STOCHASTIC MODELLING OF GPS PHASE OBSERVATIONS

Neil Brown
PhD Candidate
Department of Geomatics
The University of Melbourne 3010
Australia
neb@sunrise.sli.unimelb.edu.au

Ian Williamson
Professor and Head of Department
Department of Geomatics
The University of Melbourne 3010
Australia
ianpw@unimelb.edu.au

Allison Kealy
Lecturer
Department of Geomatics
The University of Melbourne 3010
Australia
akealy@unimelb.edu.au

ABSTRACT

The precision requirements of data collected for use within spatial data infrastructure (SDI) depend on the intended
application of the data. Regardless of the coarseness of the data, it is still important to know the data quality.
Knowledge of data quality is essential to assess data’s fitness for purpose. Mis-specification of this quality
parameter may result in inappropriate use, reduced utility and market value of the data.

At present the measures of quality for GPS derived coordinates given by commercial software packages tend to be
unrealistic because unmodelled errors remain unaccounted for. This paper reviews current theory and processes used
to calculate precise positions using GPS. It outlines the ability of these methods to handle errors present in the GPS
observables. To illustrate how current processes may be improved some empirical results are presented and analysed
for static baselines of medium length (~100km).
 INTRODUCTION

The Global Positioning System (GPS) has established itself as the primary means for spatial data capture and is
increasingly being used in a wide range of applications from personal navigation to geodesy. In recent years there
has been a lot of discussion regarding GPS data quality. Much of the research in this area has focused on the
stochastic models used in GPS processing and their importance to ambiguity resolution (eg Han, 1997, Teunissen,
1997, Wang, 1999). However, GPS data quality has wider importance than simply to the geodetic and surveying
professions due to its increasing use and developments in spatial data infrastructure (SDI). The precision
requirements of data collected for use within SDI depend on the intended application of the data. Regardless of how
coarse the data are there is a requirement to have a measure of the data quality. Knowledge of data quality is
essential to assessing data’s fitness for a particular purpose. It is recognised that for spatial data to be used correctly
and wisely metadata that describes data quality, in addition to other attributes, should accompany the spatial data
(Nebert, 2000, GI2000, 1996). Depending on the particular data and its possible uses this may mean needing
metadata down to the individual feature, object or record level (Nebert, 2000). Data quality for positional
information is generally given in terms of a standard deviation resulting from the data capture process. Mis-
specification of this quality parameter may result in inappropriate use, reduced utility and market value of the data.
At present, the measures of GPS quality given by commercial software packages are either over-optimistic, or
conversely, are overly conservative and therefore have low fidelity (Keenan and Cross, 2001, Barnes et al., 1998,
Wang, 1999). This lack of fidelity can potentially impact SDI development starting with the geodetic infrastructure
and at every jurisdictional level.

This paper examines current processing algorithms and models to determine their ability to manage errors in GPS
and to highlight the need for further research in stochastic modelling. Therefore, the focus of the paper is on the case
of GPS processing with static baselines in the order of 100km in length using the best available software and
modelling techniques. First the theory is explained and then some empirical results are presented to demonstrate
how this theory applies in practice.

GPS OBSERVABLES

The GPS satellites transmit signals on two frequencies, 1575.42 MHz and 1227.60 MHz, designated L1 and L2.
Modulated onto L1 are two pseudo-random noise (PRN) codes, C1 and P1 in RINEX (Gurtner, 2000) notation. A
second P-code (P2) is modulated onto the L2 frequency. Under anti-spoofing conditions the P-codes are encrypted by
the W-code to create the Y-code. The GPS receiver must use a codeless technique to reconstruct the P-code.

By measuring the shift between the code transmitted by a satellite and an exact copy generated internally, a GPS
receiver is able to determine the distance between the satellite and the receiver. This distance is called a
pseuodorange due to errors introduced by the satellite and receiver clocks and other sources. It is possible to
measure the carrier beat phase used to transmit the codes more precisely than the codes themselves, however the so-
called integer ambiguities must be resolved for centimetre and better positioning (Hoffman-Wellenhof et al., 1994).
Both phase and code measurements can be written in the form of pseudoranges.

