SAR Algorithms

Recap: What is SAR processing?

• SAR processing algorithms model the scene as a set of discrete point targets that do not
interact with each other (aka Born approximation)
• No multibounce
• The electric field at the target comes only from the incident wave and not from surrounding scatterers
• The target model is linear because the scattered response from point target P1 and point target P2 is
modelled as the response from point target P1 by itself + response from point target P2 by itself
• We can apply the principle of superposition!!!
• SAR processing is the application of a matched filter for each pixel in the image where the
matched filter coefficients are the response from a single isolated point target
• We will assume noise is whitened (decorrelated)
• Equivalently, we can say:
• SAR processing is a correlation filter between a single isolated point target response and the raw data
• SAR processing is an inner product between our model of a single isolated point target and the raw data
Recap: What is SAR processing?
• SAR processing algorithms model the scene as a set of discrete point targets that do not
interact with each other (aka Born approximation)
• No multibounce
• The target’s electric field is only from the incident wave and not from surrounding scatterers
• The target model is linear because the scattered response from point target P1 and point target P2 is
modelled as the response from point target P1 by itself + response from point target P2 by itself
• We can apply the principle of superposition!!!
• SAR processing is the application of a matched filter for each pixel in the image where the
matched filter coefficients are the single isolated point target response
• We will assume noise is whitened (decorrelated)
• Equivalently, we can say:
• SAR processing is a correlation filter between a single isolated point target response and the raw data
• SAR processing is an inner product between our model of a single isolated point target and the raw data
Recap: What is SAR processing?
• SAR processing algorithms model the scene as a set of discrete point targets that do not
interact with each other (aka Born approximation)
• No multibounce
• The target’s electric field is only from the incident wave and not from surrounding scatterers
• The target model is linear because the scattered response from point target P1 and point target P2 is
modelled as the response from point target P1 by itself + response from point target P2 by itself
• We can apply the principle of superposition!!!
• SAR processing is the application of a matched filter for each pixel in the image where the
matched filter coefficients are the single isolated point target response
• We will assume noise is whitened (decorrelated)
• Equivalently, we can say:
• SAR processing is a correlation filter between a single isolated point target response and the raw data
• SAR processing is an inner product between our model of a single isolated point target and the raw
data
Recap: What is SAR processing?
• So… SAR processing is a matched filter and the filter is linear
• If the filter was also space invariant we could apply it in the frequency
domain
• But: the filter is not space invariant. The point target’s shape changes
depending on the range to the radar.
Why do we care that it is not space invariant?
• Recall linear time invariant (LTIV) systems have complex exponentials as their
Eigenfunctions. A change of basis of the input and output to complex exponentials
means that a simple component-wise multiply is all that is needed to apply the
filter. A change of basis to complex exponentials can be efficiently implemented
using a Fast Fourier Transform (FFT) assuming data are uniformly sampled.
• Without Fourier method, O(N2M2) operations are required instead of O(N*log2(N)
M*log2(M)) where N and M are the dimensions of the image and are usually on
the order of thousands of pixels each. The direct application of “slow” convolution
could be more than 100x slower than “fast” or Fourier based convolution.
• Good news: we can exploit the structure of the signal to transform (usually
through interpolation) the data into a domain where the signal is space invariant!
To do this, we require properly sampled raw data and image pixels.
Principle of Stationary Phase (PSOP)
• PSOP is used to approximately solve integrals of the form
𝐵

𝐼 =∫ 𝐹 ( 𝑥 ) 𝑒
− 𝑗  (𝑡 )
𝑑𝑥
𝐴

where the phase function, , is rapidly varying over the range of integration except
for a few points where the derivative is zero (aka stationary points) AND is a
slowly varying function by comparison.
• With A and B equal to - and , the integration looks a lot like a 1-D Fourier
integral
• SAR chirp signals are similar to quadratics. Quadratic functions vary quickly
everywhere and have a single stationary point.
• The envelope of a SAR signal varies slowly with time.
 Remember:

 (𝑡 )
must include your original phase function
being integrated AND the Fourier term:
−2 𝜋 𝑓𝑡
1. Write out envelope and phase function
2. Determine derivative of phase function.
3. Solve for the stationary point, ts, in
terms of f. This is the first messy part…
𝑡 𝑠 ( 𝑓 )=…
4. Determine second derivative of phase
function. IGNORED IN OUR
DERIVATIONS!
Complex Gaussian has a 5. Plug t(f) into (4) wherever the stationary
closed form solution! point occurs.
6. Simplify! This is the second messy part…

