assisted adaptive extended kalman filter for low-cost single-frequency gps/sbas kinematic...
TRANSCRIPT
ORIGINAL ARTICLE
Assisted adaptive extended Kalman filter for low-costsingle-frequency GPS/SBAS kinematic positioning
Shiou-Gwo Lin
Received: 4 March 2013 / Accepted: 23 April 2014
� Springer-Verlag Berlin Heidelberg 2014
Abstract Using a low-cost single-frequency global
positioning system (GPS) receiver for kinematic posi-
tioning, the ambiguity resolution requires longer data
accumulation time compared to survey-grade dual-fre-
quency receivers. A satellite-based augmentation system
(SBAS) can provide added measurement to help solve this
problem. However, the SBAS signal strength is weaker
and the satellite orbit and clock errors are greater than
those of GPS satellites. The difference in phase mea-
surement quality between GPS and SBAS satellites must
be considered to prevent an unstable positioning result or
diverging position solution. This study proposes using the
assisted adaptive extended Kalman filter (AAEKF)
method to address this problem. The concept of AAEKF
involves using measurements from the reference station to
estimate the errors of each satellite. This information is
then employed to dynamically adjust the corresponding
measurement model of the extended Kalman filter. The
proposed method was validated with 24 h of experiment
data from four different baselines obtained using a con-
sumer-grade L1 GPS receiver. The experimental results
show that AAEKF can be successfully employed for GPS/
SBAS kinematic positioning. The ambiguity resolution
success rates of 2, 5, and 10 min of measurements
improved by about 3.2, 2.4, and 1.6 times, respectively,
and the positioning accuracy of the north, east, and height
directions improved by 14–44, 17–56, and 9–53 %, when
adding the SBAS measurement.
Keywords GPS � SBAS � Adaptive extended Kalman
filter � Kinematic positioning
Introduction
Centimeter-level accuracy can be achieved by using real-
time kinematic (RTK) positioning with low-cost, single-
frequency global positioning system (GPS) receivers and
open source software. Using such a RTK system could
considerably reduce the total project costs of many appli-
cations (Takasu and Yasuda 2008). However, without
precision code (P code) and dual-frequency carrier phase,
the baseline length is limited typically to \10 km, and the
ambiguity resolution requires a longer data accumulation
time. The disadvantage of the limited baseline length can
be improved with multiple reference station technology
(Odijk et al. 2000; Fotopoulos and Cannon 2001; Zhang
and Lachapelle 2001). The objective of this study is to
improve the ambiguity resolution efficiency. The ambigu-
ity resolution efficiency is related to the satellite geometry
and measurement accuracy. Generally, for dual-frequency
receivers, the initialization period of ambiguity resolution
is about 1 min or less (Tiberius and de Jonge 1995), and the
initialization period for single-frequency instruments is
about 15–45 min (Takasu and Yasuda 2008).
Many studies have adopted assistance from satellite-
based augmentation system (SBAS), GLONASS, or iner-
tial navigation system (INS) to increase the number of
measurements and improve the ambiguity resolution effi-
ciency of single-frequency GNSS receivers (Kozlov and
Tkachenko 1998; Skaloud 1998; Wirola et al. 2006; Bo-
riskin et al. 2007). Of these methods, SBAS assistance is a
convenient solution, because many low-cost single-fre-
quency receivers support SBAS. SBAS and GPS satellites
S.-G. Lin (&)
Department of Communications, Navigation and Control
Engineering, National Taiwan Ocean University,
2 Pei-Ning Road, Keelung 20224, Taiwan
e-mail: [email protected]
123
GPS Solut
DOI 10.1007/s10291-014-0381-9
have the same L1 carrier frequency and the same family of
pseudorandom noise (PRN) code. This intentional design
means that GPS receivers can easily support SBAS sys-
tems, and the corresponding pseudorange and carrier phase
measurements equal GPS L1 measurements. The most
substantial difference between SBAS and GPS L1 is the
space segment. The SBAS signal is broadcast by multi-
functional satellites; for example, the European Geosta-
tionary Navigation Overlay Service and Wide Area Aug-
mentation System are primarily broadcast by multiple
geostationary communication and navigation satellites and
the Multi-functional Satellite Augmentation System is
broadcast by multiple geostationary weather and naviga-
tion satellites. These satellites were not specifically
designed for navigation, nor are navigation their primary
function. The space-borne clock errors and orbit errors are
greater than those of GPS satellites. Analyzing the received
measurements of SBAS satellites shows that the signal
strength of SBAS measurements is weaker compared with
GPS satellites.