A phase pseuodorange (in metres) can then be expressed as (vad der Marel, 1998)

λΦ sr = ~
xs − ~
x r + cδt s − cδt r + Isr + Trs + M sr + D r + λN sr + esr

where
s
δt , δt r are the vectors of satellite and receiver coordinates at time of transmission (t)
s are the satellite and receiver clock offsets from GPS time
x ,xr
s is the apparent error in the range caused by the ionosphere
Ir
s is the apparent error in the range caused by the troposphere
Tr
s is the multipath error
Mr
 Dr is the antenna phase centre variation
s is the number of full cycles of the carrier wave between the satellite and receiver at the time lock was acquired
Nr
s is the measurement noise
er

The influence of each of these error sources must be removed from the observations before a precise position for the
antenna can be calculated. The techniques that are used to handle these errors fall into three categories: elimination,
modelling and estimation.

ELIMINATION OF ERRORS

Some errors can be corrected using a mathematical (or functional) model that relates the raw measurements to the
parameters (errors and coordinates). The two techniques commonly used in GPS processing algorithms are
differencing and linear combinations.

Linear Combinations
Receivers that are capable of measuring the observations on both L1 an L2 frequencies (so-called dual-frequency
receivers) allow for a number of combinations of the phase and/or code measurements to be constructed. A variety
of linear combinations can be used to reduce the influence of the ionosphere, geometry, satellite and receiver clocks
and the troposphere. The two most widely used linear combinations (using Bernese notation) are the L3 ionosphere-
free and the L5 wide-lane linear combinations.

The L3 combination is calculated as (Hugentobler et al., 2001)

1
L3 = (f12L1 − f 22L 2 )
f12 − f 22

and is very useful for ambiguity resolution because it eliminates almost all of the ionospheric delay. However, the
noise on L3 is approximately three times greater than that of L1. Also the ambiguities lose their integer nature when
L3 is formed meaning that L3 cannot be directly used to resolve L1 and L2 ambiguities. Instead L3 is used in
conjunction with another combination such as L5.

Calculated as (Hugentobler et al., 2001)

1
L5 = (f1L1 − f 2L 2 )
f1 − f 2

the wide-lane linear combination is named because of its long (~86cm) wavelength. Whilst the noise is greatly
magnified the influence of the ionosphere is relatively small and the ambiguities retain their integer nature. If the
wide-lane ambiguities are resolved the determination of the L1 (and hence L2) ambiguities is greatly simplified.

Differencing
Whilst the previously mentioned errors are unique to a particular receiver-satellite combination, errors between
another receiver and the same satellite will be similar if the receivers are close to one another (see Figure 1). The
same is true for one receiver and two satellites, or between one receiver and one satellite at different times. These
spatial and temporal correlations in the observations are a function of baseline length and arise from errors caused
by the satellite and receiver clock, troposphere, ionosphere and orbits. The spatial correlation between the
measurements is the basis for the double-differencing procedure.

In order to reduce the influence of many sources of error in GPS, a technique of relative positioning is employed.
The technique most commonly used in GPS software is the double difference. Double differencing has the effect of
removing spatially correlated errors in the GPS measurements. The effectiveness of differencing is closely related to
the baseline length. Over longer baselines differencing is less able to remove error and ambiguity resolution is more
difficult.
 Source: (Barnes and Cross, 1998)

The double difference for observations taken simultaneously at two stations (i and j) from two satellites (m and n) is
given by:

∇∆ϕ ijmn = (ϕ im − ϕ mj ) − (ϕ in − ϕ nj )

where ∇∆ is the double difference operator and ϕ is the code or phase observation. Therefore each double difference
requires four one-way measurements. Typically the between receiver single differences are formed first as

∆ϕijm = ϕ im − ϕ mj , i ≠ j, m=1, 2, …, s

Two between receiver single differences are then used to create a double difference. The double differencing process
can be performed for single baseline using a matrix multiplication ∆∇ϕ = D.ϕ where ϕ is a vector of code or
phase measurements and D is the differencing matrix.

