Professional Documents
Culture Documents
January 3, 2002
Content
1 Introduction ................................................................................................. 3
2 Physical model of the radio occultation signal ........................................ 4
2.1 Characteristics of the Doppler signal: Scintillations and multipath ......... 5
3 Error analysis in the reconstruction of the refractive index ................... 9
4 Optimal estimation of the occultation parameters................................. 12
4.1 Preprocessing of narrow-banded signals with additive white noise. .... 13
4.2 Models for estimation of the impact parameter .................................... 14
4.2.1 Signal processing with a phase detector. ..................................... 15
4.2.2 Maximum likelihood estimation of the Doppler frequencies .......... 16
4.2.3 Global measurement of the occultation parameters ..................... 19
4.3 Comparison of results and numerical computation .............................. 24
4.4 Signal processing conclusion............................................................... 26
5 Appendices................................................................................................ 28
5.1 Appendix I: Band pass filtered noise.................................................... 28
5.2 Appendix II: Intrinsic phase and frequency noise................................. 29
5.3 Appendix III: Maximum likelihood parameter estimation ...................... 30
5.3.1 The Fisher information .................................................................. 33
5.4 Appendix IV: Fourier transformation of the radio occultation signal ..... 34
5.4.1 Verification of assumptions ........................................................... 36
6 References................................................................................................. 37
1 Introduction
This report is deliverable WP6300(1) and covers the work outlined in the work-
packages WP6300 and WP6500. The main issues here are analyses of the sig-
nal processing possibilities for radio occultation signals, development of signal
processing algorithms, characterisation of the signal type, and error analyses.
These tasks are all connected and constitutes a complex survey of signal proc-
essing of radio occultation. Essential aspects of signal processing are covered in
the report including the traditional phase locked loop, time sliding Fourier analy-
ses and a new method based on a Fourier transform of the total radio occultation
signal.
The basics for a development of a signal processing method are the physical
model of the signal to be processed. The physical model of the radio occultation
signal is discussed in Sec. 2 together with considerations about what validity
simulations have in the real world. An essential problem here, is that recordings
of real signal hardly exist, which mean that the models based on physical theory
can not, or have not, been rigorously compared with the real world. Another as-
pect of these problems is addressed in Sec. 2.1. Here, an attempt to characterise
the structure(s) of the Doppler signal is made, when the atmosphere is in various
states. The problems, which emerge, are that a lot of non-ideal conditions pro-
duce Doppler signals with the same structure as the signals from an ideal atmos-
pheric, and only the ideal situation can give a correct reconstruction of the refrac-
tive index profile. This means that an accuracy of an occultation measurement
has an additional uncertainty, which only can be given as a statistical value. This
value has to be determined from a statistical knowledge of atmospheric condi-
tions.
An analytic error analyse of the inversion of the bending angle to the refractive
index is performed in Sec. 3. This development is a useful tool for analysing how
an error in the Doppler frequency or impact parameter propagates in the Abel
transform. It is demonstrated how gradients in the refractive index can amplify
small uncertainties in the impact parameter. This can also be used to a general
error analyse of occultation measurements if a statistical database of refractive
index features is available.
In the main part of this report (Sec. 4) optimal estimation of the occultation pa-
rameter is considered. The phase locked loop and the time sliding Fourier
method are analysed and compared. Both methods have their advantages, but
the phase locked loop fails when the signal exhibits multipath. Another, and more
advanced method is described. The basic idea in this method is to use the path
traversed by the LEO satellite as a synthetic antenna and the fact that the Dop-
The general philosophy used in this document is an attempt to make the occulta-
tion measurement conceptual and operative simple despite the complex nature of
the measurement method. A full operative simulator, which can perform a 3D
wave propagation in an arbitrary medium, detect the Doppler frequency and con-
struct the bending angle, is a serious challenge. The time for a single run on a
computer will be long and if a full Monte Carlo analysis has to be done, the com-
puting time would be extremely time consuming. Instead a description based on
a careful error analyses and general understanding of the physic of the problem
is preferable.
The diffracting correction models cannot describe the full-diffracted field, but in-
clude a physical approximation of the effect of the refractive index (phase
screens). Their validity are verified though simulation and compared with the re-
sults from the GO method.
Philosophically the value of simulations can be difficult to evaluate. Using the fol-
lowing model for a simulation;
Input ⇒ Physical ⇒ Signal ⇒ Output
model processing
⇑ ⇓
⇐ Reconstruction ⇐
It can be argued that, If the reconstruction of the output resemblance the input
does it mean that the physical model is correct, or does it mean that the signal
processing is optimal for the physical model? The last must be the correct inter-
pretation. The only way to verify a physical model is to compare real signal with
simulated signal from the physical model. Since only incomplete real data for oc-
cultation measurements exists, it can be difficult to choose the most accurate
physical model. In the conclusion of this report it is stated that an incorporation of
a scientific mode in the LEO satellite, which can produces real signal with a rela-
tive large bandwidth, would be most wanted in order to resolve this question and
other signal processing issues.
