IEEE Geoscience and Remote Sensing Magazine - June 2021 - 39

2) Original domain residual architectures: If a residual architecture
is used in the original domain, the input image is
regarded as the sum of a clean image (ideally, the output
of the network) and noise. Therefore, an additive noise
model is still employed, but now the noise has signaldependent
variance: it is more intense in regions with
high reflectivity and less intense in regions with low
reflectivity. This may not be a problem if the network
operates on small, homogeneous patches of the image.
On the contrary, heterogeneous patches will be characterized
by additive noise of spatially varying intensity,
with unpredictable effects.
3) Original domain no-residual architectures: In this case, the
network operates on the image as is; hence, it has to deal
with truly multiplicative noise. Note that the method
proposed in [40] fits here since it uses a residual architecture
based on a ratio of images rather than a difference,
which is fully consistent with the multiplicative model.
We now analyze these groups of methods. Unless explicitly
stated, we consider the fully developed speckle model
to hold and refer to the most challenging case: single-look
speckle.
LOG DOMAIN METHODS
The SAR-CNN [56], one of the first CNN-based SAR despeckling
methods proposed in the literature and often considered
to be a baseline, follows this approach. The log of the input
image feeds a 17-layer CNN, which is a straightforward adaptation
of the DnCNN proposed in [24] for the AWGN case
based on a residual architecture [36]. Therefore, the CNN extracts
the log domain noise, which is then subtracted from the
original image before compensating for the nonzero mean
and taking the exponent of the result. In the SAR-CNN, the
network is trained on simulated single-look SAR images.
However, to ensure better fidelity to the actual statistics of the
SAR signal and speckle, it is retrained on real SAR data, using
multilooked images as approximate clean references.
The reliance on AWGN filters is made fully explicit in
methods based on the multilevel logarithm with Gaussian
denoising (MuLoG) [91] paradigm. Considering the scarcity
of reliable training data for SAR image despeckling, the goal
of MuLoG is to use AWGN denoisers just as they are, with a
standard adaptation procedure to fit them to the Fisher-Tippett
distribution of log-transformed SAR speckle. Therefore,
the approach works with conventional and deep learningbased
denoisers, the latter trained on unlimited AWGN
data. In [57], MuLoG-based despeckling is performed using
a model-based AWGN denoiser (BM3D [28]) and the pretrained
DnCNN. Experiments prove the importance of the
adaptation phase (that is, the unsuitability of straight AWGN
denoisers for log domain SAR data) and the superiority of
the deep learning denoiser. Yet further analyses carried out
in [58] show that the performance remains below the level
of the SAR-CNN (a slightly improved version implemented
by the authors), underlining the importance of training or at
least fine-tuning the network on real SAR data. The MuLoG
JUNE 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
approach is also followed in [59], where the DnCNN engine
is replaced with the newer and more effective fast and flexible
discriminative CNN denoiser (FFDNet) [92], with some
beneficial effects on performance.
On the opposite side of the spectrum, [60] models not
only the speckle but also the signal itself as a random process
so as to better account for the homogeneous/heterogeneous
nature of the observed cell. Working in the log
domain, the pdf of the observed signal can be regarded as
the result of a convolution between the pdfs of the clean
signal (unknown) and speckle. By means of an elaborate
procedure, a rather shallow CNN is trained to predict the
pdf and mean of the clean signal. Experiments on synthetic
data provide some support to this approach, but results on
real-world SAR images are still inconclusive.
ORIGINAL DOMAIN RESIDUAL ARCHITECTURES
This is a fairly large group, as the residual architecture is
quite widespread. The first method to adopt this setting,
the SAR-dilated residual network (DRN) [61], is now one
of the most popular in the field, and it is often used as a
baseline. An appealing feature is its lightweight architecture
with only seven convolutional layers. This choice reduces
the training complexity without impairing performance,
thanks to the use of dilated convolutions, which ensure
that a large aggregate receptive field is obtained. In addition,
skip connections are incorporated to implement residual
blocks. The network is exclusively trained on simulated
data. Although questionable, this is such a common trait
that, from now on, we point out only exceptions to this rule.
The hybrid dilated residual attention network proposed in
[62] also uses dilated convolutions and skip connections in
a seven-layer architecture. The main innovation lies in the
introduction of attention modules that operate in space and
across channels. Channel attention modules redefine the
y
y˜
log
Network
(a)
y
x
Network
+
y
(b)
y
x
Network
(c)
FIGURE 5. Network architecture designs. (a) Direct deep learningbased
despeckling in the log domain. (b) Direct deep learningbased
despeckling with a residual architecture. (c) Direct deep
learning-based despeckling with a plain architecture.
39
x˜
x
Exponent
"
"
"
"
IEEE Geoscience and Remote Sensing Magazine - June 2021

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