The vector of observations, l, can then be formed as

 ∆∇1 .ϕ   a 1 .ϕ 
 ∆∇ .ϕ  a .ϕ 
l = ∆∇ ϕ =  2  =  2 
 !   ! 
   
∆∇ n .ϕ  a n .ϕ 

for n independent baselines.

The observations in l are ordered firstly by baseline and secondly by epoch if the mathematical correlations are to be
handled rigorously (Howind et al., 1999). Single and double differences may be calculated before or after the
formation of a linear combination.
 ERROR MODELLING

Examples of models used in GPS processing are the Hopfield and Saastamoinen atmospheric models that are based
on a standard atmosphere (fixed temperature, pressure and humidity) and relate the delay in the signal to the height
of the observer. Obviously, such models do not necessarily reflect reality and therefore care must be taken in their
use. Such models are most often used over short baselines or when using code-based point positioning.

ERROR ESTIMATION

Errors may be carried explicitly as additional parameters in the functional model. Estimation is normally only used
when the elimination or accurate modelling is not possible or the error estimates are required for later adjustments or
other applications. For example, tropospheric estimates may be estimated and used in subsequent adjustments or
used in atmospheric studies. Other types of errors, such as receiver and satellite antenna offsets and code differential
biases, may be estimated but so weaken the solution that they are not estimated with coordinates.

ERROR HANDLING IN GPS

Satellite and Receiver Clocks

As the GPS system is based on time, it is essential to know the time of transmission and time of reception of the
signals precisely and in the same time system. Errors in the satellite clocks cancel through the double differencing
process. Whilst receiver clock errors are cancelled by between-satellite differencing, residual error remains because
the receiver clock offset must be estimated in order to calculate the time (GPS) of signal reception. This time is used
to calculate the geometric distance between the receiver at reception time and the satellite at time of transmission. So
long as the receiver clock is estimated to within 1µs the error in the geometric distance will be less than 1mm
(Hugentobler et al., 2001). This is simple to achieve as code measurements need only to be measured to better than
300m.

Orbit Errors
The paths taken by satellites as they orbit the Earth are commonly described using Keplerian orbits. Deviations, or
perturbations, from the Keplerian orbits are caused by the non-sphericity of the earth, gravitational attraction of
other planetary bodies, Earth tide potential, ocean tide potential, solar radiation pressure, Earth albedo, atmospheric
drag and residual acceleration of the satellites (Govind, 1994). The GPS broadcast ephemeris models these
perturbations accurately enough to allow precise positioning for baselines of several tens of kilometres in length
over which residual orbit errors from the differencing process will be small. For longer baselines, precise
ephemerides are available from a number of organisations (IGS, JPL, NGS, NIMA, CODE) in predicted or post-
processed forms (Roulston et al., 2000). The influence of errors in the satellite orbits is relative to the baseline length
δb δr
and can be approximated by = where δb is the error in the baseline of length b caused by an error in the orbit
b r
of δr by a satellite at a topocentric distance of r (Leick, 1990). Using the Final Orbits product of the IGS, which are
accurate to better than 5cm (IGS, 2001), the error in a baseline of several hundreds of kilometres will be less than a
millimetre.

