IEEE Geoscience and Remote Sensing Magazine - March 2023 - 10

SCIENTIFIC EXPLAINABILITY
ENSURES THE PHYSICAL
CONSISTENCY OF AI OUTPUT
AS WELL AS THE LEARNING
OF TRUSTWORTHY RESULTS
WITH PHYSICAL MEANING.
spanning from the smooth vertical bridge facets facing the
sensor and the water surface. In addition, the triple-bounce
reflection and, maybe, some five-path scattering would happen
between the horizontal plane of the bridge and water
surface. Thus, an SAR image
implies the causality of multipath
scattering phenomena
and object characteristics.
This positions the load of SAR
image understanding and the
outmost challenge of data science
as a new and particular
paradigm of AI.
So far, some researchers
have discussed the paradigm
that attempts to bring scientific
knowledge and data science models together, applying
them to a broad range of research themes, such as partial
differential equation solving [13] and Earth sciences
[14], [15], [16]. Particularly for the SAR community, however,
this topic has rarely been systematically analyzed
and illustrated. Thus, we aim to propose a hybrid modeling
paradigm for intelligent SAR image understanding,
where deep learning is integrated and interacts with SAR
physical models and principles to achieve explainability,
physics awareness, and trustworthiness.
XAI is a broad concept. A scientific understanding of
explainability is the capacity to clarify the results in the
context of domain knowledge. The algorithms still remain
a black box. A different approach is algorithmic explainability.
This is constructed such that the results of the used
model can be described algorithmically. To obtain a higher
degree of explainability, we aim for the synergy of the
paradigms: algorithmic and scientific explainability. Algorithmic
explainability lies in the guarantee of transparency
to understand how the machine learning (ML) algorithm
works by the participation of SAR physical models and
principles. Scientific explainability ensures the physical
consistency of AI output as well as the learning of trustworthy
results with physical meaning.
To ground this, we first lay out a representation of the
SAR physical layer in the context of SAR domain knowledge,
as presented in the " SAR Physical Layers " section.
Further, we describe how to integrate them and have them
interact with popular neural networks to build a hybrid
and translucent model for SAR applications using illustrative
examples, demonstrated in the " Hybrid Modeling With
SAR Physical Layers " section. The perspective of trustworthy
models and supplementary explanations for the SAR
community are discussed in the " Trustworthy Modeling "
and " Supplementary Explanations " sections. The conclusion
and perspectives are given in the final section.
Chirp
Signal
Azimuth
Ground Range
SAR PHYSICAL LAYERS
Other than the neural network layers
equipped with a number of
learnable parameters, SAR physical
layers are ones embedded with
physical knowledge of SAR, well
established, interpretable, and supported
by domain theories. The concept
of a " physical layer " apart from
a " neural network layer " arose in the
literature [16] to make the model
more physically realistic. As shown
in Figure 1, three SAR physical layers
are highlighted specifically for
the SAR applications in this article,
i.e., 1) the sensor and platform, referring
to antenna characteristics
and moving satellite/aircraft; 2) the
imaging system, figuring image formation
with the focusing process;
and 3) the scattering signature, reflecting
the physical properties of
terrain and objects.
FIGURE 1. A simple illustration of how SAR images the world (stripmap mode). The SAR community
is facing the big data challenge but with limited ground truth. Meanwhile, knowledge
of SAR is equally important. This is also the motivation of the physical layers in this article.
10
SENSOR AND PLATFORM
Figure 3 shows the physical layer of
the sensor and platform that indicates
the physics behind the SAR acquisition
principle, such as aperture
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE MARCH 2023
Data
Echo
Synthetic Aperture
Doppler Variation
Slant Range
Sensor and Platform
Imaging System
Echo Processing
Azimuth Focusing
Terrain and Object
Dielectric Property
Surface/Geometry
Knowledge
Wavelength/Bandwidth
Polarization/Incidence Angle

IEEE Geoscience and Remote Sensing Magazine - March 2023

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