Perspectives on SAR Processing

1
Perspectives on SAR Processing

• Ian Cumming
• Professor Emeritus
• Radar Remote Sensing Group
• Dept. of Electrical and Computer Engineering
• University of British Columbia

2
Perspectives on SAR Processing

1. Some SAR processing history
2. Review of current SAR proc. algorithms
3. Which algorithms are used today?
4. Some image examples
5. Summary and final thoughts

Sept. 11, 2007
2
Cumming ASAR 2007
3
Some Processing History

• In 1976, SAR data were processed by coherent
  optics
• only one book available
• very hard to understand for a DSP engineer
• used laser beams and lenses to focus image
• fast, but limited dynamic range
• must be a better way digital processing!
• Challenge of digital processing
• modest computing resources (memory speed)
• processing algorithms not known
• received signal model was needed (geometry)
• limited satellite orbit knowledge

Sept. 11, 2007
3
Cumming ASAR 2007
4
Early Airborne Processing

• In 1970s, CCRS obtained an airborne SAR from
  ERIM
• Installed on a Convair-580
• Previous processor was optical
• MDA contracted to built an on-board real-time
  processor
• First RTP delivered in 1979 (additional units in
  1985)

Sept. 11, 2007
4
Cumming ASAR 2007
5
Early Airborne Processing

• Processor characteristics
• time-domain correlation
• X-band, no RCMC needed
• built from discrete components
• multiplier-accumulator chips, small memory chips
• 4 giga ops per second in one chassis
• performed azimuth compression first, then range
  compression
• real-time waterfall display on-board

On-board real-time processor on CCRS Convair-580
Sept. 11, 2007
5
Cumming ASAR 2007
6
Early Satellite Processing

• SEASAT (October 1978)
• developed detailed model of satellite motion
• led to received signal model
• matched filtering done by FD correlation process
• had to separate range and azimuth processing
• to fit in to computer resources
• range cell migration correction a challenge
• needed interpolator to eliminate paired echos

Sept. 11, 2007
6
Cumming ASAR 2007
7
First Digital SEASAT Image
Featured in Aviation Week, Feb. 26, 1979
Sept. 11, 2007
7
Cumming ASAR 2007
8
Typical product 1979-1981

• Niagara Falls
• 40 x 40 km scene
• 25 m resolution
• 4 looks
• 40 hours to process
• big minicomputer
• corner turning on disc
• needed autofocus for L-band

Sept. 11, 2007
8
Cumming ASAR 2007
9
Review of SAR Processing Algorithms

• Developed primarily for satellite SAR processing
• Range/Doppler (1978) JPL MDA
• Time-domain (1978)
• SPECAN (1980) MDA
• Omega-K (1988) Polimi, Italy
• Chirp scaling (1992) CCRS, DLR , MDA

Sept. 11, 2007
9
Cumming ASAR 2007
10
SAR Signal Model

• Range and azimuth extent of signal received from
  a point target (shown in red)
• The signal in the two axes is not orthogonal
  because of the curvature if it were, the
  processing would be quite simple
• The processing is achieved by a matched filtering
  (correlation) operation, but what to do about the
  curvature, if we are restricted to 1-D
  operations?

Sept. 11, 2007
10
Cumming ASAR 2007
11
Range/Doppler Algorithm 1

• Developed at MDA and JPL
• 1977-79 plus later refinements
• Used 1-D frequency domain correlation
• processing separated in range and azimuth
  dimensions
• for computing efficiency memory limitations
• Had to deal with range migration
• MDA recognized importance of accurate
  interpolation
• JPL developed secondary range compression to deal
  with range-azimuth coupling when migration large
  1984