The GPS provides global coverage with four to eight
simultaneously observable satellites at an elevation above
15� at any time of day. If the elevation mask is reduced to
10�, four to ten satellites are visible (Hofmann-Wellenhof
et al. 1994). Hence, it was possible that five or fewer sat-
ellites could be visible in certain periods in the survey
campaign, because of poor satellite geometry or because of
obstruction by buildings or trees. Most regions in the world
can receive at least two SBAS satellite signals. When the
GPS satellite geometry is poor, the SBAS satellite data can
provide a significant improvement, which can reduce the
required ambiguity resolution time. Compared to GPS
measurement, the signal-to-noise ratio of SBAS measure-
ment is relatively low, and the satellite orbital and clock
errors are larger and unstable. When using SBAS mea-
surements for relative positioning, SBAS satellite errors
must be appropriately considered to obtain a less biased
positioning solution.
The extended Kalman filter (EKF) is the nonlinear
version of Kalman filter (KF), and it has been widely used
for GNSS real-time kinematic positioning (Leick 2004).
However, one of the disadvantages of the EKF is its high
dependence on a priori knowledge of the potentially
unstable process and measurement noise statistics. Using
incomplete a priori statistics when designing a KF can lead
to significant estimation errors and even divergence prob-
lems. The purpose of an adaptive filter is to reduce or limit
these errors by modifying the KF to real data (Beutler et al.
1988).
The GNSS relative positioning technique requires
high-quality GNSS receivers at reference stations. The
reference station collects GNSS measurements for each
satellite in view. Measurements from the reference sta-
tion can be combined with measurements from the rover
station to form the double-difference (DD) measurements
used as the basis for relative positioning. The measure-
ments from the reference station can also be used to
estimate the error components of each satellite (Loomis
et al. 1989; Jin 1996; Farrell and Givargis 2000). The
value of error components and the distance between the
rover and reference stations can be used to estimate the
statistical information of DD measurement noise, which
is called ‘‘assisted statistical knowledge.’’ To prevent
divergence problems caused by errors from the SBAS
satellite when using the EKF approach, assisted statistical
knowledge can be applied to modified adaptive filter.
This method is called assisted adaptive extended Kalman
filter (AAEKF) and was used in this study for GPS/SBAS
kinematic positioning.
Theory background
This study modifies and uses an adaptive KF for GPS/
SBAS kinematic positioning. The GPS measurement
equations, the KF, the adaptive KF, and the proposed
AAEKF are briefly described below.
GPS measurement
The measurement equation of the L1 carrier phase and
pseudorange can be written as (Leick 2004)
ug1p ¼
f1
cRg
p þ Tgp � Ig
p þMPug
1p
� �� f1 dtp � dtg
� �� N
g1p
þ mug
1p
ð1Þ
qg1p ¼ Rg
p þ Tgp þ Ig
p þMPqg
1p� c � dtp þ c � dtg þ mqg
1pð2Þ
where ug1p is the phase measurement from the p receiver to
the g satellite, qg1p is the pseudorange, Rg
p is the geometric
range, Tgp is the tropospheric delay, dtp and dtg are the p
receiver and g satellite clock errors, Igp is the ionospheric
delay, Ng1p is the phase ambiguity, MPug
1pis the phase
multipath, MPqg
1pis the pseudorange multipath, mug
1pis the
phase noise, and mqg
1pis the pseudorange noise.
Linear DD phase combinations between two satellites,
reference receivers, and rover receivers are highly effective
for eliminating common-mode orbital, timing errors, and
atmospheric errors. A DD model was employed for this
study, and its measurement equation can be written as
(Leick 2004)
GPS Solut
123
ugh1pq ¼
f1
cRgh
pq þ Tghpq � Igh
pq þMPugh
1pq
� �� N
gh1pq þ mugh
1pqð3Þ
qgh1pq ¼ Rgh
pq þ Tghpq þ Igh
pq þMPqgh
1pq
þ mqgh
1pq
ð4Þ
The multipath is primarily caused by reflecting surfaces
near the receiver; this effect depends on the specific
receiver–satellite–reflector geometry (Hofmann-Wellenhof
et al. 1994). It does not cancel in the DD model.
