IEEE Geoscience and Remote Sensing Magazine - March 2023 - 21

localization uncertainty of two large ships visualized with
circles around the corner of the predicted bounding box is
relatively high. Intuitively, we can infer that the reason for
the weak capability of the trained model in detecting such
kinds of targets is probably the lack of large-size ships in the
training set. The feedback from the uncertainty estimation
should further inspire follow-up studies to improve the algorithm
and build more trustworthy models.
SUPPLEMENTARY EXPLANATIONS
Beyond the hybrid and trustworthy modeling, extra explanations
and other interpretable models are required as
well to assist with developing a more transparent AI model
for SAR. The XAI techniques, such as gradient-based, attention-based,
and occlusion-based explanation methods,
are helpful to demonstrate the effectiveness of integrating
physical layers to achieve explainability.
The transparent ML models, such as linear regression,
decision trees, and Bayesian models, are interpretable
[68]. The algorithm itself provides explanations; for example,
latent Dirichlet allocation (LDA) builds a threelevel
hierarchical Bayesian model to describe the underlying
relationship among document-topic-word. That is,
the document can be explained with a set of topics, where
each topic, in turn, is represented by a distribution over
words. Karmakar et al. [69] used the LDA model for SAR
image data mining to generate the topic compositions
and group them into semantic classes, which were fused
with domain knowledge obtained by active learning from
experts. The transparent model can also be integrated in
a deep learning framework to approach explainability.
Huang et al. [42], [43] applied the LDA model to generate
the physical attributes representation as the guided physics
signals rather than directly using the physical scattering
characteristic labels to train the physics-guided network.
That is because the learned physics-aware features
are expected to benefit semantic label prediction, but the
semantic gap actually exists between the physical scattering
characteristics and the semantic annotation. Consequently,
the LDA model enables the guided signals to
gain the abstract semantics and be explained with physical
scattering properties.
The other purpose for approaching explainability lies in
the applications of transfer learning. The manual annotation
in the SAR domain is difficult, and the deficiency of
labeled data basically restricts the development of datadriven
methods. Facing a wide variety of launched SAR platforms
with various frequency bands and resolutions as well
as other multispectral, hyperspectral, and optical remote
sensing sensors, it is of vital importance for elucidating the
transferability of ML models among inhomogeneous data.
BarredoArrieta et al. [68] indicated that transferability is
one of the goals toward reaching explainability. Although
many researchers have explored different deep transfer
learning methods in the SAR domain [46], [70], [71], the
inner transfer mechanisms of the deep learning model
MARCH 2023 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
still need an explanation of insight. An insufficient understanding
of the model may mislead the user toward inappropriate
design of the algorithm and fatal consequences,
i.e., negative transfer. Based on SAR target recognition, we
proposed analyzing the transferability of features in the
DNN, which contributed to explaining what, where, and
how to transfer more effectively for SAR images [72]. The
inspiration also motivates the follow-up studies, including
the SAR-specific pretrained model [73], application in the
detection task [74], and the interpretability analysis of the
deep learning model in a radar image [75].
CONCLUSION AND PERSPECTIVES
In this article, we propose an AI paradigm shift for SAR
applications that is explainable, physics aware, and trustworthy.
To ground this, SAR physical layers embedded
with domain knowledge are
introduced, which are supposed
to be integrated and
interact with neural networks
for hybrid modeling.
Some illustrative examples
are provided to demonstrate
the general patterns, showing
algorithmic and scientific explainability.
In addition, we
emphasize the importance
and approaches of trustworthy
modeling with Bayesian
deep learning as well as illustrating
some other techniques,
such as the interpretable ML method, explainable
techniques, and model transferability, that would
assist with developing more transparent AI models for
SAR. In fact, this field belonging to interdisciplinary research
is still largely undeveloped. To our best knowledge,
such approaches have not been formulated in the
past years. So far, only some plain attempts have been
made. Significant questions and challenges remain, e.g.,
the feasible representation of the SAR physical layer, the
optimized form of physical constraint, and hybrid modeling
optimization.
Currently, there are several smart sensing techniques
in the SAR community that can be exploited as preprocessing
steps of data fed into DNNs, e.g., multiaperture focusing
in bistatic configurations [76], monostatic/bistatic
tomography, polarimetric decomposition, and deformation
time series. The outputs of these techniques can expose
features that probably cannot be directly extracted
by a DNN, especially when using a small training dataset.
The newly introduced AI paradigms can apply to the
broad class of coherent imaging systems. A few examples
can be enumerated: computer tomography, terahertz imaging,
echographs in medicine or industrial applications,
sonar or seismic observations in Earth sciences, or radiotelescope
data in astrophysics.
21
SIGNIFICANT QUESTIONS
AND CHALLENGES REMAIN,
E.G., THE FEASIBLE
REPRESENTATION OF THE
SAR PHYSICAL LAYER, THE
OPTIMIZED FORM OF
PHYSICAL CONSTRAINT,
AND HYBRID MODELING
OPTIMIZATION.
IEEE Geoscience and Remote Sensing Magazine - March 2023

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