Prediction and Detection of Faraday Rotation in ALOS PALSAR Data

You might also like

Download as pdf
Download as pdf
You are on page 1of 4

Prediction and detection of Faraday rotation in ALOS

PALSAR data
Jeremy Nicoll, Franz Meyer Michael Jehle
Alaska Satellite Facility (ASF) Remote Sensing Laboratories (RSL)
University of Alaska Fairbanks University of Zurich
Fairbanks, Alaska Zurich, Switzerland
jnicoll@asf.alaska.edu michael.jehle@geo.unizh.ch

Abstract—Faraday rotation can degrade the quality of low-


frequency spaceborne SAR data, making an estimation and II. PREDICTION OF FARADAY ROTATION
correction of these effects a prerequisite for data quality A simplified equation for Faraday rotation is
continuity. In this paper, methods for predicting and estimating
Faraday rotation are presented and tested on ALOS/PALSAR  
data. A first example for unambiguous detection of Faraday Ω =  K 2  B cos(θ ) sec(ϕ )TEC (1)
 f 
rotation in SAR is shown. In addition, the improvement after
correcting for FR is proven using a real data example. where Ω is the one-way Faraday rotation, B is the local
geomagnetic field, f is the radio center frequency, θ is the angle
Keywords – polarimetric SAR calibration, Faraday rotation, between the magnetic field and the satellite pointing vector, φ
total electron content. is the off-nadir angle of the satellite, TEC is the total electron
content, and K is a composite constant [5]. Equation (1) is
I. INTRODUCTION based on the assumption that the magnetic field and electron
density are constant and fixed at 350 km elevation.
As radio waves propagate through the ionosphere, they
undergo Faraday rotation (FR), i.e. rotation of the polarization The geomagnetic field was calculated from geomagnetic
vector. Anisotropy in the ionosphere due to charged particles data provided by the National Geophysical Data Center
in the presence of a persistent magnetic field causes this (http://www.ngdc.noaa.gov/), based on the International
rotation [1]. Polarized synthetic aperture radar (SAR) data Geomagnetic Reference Field [6]. Calculations were generated
quality can be significantly impacted if the effect is not at 1º increments in latitude and longitude for late Fall 2006.
corrected [2], [3]. The effects can hamper or complicate The off-nadir angle and the angle between the magnetic field
calibration efforts, where Faraday rotation can mask cross-talk. and the satellite pointing was derived from PALSAR catalog
Image interpretation, especially decompositions relying on metadata located at the Alaska Satellite Facility’s (ASF’s)
channel ratios, can be seriously in error [4]. Americas ALOS data node (AADN)..
Faraday rotation is expected to be much more severe for L-
band than for C-band under the same ionospheric conditions
[4]. However, the extent of the effect on the backscatter
signature for a spaceborne L-band polarimetric instrument is
not well-known experimentally, due to a lack of data. With the
recent (January 2006) launch of the Japan Aerospace
Exploration Agency’s (JAXA’s) Advanced Land Observing
Satellite (ALOS) PALSAR mission, a rich archive of L-band
SAR data is now available.
One of the potential limitations to using this dataset is the
difficulty in detecting and correcting for Faraday rotation,
especially for dual polarization and single polarization data. A
listing of estimated Faraday rotation for all PALSAR image
frames could greatly aid in data selection for researchers
wishing to avoid FR effects or for those deliberately targeting
these effects. Figure 1: TEC values in 1016 units measured at CODE for October 31st, 2006
at 22:00 UTC. Highest TEC are seen on the side of the earth facing the sun.
TEC values were calculated at Center for Orbit where S is the scattering matrix, R F corresponds to the one-
Determination in Europe (CODE) using data from over 200 way Faraday rotation matrix, R and T are the receive and
GPS/GLONASS sites of the International GNSS service (IGS) transmit distortion matrices, A represents the overall gain of the
and other institutions. The interval between the maps published
on the internet (at www.aiub.unibe.ch/ionosphere/) is 2 hours. radar system, e jφ corresponds to the round-trip phase delay,
Figure 1 shows an example of a vertical TEC map generated by and N is an additive noise term.
the IGS for October 31st, 2006. These TEC maps provide a It has been shown that the calibration techniques applied to
grid of values spaced 2.5° in latitude and 5° in longitude ALOS/PALSAR data can successfully estimate and largely
denoting TEC levels [7]. The accuracy of these vertical TEC correct for the system-dependent terms A, R, and T. For targets
maps is highest where many GNSS stations contribute to the of sufficient SNR, N can be ignored as well. Under these
measurements and lowest over areas such as the ocean where assumptions, the measured scattering matrix M reduces to:
the density of the GNSS network is lower. In the latter case, the
accuracy of the vertical TEC is still within 5 TECU.  M hh M vh   cos Ω sin Ω 
M  =  ⋅
Figure 2 is a histogram of the predicted FR for each ALOS  hv M vv  − sin Ω cos Ω (3)
PALSAR granule in the AADN archive as of January 23, 2007.  S hh S vh   cos Ω sin Ω 
It is readily apparent the difficulty in finding a product with a S ⋅ 
large enough Faraday rotation to detect unambiguously. While  hv S vv  − sin Ω cos Ω
there are many thousands of images available in the archive, Nonzero Faraday rotation causes the cross-pol measurements
there are only a few (18) fully-polarimetric images with a M vh and M hv to be non-reciprocal, an effect that can be
predicted FR larger than 3º, all contained in a handful of
exploited for Faraday rotation retrieval. In this paper two
acquisitions. Current TEC values are at a low in the 11-year
different Faraday rotation estimation methods are applied. One
solar cycle, and most PALSAR acquisitions occur at a local
of the approaches extracts the Faraday rotation angle Ω by
time of 2200, when TEC activity is typically low. Finding a
solving the equation system in equation 2 directly (see [8]):
dataset with appreciable Faraday rotation would be very
difficult if approached only from a random statistical sampling
1  (M − M hv ) 
of the archive. Use of a predictive approach is proven Ω= tan −1  vh  (4)
invaluable in finding such a dataset. 2  (M hh + M vv )
10000
A more robust version of this approach uses spatial
averaging to reduce the influence of speckle on the estimated
Ω angle. The second approach introduced by Bickel and Bates
1000
[9] transforms M to a circular basis Z via

