IEEE Geoscience and Remote Sensing Magazine - June 2021 - 30

SPECKLE REJECTION
SAR data are generated through an active and coherent sensing
process whereby radar echo returns of ground targets are
acquired (the so-called raw data) and later focused into a fully
developed image. SAR images represent an important complementary
source of information with respect to optical images.
While the latter are more easily interpretable and gather data
in many spectral bands, enabling, e.g., a detailed classification
of ground covers, SAR images provide information about the
geometry of the observed scene, facilitating, in interferometric
and polarimetric modalities, the accurate retrieval of topography
and 3D structures. Moreover, SAR image acquisition does
not depend on sunlight and climatic conditions, ensuring
continuous all-day, all-weather coverage.
However, as in other coherent imaging systems, such as
medical ultrasound and tomography, the coherent sensing
process means that the intensity of SAR images is affected
by a particular type of noise known as speckle. Unlike the
Gaussian noise typically affecting optical images, speckle
noise is spatially correlated and signal dependent, and it appears
as a grainy texture superimposed onto the image (see,
e.g., Figure 1), which greatly affects interpretability and scientific
exploitation. Therefore, improving SAR image quality
by reducing noise has a significant positive impact on
downstream tasks, from visual inspection to environmental
monitoring, surveillance, and the detection of anomalies
and hazards, just to mention a few.
The interest in speckle rejection has been intense since
the early days, as evidenced by papers on multilooking [1]
and Wiener filtering [2] dating to the 1970s. From 1980 on,
several effective methods [3]-[6] based on spatial-adaptive
filtering were proposed. The key idea is to adapt the
estimation window based on local image statistics to deal
with signal nonstationarity. Even so, speckle reduction is accompanied
by a clear loss of resolution. A further class of
methods [7]-[10] appeared in the 2000s with the advent of
the wavelet transform. By leveraging the good separation of
the signal and noise in the wavelet domain, the approaches
combine wavelet shrinkage with models and tools typical of
image processing. Despite the improved speckle rejection,
visible filtering artifacts, such as ringing, are frequent.
Variational methods [11]-[13] formulate image despeckling
as an optimization problem, looking for the latent image
that best explains the observed noisy one. To remove speckle,
an objective function is minimized; this compounds
a data-fitting term with a regularization term, typically
based on total variation (TV). Again, significant artifacts
are observed, such as signal oversmoothing and staircasing.
Nonlocal methods [14]-[17] represent what is arguably
the current state of the art. The objective is to select the best
predictors of the target pixel through a suitable measure of
similarity based on the local context [18]. By combining
the nonlocal approach with other image processing tools,
good speckle rejection is obtained with limited (though not
absent) artifacts and little loss of resolution. For brevity, we
mention only in passing other less popular, yet interesting,
approaches, such as those relying on Markov random field
priors [19], [20] and sparse representations [21], [22].
The advent of deep learning has revolutionized the approach
to many image processing tasks. After the stunning
(a)
(b)
FIGURE 1. (a) A single-look TerraSAR-X image acquired over Rosenheim, Germany. (Image courtesy of DLR.) (b) An optical image of Rosenheim
downloaded from Google Earth Pro, coregistered with the SAR master. The noise intensity is extremely large for the SAR image and
virtually zero for the optical image.
30
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