IEEE Geoscience and Remote Sensing Magazine - March 2023 - 15

or even single-polarized (single-pol) SAR image. Thus, the
physical layer interacting with the DNN takes effect.
TARGET CHARACTER IDENTIFICATION
Some researchers have analyzed how the energy response
pattern in the frequency dimensions of a target vary in
SAR images and discussed nonstationary targets [18], [36].
Spigai et al. [23] pointed out four canonical targets with a
rough definition, shown in Figure 4(a). However, this remains
unknown for many complicated scenes and objects.
Figure 7 shows our related work of using the physical layer
and DNN to compensate for imperfect knowledge. The first
is the unsupervised hierarchical deep embedding clustering
based on the TFA (HDEC-TFA) [29], which was proposed
to automatically characterize the radar spectrogram
(or the subband scattering pattern defined in [29]) basically
in the urban area, discovering more various scattering patterns
than the four specific classes defined in [23]. It offered
a new perspective to describe the physical properties
of single-pol SAR. Furthermore, we used two stacked physical
layers to obtain the polarimetric and time-frequency
patterns and analyzed them with a DNN in [37].
Figure 8 demonstrates the result compared with the
polarimetric physical model. The SOLEIL synchrotron in
France, shown as the round building in the Google Earth
remote sensing image, is surrounded by three different
shapes of buildings. The HDEC-TFA method can capture
the special characteristics of the architectures even in a
single channel SAR image, as much as the physical model-based
method GD-Wishart [38] on quad-polarization
SAR. Some other human-made target examples characterized
by the time-frequency model with neural networks
are given in [39]. Our experiments in [29] demonstrated
that the trained model varies with different imaging conditions
since the subband scattering pattern is influenced
by several imaging parameters, which should be taken
into consideration when transferring the AI model to
other situations.
POLARIMETRIC PARAMETER EXTRACTION
By transmitting and receiving waves that are both horizontally
and vertically polarized, the full-polarized (full-pol)
SAR image captures abundant physical characteristics of
the imaged objects that can lead to various physical parameters.
In contrast, single-pol and dual-pol SAR data are less
informative for physical feature extraction. If only one polarization
channel is obtained, one cannot derive the other
polarization channels in principle. Once the objects are
known, i.e., once the characteristics of targets, such as geometry,
surface roughness, etc. are identified, deep learning
can be employed to transfer the knowledge learned from
physical models to reconstruct the physical parameters of
objects. As shown in Figure 7(b), Zhao et al. [30] proposed a
complex CNN model to learn physical parameters (entropy
H and angle a) with transfer learning from single-pol and
dual-pol SAR data, supervised by features obtained with a
physical layer. Some similar studies include but are not limited
to [40] and [41]. Song et al. [40] addressed the " radar
image colorization " issue to reconstruct the polarimetric
covariance matrix with a designed DNN, where the supervision
was also generated with a physical layer.
When training a data-driven DNN, some physical consistencies
may not be guaranteed. The authors pointed out
that the reconstructed covariance matrix may not be positive
semidefinite [40], and they proposed an algorithm to
correct it. In this case, the additional physical layer embedded
prior constraint acts as postprocessing to revise the
physically inconsistent result of DNNs. Furthermore, this
type of physical layer is suggested to provide feedback during
training, as demonstrated in the red part of Figure 7(c).
The feedback of the physical layer aims to prevent the model
from learning the physical inconsistency.
Intensity Image
Deep
Network 1
ComplexValued
SAR
Image
Deep
TFA
2D Projection in
Frequency Domain
Physical Layer
Physical Layer
FIGURE 6. Our recent work, DSN [28], for SAR image classification can be regarded as a typical example of inserting the physical layers into
a deep model.
MARCH 2023 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
15
Network 2
Frequency
Features
Target Position
Constraint
Spatial
Features
Semantic Labels
Feature
Fusion
Deep
Network 3

IEEE Geoscience and Remote Sensing Magazine - March 2023

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