IEEE Geoscience and Remote Sensing Magazine - March 2023 - 20

and orientated buildings with a larger z The polarization
2 .
orientation angles of ortho buildings are obviously smaller
than those of oriented buildings. Ortho built-up areas
mainly depict double scattering (DS) and mixed scattering
where the DS dominates. The oriented buildings are with
volume scattering (VS). Figure 12(b) shows the limited robustness
of recognition performance, as the angle of test
data varies when training with a small range of angles. A
trustworthy model is expected to perceive SAR scattering
variations with a variety of physical parameters and be perturbation
tolerant.
TRUSTWORTHY MODELING WITH
UNCERTAINTY QUANTIFICATION
The development of Bayesian deep learning [59] has caught
much attention in recent years, where the posterior distribution
over parameters is obtained instead of the point estimation.
A crucial property of the Bayesian method is its
ability to quantify uncertainty to the benefit of constructing
a trustworthy model.
In the case of Figure 12(b), the performance of DNNs
drops dramatically when testing SAR targets of very different
orientation angles with training data. The model is overconfident
about some uncertain data that cannot be perceived
by the frequentist method. The Bayesian DNN, instead, is
able to calibrate the output score and measure the uncertainty
of the prediction. Some recent studies applied Bayesian
deep learning for SAR sea ice segmentation [60], [61], [62] as
well as target discrimination [63]. The generated uncertainty
map can serve as a guideline for experts in annotation and
improve trust between users and the model. Some approximation
strategies of the Bayesian DNN, such as Monte Carlo
(MC) Dropout [64] and Deep ensembles [65], are promising
for different SAR applications.
We give an example of SAR ship detection for a demonstration.
The limited labeled training data and the interference
of complex scattering from the target itself or the
inshore background would strongly restrict the detection
performance. The ship detection result on some selected
SAR images from the AIR-SARShip-1.0 dataset [66], obtained
by the fully convolutional one-stage object detection
(FCOS) detection algorithm [67], are shown in the
first row of Figure 13. Compared with the ground truth annotation
in the third row, the detection result yields many
false alarms. It is crucial to estimate the model uncertainty,
which is basically brought by inadequate training data, to
evaluate the reliability of the SAR ship detection model and
provide more trustworthy predictions.
When we apply the MC Dropout training strategy to
approximate the Bayesian inference [64], it captures the
uncertainty from the existing deep model for SAR ship
detection. The results with high uncertainty and very low
classification scores are discarded, with only the trustworthy
predictions preserved. The results are shown in
the second row of Figure 13, where the false alarms are
evidently reduced. In the fourth and fifth SAR images, the
FIGURE 13. The SAR ship detection result on the AIR-SARShip-1.0 dataset [66], obtained by the detection deep learning algorithm FCOS [67].
Many false alarms appear in the detection result due to the limited training data and the interference of complex scattering. The prediction
uncertainty is estimated by MC Dropout [64] and the uncertain results are discarded to achieve a better performance.
20
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE MARCH 2023
Ground Truth
FCOS-MC Dropout
FCOS Result

IEEE Geoscience and Remote Sensing Magazine - March 2023

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