IEEE Geoscience and Remote Sensing Magazine - June 2021 - 33

be noted that SAR speckle is not spatially white. In fact, the
SAR system impulse response is such that it induces spatial
correlation. Further filtering techniques, such as apodization
to prevent strong scatterers from leaking signal in
neighboring cells, only increase the amount of correlation.
This is obviously neglected by AWGN-oriented methods
but may well be exploited to improve SAR despeckling
performance [31].
PRIMER ON DEEP LEARNING
Recent years have witnessed the rise of machine learning
methods to address a number of problems in the image
processing and computer vision fields. Such data-driven
approaches typically rely on deep neural networks to act
as universal function approximators, using training data to
learn a mapping among an input and the corresponding desired
output. Within this general framework, a distinction
must be made among supervised, self-supervised, and unsupervised
learning methods. Supervised methods are the
most common and rely on having access to labeled data,
i.e., information for which both the input and the desired
ground truth output, e.g., a class label in a classification
problem or a clean image in a denoising one, is available.
While the supervised training of deep neural networks can
provide excellent results due to the systems' ability to learn
very complex mappings, it is ultimately limited by the need
for large amounts of data with accompanying ground truth
labels. Unsupervised methods do not rely on ground truth
labels and seek to uncover latent data properties by analyzing
their features. Self-supervised learning can be regarded
as a special case of both supervised and unsupervised
learning in which the ground truth labels are not available
but it is still possible to learn a mapping to a (unknown)
desired output, as in the supervised setting, by generating
labeling information from the data themselves.
A deep neural network essentially equates to a sequence
of linear vector operations parameterized by weights and
interleaved by nonlinear functions. Training a neural network
amounts to finding the values of its parameters that
minimize the loss function specified by the setting. While
all neural networks can approximate arbitrary continuous
and differentiable functions, different architectures provide
various priors and inductive biases that can be useful for
specific applications. In this sense, the CNN proved to be
valuable for problems concerned with visual data, such as
images, video, and more. A CNN is composed of a stack of
convolutional layers, each implementing a filter bank and
a nonlinear activation function. The filter bank produces
a number of output feature maps by means of the spatial
convolution of the input feature maps with a set of kernels.
The spatial resolution of the feature maps is typically unchanged
or reduced by striding and pooling.
CNNs have been very successful at processing visual
data because they enforce, by design, some priors that are
true for natural images. In particular, the kernels have a
small spatial extent K, and this induces a localized receptive
JUNE 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
field whereby the output of a layer at a given pixel is affected
only by neighboring ones. The fact that the same kernel
weights are reused across the whole image captures the
stationarity property, where the characteristics of a feature
do not depend on its spatial location. The small kernels and
the spatial reuse due to the convolution operation also conveniently
reduce the number of trainable parameters with
respect to fully connected neural networks, increasing efficiency
and reducing the risk of overfitting. Finally, stacking
many such layers creates a compositional representation,
i.e., a hierarchy of features in which higher-level ones can
be built by combining lower-level ones.
Another common architecture in visual problems is the
generative adversarial network (GAN) [32]. GANs are usually
employed to learn to generate new data samples whose
distribution approximates that of the training data. The architecture
of a GAN is composed of two networks: a generator
G that learns to capture the distribution of the training
data and a discriminator D that learns to distinguish between
real training data and samples generated by G. The
key aspect of GANs is that they treat the training process as
a game between the two networks, where the goal of G is to
generate data samples that can fool D and the objective of D
is to distinguish between the real and fake samples. GANs
are commonly employed in image restoration problems to
ensure that the data distribution of reconditioned images
matches that of the clean images. Finally, the minimization
of the loss function required to train deep neural networks
is performed by means of first-order optimization methods,
such as stochastic gradient descent (or momentumaccelerated
variants [33]). Training is a computationally
intensive process that greatly benefits from the use of highparallel
hardware architectures, such as GPUs.
COMMON ISSUES WITH DEEP LEARNING MODELS
The remote sensing expert who is freshly approaching deep
learning techniques should be aware of a number of issues
that must be taken into account when designing deep models.
The vanishing gradient, a historically challenging issue
that inhibited the creation of neural networks with a large
number of layers, is now successfully mitigated by normalization
layers (e.g., batch normalization [34]), nonsaturating
activation functions [35], and residual connections [36].
When opting for a deep learning model, one must always
consider the matter of data for model training and model
testing. Training a neural network may require a large
amount of data so as to avoid overfitting due to the large
model capacity (i.e., the high number of trainable weights),
and one must make sure to have them at one's disposal,
especially if manual annotation is required. Techniques
to reduce this requirement exist and enjoy some success,
including transfer learning [37], i.e., starting from weights
gleaned from a similar task rather than random values; selfsupervised
pretraining [38]; and more.
The data issue is also important for the practitioner who
wants to use pretrained deep models. In fact, a matter of
33
IEEE Geoscience and Remote Sensing Magazine - June 2021

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