Sept. 11, 2007
11
Cumming ASAR 2007
12
Range/Doppler Algorithm 2
SRC not shown, as it can be applied in different
places
So named because the key operations of RCMC and
azimuth compression are performed in the range
time/azimuth frequency (i.e., Doppler) domain.
Sept. 11, 2007
12
Cumming ASAR 2007
13
Range/Doppler Algorithm 3
A one-dimensional interpolator can correct
multiple targets in the range/Doppler
domain. After the interpolation, the locus of
energy is corrected, but a phase error persists,
which can defocus the image. SRC was implemented
to correct the phase error.
Sept. 11, 2007
13
Cumming ASAR 2007
14
Range/Doppler Algorithm 4
When the squint is high or the aperture wide, SRC
is needed to ensure accurate focusing. SRC is
mainly a function of azimuth frequency, then
range frequency.
Therefore, it is best to apply SRC in the 2-D
frequency domain. It is applied efficiently by a
phase multiply after an azimuth FFT is done.
Sept. 11, 2007
14
Cumming ASAR 2007
15
Range/Doppler Algorithm 5
Processing with SRC can be done by modifying the
range FM rate in rangcomp.
This is the approximate approach.
Sept. 11, 2007
15
Cumming ASAR 2007
16
Range/Doppler Algorithm 6
This slide shows how the target focus
deteriorates when the squint angle is increased.
With no SRC, the focus deteriorates quite fast
with squint. Approx SRC allows a fair amount of
squint, but the accurate SRC (not shown) allows a
large squint.
Sept. 11, 2007
16
Cumming ASAR 2007
17
Range/Doppler Algorithm 7

• This slide shows the RDA with
• No SRC
• Accurate SRC
• Approximate SRC

The Option 3 is very simple to apply, and is most
frequently used. Which one is needed depends on
the squint angle and the width of the
aperture. Option 3 can be applied when needed.
Sept. 11, 2007
17
Cumming ASAR 2007
18
Range/Doppler Algorithm 8
Advantages
Disadvantages

• Easy to understand program
• uses only 1-D operations
• Accommodates range-variant parameters easily
  (except SRC)
• Doppler centroid
• azimuth FM rate

• Interpolator introduces a small but controllable
  error
• Cannot handle very wide apertures or high squint
• unless SRC is applied as needed

19

• City of Vancouver
• RDA processing
• RADARSAT-1 FINE mode
• 8 m resolution

20

Convair-580 Data processed using RDA
21
Chirp Scaling Algorithm 1

• Want to improve accuracy by replacing the RCMC
  interpolator with a more accurate DSP operator
• The chirp scaling concept ? if a linear FM signal
  is shifted in frequency, it will be shifted in
  time after pulse compression
• same concept that causes a railway train to be
  imaged off the tracks ? the trains Doppler shift
  moves the train in azimuth after azimuth
  compression
• Chirp scaling is used to perform differential
  RCMC in the range/Doppler domain, to equalize the
  curvature over range

22
Chirp Scaling Algorithm 2
Illustrates how a chirp (Panel 1) multiplied by a
sine wave (Panel 2) causes a registration shift
in the compressed pulse (Panel 4). Because the
matched filter (not shown) has a zero centre
frequency, the compressed pulse is registered at
the zero frequency point of the scaled signal
(Panel 3) . This shift performs RCMC in the
range/Doppler domain, assuming range compression
is not done yet (because the amount of shift is
limited by the range oversampling, only
differential RCMC is done with chirp
scaling). The sine wave scaling is applied in
the range direction, with a different frequency
at each azimuth frequency. Note that the
resulting shift is constant with range.
23
Chirp Scaling Algorithm 3
But the shift needs to be varied with range (as
well as with azimuth frequency). As the change
with range is small, this can be achieved by
making the frequency of the scaling function
change slowly with range. A linear FM scaling
function makes the range shift vary linear with
range (called linear chirp scaling). This linear
form is adequate for most satellite SAR
parameters.
24
Chirp Scaling Algorithm 4
All operations are FFTs and phase multiplies
Chirp scaling is applied in the RD domain
The 2-D freq domain can be used to perform the
remaining RCMC without an interpolator. Range
compression, SRC also done. SRC is accurate as
it is range and az freq dependent.
Subsequent ops are the same as the RDA
25
SRTM Image Processed by CSA
26
Chirp Scaling Algorithm 5
Advantages
Disadvantages

• More accurate RCMC
• no interpolator
• The 2-D frequency domain is available to apply a
  more accurate form of SRC
• better phase accuracy
• Can handle slightly wider apertures and squint
  angles

• Requires 2-D processing
• Additional complexity if the Doppler centroid
  changes quickly with range
• Assumes SRC is independent of range
• Inefficient if range compression already done

For many satellite SAR parameters, the advantages
over the RDA tend to outweigh the disadvantages,
but RDA preferred for zero Doppler cases.
27
Omega-K Algorithm 1