Process is the same for inverse Fourier

transform except replace eqns above with:
( 𝑓 ) 2 𝜋 𝑓𝑡 𝑓 𝑠 ( 𝑡 )=…
Good online SAR Resource
• https://saredu.dlr.de/unit
Satellite and Low Squint Airborne SAR
Algorithms
• Lower squint (often <4-5 deg)
• Narrow azimuth bandwidth (usually 0.5 deg to 10 deg azimuth beamwidth)
• Range Doppler Algorithm
• Used by the Canadian Space Agency to process RADARSAT-1 and RADARSAT-2 satellite
SAR data
• Chirp Scaling Algorithm
• Used by the European Space Agency and the German Aerospace Center (DLR) to
process TerraSAR-X satellite SAR data
• These two algorithms (RDA and CSA) are very similar with the primary
difference being how range cell migration correction is done.
• RDA works with any waveform, CSA requires the use of a chirp waveform
Satellite and Low Squint Airborne SAR
Algorithms
• The SAR filter is azimuth-space-invariant but it is range-variant
• The primary structure exploited by these two algorithms is that the 2-
D energy from the point target lies along a 1-D contour. This energy
will be interpolated or scaled/shifted to lie on a 1-D line that does not
cross range bins. By converting the range varying dimension to lie on
a single range bin, convolution will no longer be required in the range
dimension.
Range Doppler Algorithm (RDA) STEP 1
• Pulse compression is a LTIV filter. It is straight forward to implement in
the Fourier domain.
• Range FFT on raw data to transform to range-frequency / azimuth-space
domain
• Apply range-domain matched filter for pulse compression
• Do not take the IFFT in the range dimension when finished.
 Range Doppler Algorithm (RDA) STEP 2
• Azimuth FFT
• Transform to range-frequency / Doppler domain 2D Fourier Domain (3 targets)
0
-15
Raw Data (single target)
740 0
-5
-10
745 -5

Relative power (dB)

Relative range (m)

-5 -10

Relative power (dB)

750 -10

Frequency (MHz)
755 -15 0 -15

760 -20
5 -20
-25
765
10
-25
-30
-0.6 -0.4 -0.2 0 0.2 0.4 0.6
Slow time (sec) 15
-30
-600 -400 -200 0 200 400 600
Azimuth frequency (Hz)
Range Doppler Algorithm (RDA): STEP 3
• Blurring occurs during the Doppler Fourier transform so that the point
target “contour” is broadened. This affect is worse for large squint
angles.
• This blurring can be approximated by a frequency chirp in the range
domain… so to correct we need to do pulse compression again.
• This process is called Secondary Range Compression
• For an approximate solution, this second range compression can be applied
during the regular pulse compression… this is suboptimal because the Fourier
transform to the Doppler domain blurs the correction so it is better to apply
in the range-Doppler domain.
Range Doppler Algorithm (RDA): STEP 3

0 0

30
0
-5 -5
35
20

40 -10 -10

Relative power (dB)

Relative power (dB)

40
Relative range (m)

Range (m)
45
-15 60 -15
50
80
-20 -20
55
100

60 -25 -25
120
65
-30 140 -30
2.89 2.895 2.9 7120 7140 7160 7180 7200 7220
Slow time (sec) Azimuth frequency (Hz)

Range Space Domain (i.e. Raw Data) Range Doppler Domain

(note the blurring)
Range Doppler Algorithm (RDA): STEP 3
• The SRC correction is derived from our range Doppler representation of the signal:

• Note that this should be (midpoint of scene) if applied in the range-frequency domain as described here.
Improved performance can be seen by applying the SRC chirp compression with the RCMC interpolating
kernel since both are range varying filters at that point. If this is done, then can be used since RCMC
interpolation is done in the range-Doppler domain.
• : Doppler frequency
• : Effective velocity (rectilinear coordinate system)
• : Baseband range frequency
• : Center frequency
• : Cosine of the squint angle,
Range Doppler Algorithm (RDA): STEP 3

0 0

0 0
-5 -5
20 20

-10 -10

Relative power (dB)

Relative power (dB)

40 40
Range (m)

Range (m)
60 -15 60 -15

80 80
-20 -20
100 100

-25 -25
120 120

140 -30 140 -30

7120 7140 7160 7180 7200 7220 7120 7140 7160 7180 7200 7220
Azimuth frequency (Hz) Azimuth frequency (Hz)