Ionosphere
The upper, ionised part of the atmosphere, referred to collectively as the ionosphere, introduces phase advance and
group delay of the GPS signals and is the dominant source of error in GPS point positioning (Hoffmann-Wellenhof
et al., 1994). For two nearby receivers the satellite signals travel through approximately the same part of the
atmosphere and most of the error is removed by differencing. As the distance between the receivers increases, the
assumption that the atmosphere affects the signals equally becomes less valid and differential processing is less
effective. Due to the dispersive nature of the ionosphere, observations on two frequencies (L1 and L2) can be used to
remove the effect to the first order at the expense of increase measurement noise. This is achieved by formation of
the ionospheric-free linear combination, or L3, as described above.
The ionospheric-free combination may be formed before or after the double differences. Thereby, over long
baselines most of the error introduced by the ionosphere can be removed. However, residual errors and the amplified
noise from the modelling process remain. Therefore, for short baselines it is better to work with the L1 and L2 phase
observations directly. A fiducial network of reference stations may be used to model the ionosphere (Odjik, 2000).
Generally an ionospheric model is of most benefit when using single frequency receivers and for real-time
positioning.

Troposphere
The lower, non-ionised or neutral, part of the atmosphere, known as the troposphere, is non-dispersive at the
frequencies of GPS and its effect on the signals must be removed through modelling, differential processing and/or
estimation. The main difficulty in modelling the troposphere, following Hopfield (1969) who shows that the
tropospheric delay can be broken into a dry and wet component, is measuring the atmospherics, especially water
vapour, along the path of the signal. However, the dry component, which constitutes about 90% of the delay, is
relatively easily modelled (Hoffmann-Wellenhof et al., 1994). Chang and Tseng (1999) show that standard
atmospheric models, when used correctly, are able to effectively deal with most of the tropospheric error, though
residual errors remain. For high precision applications the tropospheric delays are typically carried as extra arc
parameters in the functional model.

Multipath
When signals from the GPS satellites are reflected by a surface and arrive at the antenna via an indirect path
multipath error results. Multipath is a function of the reflective properties of the receiver’s environment and the
satellite constellation and is independent of the baseline length. As such multipath is not cancelled by between
receiver or between satellite differencing and is difficult to model. At present, GPS receivers/antennas are not fully
able to filter out reflected code or carrier phase signals. For short baselines, where (after differencing) multipath is a
dominant source of error, most benefit is gained by observing over long periods (several days) with static receivers
over which time the repeating nature of multipath can be used to create a multipath model (eg Radovanovic, 2000).
Signal to noise ratios in conjunction with antenna gain patterns have also been shown to be effective, though
difficult to apply (eg Barnes et al., 1998). In precise applications, multipath is normally mitigated though averaging
of observations taken over a long period of time (generally 24 hours).

Phase Centre Variations

The electrical phase centre of many GPS antennas does not coincide with the physical centre of the antenna. The
vector between the physical antenna reference point (ARP) and the phase centre is known as the phase centre offset.
If ignored an error of up to a decimetre (primarily in the height component) may be introduced. Additionally, the
position of the phase centre varies with the elevation and azimuth of the signal. The phase centre offset is then really
an average offset to which corrections should be applied. Corrections and offsets are determined empirically by a
number of organisations including the NGS (Mader, 1999). The variations are described using either an
elevation/azimuth grid or spherical harmonics coefficients.

Measurement noise
Modern receivers are able to measure the carrier phase to better than 0.01 cycles, which corresponds to millimetre
precision (Hoffman-Wellenhof et al., 1994). However, in addition to this measurement noise, other small errors and
correlations occur during the measurement process that are generally treated as noise.

When anti-spoofing is enabled, receivers must apply reconstruction techniques to enable measurement of the P-code
and carrier phase on L2. This can result in a correlation between the observables on the two frequencies (Bona,
2000). Some receivers also filter the measurements in an effort to reduce noise introducing a time correlation (Bona,
2000). Receivers have also been seen to have inter-channel biases and L1/L2 differential delays (Gao et al., 1999, Lin
and Rizos, 1996).
Residual Errors
If the most appropriate techniques are employed with a long high quality data set the residual errors from the above
sources can be quite small, generally at the millimetre level. As such they have very little impact on the values of the
parameters being estimated, rather they cause unrealistic estimates of precision for the parameters. These residual
errors are very difficult to model within the functional model and therefore must be handled by the stochastic model
even though some are not strictly random in nature (Keenan and Cross, 2001). Blewitt (1998) shows that there is
model equivalence between the functional and stochastic models. Therefore the influence of an error can either be
addressed by introducing an extra parameter into the functional model or by augmenting the stochastic model. Both
approaches will lead to reduced data strength and an increased covariance matrix (decreased precision) for the
estimated parameters (Blewitt, 1998).