Kalman filters
Kalman filters are widely adopted for GNSS kinematic
positioning; their standard formalism can be written as
follows (Gelb 1974):
xk ¼ Tk�1xk�1 þ wk�1;wk�1�Nð0;Qk�1Þ ð5Þzk ¼ Hkxk þ mk; mk �Nð0;RkÞ ð6Þ
Eq. (5) is a dynamic equation, and (6) is a measurement
equation, where k represents the epoch of time, x is the
system state vector, T is the parameter transformation
matrix, and wk�1 is the noise of the dynamic model
N 0;Qk�1ð Þ with Gaussian distribution (mean 0 and
covariance matrix Qk�1), Hk represents the design matrix,
and mk represents the noise vector of measurements
Nð0;Rk�1Þ with Gaussian distribution (mean 0 and
covariance matrix Rk�1). KF estimates are divided into the
following two steps: the time update and the measurement
update. The equation for the time update is as follows:
xk �ð Þ ¼ Tk�1xk�1ðþÞ ð7Þ
Pk �ð Þ ¼ Tk�1Pk�1 þð ÞTTk�1 þQk�1 ð8Þ
where Pk represents the covariance matrix of the system
state, (2) represents the predicted results, and (?) repre-
sents the updated results. Thus, the measurement update
equation is
xk þð Þ ¼ xk �ð Þ þKk½zk �Hkxk �ð Þ� ð9ÞPk þð Þ ¼ I�KkHk½ �Pkð�Þ ð10Þ
Kk ¼ Pk �ð ÞHTk ½HkPk �ð ÞHT
k þ Rk��1 ð11Þ
where Kk is the Kalman gain matrix. The KF algorithm
begins with the initial condition values x0ð�Þ and P0ð�Þ.When new measurements of zk become available, the time
update and measurement update steps follow.
Adaptive Kalman filters
Kalman filters require a priori knowledge of the process
and measurement noise statistics. Achieving full compli-
ance of this requirement under all circumstances is diffi-
cult. However, using incomplete information may result in
significant errors and even divergence problems. The
application of adaptive filters is one of the common strat-
egies for preventing this problem (Mehra 1972; Gelb
1974). The main function of adaptive filters is to estimate
the difference between the actual filter residuals and the
estimated noise statistics and to correct the estimated sta-
tistics using the filter residuals. Adaptive filters typically
employ the covariance matching technique. Innovation
sequences are also used by adaptive filters for estimating
the noise covariance. Innovation sequences represent the
differences between obtained measurements and the esti-
mated state, as expressed in the following formula (Mehra
1972):
gk ¼ zk � zkð�Þ ð12Þ
where zk �ð Þ ¼ Hkxkð�Þ. The innovation covariance is
Ck ¼ E gkgTk
� �¼ HkPk �ð ÞHT
K þ Rk
¼ Hk Tk�1Pk�1 þð ÞTTk�1 þQk�1
� �HT
k þ Rk
ð13Þ
where Ck is considered the predicted innovation covari-
ance. Innovation sequences are easily affected by actual
unaccounted errors, such as unmodeled dynamics,
unknown initial conditions, and increased measurement
covariance. The actual innovation covariance can be esti-
mated using (Mehra 1972)
�Ck ¼1
M � 1
Xk
i¼k�Mþ1
gigTi ð14Þ
where M is a window size, and the value is determined
empirically. A larger M value results in less noise after
averaging, and a smaller M provides a superior capability
of tracking dynamics. The actual innovation covariance
shows the effect of unaccounted errors. The basic concept
of the covariance matching technique is to achieve con-
sistency between the predicted innovation covariance and
the actual innovation covariance. For example, if the actual
innovation covariance is significantly larger than the pre-
dicted innovation covariance, the process noise covariance,
or the measurement noise covariance, should be increased.
This ensures that the predicted innovation covariance is
consistent with the actual innovation covariance. However,
using (13) and (14) does not provide a unique solution
when the true values of Q and R are unknown. For cases
where the Q is known but R is not known exactly, R is
estimated as
�Rk ¼ �Ck � �HkPk �ð ÞHTk ð15Þ
In this case, the measurement covariance is modified to
improve the consistency of the predicted and actual inno-
vation covariance. Subsequently, the second case is con-
sidered, in which the value of R is known but the value of
Q is not exactly known. Here, the Pkð�Þ is increased to
GPS Solut
123
compensate for the difference between the predicted and
actual innovation covariance.
Assisted adaptive extended Kalman filter (AAEKF)
During applications in surveying and geodesy, the GNSS
receivers are typically static or move smoothly. In such
situations, the process noise covariance Q can be applied
with a priori value, then Eq. (15) can be used to adap-
tively correct the measurement noise variance and
achieve the objective of the adaptive filter. Generally, the
adaptive covariance of measurement �Rk is calculated
using (12)–(15). Because the noise of phase measurement
is approximately 1–2 mm, it is less than the noise of the
predicted measurement zkð�Þ in (14). The use of (14)
produces a �Ck with substantial noise, and the obtained�Rk cannot effectively determine the measurement noise.
In this study, a new method for estimating �Rk is pro-
posed. The concept of this new method is to use the
reference station measurements to estimate the GNSS
measurement errors. These errors are then analyzed to
determine the noise of DD measurement �Rk. Thus, the
adaptive filtering technique can be employed to deter-
mine the solution.