Range Doppler Domain Range Doppler Domain

(After Secondary Range Compression) (note the blurring)
Range Doppler Algorithm (RDA): STEP 4
• Range IFFT
• Transform to range / Doppler domain
Range Doppler Algorithm (RDA): STEP 5
• Range Cell Migration Correction (RCMC) in Doppler domain
• SAR processing is a 2-D filter, but the energy is focused along a single
hyperbolic contour.
• Contour is range dependent
• The idea is to flatten the contour using a process called RCMC
• Example point target response: 740 0

-5
• RCMC easy to apply for a single 745

Relative power (dB)

Relative range (m)
-10
point target. 750

755 -15

760 -20

-25
765
-30
-0.6 -0.4 -0.2 0 0.2 0.4 0.6
Slow time (sec)
Range Doppler Algorithm (RDA): STEP 5
• Example of two point targets at the same range and next to each
other. Envelope is about the same for both but the phases are offset
(think of two tones and what you see is the beat frequency… double
side band suppressed carrier).
• Could apply RCMC for this case as well.
0
495
-5

Relative power (dB)

500
Relative range (m)

-10
505
-15
510
-20
515
-25
520
-30
-0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8
Slow time (sec)
Range Doppler Algorithm (RDA): STEP 5
• Example of two point targets far apart from each other… RCMC not
possible because each target needs a different correction.
490 0

-5
495

Relative power (dB)

Relative range (m)

-10
500
-15
505
-20
510
-25

515 -30
-0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6
Slow time (sec)
Range Doppler Algorithm (RDA): STEP 5
• Example of two point targets far apart from each other:

480 -5

Relative power (dB)

Relative range (m)

490 -10

500 -15

-20
510
-25
520
-30
-1 -0.5 0 0.5 1
Slow time (sec)
Range Doppler Algorithm (RDA): STEP 5
• RCMC cannot be applied in the range-space domain because RCMC is
dependent on the relative along-track position rather than the
absolute along-track position.
• Hmmm… we know that the range cell migration is a function of
incidence angle (i.e. Doppler).
• RCMC can be applied in the range-Doppler domain because RCMC
depends on the absolute Doppler.
• Every target at the same range has the same envelope in the range-
Doppler domain!!!
 Range Doppler Algorithm (RDA): STEP 5

450 0 450 0

460 460
-10 -10

470 470
-20 -20
480 480

Relative power (dB)

Relative power (dB)
-30 -30
490 490
Range (m)

Range (m)
500 -40 500 -40

510 510
-50 -50
520 520
-60 -60
530 530

-70 -70
540 540

550 -80 550 -80

-1500 -1000 -500 0 500 1000 1500 -1500 -1000 -500 0 500 1000 1500
Azimuth frequency (Hz) Azimuth frequency (Hz)

Single Target Both Targets… envelope has not

changed, but interference pattern
has.
Range Doppler Algorithm (RDA): STEP 5
• We need to remove this much delay (this turns out to be simple
geometry):

• : Doppler frequency
• : Effective velocity (rectilinear coordinate system)
• : Cosine of the squint angle
 Range Doppler Algorithm (RDA): STEP 5
• Use the truncated and windowed sinc interpolation method to do the
time shift. Example of 3 deg squint:
-100 0 -100 0

-50 -5 -50 -5

0 -10 0 -10

Relative power (dB)

Relative power (dB)
Relative Range (m)

Relative range (m)

50 -15 50 -15

100 -20 100 -20

150 -25 150 -25

200 -30 200 -30

-2800 -2600 -2400 -2200 -2000 -1800 -1600 -1400 -2800 -2600 -2400 -2200 -2000 -1800 -1600 -1400
Azimuth frequency (Hz) Azimuth frequency (Hz)
 Range Doppler Algorithm (RDA): STEP 5
• Use the truncated and windowed sinc interpolation method to do the
time shift. Example of 10 deg squint:
0 0

-200 -200

-5 -5

-100 -100

-10 -10

Relative power (dB)

Relative power (dB)

Relative range (m)

0 0
Range (m)

-15 -15

100 100

-20 -20
200 200
-25 -25
300 300
-30 -30
-7600 -7400 -7200 -7000 -6800 -6600 -6400 -6200 -7600 -7400 -7200 -7000 -6800 -6600 -6400 -6200
Azimuth frequency (Hz) Azimuth frequency (Hz)
Range Doppler Algorithm (RDA): STEP 6
• All targets have been interpolated so that they occupy a single range
bin in the range-Doppler domain.
• Originally the problem was that the range cell migration changed as a
function of range  This prevented a simple application of Fourier
methods since the response was space-variant.
• Now it is no longer a 2-D filter so the space variance does not matter
and we only need to apply a 1-D azimuth filter.
Range Doppler Algorithm (RDA): STEP 6
• Using the range-Doppler representation of the signal after RCMC, the
azimuth compression filter is:

