IEEE Geoscience and Remote Sensing Magazine - June 2021 - 44

(a)
(b)
(c)
(d)
(e)
(f)
(g)
FIGURE 6. A comparison of state-of-the-art, model-based methods and techniques based on supervised deep learning. (a) A single-look
COSMO-SkyMed input image. (b) A 25-look multitemporal reference. Despeckled images provided by (c) PPB, (d) SAR-BM3D, (e) NL-SAR,
(f) the SAR-CNN, and (g) the CNN-NLM.
be hitting a performance bound and calls for the use of new
solutions, including multitemporal and multisource data.
SELF-SUPERVISED MODELS
The previous sections described the traditional avenue taken
by researchers investigating deep learning despeckling
algorithms, i.e., supervised learning. As discussed, while the
conceptual simplicity of the supervised setting is appealing,
several issues stem from the lack of speckle-free ground
truth data. For this reason, the past year has witnessed
y(i)
Network
Noisy
Image i
(a)
y(i )
Noisy
Image i
y(i )
Noisy
Image i
(c)
FIGURE 7. Different approaches to model training. (a) Supervised
training requires a clean, speckle-free target. Self-supervised training
uses noisy images as targets: either (b) different realizations
of the speckle process for the same scene (multi-image) or (c) the
same image (single-image).
44
Network
(b)
x(i )
Network
Loss
y(i )
Noisy
Image i
x(i )
Loss
Noisy
Image j
y(j )
x(i)
Loss
Clean
Image i
x(i)
growing interest in self-supervised methods, i.e., those
that can directly exploit noisy images without the need for
clean data. This is an extremely important direction for the
field, as it enables us to better exploit real SAR data and to
do without optical image data sets and their associated issues.
However, this approach is in its infancy, and some of
the works that will be presented are still undergoing peer
review, so one must be careful with the information available
at this stage. Nevertheless, the numbers of publications
and preprints [41], [103]-[109] that appeared in a short time
signals a strong interest in these techniques and calls for further
research. Self-supervised models for SAR despeckling
can be broadly categorized in two approaches, as shown
in Figure 7: multi-image models and single-image models.
These are derived from recent techniques developed for traditional
image denoising, which we now briefly review.
SELF-SUPERVISED MODELS FOR IMAGE DENOISING
Multi-image models follow the approach begun by Noise2Noise
[110] whereby both the input and the training target
are noisy images, representing the same scene with different
realizations of the noise process. In standard supervised
methods for image restoration, the model is trained
by minimizing the distortion, usually measured using the
2
, norm, between the output and the ground truth image.
In [110], the authors observe that the estimate of the 2
, loss
remains unchanged if we replace the ground truth images
with noisy observations whose expectation matches the
clean image. Consequently, as long as the expected value
of the noisy images is equal to the ground truth image, it
is possible to train a neural network by using pairs of images
with the same content but independent realizations of
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2021
"
"
"
IEEE Geoscience and Remote Sensing Magazine - June 2021

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