IEEE Geoscience and Remote Sensing Magazine - March 2023 - 14

insert the physical layer into
the input and the intermediate
position of the deep
model. As shown in Figure 6,
DSN was proposed for classifying
SAR images with complex
values. Instead of the
entire data-driven method,
i.e., CV-CNN, the designed
DSN encompasses three shallow
neural network modules
and two physical layers. The
first physical
layer generates
the high-dimensional radar
spectrogram based on a TFA.
The second one handles the
features of the 2D projection
along the frequency axis [22]
to maintain the location constraint,
making it possible to
be fused with spatial features
from the intensity image.
DSN outperformed CV-CNN,
especially with limited labeled
training data, and had
a remarkable performance in
discriminating the humanmade
target scenes compared
with the traditional CNN. It
demonstrates that the Fourier
process on single-look complex
SAR image-embedded
knowledge, like synthesizing
the antenna, well characterizes
the physical property of
the SAR target, and the usages
of the physical layer cut down
unnecessary parameters in
neural network layers to improve
the model performance
with a limited ground truth.
COMPENSATION
FOR IMPERFECT
KNOWLEDGE WITH
FEEDBACK
In the condition of unknown/
inconclusive physical models
or incomplete knowledge, it
is difficult to extract perfect
physical parameters or physical
scattering
characteristics
of SAR via model-based methods-for
instance, obtaining
the polarimetric features from
a dual-polarized (dual-pol)
14
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE MARCH 2023
α (q)
Range Direction
Scattering
Single
Double
Scattering
Layover
Range Direction
Secondary Scattering
Shadow
Multiple
Scattering
Bridge
Single Scattering
10
20
30
40
50
60
70
80
90
Building
Scattering
Phenomenon
(a)
Scattering Center,
Scattering Mechanism, etc.
(c)
FIGURE 5. Physical layer 3: scattering signatures of objects. (a) The Golden Gate Bridge, revealing multipath scattering characteristics in a Gaofen-3 quad-polarization SAR image, and a typical single
building representing different scattering regions in a high-resolution (1-m) SAR image [24]. (b) The scattering mechanisms indicated by the H/a plane for full-polarized SAR data point out the land
use and land cover classes [25]. (c) The physical layer describes the relationship and reasoning between the scattering characteristics seen in the SAR image and the object's features, such as its
shape, structure, or even semantics.
Relation and
Reasoning
Object
Property
0.5
Entropy (H)
(b)
0.9 1
Scattering Mechanisms
Low-Entropy Scattering
Low-Entropy Dipole
Low-Entropy Surface
Medium-Entropy Surface Scattering
Medium-Entropy Dipole Scattering
Medium-Entropy Multiple Scattering
High-Entropy Double Bounce
High-Entropy Multiple Scattering
Forest
Water Landslide Farmland
Land Cover Types
Snow
...
Shape, Orientation,
Structure, Position, etc.
Azimuth Direction
Azimuth Direction

IEEE Geoscience and Remote Sensing Magazine - March 2023

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