IEEE Geoscience and Remote Sensing Magazine - September 2021 - 4

to capture spatial-spectral information in remote sensing
data, but at a high computational cost. APs are a more
generalized method, working on connected components
rather than individual pixels, representing images as hierarchical
tree structures having the connected components
as nodes. This allows for the exploitation of spatially contextual
information in conjunction with the spectral characteristics
of pixels.
Although APs have been used mostly for the supervised
classification of optical images, unsupervised classification
and change detection, and application to other
sensor data, have also been studied. This review article
is organized according to the computational steps in the
application of APs, with an experimental comparison of
different AP approaches and the impact of various attributes.
Two publicly available airborne data sets are
used for these comparisons, and all the source code is
made available.
The article concludes with a discussion of the generOUR
FIFTH FEATURE
ADDRESSES ANOTHER
PROBLEM IN THE
POSTPROCESSING OF
REMOTE SENSING DATA:
THE RECONSTRUCTION
OF LAND SURFACE
TEMPERATURE
PRODUCTS.
alization of APs to multiple-image data sets, and various
challenges such as the distribution
of training data and the
image-quantization level. Some
possible future directions for research
on APs are proposed.
The next feature continues
on the important theme of
classification, addressing the
band selection (BS) problem,
specifically for land cover in
hyperspectral images (HSIs).
Given the high dimensionality
of HSIs, proper BS is required
to remove redundancy. In general,
BS methods can be categorized
as spectral, spatial, or hybrid, and as supervised
versus unsupervised. The article focuses on the more extensively
used unsupervised techniques, with the goal
to provide a framework for scientists to properly select
and use BS methods without specialized knowledge
of the theory.
The article provides a thorough review of various approaches
to BS. Starting with a comparison of correlationand
information-theory-based measures, then reviewing
grouping and ranking strategies, and finally, discussing
search methods and selection strategies. Various state-ofthe-art
BS approaches are evaluated using three benchmark
data sets.
The article concludes with some open questions about
selecting the number of bands, the use of pixels versus superpixels,
and approaches to validation. Deep learning is
suggested as a novel approach to explore further.
Our fifth feature addresses another problem in the postprocessing
of remote sensing data: the reconstruction of
land surface temperature (LST) products. LST is a crucial
variable in a number of different fields, including moni4
toring
environmental and ecological processes, the study
of heat-related health issues and urban heat islands, and
evaluating long-term (climate related) surface temperature
changes. Both thermal infrared and passive microwave
methods have been developed for LST retrievals, with
widely different resolutions, sensitivity to atmospheric conditions
(e.g., clouds), and spatial continuity.
Recent research to correct these discontinuities is reviewed
in the article. Starting with the fundamental theoretical
basis for LST measurements and a review of current
satellite-based LST products, the article describes
the related problems of reconstruction (combining
complementary information from different sensors to
produce a continuous LST time series) and spatiotemporal
fusion (enhancing the resolution of LST products
by combining multiple sensors). Several approaches to
validation of the reconstructed or fused products are
also explained.
The article concludes with suggestions on the future
prospects of this complicated and often ill-posed problem.
These include a better integration of physical properties
into the signal processing, the use of data-assimilation
methods, and a better understanding of the synergies between
reconstruction and spatiotemporal fusion.
Our final feature for this month's issue reviews techniques
for deriving ocean surface current from the Geostationary
Ocean Color Imager by tracking the motion of surface
features between subsequent images. One objective of
this article is to develop a more reliable operational procedure
based upon a satellite image-matching system. A selection
of the appropriate features to detect and track, and the
effects of clouds, quality control, and accuracy assessment
are all addressed. These methods are compared using three
example data sets.
Beginning with the columns in this issue, we have a report
from the IGARSS conference, held virtually from Brussels,
Belgium, this year, starting on page 157.
The " Technical Committees " column introduces three
new working groups (WGs) in the Image Analysis and Data
Fusion (IADF) Technical Committee, created to accommodate
recent advances and trends in the field. The new WGs
include Benchmarking, which addresses the need for wellcurated
and annotated data sets for the training and validation
of new methods. The Image and Signal Processing WG
was formed to promote the development and application of
advances in those fields to remote sensing problems. The
Machine/Deep Learning for Image Analysis WG fosters
advancements in these fields as they relate to the analysis
of geospatial and remote sensing images in ethical, understandable,
and trustworthy ways.
The Oman Section, formed in 2020, is featured in the
" Chapters " column. This new section began with the
nomination of an IEEE Geoscience and Remote Sensing
Society (GRSS) ambassador from that country a few years
earlier and has already organized a number of workshops
and training sessions as well as a COVID-19 monitoring
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE SEPTEMBER 2021
IEEE Geoscience and Remote Sensing Magazine - September 2021

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