The relatively slow variation in the atmosphere compared to the GPS observation intervals introduces a time
correlation in the errors. To prevent undue weakening of the solution by introducing too many unknowns,
atmospheric parameters are generally assumed constant over a period of time. For example, a single estimate (called
an arc parameter) of the tropospheric delay will often be used for a period of one or more hours. The existence of
spatial correlation is the basis for the differencing process and is directly related to the baseline length. Other
correlations may occur from unmodelled errors (eg L1/L2 differential delay). It is this correlation in the
measurements that is currently not fully modelled and is the focus of research activity around the world.

STOCHASTIC MODELLING

What is a stochastic model?

Typically GPS data are processed using a least squares or Kalman filter based algorithm in which the relationship
between the measurements and the parameters is described by the functional (or mathematical) model and
represented by the design (A) matrix. The statistical properties of the measurements, in the context of the functional
model, are described by the stochastic model and are generally given in the form of the variance-covariance matrix
(Cl). The diagonal terms of the covariance matrix are the variances and the off-diagonal terms are the covariances
(or correlations) between the measurements. The inverse of the covariance matrix is the weight matrix. The
functional models used in GPS positioning are well defined and documented and are generally accepted (Bock et al.,
1986, Hoffman-Wellenhof et al., 1994, Strang and Borre, 1997), though Euler (2000) has shown that slight
modifications can help to better model the ionosphere.

Why are better stochastic models needed?

The least squares and Kalman filter algorithms rely on correct weighting of the measurements to provide unbiased
estimates of the parameters. Residual and unmodelled systematic errors in the functional model and temporal and
spatial correlation between the GPS observables mean that determination of the covariance matrix (and hence the
weight matrix) is difficult (Barnes et al, 1998, Wang, 1999, Keenan and Cross, 2001). It is common practice
(Hoffman-Wellenhof, 1994, Strang and Borre, 1997) to simply assume that the covariance matrix can be represented
by a block identity matrix appropriately modified to account for the mathematical correlation between the double
differenced observations used for relative positioning. I.e. Cll = DCl D T where Cl and Cll are respectively the
variance-covariance matrices of the zero and double difference observations.

Some software use a slightly more sophisticated process that calculates individual weights for each satellite based on
its elevation, signal to noise ratio (SNR) or the baseline length, but which also ignore correlations between the
measurements (Barnes et al., 1998, Bock et al., 1986). Even high-end software use these stochastic models, though
they may also have the option to estimate a stochastic model by post-fitting the residuals from the processing (MIT
and Scripps, 2000).

Whilst satellite elevation and SNR can be used as quality indicators, they do not always reflect reality because of
signal diffraction and other influences (Satirapod and Wang, 2000, Brunner et al., 1999, Gerdan, 1995). Barnes et al.
(1998) demonstrates the inadequacies of these approaches and the increased precision and fidelity obtained by using
the correct stochastic model which they calculate using a reverse engineering approach. In general, an incorrect
stochastic model will result in unreliable results, over-optimistic measures of precision and a reduced ability to
detect errors (Nicolai, 1988). Improving the stochastic model also has benefits in faster and more reliable integer
ambiguity determination (Euler, 2000, Han, 1997).

EMPIRICAL ANALYSIS

To demonstrate how the above theory relates to the real-world a data set collected over a 100km baseline between
two base stations in the Victorian reference network (GPSnet), Australia is examined. Both stations have Trimble
4000SSI receivers, one with a Choke Ring and the other a Permanent L1/L2 antenna.

