IEEE Geoscience and Remote Sensing Magazine - March 2023 - 13

SCATTERING SIGNATURES OF OBJECTS
Third, we introduce the physical layer regarding the scattering
signatures of objects, in which the causality of target
characteristics and scattering behaviors is involved.
For optical images, what you see is what you receive; that
is, the objects depicted on the optical image are in accord
with human cognition. Targets in SAR images are reflected
by scattering characteristics, yet they include a wealth of
physical information that the human eye cannot immediately
identify. Figure 5(a) shows an example of two typical
SAR targets: a bridge and a building. The scattering phenomenon
that shows several parallel lines over the river can
be interpreted as single, double, and multiple scattering of
the bridge based on the domain knowledge. The building,
with scattering signatures of the layover, shadow, and single
and secondary scattering in the high-resolution SAR image,
can also be reflected as only the layover and shadow [24],
depending on the building orientation and shape. Similar
research by Ferro et al. [26] investigated the relationship
between double bounce and the orientation of buildings
in VHR SAR images. Figure 5(b) demonstrates the relations
between the scattering mechanism of the H/a plane and
the semantics of land cover and land use classes [25]. Likewise,
one can deduce the scattering center position and the
specific shape of a distributed target from an SAR image by
applying some scattering models [27].
The conventional data-driven convolutional neural network
(CNN) can capture the image contents as we " see " in
the SAR image, whereas it is not equipped with the ability
to " interpret " the scattering phenomenon, as we discussed
before. This indicates the knowledge gap between SAR scattering
signatures and human vision cognition. The physical
layer delivering semantic understanding behind the SAR
scattering signature permits a more thorough interpretation
of the SAR image. As shown in Figure 5(c), the physical layer
defines the association between the scattering characteristics
of an SAR image and the object's qualities, such as its
shape, structure, or semantics. It can be written as an objective
function or a regularization term that constrains the
training of neural networks. This will improve the intelligence
of the AI model such that it can master some causality
between scattering signatures and the object nature.
HYBRID MODELING WITH SAR PHYSICAL LAYERS
The integration and interaction of neural network layers
and physical layers construct the hybrid modeling for SAR
image interpretation. In view of algorithmic explainability,
the explainable physical models and domain knowledge
improve the transparency. For scientific explainability,
hybrid modeling ensures that the physical meaning of the
output in physical layers and the prediction can maintain
physical consistency. In this section, we demonstrate several
hybrid modeling approaches with the SAR physical layer
to achieve explainability and physics awareness.
INSERT FOR SUBSTITUTION
The introduced physical layer can be inserted in a deep neural
network (DNN) for substitution, extracting explainable
and meaningful features, either as the input of a DNN or
fused with DNN features in intermediate layers. A common
way is to insert a physical layer into the input layer to obtain
the polarimetric features for PolSAR image classification,
including the elements of coherency matrix, Pauli decomposition
features, etc. [31], [32]. Similarly, the subaperture
images are generated as the input for target detection [33].
The other usage of the physical layer is for feature fusion,
where the features obtained by a well-established physical
model and DNNs are combined [34], [35].
Our recent work, a deep learning framework named
Deep SAR-Net (DSN) [28], addressed both aspects that
Optical
Image
Stable Targets
Unstable Targets
Azimuth Variant Targets
Range Variant Targets
(a)
SLC
Data
Data Transform:
Wavelet, Fourier,
Wigner-Ville
Physical Parameters:
Frequency, Angle,
Wavenumber, etc.
SAR
Image
Radar
Spectrogram
4D Hyperimage
or
...
Frequency-Anguler
Energy Response
or
(b)
FIGURE 4. Physical layer 2: image formation. (a) Targets are characterized by sliding bandpass filtering in the Fourier domain.
(b) On the basis of the image formation principle and target scattering model, the physical layer generates the rich target description
with physical meaning.
MARCH 2023 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
13
...
Target Description With Physical Meaning

IEEE Geoscience and Remote Sensing Magazine - March 2023

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