• : Doppler frequency
• : Effective velocity (rectilinear coordinate system)
• : Cosine of the squint angle
• : Center frequency
• : Speed of light
Range Doppler Algorithm (RDA): STEP 7
• Azimuth IFFT
• Transform into range / azimuth-space domain
 Range Doppler Algorithm (RDA): STEP 7
• Example (side note: range dependent Doppler centroid correction and
relative range cell migration correction when there is squint).
0 -140 0

-50
-5 -120 -5

0
-100
-10 -10

Relative power (dB)

Relative power (dB)

Relative range (m)

Relative range (m)

50 -80
-15 -15

-60
100
-20 -20
-40
150
-25 -25
-20

200
-30
-30 -40 -20 0 20 40 60
-60 -40 -20 0 20 40 60
Along-track (m)
Along-track (m)

No squint: Position is perfect 3 deg squint: range is correct, but azimuth is off by one pixel
Range Doppler Algorithm (RDA): STEP 7
Azimuth correction ends with smeared
10 deg squint (RCMC not perfect) range bins
40 0 0

-1520
-5 -5
50

-1500
-10 -10

Relative power (dB)

Relative power (dB)

Relative range (m)

Relative range (m)

60

-1480
-15 -15
70
-1460
-20 -20
80
-1440
-25 -25
90
-1420
-30 -30
-7500 -7000 -6500 1.574 1.576 1.578 1.58 1.582
Azimuth frequency (Hz) Along-track (m) 4
x 10
Chirp Scaling Algorithm (CSA)
• The problem with RDA is that the RCMC interpolation is slow and
requires SRC.
• Chirp scaling does the same thing as RDA, but does the RCMC with
chirp scaling which also makes the blurring from the Doppler Fourier
transform smaller.
• Greater efficiency + range/azimuth decoupling built into range compression
(analogous to range Doppler algorithms secondary range compression)
Chirp Scaling Algorithm (CSA): Step 1
• Azimuth FFT
• Transform to range / Doppler domain
-1500 0
-1500 0

-5
-1000 -1000 -5

-10
-500 -500 -10

Relative power (dB)

Relative Range (m)
Relative power (dB)
Relative range (m)

-15

0 -20 0 -15

500 -25
500
-20
-30
1000 1000
-25
-35

1500
-40 1500
-0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 -30
-600 -400 -200 0 200 400 600
Azimuth time (sec)
Azimuth frequency (Hz)
Chirp Scaling Algorithm (CSA): Step 2
• Apply chirp scaling… multiply by:

• : Doppler frequency

• : Effective velocity (rectilinear coordinate system)

• : Cosine of the squint angle
• : Time
• : Speed of light
Chirp Scaling Algorithm (CSA): Step 2
• Continued…

• : Range chirp rate

• : Doppler frequency
• : Effective velocity (rectilinear coordinate system)

• : Cosine of the squint angle

Chirp Scaling Algorithm (CSA): Step 3
• Range FFT
• Transform to range-frequency / Doppler domain
0
-15

-5
-10

-5 -10

Relative power (dB)

Frequency (MHz)

0 -15

5 -20

10
-25

15
-30
-7600 -7400 -7200 -7000 -6800 -6600 -6400 -6200
Azimuth frequency (Hz)
Chirp Scaling Algorithm (CSA): Step 4
• Range Compression (including range/azimuth decoupling) + bulk range cell
migration correction

• : Doppler frequency

• : Effective velocity (rectilinear coordinate system)

• : Cosine of the squint angle
• : Baseband range frequency
• : Speed of light
• : From before but evaluated at
Chirp Scaling Algorithm (CSA): Step 5
• Range IFFT
• Transform to range / Doppler domain
0

-50

-5

-10

Relative power (dB)

Relative Range (m)

50

-15

100
-20

150
-25

200
-30
-400 -200 0 200 400 600
Azimuth frequency (Hz)
Chirp Scaling Algorithm (CSA): Step 6
• Azimuth compression and phase correction. Multiply by…

• : Doppler frequency
• : Effective velocity (rectilinear coordinate system)
• : Cosine of the squint angle
• : Center frequency
• : Speed of light

• : From before
 Chirp Scaling Algorithm (CSA): Step 7
• Azimuth IFFT
• Transform to range / azimuth-space domain
0 0
-1520
-1520
-1510
-5 -5
-1510
-1500