The data has been processed using the Bernese software developed by the University of Bern (Hugentobler et al.,
2001). The Bernese software has a number of characteristics that separate it from standard commercial processing
packages: complete user control over the error models, linear combinations and strategies that are used,
sophisticated orbit models and data screening. Standard commercial software was also used, however whilst it was
able to resolve ambiguities the coordinates were off at the metre level from the known values. In both instances the
IGS Final Orbit Product is used. The exact algorithms and models used by the commercial software and very limited
and poorly documented compared with Bernese. In the processing with Bernese an elevation-dependant (1/coz z)
weighting model was used. The wide-lane ambiguities were resolved first after which the L1 ambiguities were
determined using the ionospheric free linear combination. IGS phase centre models were used and the troposphere
was estimated using the Dry Niell mapping function.

Using Bernese the formal error estimates (RMS) of the estimated coordinates are 1.2mm in height and 0.2mm in
latitude and longitude. The repeatability of daily solutions is in the order of 4-5mm in height and 2mm in latitude
and longitude. However, by looking at the double difference residuals in Figure 1 we can see that relatively large
errors remain in the double differences. Over the 24 hours of the data set these errors mostly cancel, but over shorter
periods more of the error remains. The residuals for two days (234 and 236) are shown. Similar patterns are seen for
other satellite pairs. In theory the multipath error should be the same for both days but offset by eight minutes.
Figure 1 shows a definite trend between the two days but it is obvious that other errors besides multipath remain in
the double difference observations. Filtering the multipath component leaves a non-random series with a larger RMS
than would be expected for L3 noise. If this residual error could be better modelled the error estimates would be
more realistic and less data would be required to attain the same level of precision.

Double Difference Residuals SV2-4

80

60

40
Residual (mm)

20 234
0 236
1

113
129
145
161
177
193
209
225
241
17
33
49
65
81
97

-20

-40

-60
Epoch (30 second)

Figure 1: Double difference residuals from two days with the same baseline and satellites.
The authors plan to analyse large amounts of data from the Victorian GPSnet reference network in order to try and
establish trends in the residuals and improve understanding of receiver behaviour. If the stochastic model is correct
and systematic errors (such as multipath) have been removed the residuals may be expected to have a constant (zero)
mean and variance. If not a modification is required to the stochastic model, perhaps requiring the addition of one or
more variables. Due to modelling restrictions the one-way observations are unsuitable for analysis at the millimetre
level, while the double difference observations have a more complicated structure. It is expected that the time
correlations mentioned earlier will be the most problematic. The aim is to separate out each error source, starting
with multipath, from the residuals. At this stage preparations are still being made to streamline the processing and
analysis of the very large data set.

CONCLUSION
Examination of current techniques used to handle the numerous error sources in GPS has shown that all error
sources can be effectively given the right circumstances (equipment, baseline length, time etc). The particular
techniques that are employed depend on the situation, however even under the most favourable conditions and using
the most sophisticated software error estimates are optimistic and residuals far from random. There is little room for
improvement of the functional model so a great deal of research is being directed towards the complementary
stochastic (variance-covariance) model.

REFERENCES
Barnes, J.B., Ackroyd, N., and Cross, P.A., (1998), Stochastic Modelling for Very High Precision Real-Time
Kinematic GPS in an Engineering Environment, Proceedings of FIG XXI International Conference, 21-25 July,
Brighton, UK, Commission 6, 61-76.

Barnes, J.B., and Cross, P.A., (1998), Processing Models for Very High Accuracy GPS Positioning, Journal of
Navigation, 51(2), 180-193.

Bock, Y., Gourevitch, S.A., Counselman III, C.C., King, R.W. and Abbot, R.I., (1986), Interferometric Analysis of
GPS Phase Observations, Manuscripta Geodaetica, 11, 282-288.

Bona, P., (2000), Precision, Cross Correlation, and Time Correlation of GPS Phase and Code Observations, GPS
Solutions, 4, 2, 3-13.