 Z11 Z12  1 j   M hh M vh   1 j 
Z = ⋅ ⋅ (5)
100
 12 Z 22   j 1  M hv M vv   j 1
18
From Equation (5) Ω can be derived by calculating
10
1
Ω= arg (Z12 Z 21

). (6)
1
4
Both approaches are applied in the following to estimate
-3.5
-2.5
-1.5
-0.5
0.5
1.5
2.5
3.5
4.5
5.5
6.5

Faraday rotation from ALOS/PALSAR data. Both point targets


of known scattering behavior and distributed targets are used
Figure 2: Histogram of predicted Faraday rotation for all ALOS PALSAR for estimation.
holdings through Jan 23 2007 at the AADN archive. Y-axis shows number of
scenes at each interval on X-axis. Solid bar is total PALSAR. Hatched bar is
fully polarimetric data only. Note the logarithmic scale.
A. Point Target Analysis
Bickel and Bates [9] introduced an approach to measure
Faraday rotation from the signal’s odd-bounce point scatterers,
III. ESTIMATION OF FARADAY ROTATION FROM THE DATA e.g. calibration targets such as trihedral corner reflectors.
In the presence of Faraday rotation, the backscatter Theoretically, calibration targets are optimal for FR estimation,
measurement reciprocity (HV=VH) is violated. This effect is as their scattering matrix is known and their cross-pol return is
used to estimate FR in the image itself from a comparison of zero in case of Ω = 0 . However, when analyzing real data sets,
the cross-polarized terms of the scattering matrix. According to the use of controlled targets has its limitations:
[8] the measured scattering matrix M of a SAR system • In case of very low Faraday rotation, such as that observed
measuring linear horizontal (H) and vertical (V) polarized in years of low solar activity, the cross-pol peak of a
signals can be written as: corner reflector is often below or on the order of the radar
cross section (RCS) of its surrounding image background.
M = Ae jφ R T R F SR F T + N (2) This causes the FR estimates to be contaminated with
noise and prevents the detection of low Faraday rotation panel displays the SAR image, the Faraday rotation is shown in
angles. the center, while the right panel includes statistic plots as well
• Targets of known scattering behavior are rare, preventing a as analyses of possible range and azimuth dependencies.
statistic analysis of their signals. The limited swath-width
The Washington DC example was selected for this paper as
of full-polarimetric PALSAR images amplifies this
it is the image with the highest predicted FR (4º) in the ASF
problem, additionally.
archive. The pixel-by-pixel FR results are averaged over an
Although these limitations reduce the merit of point targets area of 10 × 10 pixels to remove the influence of speckle and
for the estimation of Faraday rotation, they are still valuable for azimuthal slopes on the estimated results. The average FR
validating FR estimates derived from distributed targets and are resulted in µ Ω = 2.83D , with a standard deviation of
well suited for controlling the quality of FR correction. Of the
data analyzed, only one image (shown in Figure 4) was σ Ω = 0.22 D . The estimated range dependence shows a trend of
available with a corner reflector present. The target had no Ω in far range. This trend is created by data processing
detectable cross-pol peak, consistent with the distributed target artifacts. No range reference function is applied during data
estimates for Figure 4 discussed in the next section. processing for the last lines in far range. For estimating FR
statistics the affected far range lines were not considered. The
B. Distributed Target Analysis azimuth analysis shows no significant trends. An inspection of
the spatial distribution of the FR estimates (center panel of
Figure 2) shows significant deviations from the mean FR in
areas with open water. These deviations increase with a
decrease in backscattered signal, increasing the contribution of
thermal noise. With the results presented in Figure 2 a first
example of unambiguous detection of FR in SAR is published.

