IEEE Geoscience and Remote Sensing Magazine - December 2021 - 147

units in recurrent GNNs (RGNNs) [23], [55] have been
shown to obtain achievements in learning from graphs.
The usefulness of GNNs in SAR is still to be properly explored,
and [56] is one of the only attempts to do so.
POSSIBLE PITFALLS
To develop tailored deep learning architectures and prepare
suitable training data sets for SAR and InSAR tasks, it is
important to understand that SAR data are different from
optical remote sensing data, not to mention images downloaded
from the Internet. In this section, we discuss the
special characteristics (and possible pitfalls) encountered
while applying deep learning to SAR. What makes SAR
data and SAR data processing by neural networks unique?
SAR data are substantially different from optical imagery in
many respects. The following points should be considered
when transferring CNN experience and expertise from optical
to SAR data:
◗ Dynamic range: Depending on the spatial resolution, the
dynamic range of SAR images can be up to 90 dB (TerraSAR-X
high-resolution spotlight data with a resolution
of roughly 1 m). Moreover, the distribution is extremely
asymmetric, with the majority of the pixels in the lowamplitude
range (distributed scatterers) and a long tail
representing bright discrete scatterers, in particular, in
urban areas. Standard CNNs are not able to handle such
dynamic ranges, and hence most approaches feature
dynamic compression as a preprocessing step. In [57],
the authors first take only amplitude values from zero
to 255 and then subtract the mean values of each image.
In [11] and [58], normalization is performed as a
preprocessing step, which significantly compresses the
dynamic range.
◗ Signal statistics: To retrieve features from SAR (amplitude
and intensity) images, speckle statistics must be
considered. Speckle is a multiplicative, rather than an
additive, phenomenon. This has consequences: while
the optimum estimator of the radar brightness of a homogeneous
image patch under speckle is a simple moving-average
operation (i.e., a convolution, such as in the
additive noise case), other optimum detectors of edges
and low-level features under additive Gaussian noise
may no longer be optimum in the case of SAR. A popular
example is Touzi's constant false alarm rate edge
detector [59] for SAR images, which uses the ratio of
two spatial averages across adjacent windows. This operation
cannot be emulated by the first layer of a standard
CNN. Some studies use a logarithmic mapping of
the SAR images prior to feeding them into a CNN [9],
[60]. This turns speckle into an additive random variable
and, as a side effect, reduces the dynamic range.
But, still, a single convolutional layer can emulate only
approximations to optimum SAR feature estimators. It
could be valuable to supplement the original logarithmic
SAR image by a few low-pass-filtered and logarithmized
versions as input to a CNN. Another approach is
DECEMBER 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
to apply a sophisticated speckle reduction filter before
entering a CNN, e.g., nonlocal averaging [61]-[63].
◗ Imaging geometry: SAR image coordinates' range and azimuth
are not arbitrary, such as east and north or x and y,
but, rather, reflect the peculiarities of the image generation
process. Layover always occurs at the near range of
an object, and shadow always results at the far range.
That means data augmentation by SAR image rotation
would lead to nonsense imagery that would never be
generated by a SAR.
◗ The complex nature of SAR data: The most valuable SAR
data information lies in its phase. This applies to SAR image
formation, which takes place in the complex signal
domain, as well as for PolSAR, InSAR, and tomographic
SAR data processing, meaning that an entire CNN must
be able to handle complex numbers. For the convolution
operation, this is trivial. The nonlinear activation
function and the loss function, however, require thorough
consideration. Depending on whether the activation
function independently acts on the real and imaginary
parts of the signal, or only on its magnitude, and
where bias is added, the phase will be distorted to different
degrees.
If we use PolSAR data for land cover and target classification,
a nonlinear processing of the phase is even desirable
because the phase between different polarimetric
channels has physical meaning and hence contributes
to the classification process. In SAR interferometry and
tomography, however, the absolute phase has no meaning;
i.e., the CNN must be invariant to an arbitrary phase
offset. Assume some interferometric input signal x to a
CNN and the output signal CNN(x) with phase
t =
z +CN ().xN
(1)
Any constant phase offset 0z does not change the meaning
of the interferogram. Thus, we require an invariance
that we refer to as phase linearity (which is valid at least
in the expectation):
CNNCNN xe
xe
() () .
00
jj
zz=
(2)
This linearity is violated, for example, if the activation
function is separately applied to real and imaginary
parts and if a bias is added to the complex numbers.
Another point to consider in regression-type InSAR
CNN processing (e.g., for noise reduction) is the loss
function. If the quantity of interest is not the complex
number itself but its phase, the loss function must be
able to handle the cyclic nature of phases. It may also be
advantageous that the loss function is independent, at
least to a certain degree, of the signal magnitude to relieve
a CNN from modeling the magnitude. A loss function
that meets these requirements is, for example,
LE e (())jx y
CNN
=
6 ++@
,
(3)
147
IEEE Geoscience and Remote Sensing Magazine - December 2021

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