series:
Benchmarking break detection methods on synthetic data sets. Earth and Space Science, 7(5). https://doi.org/10.1029/2020EA001121
Venema, V. K. C., Mestre, O., Aguilar, E., Auer, I., Guijarro, J. A., Domonkos, P., et al. (2012). Benchmarking homogenization algorithms
for monthly data. Climate of the Past, 8(1), 89–115. https://doi.org/10.5194/cp-8-89-2012
Vey, S., Dietrich, R., Fritsche, M., Rülke, A., Steigenberger, P., & Rothacher, M. (2009). On the homogeneity and interpretation of pre-
cipitable water time series derived from global GPS observations. Journal of Geophysical Research, 114(D10), D10101. https://doi.
org/10.1029/2008JD010415
Wang, J., Zhang, L., Dai, A., Van Hove, T., & Van Baelen, J. (2007). A near-global, 2-hourly data set of atmospheric precipitable water from
ground-based GPS measurements. Journal of Geophysical Research, 112(D11), D11107. https://doi.org/10.1029/2006JD007529
Wang, X., Zhang, K., Wu, S., Fan, S., & Cheng, Y. (2016). Water vapor-weighted mean temperature and its impact on the determina-
tion of precipitable water vapor and its linear trend. Journal of Geophysical Research: Atmospheres, 121(2), 833–852. https://doi.
org/10.1002/2015JD024181
Wang, X., Zhang, K., Wu, S., Li, Z., Cheng, Y., Li, L., & Yuan, H. (2018). The correlation between GNSS-derived precipitable water vapor
and sea surface temperature and its responses to El Niño–Southern Oscillation. Remote Sensing of Environment, 216, 1–12. https://doi.
org/10.1016/j.rse.2018.06.029
Wang, X. L. (2008a). Accounting for autocorrelation in detecting mean shifts in climate data series using the penalized maximal t- or F-test.
Journal of Applied Meteorology and Climatology, 47(9), 2423–2444. https://doi.org/10.1175/2008JAMC1741.1
Wang, X. L. (2008b). Penalized maximal F-test for detecting undocumented mean shift without trend change. Journal of Atmospheric and
Oceanic Technology, 25(3), 368–384. https://doi.org/10.1175/2007JTECHA982.1
Wang, X. L., Wen, Q. H., & Wu, Y. (2007). Penalized maximal t-Test for detecting undocumented mean change in climate data series. Jour-
nal of Applied Meteorology and Climatology, 46(6), 916–931. https://doi.org/10.1175/JAM2504.1
Wu, D., Yan, H., & Yuan, S. (2018). L1 regularization for detecting offsets and trend change points in GNSS time series. GPS Solutions,
22(3), 88. https://doi.org/10.1007/s10291-018-0756-4
Zhang, W., Lou, Y., Haase, J. S., Zhang, R., Zheng, G., Huang, J., et al. (2017). The use of ground-based GPS precipitable water measure-
ments over China to assess radiosonde and ERA-interim moisture trends and errors from 1999 to 2015. Journal of Climate, 30(19),
7643–7667. https://doi.org/10.1175/JCLI-D-16-0591.1
Zhao, Q., Ma, X., Yao, W., Liu, Y., & Yao, Y. (2020). A drought monitoring method based on precipitable water vapor and precipitation.
Journal of Climate, 33(24), 10727–10741. https://doi.org/10.1175/JCLI-D-19-0971.1
Zhao, Q., Yao, Y., & Yao, W. (2019). Studies of precipitable water vapor characteristics on a global scale. International Journal of Remote
Sensing, 40(1), 72–88. https://doi.org/10.1080/01431161.2018.1492177

ZHU ET AL. 21 of 21

You might also like