IEEE - Aerospace and Electronic Systems - May 2022 - Tutorial XV - 26

A Survey on the Applications of Convolutional Neural Networks for Synthetic Aperture Radar: Recent Advances
Figure 7.
Proposed CNNby [88] for speckle removal in SAR images.
encoder-decoder CNN, for SAR image. despeckling. Li
et al. [104], incorporated the convolutional block attention
module (CBAM) [105] in their model to suppress the
speckle noise in SAR images. CBAM has channel attention
modules to concentrate on " what " is meaningful and
spatial attention modules to concentrate on " where " the
information is.
IMAGE SEGMENTATION
Image segmentation refers to the task of decomposing an
image into different segments (regions) based on their
characteristics. In semantic segmentation, meaningful
labels are associated to each pixel of the image. In the RS
community, the term " classification " is often preferred to
the term " semantic segmentation " [106]. Moreover, terrain
surface classification in SAR is very similar to the
task ofimage segmentation in computer vision [37].
FCN [107] is the most remarkable architecture used
for semantic segmentation. Instead of using one or more
fully connected layers at the end of networks, FCNs
exploit deconvolution layers. Deconvolution, in the context
of CNN, means upsampling, therefore, for the sake of
clarity the term " transposed convolution " is often preferred.
Henry et al. [108] employed three CNN architectures
of FCN-8s [107], DeepLabv3 [109], and Deep
Residual U-Net [110] for road segmentation in SAR
satellite images. The result of road segmentation with
FCN-8 in [108] is illustrated in Figure 8. Roads, as they
are seen as narrow objects in SAR images, are likely to be
immersed in heavy clutter. Pixel-wise labels (1) are
assigned to roads and (0) to the background. Mohammadimanesh
et al. [111] proposed a new FCN architecture for
semantic segmentation of polarimetric SAR imagery,
especially for the classification of wetland complexes.
Using inception modules [30] and skip connections in
their suggested framework, input data is fed into a stack of
convolutional filters (encoder) to extract high-level
abstract features and a stack of transposed convolutional
filters (decoder) to gradually up-sample the low-resolution
output to the spatial resolution of the original input image.
Inception modules [30] make parallel use of different kernel
size (1 1, 3 3, and 5 5) of a feature map to
improve the adaptability of the network to different scales.
U-Net [112] is an extension of the FCN and the letter
" U " refers to symmetric encoder-decoder architecture.
U-Net was initially proposed for biomedical image segmentation.
DeepLabV3 segmentation model is also an
encoder-decoder architecture that utilizes the dilated convolutions.
Krestenitis et al. [113] addressed the applications
ofU-Net and DeepLabV3 for the segmentation of oil
spills in satellite SAR images. A pixel with an oil label in
natural image is somehow uncommon since large oil spills
do not frequently occur. Cantorna et al. [114] have also
discussed the problem of oil spills in satellite SAR images.
Figure 8.
Road segmentation in [106]. (a) SAR image. (b) FCN-8 segmentation result.
26
IEEE A&E SYSTEMS MAGAZINE
MAY 2022
IEEE - Aerospace and Electronic Systems - May 2022 - Tutorial XV

Table of Contents for the Digital Edition of IEEE - Aerospace and Electronic Systems - May 2022 - Tutorial XV
