IEEE Geoscience and Remote Sensing Magazine - December 2021 - 152

the input TerraSAR-X image of Berlin and Figure 5(b) is
the predicted building mask. Another major contribution
made in that paper addresses the lack of training data by
introducing an automatic annotation technique, which
annotates the SAR tomography data using Open Street
Map (OSM) data.
As an extension of the preceding work, Sun et al. [123]
tackled the problem of individual building segmentation in
large-scale urban areas. They proposed a conditional geographic
information system (GIS)-aware network (CG-Net)
that learns multilevel visual features and employs building
footprint data to normalize these features for predicting
building masks. Thanks to the novel network architecture
and the large number of building labels automatically generated
from accurate digital elevation model (DEM) and
GIS building footprints, this network achieves an F1 score
of 75.08% for individual building segmentation. With
the predicted building masks, large-scale level-of-detail 1
building models are reconstructed, with a mean height error
of 2.39 m.
Overall, deep learning has shown very good performance
in existing SAR object detection tasks. There are two
main challenges that the algorithm designer needs to keep
in mind when tackling any SAR object detection tasks. The
first relates to identifying characteristics of SAR imagery,
such as imaging geometry, the size of objects, and speckle
noise. The second and bigger difficulty concerns the lack of
good quality standardized data sets. As we observed, the
most popular data set, MSTAR, is too easy for deep nets,
and, for ship detection, the majority of authors create their
data sets, which makes it very hard to judge the quality of
the proposed algorithms and even harder to compare different
algorithms. An example of a difficult-to-create data
set can be found in global building detection. The shape,
size, and style of buildings change quite drastically from
region to region, and so a good data set for this purpose re -
quires training examples taken from buildings from around
the world, a task that requires significant effort to produce
high-quality annotations of enough structures that deep
nets can learn from them.
PARAMETER INVERSION
Parameter inversion from SAR images is a challenging field in
SAR applications. As one important branch, ice concentration
estimation is now attracting great attention due to its
importance to ice monitoring and climate research [124].
Since there are complex interactions between SAR signals
and sea ice [125], empirical algorithms face difficulties
with interpreting SAR images for accurate ice concentration
estimation. Wang et al. [8] resorted to a CNN for
generating ice concentration maps from dual-polarized
SAR images. Their method takes image patches of intensity-scaled
dual-band SAR images as inputs and directly
outputs ice concentrations. In [126] and [127], Wang et al.
employed various CNN models to estimate ice concentrations
from SAR images during the melt season. Labels
152
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE DECEMBER 2021
Gabor
Filter
TPLBP
LSTM LSTM LSTM
Softmax
T72 : 0.98
BMP2 : 0.01
BRDM2: 0.01
Gabor
Filter
TPLBP
LSTM LSTM LSTM
Softmax
T72 : 0.92
BMP2 : 0.02
BRDM2: 0.06
Gabor
Filter
Original Images
Multiaspect
Image Sequence
Construction
Multiaspect
Sample
TPLBP
LSTM LSTM LSTM
Feature Detection
Dimensionality
Reduction
Multiaspect Feature Learning
Softmax
Classification
T72 : 0.9
BMP2 : 0.03
BRDM2: 0.07
FIGURE 4. A flowchart of the multiaspect-aware bidirectional approach for SAR automatic target recognition proposed in [95]. TPLBP: three-patch local binary pattern.

IEEE Geoscience and Remote Sensing Magazine - December 2021

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - December 2021