-1500 -10 -10

Relative power (dB)

Relative power (dB)

Relative range (m)

-1490

Relative range (m)

-1490
-15 -1480
-15
-1480
-1470
-20 -20
-1470 -1460

-1460 -25 -1450 -25

-1450 -1440
-30 -30
-20 -10 0 10 20 30 -30 -20 -10 0 10 20 30 40 50
Along-track (m) Along-track (m)
Wide Aperture (Airborne and Ground based)
Algorithms
• f-k migration (AKA -k migration as in omega-wavenumber migration)
• Handles strip map mode data collection with very wide apertures
• Disadvantage is that time and space variant modifications are not handled well because processing is done in the
f-k domain.
• Time domain correlation (TDC): not covered
• Fast factorized TDC is a good and fast implementation of TDC which keeps most of the desirable properties of
TDC
• Lars M.H. Ulander et al., Synthetic-Aperture Radar Processing Using Fast Factorized Back-Projection, Transactions
on Aerospace and Electronic Systems, vol. 39, no. 3, July 2003.
• Polar Format Algorithm (PFA) : not covered
• Armin W. Doerry, Synthetic Aperture Radar Processing with Tiered Subapertures, Sandia Report SAND94-1390,
1994.
• Very complete description of PFA
• Jack L. Walker, Range-Doppler Imaging of Rotating Objects, IEEE Transactions on Aerospace and Electronic
Systems, vol. 16, no. 1, Jan 1980.
• Original reference.
F-k migration
• Exploding reflector model
• The linear target model is equivalent to the exploding reflector model
• Rather than the radar transmitting a pulse at time zero, each target is
replaced by an isotropic source that radiates a pulse starting at time zero and
the velocity of propagation is halved.
F-k migration: Step 1
• Two-dimensional FFT
• Transform to range-frequency / wavenumber domain
• (Wavenumber has a one to one mapping with Doppler domain)
0
-15

-5
-10

-5 -10

Relative power (dB)

Frequency (MHz)

0 -15

5 -20

10
-25

15
-30
-0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8
Wavenumber (rad/m)
F-k migration: Step 2
• Reference frequency multiply (RFM)
• Applies the 2-D filter for the reference range (i.e. determine the response
from a point target at the reference range and then use that as a
correlation/matched filter)
• This will apply both range and azimuth compression
• We know that this will perfectly focus the reference range, but slowly degrade
away from that range because the filter needs to be space variant to perfectly
focus the targets
F-k migration: Step 2
•T

• : Speed of light

• : Baseband range frequency

• : Doppler frequency
• : Effective velocity (rectilinear coordinate system)
• : Chirp rate
 F-k migration: Step 2
• Examples of reference range and away from reference range
4
x 10
0 0
-50 0.998

-5 -5

0 1

-10 -10

Relative power (dB)

Relative power (dB)

Relative range (m)

Relative range (m)

1.002
50

-15 -15
1.004
100
-20 1.006 -20

150
-25 1.008 -25

200
1.01
-30 -30
-100 -50 0 50 100 150 -150 -100 -50 0 50 100 150
Along-track (m) Along-track (m)
F-k migration: Step 3
• Stolt Interpolation
• First we note the residual phase after reference frequency multiply (RFM) filter is:

• : Speed of light

• : Baseband range frequency

• : Doppler frequency
• : Effective velocity (rectilinear coordinate system)
F-k migration: Step 3
• Stolt Interpolation
• Data start uniformly sampled in
• Define a new variable :

• We note that there is a one to one mapping between to and we can solve for in terms of :

• If we do a change of variable to and resample the range frequency axis so that is uniformly
sampled (instead of ), then we end up with:

• Now the IFFT of this signal will produce a focused point at which is just what we want!
• Resampling usually uses sinc interpolation for best results, but sometimes other
interpolators are used such as linear interpolation with oversampling
 F-k migration: Step 4
• Two-dimensional IFFT
• Transform to range-space domain
• Before and after Stolt interpolation for target a long way from the reference
x 10 range.
4 4
x 10
1.996 0 0
1.994

-5 1.996 -5
1.998

1.998
-10 -10

Relative power (dB)

Relative power (dB)

Relative range (m)

Relative range (m)

2
2

-15 -15
2.002 2.002

-20 2.004 -20

2.004
2.006
-25 -25
2.006 2.008

-30 -30
-100 -50 0 50 100 -20 -10 0 10 20
Along-track (m) Along-track (m)