The physical model chosen here for the estimation of the signal parameters is the
geometrical optical model. It describes all relevant phenomena such as
singlepath, multipath and critical refraction. The geometrical set-up used (Sec. 6)
in this report, is throughout a configuration where the GPS, the LEO satellites
and the centre of the Earth are in the same plane during all the occultation
measurement. Further it is assumed that the satellites are moving in circles. For
an operative use of the results derived in this report corrections to real satellite
geometry must be made. Numerical examples make use of the values stated in
Table 1 and 2 in Sec. 4.3.
The condition for a reliable reconstruction of the refractive index and thereby
temperature and pressure, are besides measurement errors, that the atmosphere
is stationary and the refractive index is spherical symmetric. These assumptions
are probably never fulfilled exactly in a real atmosphere. The questions are then,
how close we are to the ideal atmosphere, what impact the derivations have on
The refractive index from a spherical symmetric atmosphere can cause multipath
rays. The multipath does not prevent a perfect inversion obtained from the Abel
transform, unless critical refraction occurs (see Sec. 6). Multipath does, however
occur within a non-spherical symmetrical atmosphere and also when the atmos-
phere is non-stationary. Signal caused by scintillation in the atmosphere can
look very much alike multipath, and must, if possible be removed from the signal
before the refractive index can be recovered. The effects of strong scintillations
can only be described using a full wave solution of the Maxwell equations,
whereas weak scintillations can be described with the thin phase screen ap-
proximation. The latter approximation will in the following be used to give a
physical description of the appearance of the occultation signal.
The concept of a thin phase screen comes from geometrical optics and can be
used to describe diffractive effects from local perturbation in a refractive medium.
A thin phase screen is a phase variation in a plane passed by the electromag-
netic field. The phase is given by the integrated perturbation of the refractive in-
dex projected down of the screen. With this, the ordinary ray pattern can be cal-
culated from the unperturbed medium after which the effect of the phase screen
can be taken into account. The structure of the unperturbed ray pattern is then
unchanged whereas the amplitude and the phase of the single rays are affected
by the perturbation in the refractive index. If the phase screen is a grating the
resulting complex field is a sum of modes which match the direction of the unper-
turbed ray, which means that the received intensity of a single ray will exhibit os-
cillation and looks like the intensity variations of multipath. This multiplicative ef-
fect from the refractive index perturbation on single ray field is interpreted as
scintillations.
In this report we will focus on the possibilities, if any, of predicting the conditions
of the non-ideal atmosphere from an analyses of the received signals. If this pre-
diction can be done, a figure of merit can be associated with actual measure-
ments. This is valuable when data are to be assimilated into weather forecast
models.
1) The effect of a prism: The ray pattern seen by the LEO satellite is bend so
they seem to appear from a virtual shifted GPS transmitter in a direction per-
pendicular to line of sight between the satellites. This means that the impact
parameter is different at the GPS and the LEO satellite. The difference be-
tween the impact parameters is to first order constant and will act as an un-
known bias (see Sec. 4).
2) The lens effect: The ray pattern will seems to come from the GPS satellite
shifted in the direction of the line of sight. Again, it must be expected that the
impact parameters are different at the receiver and transmitter.
3) The effect of a grating: At the phase screen the directions of the incoming
rays are split into an un-shifted zero order ray, two symmetrical shifted first
order rays, and possibly also higher order modes. The grating effect is very
different form the prism and lens effect; the same Doppler frequency will ap-
pear several times, and could be denoted as frequency multipaths.
Since 1) and 2) cannot be distinguished from the ideal case (Case 1), it is impor-
tant to analyse the statistical nature of the vertical perturbation and to get an es-
timate of the bias of the impact parameter. If the vertical perturbations are pre-
sent at several places, along the optical path, these effects can be enhanced or
averaged out depending on the structure of the perturbations.
Non-stationary conditions
A non-stationary atmosphere means that the refractive index is changing during
an occultation. Introducing the sampling rate ∆f s , the measuring time T and the
time scale of the atmospheric changes τ , their distinctive cases can be consid-
ered:
a) τ ≤ 1 The changes are faster than the sampling time.
∆f
In a) the condition depends on the nature of the time changes. Most likely the
time changes are fast fluctuations in the atmosphere with short temporal and
spatial coherence, and the received intensity will exhibit scintillations. If a sliding
Fourier transform in a short time window is performed, it should be possible to
extract the unperturbed Doppler frequency as function of time, due to the as-
sumption of fast fluctuations of the refractive index. If the atmosphere really is
changing (no fluctuations), which is unlikely, the occultation measurements can-
not be used to reconstruct the refractive index.
Some of the cases described above can be detected by analysing the received
signal, while others are not identifiable. This means, that even if an actual occul-
tation measurement looks correct, the reconstructed refractive index could have
a serious bias. This lead to a figure-of-merit assignment to actual measurements
based on feature analysis of a real refractive index.
