IEEE Geoscience and Remote Sensing Magazine - March 2023 - 19

be young ice. The explanation will motivate us to improve
the algorithm by, for instance, relaxing the physical constraints
between the two classes.
Similarly, a very recent work [57] was proposed for
SAR target recognition inspired by our work [43]. The
authors proposed a CNN under the guidance of an SAR
target physical model, attributed scattering center, to extract
significant target features, which were successively
injected into the classification network for more robust
and interpretable results.
TRUSTWORTHY MODELING
WHY TRUSTWORTHY MODELING IS NEEDED
The results obtained by applying AI techniques in SAR
processing can be validated using in situ measurements of
known targets. For example, a common approach for the
calibration/validation of SAR data is to employ an electronic
target (transponder) that receives a signal, applies a
controllable time delay, and transmits the delayed signal toward
the receiver of the bistatic/monostatic system. Such a
target can be used to validate results related to deformation
measurements (e.g., atmospheric corrections) or polarimetric
analysis.
Some real-world applications of SAR require the measurement
of reliability and uncertainty. One example
is the sea ice classification in the untraversed polar regions
where the ice is always promptly changeable, which
would result in difficulty for annotation and the lack of
reference data. In this case, the predictions in unknown
polar areas obtained by an AI model need to be trusted
by humans. Strong robustness and a plausible degree of
confidence in ML system prediction are equally as important
as its accuracy.
Figure 12(a) indicates that building orientations have
a great impact on polarization orientation angles [58]
and scattering mechanisms [38]. The zoomed-in region
mainly contains ortho buildings where 1
z is close to 0°
PoISAR Image
Built-Up Area
Scattering Mechanism
SS
DS
MS
VS
φ2
φ1
Range
Polarization Orientation Angle Map
45
30
15
Building Orientations
(a)
-30
-60
Known Angle
+30°
+60°
+90°
-120
-90
-150
±180
-150°
±180°
+150°
+120°
(b)
FIGURE 12. A trustworthy model should perceive the SAR scattering variations with a variety of physical parameters and be perturbation tolerant.
(a) Differently oriented buildings reflect various polarization orientation angles and scattering mechanisms in a PolSAR image. (b) SAR
targets vary violently with orientation angles. When training with a small range of angles, limited robustness of recognition performance is
observed as the angle of the test data varies. DS, double scattering; MS, mixed scattering; SS, single scattering; VS, volume scattering.
MARCH 2023 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
19
Test Performance of Different Angles
-150
150
-30
30
60
90
120
1
6
ZSU234
ZIL131
D7
BRDM2
2S1
BTR60
T72
BTR70
BMP2
-15
-30
-45
Azimuth

IEEE Geoscience and Remote Sensing Magazine - March 2023

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