GNSS errors, as discussed in (1)–(4), can be classified
into the following categories: (1) satellite errors, that is,
errors from the satellite ephemeris and clock; (2) atmo-
spheric effects, that is, errors from ionospheric and tro-
pospheric effects; and (3) receiver errors, that is, errors
from receivers and the reception environments, such as
the multipath effects, receiver clock errors, and thermal
noise. Satellite errors are strongly correlated with dis-
tance and vary slowly over time; atmospheric effects are
also correlated with distance and time under normal
weather conditions. Receiver errors are nearly completely
common to different satellites. The DD model is a
between-satellite and between-receiver process. The
between-satellite difference can eliminate receiver clock
errors. The between-receiver difference can eliminate
most satellite and atmospheric errors; the extent of this
elimination is reduced as the baseline length increases;
these errors are distance-correlated errors. Hence, using
the value of distance-correlated error and baseline length,
the remaining DD errors can be estimated. Many studies
have applied measurements taken at reference station to
estimate the GNSS measurement errors (Loomis et al.
1989; Jin 1996; Farrell and Givargis 2000), and these
estimation methods are widely employed in DGPS
applications. Following these studies, a KF was used to
estimate the errors of a reference station. The estimation
parameters included the tropospheric, ionospheric, and
satellite errors. With these errors, the remaining errors of
DD measurements can be estimated using the formula
below. The single-difference (SD) tropospheric errors
Tgpq, ionospheric errors Ig
pq, and satellite errors Sgpq are
defined as (Parkinson and Spilker 1996; Beutler et al.
1988):
Tgpq ffi Tg
p 0ð Þ 1
sin hgq
� 1
sin hgp
!ð16Þ
Igpq ffi Ig
p 0ð Þ 1
sin agq� 1
sin agp
� ð17Þ
Sgpq ffi
lpq � Sgp
Rgp
ð18Þ
where 1sin ag
q¼ 1
cos sin�1 0:94792 cos hgqð Þ½ � ffi 1þ 2
96�hgq
90
� �3
. The
symbol hgp denotes the elevation angle from reference sta-
tion p to satellite g, Tgp 0ð Þ and Ig
p 0ð Þ are the vertical tro-
pospheric and ionospheric errors, respectively, a is the
elevation angle at the ionospheric point (Klobuchar 1987),
R is the distance from reference station p to satellite g, I is
the baseline length from the reference station to the rover
station, Sgp is the satellite error from reference station p to
satellite g, the relationship between Sgp and Sg
pq in (18) is the
worst case situation. The overall estimated SD distance-
correlated error dE is
dEgpq ¼
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiS
g2
pq þ Tg2
pq þ Ig2
pq
qð19Þ
In addition to the distance-correlated errors, the uncorre-
lated errors should also be considered. Uncorrelated errors
include multipath effects and thermal noise. The multipath
effects are generally not estimated and allowed to impose
their signature on the residuals. Considering the remaining
uncorrelated error, receiver thermal noise, and overall
estimated distance-correlated error, the estimated SD
measurement error of satellite g is dEg2
pq þ rg2
pq
� �1=2
, where
rgpq ¼ rg2
pq þ rg2
pq
� �1=2
. Additionally, rgp and rg
q represent
the receiver thermal noise from satellite g to receivers
p and q; their values are a function of the carrier-to-noise
ratio. The relationship between the receiver thermal noise
and carrier-to-noise ratio is discussed in a later section.
The adaptive factor can be applied to KF to suppress
the contribution of measurements with comparatively
larger errors (Yang et al. 2001). The adaptive factor of
each measurement changes according to the standardized
estimated error. Thus, the influence of measurements
with larger errors can be reduced by increasing the
variance to avoid a significant rise in standard errors.
The adaptive factor can be obtained using
GPS Solut
123
ag ¼cg; cg\k
k � ec2g
k2�1; cg� k
(; g ¼ 1; 2; . . .;m ð20Þ
cg ¼
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffidE
g2
pq þ rg2
pq
q
rgpq
ð21Þ
where cg is the standardized estimated error and k is
determined according to the confidence level. The value for
k is set as 3.29 in this study (0.1 % confidence level).
Considering the adaptive factor, the SD covariance matrix
CSD ¼ diag ðr1pqÞ
2; ðr2pqÞ
2. . .; ðrmpqÞ
2� �
can be rewritten as
CSD ¼ diag ða1 � r1pqÞ
2; ða2 � r2pqÞ
2. . .; ðam � rmpqÞ
2� �
.
DD measurements are linear functions of the SD mea-
surements. By applying the law of variance–covariance
propagation and by taking the SD covariance matrix into
account, the DD covariance matrix is
CDD ¼ B � CSD � BT ð22Þ
where B is the coefficient matrix. If the first satellite is
selected as the reference satellite, then B is
B ¼
�1
�1
..
.
�1
1
0
..
.
0
0
1
..
.
0
� � �� � �. .
.
. . .
0
0
..
.