Figure 3. Example of Faraday rotation estimation from a quad-pol PALSAR


data set over the area of Washington DC. The left panel displays the SAR
image, the Faraday rotation is shown in the center. Statistic plots and range as
well as azimuth dependences are presented in the right.

The only requirement observed targets need to meet to


allow for a successful estimation of FR from full-polarimetric
SAR observations is backscatter reciprocity. Targets having Figure 4: Example of Faraday rotation estimation from a quad-pol PALSAR
data set over the area of Delta Junction, AK.
significant azimuth slopes will violate this requirement.
Although such targets are present in every SAR image, their
The second example was chosen as it shows mountainous
influence on the FR estimates will disappear if the FR results
areas with a high number of slopes varying in inclination and
are averaged over large areas [4].
orientation. From the FR prediction and from measurements
At the Alaska Satellite Facility (ASF), a tool has been taken on a corner reflector in the image we expect the FR to be
developed that provides FR estimates over an entire full- between -0.5 and 1D . After averaging over 10 × 10 pixels, the
polarimetric PALSAR image. In a first step, Ω is estimated on mean FR of the second example is estimated to be µΩ = −0.15D ,
a pixel-by-pixel basis applying the method introduced in [9].
which corresponds well to the predicted value. Its standard
This method was chosen over the method introduced by
Freeman (see [8]) because of its lower sensitivity to noise and deviation of σ Ω = 0.20 D confirms that the estimated FR does
calibration deficiencies. In a second step, the results are not differ significantly from zero and that terrain slopes did not
averaged over areas large enough to remove the effects of introduce a bias. No significant trends in either range or
azimuthal slopes while small enough to retain possible low azimuth are present.
frequency FR variations over the image. The Figures 3 and 4
show two examples of Faraday rotation estimates for PALSAR
data over Washington DC and Delta Junction, AK,
respectively. Each figure is divided into three panels. The left
IV. COMPARISON BETWEEN ESTIMATION AND PREDICTION data. This image demonstrates the efficacy of screening for FR
OF FARADAY ROTATION with the TEC prediction method, estimation from the Bickel
In addition to the two datasets analyzed in detail in the approach, and removal by inversion. Application of the Bickel
previous section, FR for 12 more images was measured. A approach to large distributed areas was in good agreement with
both the measurements from point targets and the FR
least-squares linear fit between TEC prediction and estimation
from data (shown in Figure 5) yields a slope of prediction.
m = 1.02 (σ m = 0.14) and an intercept of b = 0.77(σ b = 0.19) .

y = 1.02x + 0.77
4 2
R = 0.82
TEC prediction

Figure 6: Magnitude of (MVH – MHV) for an image with a) ~3º Faraday


-1
rotation, and b) Faraday rotation removed. The image is an urban area from a
-1 0 1 2 3 4 5 subset of the Washington DC image in Figure 2.
Bickel Estimation