The noise and uncertainties for the detection of the impact parameter a , is de-
noted ê which is supposed to be a sum of various contribution, thermal noise in
the receiver, processing uncertainties, numerical errors and so on. This yields:
eˆ = ∑ ∆aˆ q (1)
The single contributions are depending on the specific processing methods ap-
plied through the whole chain. Each term, Ɖq can correlated with other terms
and be biased, or what is equivalent, having a non-zero mean value. A bias is in
general unwanted and should be avoided if possible, i.e. numerical computation
should be done with random samples and processing methods with an intrinsic
bias should be considered with care.
The measurement of the impact parameter at a certain time, gives the bending
angle as function of the impact parameter. The bending angle is then Abel trans-
formed giving the logarithm to the refractive index as expressed below.
∞
1 α ( x)
ln(n) = Aˆ (α ) = ∫ dx (2)
π a (x 2 − a 2
In α (a) each value of a has an uncertainty of eˆ(a) . If the true value of a is de-
noted a~ the bending angle and its related error eˆ(a ~) , can be expressed as:
~) dα
~) + eˆ(a (3)
α (a) = α (a ~
da
This will generate the expression for the measured and the true value of the loga-
rithm to the refractive index:
~ 1
∞
1 dα (4)
ln(n) = ln(n ) + ∫ eˆ( x )dx
π a ( x 2 − a 2 dx
where n~ denotes the true value of the refractive index.
From the integral above, the mean value and the variance or the statistics of the
refractive index can be found. The integrals cannot in general be expressed ex-
plicit for operative use. However, an asymptotic expansion in 1/ a can be ob-
It is seen that the bias of the impact parameter gives a bias on refractive index
and that the bias depend on the gradient of the refractive index. The condition in
Eq.6 for small gradient is fulfilled for a simple exponential decaying refractive in-
dex typical for the atmosphere, which means that only in special cases the area
of large gradients comes into effect.
In cases where the gradients are small (in the sense defined above) small bias
terms in the impact parameter will not affect the measurement of the refractive
index, whereas in cases with large gradients the induced bias on the refractive
index can be large, especially in the upper part of the atmosphere. At the surface
of the Earth a ≈ 6200 km and ln(n) ≈ 2.5 ⋅ 10 −4 which means that the product
a ln(n) ≈ 1.5 km . An allowable bias in the impact parameter should therefore be
smaller than 10m (if the relative bias as a goal should be smaller than 1%). The
10m is equivalent to a relative bias on the impact parameter of 10 −4 . At higher
altitudes the refractive index falls of exponentially and the demand on the bias
becomes critical.
The impact of the first term for the variance is similar to the analysis performed
above for the bias term. The second term in Eq. 7 involves the second order de-
rivative of the correlation function of the error function. The correlation function
can depend explicitly on the impact parameter, but in general, it will depend ex-
plicitly on time and implicit on the impact parameter. For the part of the error
function which depend explicit on time we can write the second order derivative
as:
d 2 c(a1 − a 2 ) d 2c da
= ( ) −2 (8)
da1da2 a =a =a
dt 1dt 2 t =t =t dt
1 2 1 2
It is seen that large gradient in the bending angle causes the correlation length to
go to zero and consequently induce a large errors in the reconstructed refractive
index. If the gradient of the bending angle is small it can be estimated that the
correlation length is in the range of 500m − 100m in the case of usual GPS/LEO
parameter. This is a low number and demands an accuracy of meters for the im-
pact parameter if the reconstruction of the refractive index should be accurate.
The present analysis is concerned about the point accuracy for the reconstruction
of the refractive index. If the relative variance at a point is large it will be difficult
to get a reliable result in the specific point, but by averaging over several points it
is possible to obtain a good estimate but with the cost of larger spatial resolution.
This calls for a reconstruction where the spatial resolution is varying.
e iqt
Band Low
pass Mixer pass
Figure 1.
General scheme for pre-processing a narrow-banded signal with additive white noise.
1. The received signal U (t ) is band pass filtered in order to reduce the addi-
tive noise as much as possible, i.e. the band pass should be as narrow as
possible without disturbing the useful information in the received signal.
2. The band pass filtered signal is feed into a complex mixer where it is mul-
tiplied with a phasor oscillating with frequency q , close to the centre fre-
quency of the received signal.
3. The complex output from the mixer is low passed filtered so that only fre-
quencies in order of the bandwidth of the signal survive. The low pass of
the mixed noise signal further reduces the residual noise from the band
pass.
Now, this produce the raw signal, that further can be processed with the purpose
of finding the phase, amplitude and the frequency spectrum or performing a pa-
rametric match. The band pass filtered noise is described in Appendix I.
Different methods for extracting the impact parameter can be applied on the radio
occultation signal. The measurement of the impact parameter and the knowledge
of the GPS and the LEO satellite position is the basis for the reconstruction of the
refractive index.
Measurement errors are of various origins. Clock errors that affect the accuracy
of the satellite positions, signal detection noise and errors or uncertainties from
the processing methods are among the most important. Detection noise and
processing uncertainties will here be considered in the selected processing
method. Both sources of errors give an uncertainty in the impact parameter and
the added effect of errors gives errors in the reconstruction of the refractive in-
dex, as described in Sec. 4.