Blewitt, G., (1998) GPS Data Processing Methodology: From Theory to Applications in Teunissen, P.J.G. and
Kleusberg, A. (eds), GPS for Geodesy, Springer-Verlag, Berlin, pp 231-270.

Brunner, F.K., Hartinger, H., and Troyer, L., (1999), GPS Signal Diffraction Modelling: The Stochastic SIGMA-∆
Model, Journal of Geodesy, 73, 259-267.

Chang, C and Tseng, C, (1999), Testing on Tropospheric Modelling for GPS Tracking Stations in Taiwan,
Geomatics Research Australasia, No 70, 77-94.

Euler, H-J., (2000), Advances in Ambiguity Resolution for Surveying Type Applications, Proceedings of ION GPS-
2000, Salt Lake City, Utah, September 19-22, 95-103.

Gao, Y., Lahaye, F., Héroux, P, Liao, X., Beck, N., and Olynik, M., (1999), Investigation of Measurement
Inconsistency in Dual-Frequency GPS Receivers, Proceedings of the 12th International Technical Meeting of the
Satellite Division of the Institute of Navigation, September 14-17, 1999, Nashville, Tennessee, 251-258.

Gerdan, G.P., (1995), A Comparison of Four Methods of Weighting Double Difference Pseudorange Measurements,
Trans Tasman Surveyor, 1, 1, 60-66.
GI2000, (1996), Towards a European Policy Framework for Geographic
Information, A working document of GI2000 homepage, Accessed May 2001,
<http://158.169.50.95:10080/gi/en/gi2000/gi2000dd.html>

Govind, R.J., (1994), Absolute Sea Level Monitoring in Australia: The Geodetic Fixing of Tide Guage Benchmarks
Using the Global Positioning System (GPS), PhD Thesis, University of Colorado.

Gurtner, W., (2000), RINEX: The Receiver Independent Exchange Format Version 2, Astronomical Institute,
University of Bern.

Han, S., (1997), Quality-control Issues Relating to Instantaneous Ambiguity Resolution for Real-time GPS
Kinematic Positioning, Journal of Geodesy, 71, 351-361.

Howind, J., Kutterer, H., and Heck, B., (1999), Impact of temporal correlations on GPS-derived relative point
positions, Journal of Geodesy, 73, 246-258.

Hoffmann-Wellenhof, B., Lichtenegger, H., and Collins, J., (1994), GPS Theory and Practice, Springer-Verlag,
New York.

Hopfield, H.S. (1969), Two-quadratic tropospheric refractivity for correcting satellite data, Journal of Geophysical
Research, 74(18), 4487-4499.

Hugentobler, U., Schaer, S., and Fridez, P., (eds), (2001), Bernese GPS Software Version 4.2, Astronomical
Institute, University of Berne.

IGS, (17 October, 2001), IGS Products, http://igscb.jpl.nasa.gov/components/prods.html

Keenan, C.R. and Cross, P.A., (2001), Stochastic Modelling: The Crucial Element in High-Fidelity DGPS Quality
Assessment?, Proceedings of Hydro'2001, Conference of The Hydrographic Society, University of East Anglia in
Norwich, 26-30 March 2001, 12 pp.

Leick, A., (1990), GPS Satellite Surveying, John Wiley & Sons, New York.

Lin, L, and Rizos, C., (1996), An Algorithm to Estimate GPS Satellite and Receiver L1/L2 Differential Delays and
its Application to Regional Ionosphere Modelling, Geomatics Research Australasia, 65, 1-26.

Mader, G.L., (1999), GPS Antenna Calibration at the National Geodetic Survey, GPS Solutions, 3, 50-58.

MIT and Scripps, (2000), Documentation for the GAMIT GPS Analysis Software, Release 10.0 December 2000.

Nebert, D.D., (ed), (2000), Developing Spatial Data Infrastructures: The SDI Cookbook, Global Spatial Data
Infrastructure Technical Working Group, Draft 1.0 June 2000.