Figure 5: FR predicted from TEC (Equation 1), and estimated from data The paper demonstrates unambiguous detection of Faraday
(Equations 5 and 6) to 14 fully polarimetric acquisitions. Solid line is the least rotation in SAR data, and demonstrates that FR effects can be
squares fit to the data. Dashed line is the ideal. removed successfully. Discovery of an image with high enough
While the prediction agrees reasonably well with estimates, FR for detection was greatly aided with the FR prediction from
there are limitations to the precision of the fit. There are few TEC. Azimuthal slopes can be accounted for in large samples
points in the curve at the extremes of the fit, which with spatial averaging. The prediction from TEC can be used to
significantly affects determination of the slope. The available screen all PALSAR data, but estimation from the data and
data characterizes a very small region (-1º to 3º) of the possible removal is possible only for fully polarimetric data.
FR values (-45º to 45º), so caution is advised when
extrapolating the correlation to much higher FR. Simulations REFERENCES
predict that results should be similar or even improved. With [1] A.R. Thompson, J.M. Moran, and G.W. Swenson, “Interferometry and
increasing FR more and more co-pol signal bleeds into the Synthesis in Radar Astronomy,” New York: Wiley, 1986.
cross-pol channels placing all four channels at comparable [2] Z.-W. Xu, J. Wu, and Z.-S. Wu, “A Survey of Ionospheric Effects on
SNR, and minimizing noise in the FR estimates. Space-based Radar,” Waves in Random Media, vol. 14, pp. 189–273,
2004.
The intercept bias is difficult to explain. It requires a cross- [3] F. Meyer, R. Bamler, N. Jakowski, and T. Fritz, “The Potential of Low-
talk of ~ -25 dB to create a 0.8º bias in the FR estimates. Frequency SAR Systems for Mapping Ionospheric TEC Distributions,”
However, corner reflector analysis limits cross-talk to about Geoscience and Remote Sensing Letters, vol. 2, no. 4, pp. 560–564,
-35 dB, which accounts for only 0.2º of bias. Therefore, the 2006.
observed bias cannot be fully accounted for by cross-talk. It is [4] A. Freeman, and S. Saatchi, “On the Detection of Faraday Rotation in
Linearly Polarized L-Band SAR Backscatter Signatures,” Trans. On
also possible that there is a bias in the TEC measurements. The Geoscience and Remote Sensing, vol. 42, No. 8, pp. 1607–1616, 2004.
objection to this is that TEC is a multiplier in Equation 1; if [5] P.A. Wright, S. Quegan, N.S. Wheadon, and C.D. Hall, “Faraday
there were errors in magnitude one would expect a bias in the Rotation Effects on L-Band Spaceborne SAR Data,” Trans. On
slope, not the intercept. A third possibility is that the dataset is Geoscience and Remote Sensing, vol. 41, No. 12, pp. 2735-2744, 2003.
too limited in range of values, and that additional observations, [6] McLean, S., S. Macmillan, S. Maus, V. Lesur, A. Thomson, and D.
especially with larger FR, would average out the bias. Dater, December 2004, The US/UK World Magnetic Model for 2005-
2010, NOAA Technical Report NESDIS/NGDC-1.
[7] Schaer, S.: Mapping and Predicting the Earth’s Ionosphere using the
V. SUMMARY Global Positioning System, Geodätisch- geophysikalische Arbeiten der
Removal of FR, once it is known, is accomplished by Schweiz, Vol. 59, 1999.
inversion of Equation (3). As long as there are not significant [8] A. Freeman, “Calibration of Linearly Polarized Polarimetric SAR Data
Subject to Faraday Rotation,” Trans. On Geoscience and Remote
FR trends in the image in azimuth or range, a single value can Sensing, vol. 42, No. 8, pp. 1617–1624, 2004.
be used to correct the whole image. Figure 6 shows the cross-
[9] S. H. Bickel and R. H. T. Bates, “Effects of magneto-ionic propagation
pol difference for a subset of the Washington DC image before on the polarization scattering matrix,” Proc. IRE, vol. 53, pp. 1089–
(left) and after (right) FR correction. For reciprocal targets and 1091, 1965.
under zero FR conditions, the differences should be identical
exception for speckle and thermal noise [8]. The removal of all
deterministic signals in the corrected image on the right proves
the successful removal of Faraday rotation effects from the

You might also like