1
2664
3775 ð23Þ
After obtaining CDD, Eq. (11) can be rewritten as
Kk ¼ Pkð�ÞHTk ½HkPkð�ÞHT
k þ CDD;k��1 ð24Þ
The effects of increased measurement covariance can be
compensated by reducing the magnitude of the Kalman
gain.
Experimental results
The experiments consist of two parts. In the first part of the
experiment, the receiver thermal noise of code and phase
measurements were analyzed. The second part involved
analysis of the impact of L1 kinematic positioning when
adding the SBAS measurements. The GPS receiver used
for our experiments was a consumer-grade GPS receiver
with ublox-4t chipset, which costs approximately $100.
The cutoff angle is 15�, and the sampling rate of ranging
measurements is 1 Hz. Three SBAS satellites can be
observed in Taiwan, that is, PRN128, PRN129, and
PRN137.
Noise analysis of the code and phase measurements
This study used a 0.5 m baseline to analyze receiver
thermal noise on the code and phase measurements. The
experiment was performed on January 13, 2013. The 24-h
data were analyzed. To combine the two satellites and two
receivers for double differences, a carrier-to-noise ratio (C/
N0) was computed using the following equation (de Bakker
et al. 2008),
ðC=N0Þij ¼ �10:0 log1
210�
C=N0i10 þ 10�
C=N0j10
� � �ð25Þ
This equation follows the inverse relationship between the
variance and the carrier-to-noise ratio when expressed in
dB–Hz. A factor of one half is added to normalize the
C=N0 to the undifferenced levels. The residual of the short
baseline DD code and phase measurements is shown in
Figs. 1 and 2, in which the horizontal axis denotes the
carrier-to-noise ratio, which was calculated using (25). The
GPS and SBAS satellite signals were plotted separately.
The figures show that the GPS code and phase accuracy
were 0.11 m and 0.015 cycle when the C/N0 = 50 dB–Hz,
and the SBAS satellites code and phase accuracy were
0.46 m and 0.021 cycle when the C/N0 = 45 dB–Hz.
During the measurement campaign, the SNRs of GPS and
SBAS measurements are approximately 50 and 45 dB–Hz,
respectively. The standard error, or root mean squared error
(RMSE), of the regression for the GPS satellite code and
the phase measurement were 0.122 m and 0.015 cycles,
respectively. The RMSE for the SBAS satellite code and
the phase measurement were 0.282 m and 0.015 cycles,
respectively.
Fig. 1 Relationship between pseudorange errors and C/N0 for a short
baseline
Fig. 2 Relationship between phase errors and C/N0 for a short
baseline
GPS Solut
123
Half-integer ambiguity problem
The value of ambiguity is essentially an integer. But, there
may be half-integer cycle slips due to the navigation mes-
sages demodulated error (‘‘bit error’’) in low-cost L1 GPS
receivers. The navigation messages are modulated onto the
carrier using the binary-phase shift keying (BPSK) method.
When a navigation data bit transition occurs, the carrier is
instantaneously phase shifted by 180�. The ‘‘bit error’’
occurs when a bit transition is incorrectly detected yet the
tracking loop continues tracking. This will cause a 180�phase shift error (half-integer ambiguity problem). The half-
integer ambiguity problem is usually recovered in a few
seconds, when the ‘‘bit error’’ is found and repaired by parity
check and analyzing of the validity of the navigation mes-
sages. The validity items include the transmission time,
broadcast ephemeris, and almanac. The probability of error
Pe for a BPSK signal with Gaussian noise can be written as
(Tsui 2005):
Pe ¼ erfc
ffiffiffiffiffiffiffiffiffiffiffiffiffiC
N0
� t
r� ð26Þ
where the erfc is the complementary error function, t is the
length of predetection integration data, which is 20 ms for
GPS navigation message. For example, at C/N0 = 30 dB–
Hz, the probability of error for GPS navigation message is
about 2.59 9 10-10. The half-integer ambiguity problem
caused by ‘‘bit error’’ is temporary, and the probability of
error for the GPS navigation data bit is very low when the
SNR thread is set as C/N0 [ 35–40 dB–Hz. Hence, the
half-integer ambiguity problem of GPS measurements is
categorized as a temporary error which is excluded by
outlier detection method (Lin and Yu 2013). The signal
strength of SBAS satellite is weaker than GPS satellite and
the symbol rate of SBAS satellite is 10 times faster than the
rate of GPS satellite. According to (26), the probability of
error for the SBAS navigation data bit is much higher than
that for the GPS satellite. In the event of an error bit, the
ambiguity takes the form of a half-integer. For this reason,
the ambiguity of SBAS satellite is regarded as half-integer.