The raw signal is processed by a phase locked loop (PLL) (this is usually in-
cluded in the mixer). The phase and time derivatives of the phase can then be
measured, averaged and recorded with a certain sampling frequency f s . If the
bandwidth of the signal is ∆f , it can be averaged in a time 1 / 2∆f without signifi-
cant loss of relevant information and with a reduction of noise before the phase is
differentiated.
The conditions for the PLL to function properly are that the amplitude of the sig-
nal and the signal to noise ratio is of a certain magnitude and that the bandwidth
of the PLL is much lager than the bandwidth of the signal. The latter can easily
be fulfilled if the bandwidth of the signal is known, but if not, bias problems and
unattended averaging of the signal will occur and erroneous results will be pro-
duced.
In the case where the signal amplitude is small (and the S/N ratio likewise is re-
duced) the PLL will have difficulties to catch the correct phase within a time
smaller than 1 / ∆f and again bias errors and erroneous result will be produced.
In general the PLL technique is a very robust signal processing method for sig-
nals with small bandwidth, but it demand a high degree of pre-knowledge of the
signal. The high degree of pre-knowledge demanded for the PLL to work prop-
erly, is a serious disadvantage, when the signals, coming from certain areas in
the lower part of the atmosphere, are fading and changing drastically. Strategies
for deciding a ‘go or no-go’ signal processing can be difficult to state for the PLL.
Furthermore, detailed information of the signal, such as actual bandwidth and
spectral form cannot be obtained with a PLL.
In the case of multipath the receiver gets a superposition of signals with different
phases and the PLL will detect a phase, which is a mixture of the incoming
phases. This will give an error in the estimation of the impact parameter. The
best way to avoid this is to reconstruct the complex signal and use methods that
can cope with multipath. The complex signal can be reconstructed if the PLL has
the proper working conditions according to the remarks above.
In Appendix III the MLE is applied on the present problem with multipath radio
occultation signals embedded in additive white noise. It is shown that the MLE
estimation is bias free and that the optimal signal processing method is equiva-
lent to a spectral analysis. The estimated values of the Doppler frequencies (or
rather the excess frequencies) are found by detecting the position of the peaks in
the spectrum.
The radio occultation multipath signal is described in Eq. 1 in Sec. 2. The main
results from the MLE analysis described in Appendix III are the computation of
the Fisher information, which produces the uncertainties for the unknown signal
parameters, the amplitude Q l , the phase φ l and the Doppler frequency (ω D ) l for
the l th optical path of a ray. The uncertainties yield:
Now, the MLE estimation has been done under certain constrains which shall be
evaluated in details. The instantaneous Doppler frequency is defined as the time
derivative of the phase, and the MLE spectrum analysis is using data in a time
window T which means that the Doppler frequency or the impact parameter is
determined with a time uncertainty of T . The parameter set (a(t ), θ(t )) , where
θ(t ) is the angle between the GPS and the LEO satellite, gives the bending angle
α as function of the impact parameter a , from which the refractive index can be
reconstructed by the Abel transform. If the spectral analysis is done with a sliding
time window the bending angle can be sampled with the actual sampling rate or
lower.
If the Doppler frequency changes during a measurement, i.e. in the time win-
dowT , the MLE spectrum analysis cannot be applied. The constraint on the time
window for a proper use of the MLE method is derived in Appendix III and yields
DD l . If the frequency resolution in the spectrum effectively
T < τ l with τ l = 2π φ
should reach the MLE uncertainty ∆ω D , the constraint on the time window has to
be sharpened. In the time widow the Doppler frequency can increase an amount
of ∆φD ≅ T 2 φ
DD , which should be smaller than or equal to the noise frequency un-
certainty ∆ω D . This is equivalent to the constrain: π T ∆ω D < τ l2 which can be
converted to a constrain for the time window:
τ 2 2 24E 1
T ≤ T0 = ( l ) 5 ( 2 0 ) 5 ,
π Ql
with the corresponding Doppler uncertainty: (2)
τ l2 − 35 24E 0 25
∆ω D = ( ) ( 2 ) .
π Ql
In the optimal time window a reasonable number of samples have to be used for
the Fourier transform if the theoretical limit of the noise induced frequency uncer-
tainty has to be reached. The number of samples is given by the time-bandwidth
product, which yields:
τ l2 52 24E 0 51
T0 ∆f = ∆f ( ) ( 2 ) (3)
π Ql
For τ l ≈ .3 sec (see Table 1 in Sec. 3.3), ∆f = 50Hz , and a signal to noise ratio of
30 dB, the time-bandwidth product is in the range of 1, first with a increase in the
value of the bandwidth to ∆f = 1000Hz will give a two-digit number for the time-
bandwidth product. Lager value of τ l , which occurs in multipath areas, will also
improve the time-bandwidth product.
With the conventional conception of the Fourier transform all time information
disappear. However, the Fourier transform can also be used as a mapping
method where, for instance, the time is mapped into a frequency space. This is a
well-known technique in optical processing where it is used as coordinate trans-
formations. With some constrains this can also be done in the time domain.