• If the range equation is hyperbolic (straight
  line sensor motion), and the effective radar
  velocity is independent of range, an exact
  solution can be obtained
• this assumption is excellent for airborne radars
  (after mocomp) and quite good for satellite SARs
• Engineers at the Polytechnic of Milano discovered
  this, adapting an algorithm known as Stolt
  interpolation from the seismic field
• All processing is done in the 2-D frequency domain

28
Omega-K Algorithm 2
The Stolt interpolation consists of scaling and
shifting operations, as illustrated by a point
target simulation. Col 1 phase plot in 2-D
FD Col 2 range slices in 2-D FD Col 3 data
after range IFFT (2-D energy 1-D azimuth
slice) The simulation starts with data after the
bulk compression, and shows the effects of the
scaling and shifting separately. SRC is done
exactly during the bulk compression the Stolt
interpolation effectively makes the signal
stationary in range and azimuth.
Sept. 11, 2007
28
Cumming ASAR 2007
29
Omega-K Algorithm 3

• SIVAM X-band airborne radar built by MDA
• Spotlight operation 1.8 m resolution

30
Omega-K Algorithm 4
Advantages
Disadvantages

• Most accurate algorithm if constant-velocity
  assumption valid (SRC exact)
• Accuracy not affected by wide aperture and high
  squint
• Fairly easy to program if Doppler centroid does
  not vary much with range

• An interpolation is needed in the 2-D FD
• Cannot handle large changes of Doppler frequency
  with range efficiently
• need range block processing

31
SPECAN Algorithm

• When the signal is a chirp, it can be deramped
  using a phase multiply, making it a sine wave
• a single FFT then focuses the data
• Single-look version was developed many years ago
• multilook version developed in 1979 under an
  ESTEC contract for on-board processing
• Has the most efficient computing, but the worst
  image quality, notably phase properties
• particularly suited to burst-mode data, such as
  ScanSAR

32

• RADARSAT-1
• ScanSAR Narrow
• 300 km wide
• SPECAN processing
• Vancouver Island

33
Doppler Centroid Estimation

• Considered a mature technology
• yet processing mistakes are still made
• Recommend a new approach
• take a global view
• use spatial diversity over a wide area
• use quality checks and filtering to eliminate bad
  regions
• use a physical geometry model for overall
  estimate
• use phase increments for primary estimator
• Can be used for specialized applications
• such as ocean current estimation

34

• More details can be found in our SAR Processing
  book
• Published by
• Artech House
• January 2005
• Also to be published in Chinese, November 2007

34
35
Summary 1

• RDA
• a well-known algorithm for general use (30 years
  experience)
• except when the aperture is wide and the squint
  angle is high
• CSA
• a little more accurate than the RDA, as long as
  the Doppler parameters do not change too quickly
  with range
• WKA
• the most accurate as long as the flight path is
  linear and the velocity does not change with
  range (well suited to airborne applications)
• SPECAN
• needs the least memory and fewest DSP operations
• ideal for low-resolution imaging

Sept. 11, 2007
35
Cumming ASAR 2007
36
Summary 2

• Many choices of algorithm are available
• Best algorithm for each application depends on
• geometry and radar parameters
• accuracy required
• efficiency required
• Most commonly-used algorithm for general
  precision processing seems to be the RDA,
  followed by WKA and the CSA
• but a systems study needs to examine the pros and
  cons for each radar situation
• SPECAN is commonly used for ScanSAR and quicklook
  processing

Sept. 11, 2007
36
Cumming ASAR 2007
37
Current Future Research

• Many people consider SAR processing for
  conventional (remote sensing) SARs to be a mature
  subject, not requiring significant further
  development
• Current work seems to be directed towards
  advanced items such as
• bistatic SAR processing
• ground moving target detection (GMTI)
• information extraction, e.g.
• classification of polarimetric SAR data
• change detection through interferometry
• polarimetric interferometry

Sept. 11, 2007
37
Cumming ASAR 2007
38
Final Word

• Ever since working with ESTEC in 1979, on-board,
  real-time, digital SAR processing on a satellite
  has been a dream
• For various technical, financial and strategic
  reasons, it has not been implemented yet
• it is becoming feasible now, although its
  necessity is not justified
• As a compromise, I would love to build a
  real-time Doppler estimator
• would give the most accurate antenna steering
  possible, and
• be a technical showcase !

Sept. 11, 2007
38
Cumming ASAR 2007