The success rate of ambiguity resolution with/without
(w/o) the SBAS satellite
In order to assess the feasibility of structural monitoring
using low-cost single-frequency receivers, four continu-
ously operating GPS (CGPS) stations located in New
Taipei City, Taiwan, were installed on flat roofs. In Tai-
wan, the ionospheric effect has large temporal and spatial
variations in the total electron content (TEC) and TEC
gradients. To eliminate most ionospheric effects in DD, a
conservative interstation distance is not greater than Ta
ble
1S
ucc
ess
rate
sin
per
cen
tas
afu
nct
ion
of
bas
elin
ele
ng
than
do
bse
rvat
ion
tim
e
Len
gth
of
tim
e
(min
)
50
cm2
.1k
m3
.3k
m4
.9k
m
GP
S-
on
ly
EK
F
GP
S?
SB
AS
EK
F
GP
S?
SB
AS
AA
EK
F
GP
S-
on
ly
EK
F
GP
S?
SB
AS
EK
F
GP
S?
SB
AS
AA
EK
F
GP
S-
on
ly
EK
F
GP
S?
SB
AS
EK
F
GP
S?
SB
AS
AA
EK
F
GP
S-
on
ly
EK
F
GP
S?
SB
AS
EK
F
GP
S?
SB
AS
AA
EK
F
23
0.5
69
2.3
69
6.5
32
3.6
57
.68
6.1
11
9.4
42
7.7
85
9.0
31
6.6
72
5.0
05
0.0
0
55
3.4
79
9.3
11
00
.04
0.2
72
.92
92
.36
29
.86
38
.19
72
.22
22
.22
28
.47
64
.58
10
70
.83
10
0.0
10
0.0
59
.77
9.8
69
5.8
34
8.6
14
1.6
77
5.0
03
8.1
93
5.4
26
7.3
6
GPS Solut
123
5.0 km. The baseline lengths of CGPS stations were 0.5 m,
2.1 km, 3.3 km, and 4.9 km. Data were collected from Day
5 of 2013. For this experiment, every 10 min over 24 h,
one datum segment was obtained. The number of tested
segments was 144. Additionally, 2, 5, and 10 min data
were used to resolve the ambiguity in each segment. The
LAMBDA method (Teunissen 1995) was used for ambi-
guity resolution. A chi-square test (confidence interval
1 - a = 0.99) and contrast statistic test (contrast value is
[4.0) were used to judge the reliability of the solution
(Counselman and Abbot 1989; Frei and Beutler 1990). The
true value of ambiguity was obtained using the known
baseline vector. A successful solution must pass the chi-
square and contrast tests, demonstrating an identical value
to the truth.
Table 1 lists the success rates of GPS-only, GPS/SBAS
EKF, and GPS/SBAS AAEKF solutions for the three time
durations. The results in the table indicate that the GPS/
SBAS AAEKF solutions provide the optimum success
rates for the four baseline lengths and three time durations.
For the 3.3 and 4.9 km experimental results, the success
rates of the GPS/SBAS EKF solutions were lower than
those of the GPS-only solutions with longer time durations.
However, the success rates of the GPS/SBAS AAEKF
solutions were consistently superior to those of the GPS-
only solutions, and the rate of improvement was approxi-
mately 3.2, 2.4, and 1.6 times for the 2, 5, and 10 min data
collection durations, respectively. The improvement shows
that adding the SBAS data was very helpful, especially the
shorter the length of data. The ionospheric effects and
SBAS satellite errors have large values and variations, and
it is difficult to model these errors during EKF initializa-
tion. However, these errors are corrected, and their effect
on the DD model is directly proportional to the length of
baseline, increasing as the baseline length. Regarding the
EKF, the imperfect error model significantly decreases the
success rate when the baseline length increases. Compared
with EKF, AAEKF yields a more gentle decrease in suc-
cess rate, proving that AAEKF more successfully models
these errors than does EKF. It is worth mentioning that
during the 24 h experiment, for three periods (each period
lasted approximately 10–30 min), the signals from only
four or five GPS satellites could be received. During these
periods, the ambiguity resolution was very difficult using
only GPS data. Additionally, if any satellite was obscured,
the ambiguity almost cannot be resolved using GPS satel-
lite data along. However, the use of three SBAS satellites
significantly benefited the ambiguity resolution.