In the Fourier transform of the occultation signal the frequency axis represent an
ordering of the impact parameter in increasing order. The information about the
refractive index is present in some form in the amplitude and in the phase υ , of
the Fourier transformed signal. If the information produced by the Fourier trans-
form is usable, the reconstruction of the refractive index can be done. This will
be shown in details in Appendix IV: But shortly, the result is that the frequency
ω , and the frequency derivative of the phase υ , are proportional to the impact
parameter and the time t , respectively, i.e.
dυ
(ω, ) ∝ (a(t ), t ) (1)
dω
This very simple result is obtained by Fourier transforming the nearly raw data
(the carrier frequency is supposed to be removed) from the antenna. The sam-
pling rate has in principle to be done with the Doppler frequency, but it can be
reduced by mixing the signal with an appropriate frequency that later can be used
to reconstruct the signal before the Fourier transform is done.
Figure 1a
Figure 1b
Figure 1a and 1b shows the impact parameter as function of time. The straight red line is the
calculated impact parameter for the neutral atmosphere (the refractive index is constant) as func-
tion of time. The black dotted curve shows the impact parameter versus time calculated from the
phase of the Fourier transform of the occultation signal from the neutral atmosphere. The spread
of the black dots reflects the wrapping noise in the phase. It has not been attempted to optimise
the unwrapping and differentiation of the phase.
In Appendix IV, the relation stated in Eq. 1 will be derived. The analysis is done
on one single path that do not exclude multipath, because of the ability of the
Fourier transform to separate frequencies.
The numerical approximations, which lead to Eq. 1, can partly be verified by ap-
plying the technique on an atmosphere with a constant refractive index. This is
shown in Figure 1.
The numerical results shown in Fig. 1a and 1b are obtained by Fourier transform-
ing the occultation signal from the atmosphere with a constant refractive index.
The signal yields:
1 ik ( L (t )−a(0 )Ωt )
U (t ) = e (2)
L(t )
where L(t ) is the distance between the GPS and LEO satellite, a(0) the impact
parameter at the time t = 0 , and k is the wave number. The signal U (t ) is Fou-
rier transformed and the operation described in Eq. 1 is performed. The result is
compared with the exact result a(t ) = rG r L sin(θ(t )) / L(t ) ( where θ is the angle
between the GPS and the LEO satellite, and rG , r L the GPS and LEO radii re-
spectively). It is seen that the results shown in Fig. 1a and 1b verify the mathe-
matical approximations used in order to derive Eq. 1. In the real atmosphere the
amplitude and phase have a more complicated structure than shown in Eq. 2, but
due to the arguments stated in Appendix IV this will not change the validity of
Eq.1.
A more realistic verification has been done. The data used for this simulation is a
wave propagation in the atmosphere based on the solution of the Helmholtz
wave equation. This approach applies a full wave forward scatter model that is
capable of predicting propagation for an arbitrary atmospheric refractivity. The
primary limitations of this technique are that it neglects the backscattered field,
and that accurate calculations are restricted to near-horizontal propagation direc-
tions.
Figure 2
The amplitude of the occultation signal as function of time.
In Fig. 2 the amplitude of the occultation signal is shown. It is seen that signs of
strong multipath (significant variations in the amplitude due interference) are pre-
sent around 23, 33 and 36 sec. Other extremes in the amplitude can be associ-
ated with variations in the defocusing factor.
Figure 3
The amplitude of the Fourier transform of the occultation signal
The results from a FFT of the restored signal are shown in Fig. 3. Within the
bandwidth of the signal the FFT spectrum is flat as predicted in Appendix IV, Eq.
IV.14.
The phase of the FFT is differentiated with respect to the frequency giving the
time (see Eq. 4). This is shown in Fig. 4 in blue colour. For comparison, the Dop-
pler frequency (the time derivative of the measured phase) is mapped versus
time in a red colour. It is seen that the two curves are overlapping by near areas
where multipaths are occurring.
Figure 4
The Doppler frequency versus time. The red curve shows the time differentiated phase of the
occultation signal. Around 23 seconds, it is seen that the reconstruction of the Doppler frequency
fails due to multipath. The blue curve show the frequency differentiated Fourier phase as function
of time. Besides, the multipath area the two curves have a nearly perfect match.
In the following we will compare results for the PLL and the MLE processing
method. But first, the basic numerical assumptions for the occultation technique
geometry will be stated. The neutral atmosphere and realistic path parameter for
the GPS and LEO satellite will be used for this purpose. This is done in Table 1
and 2.
Distance GPS-LEO
L(t ) = r L2 + rG2 − 2r L rG cos(θ )
2.6 2.7 2.8 2.8 2.9 3.0 3.0 3.1 3.1
10 −4 Km
Impact parameter
r r
a0 = L G sin(θ )
L
7213 7041 6814 6538 6216 5853 5455 5023 4563
Km
Doppler frequency
f D0 = kLD(t ) = kΩa0 (t )
kHz 36 35.2 34.0 32.6 31.0 29.2 27.2 25.1 22.8
Table 2.