Nicolai, R., (1988), Statistics in Positioning Quality Control, Proceedings of Hydro ’88, 15-17 November 1988,
Amsterdam, The Netherlands, pp 97.

Odijk, D., (2000), Weighting Ionospheric Corrections to Improve Fast GPS Positioning Over Medium Distances,
Proceedings of ION GPS-2000, Salt Lake City, Utah, September 19-22, 1113-1123.

Radovanovic, R.S., (2000), High Accuracy Deformation Monitoring via Multipath Mitigaiton by Day-to-Day
Correlation Analysis, Proceedings of ION GPS-2000, Salt Lake City, Utah, September 19-22, 35-44.

Roulston, A., Talbot, N., and Zhang, K., (2000), An Evaluation of Various GPS Satellite Ephemerides, Proceedings
of ION GPS-2000, Salt Lake City, Utah, September 19-22, 45-54.
Satirapod, C. and Wang, J., (2000), Comparing the Quality Indicators of GPS Carrier Phase, Geomatics Research
Australia, 73, December 2000, 75-92.

Strang, G., and Borre, K., (1997), Linear Algebra, Geodesy and GPS, Wellesley-Cambridge Press, Wellesley.

Teunnisen, P.J.G., (1997), A Canonical Theory for Short GPS Baselines Part IV: Precision and Reliability, Journal
of Geodesy, 71, 513-525.

van der Marel, H., (1998), Virtual GPS Reference Stations in the Netherlands, Proceedings of ION GPS 98,
Nashville, 49-58.

Wang, J., (1999), Modelling and Quality Control for Precise GPS and GLONASS Satellite Positioning, PhD Thesis,
School of Spatial Sciences, Curtin University of Technology.

ACKNOWLEDGEMENTS

The authors would like to gratefully acknowledge the support of Land Victoria and the Centre for Spatial Data
Infrastructures and Land Administration, The University of Melbourne in preparing this paper.

BIOGRAPHICAL NOTES

Neil Brown
Neil graduated from the University of Melbourne in 1999 with a Bachelor of Arts and a Bachelor of Geomatics
(Honours). He is currently undertaking a Doctor of Philosophy degree in the Department of Geomatics, University
of Melbourne on stochastic modelling of the GPS observables.

Allison Kealy
Allison Kealy is currently a lecturer in the Department of Geomatics at the University of Melbourne, specialising in
the research areas of GPS, GLONASS and integrated systems. Allison received her PhD in Geodesy from the
University of Newcastle upon Tyne, UK in 1996, after which she spent 2 years in industry providing technical
support for GPS/GLONASS manufacturers Ashtech Ltd.

Professor Ian Williamson

PhD (NSW), DrHC (Olsztyn), FTSE, FISAust, FIEAust, HonMFIG, HonFMSIAust, LS, CPEng

Ian Williamson is Professor of Surveying and Land Information, Head, Department of Geomatics and Director,
Centre for Spatial Data Infrastructures and Land Administration at the University of Melbourne. His teaching and
research is concerned with designing, building and managing land administration, cadastral, and land and
geographic information systems in both developed and developing countries. He is a Licensed Land Surveyor and
Chartered Professional Engineer.

He has undertaken research and consultancies worldwide including for Australian governments, AusAID, many
individual country governments, the United Nations and the World Bank. He was Chairperson of Commission 7
(Cadastre and Land Management) of the International Federation of Surveyors 1994-98, and is currently Director,
FIG/UN Liaison 1998-2002. He is an Honorary Member of the FIG. At the University of Melbourne he has been
President of the Academic Board and Pro-Vice-Chancellor. He is currently Chairperson of the Victorian
Government’s Geospatial Information Reference Group and Chairperson of Working Group 3 (Cadastre) of the
United Nations sponsored Permanent Committee for GIS Infrastructure for Asia and the Pacific (2001-2004).