IEEE Geoscience and Remote Sensing Magazine - March 2023 - 12

complex-valued (CV) SAR image is first transformed to
the azimuth spectral domain by a 1D Fourier transform.
Then, the full Doppler spectrum is equally split into three
intervals, named subapertures or sublooks, each containing
one third of the range of the azimuth angles. Finally, the
three subapertures are transformed back to the time domain
using an inverse Fourier transform, coded as the red
(R), green (G), and blue (B) channels, respectively. The R,
G, and B targets respond mainly on the first, second, and
third sublooks, respectively, while gray targets indicate
that they respond equivalently in different sublooks. The
pseudocolored image well demonstrates the particular behavior
of some targets. Given precise knowledge of the parameters
related to Doppler variations (e.g., the orbit, azimuth
steering rate, radiation pattern, and incidence angle),
the physical layer can generate sublook data deterministically,
and there is no need to design a neural network that
should learn to create sublooks from various types of SAR
training data.
Sensor characteristics, such as polarization, interferometry,
and tomography, construct the physical layer as well.
Figure 3(b) presents a Pauli pseudo-RGB image, where the
R, G, and B channels are formed with |HH−VV|2, |HV|2 and
|HH+VV|2, respectively, indicating the polarimetric relation.
Several physical layers can be stacked to represent the rich
Full Aperture
p
Subaperture
Azimuth Direction
Subap
Azimuth
physics of the SAR sensor and platform. Early in the literature
[18], the diversity in the polarimetric features with the
azimuthal look angle was exploited. Thus, the moving platform
and polarimetric sensor are both characterized. Similarly,
the stacked physical layers can represent polarimetric
and interferometric properties of polarimetric SAR interferometry
data or any other combinations.
IMAGING SYSTEM
The second physical layer we suppose delineates the physics
behind SAR image formation with an imaging system. The
selected exemplars are illustrated in Figure 4.
A pulse-based radar or a frequency-modulated continu......
Subaperture
Features
R: Sublook 1
G: Sublook 2
B: Sublook 3
Target
* Moving Platform
Transmitter
Horizontal
Vertical
Receiver
Horizontal
Vertical
HH
VH
HV HH HV
VV VH VV VH
HV HH HV HH
VV VH VV VH
or
Polarimetric Features
R: |HH - VV|2
G: 2|HV|2
B: |HH + VV|2
or
......
Physical Params:
Doppler Centroid, Beamwidth,
Entropy, Height, Incidence
Angle, etc.
ous-wave (FMCW) radar is usually used in an SAR system,
where a range profile is obtained for each transmitted/received
waveform, either by range compression in the case
of a pulse-based radar or by applying a Fourier transform to
the beat signal in the case of an FMCW radar [19]. By coherent
processing of the range profiles, the azimuth-focusing
process outputs an SAR image representing a 2D complex
reflectivity map of the illuminated area. SAR processing, taking
a simple point target as an example, aims to collect the
dispersed signal energy in the range and azimuth into a single
pixel. Many traditional imaging algorithms are in terms
of a Fourier synthesis framework [20]; as such, the Fourier
transform provides a specific physical
meaning for the SAR image. This kind
of physical layer helps the AI model
to better depict the target scattering
beyond the " image " domain.
Figure 4(a) first shows a simple
* Polarimetric Sensor
(a)
......
(b)
FIGURE 3. Physical layer 1: the sensor and platform. (a) The moving platform creates Doppler
variations and synthesizes a large virtual aperture; PolSAR transmits and receives diverse
polarized waves, and SAR polarimetric characteristics are depicted. (b) Based on the physics
behind the platform and sensor, the physical layer produces SAR-specific representations,
such as a subaperture synthesis image and polarimetric feature, with specified physical
parameters. B, blue; G, green; H, horizontal; R, red; V, vertical.
12
time-frequency analysis (TFA) of
the target with a short-time Fourier
transform [21], [22], characterizing
the backscattering intensity variations
in the 2D range and azimuth
frequency domain. Four kinds of
backscattering behaviors observed
in SAR were defined in the literature
[23], related to different objects
shown in Figure 4. In the high-resolution
case (a wide bandwidth chirp
signal and broad angular aperture),
the complex amplitude of a target is
frequency and aspect dependent [17].
Thus, the image formation can be extended
to four dimensions (called a
hyperimage) with the wavelet transform,
providing a concise, physically
relevant description of target scattering.
This frequency and angular
energy response pattern has proved
useful for discriminating different
scatterers, offering valuable prior information
to the AI model, depicted
in Figure 4(b).
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE MARCH 2023
Complex
Image

IEEE Geoscience and Remote Sensing Magazine - March 2023

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - March 2023