To compare the performance of the phase locked loop and the spectral analysis
signal processing method, bias and uncertainties for the impact parameter are
computed as function of the signal-to-noise ratio obtained from Eq. 1 Sec.
4.2.1.1. and Eq. 2 Sec. 4.2.2.1. The results are shown in Table 3. The signal to
noise ratio is here defined in the context of the phase locked loop, i.e.
2
S N =Q where Q is the signal amplitude and 2E 0 ∆f the total noise from
2E 0 ∆f
the band pass.
The comparison is not quite fair with respect to the PLL method since the results
from the MLE have been chosen from optimal values. The results have been cal-
culated for a sampling rate of ∆f = 50Hz . If the sampling rate is increased, the
PLL results become worse whereas the MLE results are unchanged. In contrast,
a decrease in the sampling rate improves the PLL result.
The uncertainties for the impact parameter in Table 3 do not include clock errors.
The bias on the PLL is computed with the assumption that the measured excess
Doppler frequency is 10Hz . In the lower part of the atmosphere this is reason-
able, but at higher altitudes the excess frequency is the range of 1 − 2Hz . For
comparison, the positional accuracy of the GPS and LEO satellite are within one
meter.
The bias on the Doppler frequency in the PLL processing scheme is in some re-
spect serious. In Sec. 3, the error analysis, it is stated that near the surface of the
Earth the bias should be smaller than 10m , in case of large gradients in the re-
fractive index (when the relative bias on the refractive index is supposed to be
smaller than 1%). At higher altitudes the conditions become worse, but is im-
proved by smaller values of the excess frequency, as mentioned above.
The global measurement technique suggested in Sec. 4.2.3 is not included in this
comparison between the described processing methods, but it can be expected
that the uncertainties will be in the order of meter.
Another method described in the literature is the ‘MUSIC’ [Schmidt, 1986] proc-
essing method, which recently has been reported to give good results when used
on CHAMP data [Wickert et al. 2001]. The method is based on Karhunen-Loève
feature extraction theory, and are capable of detecting multipath. Its performance
is claimed to be better than the MLE method, but only marginally. Computation-
ally ‘MUSIC’ demands calculation of a covariance matrix for the signal and of the
eigenvalues of that matrix. The expected numbers of single paths have to be in-
cluded in the method. So, compared with the MLE method it is more computa-
tional intensive and not so flexible, but delivers marginal better results.
The aims in the signal processing part of this report have been to analyse the
phase locked loop (PLL) and the maximum likelihood estimation (MLE) method.
The results are viewed in the light of the error analysis presented in Sec. 4. The
error analysis is very important because it gives the requirements for the signal
processing method.
During this investigation is has been suggested to use a global method utilizing
the whole synthetic aperture. This is very promising, but the technique is pres-
ently not verified in details, so it will not be a part of the comparison of the differ-
ent methods.
In Sec. 4.3 numerical computation is done for the bias and uncertainties of the
processing method. Based on this and the analysis done throughout this report, a
descriptive comparison of the PLL, the MLE and the Karhunen-Loève feature
extraction method, ‘MUSIC’ have been shown in Table 1.
Table 1
The preferable signal processing method according to Table 1, the MLE method,
mostly because of its low demand of pre-knowledge. The MLE method is easy to
implement onboard the satellite, though the optimal performance can be difficult
to reach in a in-flight situation. A high sampling rate is here preferable and the
data can be reduced and packed so that only a limited bandwidth of a telemetric
connection to a ground station is necessary.
5 Appendices
In order to describe the impact of the noise we will first define and describe the
properties of the band pass filtered noise, which is the result of the pre-
processing. The noise is passing in the band pass ± 1 2 ∆ω around the centre
frequencies q and − q . The noise contribution around q is denoted by N + and
around − q , N − .
1
q + 12 ∆ω 1 ∆ω
2 (I.1)
2π q − 1∫2 ∆ω ∫
+ iωt iqt iωt
N (t ) = n (ω )e dt = e n (ω + q )e dω
− 12 ∆ω
− q + 12 ∆ω 1 ∆ω
1 2
The total noise N(t ) = N + + N − can be expressed as a sum of cosine and sine
functions, as shown below:
N(t ) = N + + N − = NC (t ) cos(qt ) − N S (t ) sin(qt ) (I.2)
where
1 ∆ω
1 2
The pair (NC , NS ) can be considered as a pair of noise signals in quadrature. The
statistics for K = NC + iNS is Gaussian with the covariance matrix T given by:
Ti , j =< K(t i )K ∗ (t j ) >= 2E0 ∆f sinc( 21 ∆ω(t i − t j ) (I.3)
where ∆f = ∆ω 2π , and E 0 is the noise energy per frequency. E 0 is constant in
the whole noise spectrum due to the assumption that only white noise is present.
The noise power from the band pass filtering, E 0 ∆f , is proportional to the band-
width ∆f , which it should be from a physical point of view.
