IEEE Geoscience and Remote Sensing Magazine - June 2020 - 65

introduce phase noise. It is well known that the phase-noise level is a spatial variant in the interferometric image domain, which is an inherent characteristic of InSAR system. The presence of noise
increases the phase unwrapping difficulty and even leads to failures in the process, seriously degrading the final interferometric results [9], [10]. Therefore, phase-noise reduction is a necessary step
during interferometric processing, and it has been developed as an important technology [11]-[20].
The original phase-noise reduction method can be traced to the multilook filter [11], which uses
a simple moving average of the neighboring pixels in a rectangular window, that is, boxcar filtering [11]. Due to its easy implementation, the multilook filter has been employed during the digital
elevation-model production of remote sensors, such as the TanDEM-X. The obvious disadvantages
of the multilook filter include resolution loss and phase-fringe distortion when dealing with hightopography and high-heterogeneity areas. Essentially, the multilook filter assumes that the interferometric phase is locally stationary and that the scene reflectivity is homogeneous in a local window
where the selected samples follow an independent and identical distribution (IID). In this case,
the multilook filter expects to perform a maximum-likelihood (ML) estimation [[12],
which is the foundation of most of the phase-filtering methods. However, this
assumption is not always true due to topography variation and reflectivity heterogeneity, especially when facing region edges, man-made
structures, and texture. In such situations, the interferometric
phase tends to exhibit nonstationary and nonhomogeneous
characteristics, conflicting with the IID assumption.
Simply put, the multilook filter is just a basic nonadaptive estimator, regardless of its interferometricphase property.
Since the invention of InSAR technology,
research on phase-noise reduction has been a
hot topic. The objective of phase denoising
is to accurately retrieve the interferometric
phase while maintaining the spatial resolution. There are numerous established
and emerging algorithms that show
great potential. Among them, recently
developed nonlocal (NL) filters exhibit
unique superiorities that have attracted
attention [21]. As opposed to traditional
local filters, NL filters measure the patch
similarity to select the most relevant
samples and perform a weighted average
on similar pixels to achieve noise reduction. The patch-wise sample-selection strategy is used instead of the pixel-wise method.
NL filters can capture phase redundancy using
nonconnected pixels, breaking the local-filter restriction. The possible application of NL filters to
the TanDEM-X is being studied; some attractive results
can be found in [22].
Furthermore, new concepts in sig[
nal processing technologies, such as sparse signal processing
and machine learning, have inspired innovate applications of
InSAR phase denoising, such as sparse regularization [23]
and sin[
gular-value decomposition (SVD) [24].
This
group
of
methods
has
shown
[
significant potential and can be treated as good candidates for a new generation
of InSAR denoising processors.
To boost InSAR development, a review of phase denoising is necessary through an examination of current research. Although some of the literature has included a short review of the topic
in their introductions, to the best of our knowledge there is still no comprehensive survey of phasedenoising technology. To fill the gap, this article attempts to provide an overview of existing methods and update the presentation of newly proposed techniques. A brief perspective on upcoming
JUNE 2020

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

IEEE Geoscience and Remote Sensing Magazine - June 2020

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - June 2020