IEEE Geoscience and Remote Sensing Magazine - June 2021 - 46

Finally, double-sided masked CNN for pixel-adaptive
multiplicative noise despeckling (DoPAMINE) [108] also
exploits a blind-spot network, not to directly compute the
denoised pixel value but to derive the parameters (the slope
and bias) of an affine regressor of each clean pixel, given
the noisy pixel. The loss function used to train the network
is an unbiased estimator of the MSE derived from SURE,
which requires only the noisy pixels and the despeckled
output. As in [41], the method heavily relies on spatially
uncorrelated noise. Unfortunately, while it could work with
real noisy images, the authors present only results trained
on synthetically speckled images from the University of
California, Merced, land use data set.
Figure 9 compares the Speckle2Void [41] self-supervised
method with traditional model-based techniques
on a TerraSAR-X image. The self-supervised technique
demonstrates improved performance over the classic
methods, showing better speckle suppression in uniform
areas as well as better preservation of edges and texture.
We refrain from showing comparisons across multiple
self-supervised methods, as the approaches have been
trained for different platforms (e.g., Sentinel-1 instead
of TerraSAR-X). Indeed, self-supervised techniques are
optimized for the characteristics of a specific platform
since they require only noisy images, and applying the
pretrained model to a different platform would be suboptimal
and result in unfair comparisons. We believe that
studying how well self-supervised representations transfer
across platforms could be an interesting research direction
for the community.
FUTURE DIRECTIONS AND OPEN PROBLEMS
The analysis of existing work on SAR image despeckling
based on deep learning has shown that the field has come
a long way from its beginnings. However, we also feel
that the available methods are not sufficiently mature,
and there is still a lot of work to be done toward methods
that can provide despeckled images of consistently good
quality. In the following, we identify several aspects that
should be the subject of further work by the community.
Some of these are related to the fact that studies of the
topic often do not adhere to the standards of reproducible
research, making it difficult to assess the specific merits
of each contribution with respect to the body of available
literature and significantly slowing down the pace of technical
advances.
RESEARCH DIRECTIONS
There are several interesting directions that can be explored
by future research into deep despeckling methods. First,
(a)
(b)
(c)
(d)
(e)
FIGURE 9. Self-supervised methods versus model-based approaches. (a) A noisy image. (b) PPB [14]. (c) NL-SAR [17]. (d) SAR-BM3D [15]. (e)
Speckle2Void [41].
46
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