All the calculations assume that the center frequency, q , is much larger than the
bandwidth. For practical purposes the covariance matrix is diagonal: The sam-
pling rate should always be smaller or equal to the bandwidth of the band pass.
The impact of additive noise at the detection of the phase of a narrow banded
signal is elaborated in the following. The detection of the phase is supposed to be
done in an ideal way, i.e. the PLL is working properly under all conditions. As a
result, the errors are only caused by the additive noise. If the signal to noise ratio
is low the description will not be adequate for the real detection process.
The signal Q(t ) cos(φ) is supposed to survive the band pass filtering without any
disturbances or changes.
The time derivative of the phase is the frequency. From Eq. II.3 the frequency
error can be found to yield:
dψ − iφD K (t ) + KD (t ) + QD (t )e iφ (t ) (II.4)
= Im( )
dt Q(t )e iφ(t ) + K (t )
Averaging the derivative of the error phase gives the mean frequency error:
dψ 2E ∆f (II.5)
< >≅ − 2 0 φD
dt Q + 2E0 ∆f
This means that the frequency estimation has a bias. The variance of the fre-
quency error can be estimated from Eq. II.4. By keeping terms to second order in
the signal amplitude it yields:
dψ 1 2Q 2 E0 ∆f (II.6)
< ∆( )2 >≈ (φ 2 + ( ∆ω) 2 ) 2 2
dt 6 (Q + 2E0 ∆f )
where ∆ω is the angular frequency noise bandwidth. It is seen that the noise
bandwidth must be reduced to a minimum in the pre-processing stage.
The mean value and the variance dependence on the frequency shows that the
initial guess on the centre frequency are important. The best guess will be φD = 0 .
The maximum likelihood method finds the most likely value of a parameter θ in
the logarithm to the conditional distribution function f ( xˆ θ ) , i.e. the distribution
function is maximized with respect to the parameter. It is supposed that the dis-
tribution function for the band pass filtered radio occultation signal is a product of
independent Gaussian distribution functions. However, this is not fully true. The
band pass noise is correlated, but only weakly in neighbour points. However
keeping the assumption the maximum likelihood function yields:
1 1 (III.1)
∑
2
M= − U (t i )− < U (t i ) >
4πE0 ∆f 4E0 ∆f
where U (t ) is the measured multipath signal and 2E0 ∆f the noise power in the
band pass (see Eq. I.3). The times t i are all in the measuring time interval
[t − T / 2, t + T / 2] .
For a general case of multipath rays, the complex band pass filtered signal can
be generalized form Eq. AII.1 to yield:
where the index p is a dummy number denoting a single path. The amplitudes
Q(t ) p are all supposed to be slowly varying quantities (compared to the variation
of the phase factors), q is the frequency, and φ p the phase for the single path.
The noise term K (t ) is defined in Appendix I.
The parameters that are to be estimated are the amplitude, the phase, the fre-
quency and eventually the derivative of the frequency for each single path signal
and the numbers of single paths appearing in the measurement. The parameters
are all time dependent quantities and we want to estimate their values at the time
t in the measuring time interval [t − T / 2, t + T / 2] . In order to do this the mean
signal is expanded in time around t . This yields:
< U (t + t ′) >= ∑ Qp (t + t ′)e p
iφ ( t + t ′ )
≅ (III.3)
i ( φ p (t )+t ′φD p (t )+ 21 t ′2φ
DD ( t ))
∑Q ≅ ∑ Qp (t )e
i ( φp (t )+t ′φD p ( t ))
p (t )e
where − T / 2 ≤ t ′ ≤ T / 2 .
In Eq.III.3 is has been assumed that the amplitudes are very slowly varying in
DD p , of the quadratic term compared
− T / 2 ≤ t ′ ≤ T / 2 . The time scale, τ p = 2π φ
with the measuring time T determines whether the derivative of the frequency
can be neglected or not. Here it will be assumed that T < τ p for all paths that
leads to the final approximation of the expected signal, as shown in Eq. III.3. This
last assumption is equivalent to suppose that the Doppler frequency is constant
during the measuring time. The parameters to be estimated are now the phases
φ p (t ) , the Doppler frequencies φ p and the amplitudes Q p . The estimation is
done by differentiating Eq. III.1 with respect to the wanted parameters and by
setting these equations equal to zero. Using Eq. III.3 the equations yields:
Re(e − iφl (t ) ∑ U (t + t i )e − iφl (t )ti ) = NQl (III.4)
Eq. III.4 and III.5 give the estimated values of the amplitude and phase for a sin-
gle pass l , with the condition that the Doppler frequency belongs to the same
single pass is known. The amplitude and phase becomes:
1 (III.7
∑
Ql = U (t + t i )e − iφl (t )t i
N
φ l = Phase(∑ U (t + t i )e −iφl (t )t i )
The result is, not surprisingly, that the amplitude of the l th single pass is equal to
the (temporal) mean value of the signal filtered with the frequency of the single
pass.
