IEEE Geoscience and Remote Sensing Magazine - March 2023 - 18

from [56], which indicated that the abstract topic mixture
on scattering properties and the high-level image
features have similar semantics. Thus, we built the relation
between the image semantics and SAR scattering
characteristics. A novel objective function was designed
to instruct self-supervised learning guided by physical
scattering mechanisms.
The advantages of this kind of learning paradigm lie in
two aspects. First, the training process takes all labeled and
unlabeled data so that the learned features generalize well
in the test set. Second, the guidance of physical information
leads to physics awareness of features learned by neural
networks; i.e., the DNN feature maintains physical consistency.
In a word, the prior physical knowledge is embedded
in the neural network. The experiments in [43] verified this
quantitatively and qualitatively.
Additionally, the outputs of the physically interpretable
deep model can be further explained, which, in turn,
inspires algorithm improvement. We illustrate with an
example of sea ice classification in a polar area [42]. The
physics-guided learning is driven by physical signals that
reflect the scattering properties of the SAR image. The
guided physical signals are visualized with t-distributed
stochastic neighbor embedding (t-sne) in Figure 11, where
different colors in Figure 11(a) represent semantic labels
of sea ice, and each color in Figure 11(b) indicates samples
with similar physical scattering properties. One characteristic
that can be seen is that young ice and water bodies
have extremely similar physical representations, which
would impede semantic discrimination. This can explain
the physics-guided learning result in [42] that about 23%
of test samples in the water bodies class were predicted to
Physical Layer
SAR
Complex
Physical Model
Physics
Information
Topic Model
x
Physical Layer
512
Physics-Guided Learning (Explainable and Learnable)
SAR
Amplitude
64
ConvBlk
ResBlk-1
64
ResBlk-2
128
ResBlk-3
256
BN + ReLU
fc 512 × Tk
Convolutional
3 × 3 - 512
Physics-Aware
Features
FIGURE 10. A physics-guided network was proposed [42], [43] where a novel deep learning paradigm and loss function were designed to
associate the SAR scattering characteristics with image semantics.
Physics Attributes
Semantic
Relation
Young Ice
Glacier
First-Year Ice
Water Bodies
Iceberg
Floating Ice
Old Ice
Phy. Scat 0
Phy. Scat 1
Phy. Scat 2
Phy. Scat 3
Phy. Scat 4
Phy. Scat 5
Phy. Scat 6
Phy. Scat 7
(a)
(b)
FIGURE 11. Visualization of physics-guided signals on test data by t-sne. (a) Different colors represent the semantic labels of sea ice. (b) The
physics-guided signals are grouped into eight clusters, where each color indicates samples with similar physical scattering properties. Phy. Scat:
physical scattering label.
18
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE MARCH 2023
IEEE Geoscience and Remote Sensing Magazine - March 2023

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