The accuracy of kinematic positioning w/o the SBAS
satellite
The purpose of this experiment was to analyze the effect that
supplementing SBAS satellite data had on the kinematic
positioning accuracy. The applied data were the same as the
data mentioned in the previous section. These data were
categorized into only GPS data and both GPS and SBAS
data. The GPS/SBAS data were solved with AAEKF. The
first 900 s data were used for ambiguity resolution, and the
remaining 85,500 s data were used for kinematic position-
ing. The positioning accuracy of the four baselines is listed in
Table 2. The positioning results and the positional dilution of
precision (PDOP) are shown in Fig. 3. The results of the four
baselines were similar, and only the 2.1 km baseline is
Table 2 Accuracy of kinematic
positioning w/o the SBAS
satellite
Units are meter
50 cm baseline 2.1 km baseline 3.3 km baseline 4.9 km baseline
GPS-
only
RMSE
GPS ? SBAS
RMSE
GPS-
only
RMSE
GPS ? SBAS
RMSE
GPS-
only
RMSE
GPS ? SBAS
RMSE
GPS-
only
RMSE
GPS ? SBAS
RMSE
E 0.0036 0.0020 0.0064 0.0040 0.0077 0.0066 0.0103 0.0087
N 0.0054 0.0024 0.0083 0.0049 0.0126 0.0091 0.0165 0.0137
H 0.0161 0.0076 0.0274 0.0132 0.0318 0.0252 0.0367 0.0334
Fig. 3 Kinematic positioning of the 2.1 km baseline vector; units are
in meter; the corresponding mean and RMS error of GPS-only data
for the east, north, and height coordinate components were 0.10 and
0.64 cm, -0.09 and 0.83 cm, and -0.09 and 2.74 cm, respectively;
the corresponding mean and RMS error of GPS/SBAS data for the
east, north, and height coordinate components were 0.12 and 0.40 cm,
-0.12 and 0.49 cm, and 0.03 and 1.32 cm, respectively
GPS Solut
123
plotted. Table 2 shows that the positioning accuracy of L1
kinematic positioning achieves 1–2 cm, and the positioning
accuracy of the used SBAS data was shown to have increased
by about 50.9, 43.4, 20.9, and 13.8 % for 50 cm, 2.1 km,
3.3 km, and 4.7 km baselines, respectively. When the PDOP
increased, the addition of SBAS satellite data significantly
enhanced the positioning accuracy.
Conclusion
Using SBAS can enhance the satellite geometry. In situa-
tions with fewer satellites, adding the SBAS measurements
is extremely beneficial. However, the SBAS satellite signal
strength is weak and the satellite clock and orbit errors are
larger and less stable compared with those of a GPS
satellite. Ignoring these factors may worsen the solution.
This phenomenon was demonstrated by the ambiguity
resolution success rate experiment. The 3.3- and 4.9-km
experimental results demonstrated that the success rate of
GPS/SBAS EKF was less satisfactory than that of GPS
EKF at longer time durations. This study proposes a
strategy to solve this problem, which can be successfully
employed for single-frequency GPS/SBAS positioning. In
this study, the addition of SBAS satellite data was shown to
enhance the ambiguity resolution success rate and posi-
tioning accuracy. The GPS/SBAS AAEKF solutions were
superior to those of GPS-only EKF solutions, the ambi-
guity resolution success rate for 2, 5, and 10 min mea-
surements were improved by about 3.2, 2.4, and 1.6 times,
respectively, and the positioning accuracy of the north,
east, and height directions were improved by 20–80,
20–125, and 20–110 %.
The virtual reference station (VRS) (Vollath et al.
2000) and multiple reference station techniques do not
currently include SBAS satellite data. Therefore, the
method proposed in this study is only applicable for short
baselines (typically \5 km). Further investigation
regarding the addition of SBAS functions to the VRS and
multiple reference station techniques is required. In recent
years, various developments of satellite navigation sys-
tems have been carried out; for example, the GLONASS
in Russia, the Compass from China, and the Galileo from
European Union. If these systems are supplemented,
problems such as varying measurement quality between
different system data can occur. However, the method
proposed in this study can be employed to resolve this
problem.
Acknowledgments The authors are grateful to a research Grant
NSC 101-2221-E-019-070-MY3 from the National Science Coun-
cil. The generous provision of GPS data for this study by the Institute
of Earth Sciences Academia Sinica is greatly appreciated.
References
Beutler G, Bauersima I, Gurtner W, Rothacher M, Schildknecht T,
Geiger A (1988) Atmospheric refraction and other important
biases in GPS carrier phase observations. In: Brunner FK (ed)
Atmospheric effects on geodetic space measurements. Mono-
graph 12, School of surveying NSW, Sydney, pp 15–44
Boriskin A, Kozlov D, Zyryanov G (2007) L1 RTK system with
fixed ambiguity: what SBAS ranging brings. In: Proceedings of
ION GNSS-2007, Fort Worth, TX, September 25–28,
pp 2196–2201
Counselman CC III, Abbot RI (1989) Method of resolving radio
phase ambiguity in satellite orbit determination. J Geophys Res
94:7058–7064
de Bakker PF, van der Marel H, Tiberius CJM (2008) Geometry-free
undifferenced, single and double differenced analysis of single
frequency GPS EGNOS and GIOVE-A/B measurement. GPS
Solut 13(4):305–314. doi:10.1007/s10291-009-0123-6
Farrell J, Givargis T (2000) Differential GPS reference station
algorithm: design and analysis. IEEE Trans Control Syst
Technol 8(3):519–531. doi:10.1109/87.845882
Fotopoulos G, Cannon ME (2001) An overview of multi-reference
station methods for cm-level positioning. GPS Solut 4(3):1–10.
doi:10.1007/PL00012849
Frei E, Beutler G (1990) Rapid static positioning based on the fast
ambiguity resolution approach ‘‘FARA’’: theory and first results.