The second equation, which determines the Doppler frequency, is found by using
Eq. III.5 and III.6 together. This gives:
∑ (t i − t j )U(t + t i )U ∗ (t + t j )e l i j = 0
− iφD ( t )( t −t )
or
1 d 2
( ∑ U (t + t i )e −iφl (t )ti ) = 0
D
i dφD l (III.6)
The Doppler frequency is then found by Fourier transforming the band pass sig-
nal, and searching for the maximum frequencies in the absolute square of the
transform. This is equivalent to a spectral analysis of the signal.
Figure III.1
The normalized noise spectrum shown for various time-bandwidth products, N=16, 32 and 64.
4 E0 1
The units are in noise power T , and in frequency T .
From Figure III.1 it is seen that the noise spectrum is reasonable flat and thereby
neutral at the peak position determination. Only at the edges of the spectrum
variations occur. To prevent bias problems, a rule of only using 90% of the spec-
tral width could be applied.
In the following the Fisher information matrix will be computed in order to evalu-
ate the performance of the maximum likelihood method.
The Fisher information matrix (for multi variable estimation) is defined as follows:
∂ 2M
(F ) i , j = − < >
∂ν i ∂ν j
where M is the maximum likelihood function, ν i the variable no. i . The brackets
denote an ensample averaged over the stochastic variables. With the form of the
maximum likelihood function given in Eq. III.1 the Fisher information matrix ele-
ment (i , j ) can be written as:
It is seen that the Fisher matrix is diagonal which means that the estimation of
the parameters can be performed independently. The matrix elements can be
considered as conditional information gains. By inverting the matrix the elements
can be considered as uncertainties. The eigenvalues of the inverted matrix gives
the main axis an uncertainty ellipsoid. In this case where the matrix is on a di-
agonal form the uncertainties of the estimated parameters becomes:
2E0 ∆f
( ∆Ql ) 2 =
N
2E0 ∆f
( ∆φl )2 = (III.9)
NQl2
24E ∆f
( ∆ωD,l ) 2 = (φ l ) 2 = 2 0 2
T NQl
t −T / 2 t −T / 2
Only terms up to third order are shown in the expansion of the phase. If the am-
plitude, and higher order phase terms, are varying slowly compared with the time
scale of the second order term τ = 2π φ DD , the Fourier integral can be approxi-
mated to:
t +T / 2
1
i (t ′−t1 )2 φ
DD ( t1 )) (IV.6)
Vˆ (ω, t ) ≅ Q(t 1 )e i (φ(t1 )−φ(t1 )t1 )
D
∫
t −T / 2
e2 dt ′
ω = φD (t 1 )
2πi
≅Q(t 1 )e i ( φ(t1 )−ωt1 ) D
DD
− φ(t 1 ) ω = φ ( t1 )
Eq. IV.8 leads then to the formula for the determination of impact parameter as
function of time:
dυ
(ω, ) = (kΩa(t ),−t ) ∝ (a(t ), t ) IV.9
dω
To verify the assumptions, which lead to Eq. IV.6 expressions, for the amplitude
and the phase are needed.
The approximate result stated in Eq. IV.6, the expression for the amplitude Eq.
IV.7 and the last result in Eq. IV.10 show the factor dθ da , the defocusing factor,
is cancelled out. The defocusing factor gives the main time scale for the ampli-
tude, so in this respect the assumption concerning the time variation of the ampli-
tude is ‘self’ fulfilled. The higher order phase term (from third order and on) de-
pends on angular velocity Ω in powers of the order, so it seems to be safe to
assume that these terms are decaying very fast ( Ω ≈ 10 −3 rad / sec )
Now, the final result can be obtained: Inserting the amplitude from Eq. IV.7, the
second derivative of the phase from Eq. IV.10 into the expression for the Fourier
transform Eq. IV.6. This can be rewritten to yield:
2πi
Vˆ (ω, t ) ≅ Q(t 1 )e i ( φ(t1 )−ωt1 ) D =
DD
− φ(t 1 ) ω = φ (t1 )
t +T / 2
aPG 1 (IV.14)
) 2 e iυ ( t 1 ) + iυ 0 ∫ K (t )e
− iωt ′
( + dt ′
rG r L sin(θ )rG cos(ψ G )r L cos(ψ L ))) ω = φD (t 1 )
t −T / 2
6 References
Born, M. and E. Wolf, Principles of Optics, Cambridge University Press, 1999.
Meincke, M. D., Inversion methods for atmospheric profiling with GPS occulta-
tions. Scientific Report 99-11, Danish Meteorological Institute, Copenhagen,
1999.
Schmidt, R. O., Multiple emitter location and signal parameter estimation, IEEE
Transaction on Antennas and Propagation, AP-34, (3), 276-280, 1986.
Wickert, J., Reiger, C., Beyerle, G., Kçnig, R., Marquardt, C., Schmidt, T., Grun-
waldt, L., Gals, R., Meehan, T. K., Melbourne, W. G., Hocke, K. , Atmosphere
sounding by GPS radio occultation: First result from CHAMP, submitted to Geo-
physical Research Letter, 2001.