IEEE - Aerospace and Electronic Systems - May 2022 - Tutorial XV - 25

Oveis et al.
Figure 5.
Flevoland image. (a) Pseudo RGB. (b) Ground truth. (c) Classification result of [79]. (d) Color code for the ground truth.
methods [89] and [90], the MSE has been used as the loss
function. As an example of regression with CNN, the
result of ice concentration in [89] is illustrated in Figure 6.
Estimating sea ice concentration from SAR images has
also been addressed by Cooke et al. [91] where they used
passive microwave data for training their CNN. Scarpa
et al. [92] proposed a CNN to estimate normalized difference
vegetation index through fusion of optical and SAR
Sentinel data. They employed mean absolute error as the
loss function. Estimation of ship orientation (angle regression)
in SAR images, using a truncated VGG16 model as
backbone, is addressed by Wang et al. [93]. All the convolutions
in VGG16 have the kernel size of 3 3 and the
number 16 in VGG16 implies that it has only 16 trainable
layers. Smooth L1 is used in [93] as the loss function.
Song et al. [94] studied regression with CNN for estimation
of rough surface parameter from SAR images.
IMAGE DENOISING
Speckle, which is typically modeled as a multiplicative
noise, is caused by constructive and destructive interference
of the coherent returns scattered by small reflectors within
one resolution cell [95]. Speckle noise hampers SAR image
interpretability and many methods of speckle filtering have
been proposed for SAR images in the past three decades.
Wang et al. [95], [96] proposed a CNN architecture to
reduce the speckle noise. By appropriate zero padding and
avoiding any pooling layer, the output size ofthe last convolution
layer is equal to that of the input. Subsequently, the
input image, using a skip-connection, is divided by the output
(speckle) to produce despeckled image. As an example
of SAR image denoising with CNN, the model in [95] is
illustrated in Figure 7.
Instead of employing residual learning to predict
noise, Pan et al. [97] proposed another approach based on
the FFDNet [98] for SAR image denoising. Fast and flexible
denoising CNN (FFDNet), which uses downsampling-
upsampling steps, takes a tunable noise level map as the
input. In fact, a 2-D noise level map is fed to FFDNet during
the training phase so the model parameters become
more robust to the noise level. Consequently, only a single
trained model is sufficient to denoise images with different
noise levels.
Considering dilated convolutions, Zhang et al. [99]
Figure 6.
Regression application for ice concentration presented in [89]
together with color map (a) ground truth (b) regression result of
CNN.
MAY 2022
proposed a CNN by using residual learning and dilated
convolutions for SAR image despeckling. In comparison
with standard 2-D convolutions, dilated convolutions can
deliver a larger receptive field with the same number of
parameters. However, aggressive application of dilated
convolution causes two problems, 1) spatial consistency
between neighboring units becomes weak and 2) local
structure cannot be extracted in higher layers [100]. Using
dilated convolutions, Liu et al. [101] proposed a CNN
framework to obtain a different denoised version of each
SAR image. Afterward, they employed a guided filteringbased
fusion algorithm to integrate them into a finaldenoised
image.
Mukherjee et al. [102] proposed a convolutional AE
architecture for InSAR denoising purposes. Lattari et al.
[103] proposed a modified version of U-Net, which is an
IEEE A&E SYSTEMS MAGAZINE
25
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