Manuscr Geod 15:325–356
Gelb A (1974) Applied optimal estimation. MIT Press, Cambridge,
MA. ISBN 0262570483
Hofmann-Wellenhof B, Lichtenegger H, Collins J (1994) Global
positioning system theory and practice, 3rd edn. Springer, Wien,
NY, pp 14, 124–127
Jin XX (1996) A new algorithm for generating carrier adjusted
differential GPS corrections. J Geod 70(11):673–680. doi:10.
1007/BF00867146
Klobuchar JA (1987) Ionospheric time-delay algorithm for single-
frequency GPS users. IEEE Trans Aerosp Electron Syst AES
23(3):325–331. doi:10.1109/TAES.1987.310829
Kozlov D, Tkachenko M (1998) Centimeter level real-time kinematic
positioning with GPS ? GLONASS C/A receivers. Navigation
45(2):137–147
Leick A (2004) GPS satellite surveying, 3rd ed. Wiley, Hoboken,
pp 172–177. ISBN: 978-0-471-05930-1
Lin SG, Yu FG (2013) Cycle slips detection algorithm for low cost
single-frequency GPS RTK positioning. Surv Rev
45(330):206–214. doi:10.1179/1752270612Y.0000000034
Loomis P, Kremer G, Reynolds J (1989) Correction algorithms
for differential GPS reference stations. Navigation
36(2):179–193
Mehra RK (1972) Approaches to adaptive filtering. IEEE Trans Autom
Control AC 17(5):693–698. doi:10.1109/TAC.1972.1100100
Odijk D, van der Marel H, Song I (2000) Precise GPS positioning by
applying ionospheric corrections from an active control network.
GPS Solut 3(3):49–57. doi:10.1007/PL00012804
Parkinson BW, Spilker JJ Jr (1996) Global positioning system: theory
and applications, vol 2. Wiley, London, pp 12–23
Skaloud J (1998) Reducing the GPS ambiguity search space by
including inertial data. In: Proceedings of ION GPS-1998,
Nashville, TN, September 15–18, pp 2073–2080
Takasu T, Yasuda A (2008) Evaluation of RTK-GPS performance
with low-cost single-frequency GPS receivers. In: International
symposium on GPS/GNSS 2008, November 11–14, Tokyo
International Exchange Center, Japan 2008. RTKLIB: An Open
Source Program Package for RTK-GPS. http://gpspp.sakura.ne.
jp/rtklib/rtklib.htm
GPS Solut
123
Teunissen PJG (1995) The least-squares ambiguity estimation
decorrelation adjustment: a method for fast GPS integer
ambiguity estimation. J Geod 70(1–2):65–82. doi:10.1007/
BF00863419
Tiberius CCJM, de Jonge PJ (1995) Fast positioning using the
LAMBDA-method. In: Proceedings of DSNS-95, Bergen, NOR,
April 22–28, paper no. 30
Tsui JB (2005) Fundamentals of global positioning system receivers:
a software approach. Wiley, Hoboken, NJ, pp 226–229
Vollath U, Buecherl A, Landau H, Pagels C, Wagner B (2000) Multi-
base RTK positioning using virtual reference stations. In: Proc.
ION GPS 2000, Salt Lake City, UT, September 19–22,
pp 123–131
Wirola L, Alanen K, Kappi J, Syrjarinne J (2006) Bringing RTK to
cellular terminals using a low-cost single-frequency AGPS
receiver and inertial sensors, IEEE/ION PLANS 2006 confer-
ence, San Diego, CA, pp 24–27
Yang Y, He H, Xu G (2001) Adaptive robust filtering for kinematic
geodetic positioning. J Geod 75:109–116
Zhang J, Lachapelle G (2001) Precise estimation of residual
tropospheric delays using a regional GPS network for real-time
kinematic applications. J Geod 75(5–6):255–266. doi:10.1007/
s001900100171
Shiou-Gwo Lin received the
B.S. degree in Physics from the
Soochow University, the M.S.
degree in Space Science, and
the Ph.D. degree in Atmo-
spheric Physics from the
National Central University. He
is currently an assistant profes-
sor of Department of Commu-
nications, Navigation and
Control Engineering at National
Taiwan Ocean University. His
research interests include GNSS
software receiver, GNSS rela-
tive positioning, and signal
denoising.
GPS Solut
123