IEEE Geoscience and Remote Sensing Magazine - December 2022 - 127

processing since as early as the 1990s. Lee et al. [95] used
a Cray-2 supercomputer to process the data of the Marine
Observation Satellite and estimate the sea surface temperature
in the Yellow Sea. Using a Cray-2s supercomputer, Yang
et al. [210] developed a remote sensing image processing
software system that realized common functions, such as
preprocessing, image enhancement, image transformation,
and classified image compression. A geometric correction
process correction algorithm was tested, and the processing
speed was increased by more than 50× compared with the IBM
3083 and VAX systems. Armstrong et al. [8] implemented
the inverse distance weighted interpolation algorithm using
a KSR1 supercomputer. When using 58 processors, the
data processing time was shortened from 2 h to 2.5 min,
and the speed increase was close to linear. These findings
confirm that supercomputers have great potential in remote
sensing data processing and analysis.
With continuous development, along with remote sensing,
big data, and other technologies, supercomputers have
been applied in more aspects of remote sensing big data processing.
The automated extraction of spectral endmembers is
an important part of hyperspectral processing, but the computational
complexity of high-dimensional datasets is very high.
Plaza et al. [146] used Thunderhead, a 256-node Beowulf
cluster supercomputer at NASA' s Goddard Space Flight Center,
to compare the efficiency of three algorithms in automatically
extracting spectral endmembers with different numbers
of processors (256 of which were relatively single processors),
and the speed-up ratio of the algorithm was between 99 and
234. Osterloh et al. [139] carried out aerosol retrieval based
on lidar data on the MareNostrum supercomputer, one of the
most powerful supercomputers in Europe, with peak performance
of 94.21 trillion floating-point operations per second.
Through the computing power of a supercomputer, the
influence of different search space dimensions on solution
quality can be understood. Fuzzy clustering is one of the
most commonly used methods to identify homogeneous
regions in remote sensing images, but the computational
cost for large-scale remote sensing images is very high. Petcu
et al. [240] used the IBM BlueGene/P supercomputer to realize
the parallelization of the fuzzy c-means algorithm,
which effectively reduced the computational cost. Because
knowledge extraction from massive remote sensing datasets
requires complex calculation flow, Kumar et al. [90] developed
a large-scale parallel and scalable parallel k-means
algorithm for superlarge datasets. Using the Cray XT5 peak
performance supercomputer, with 2.3 trillion floating-point
operations per second, the algorithm successfully identified
complex abnormal forest events, such as a mountain pine
beetle outbreak in Colorado, USA.
It is very important to study the imaging mechanism
of a simulated satellite to understand the imaging effect
before the satellite enters orbit, but the research process
needs a lot of computational support to make use of complex
physical processes and mathematical models. Wu et
al. [192] proposed a highly scalable physics-based ray tracer
DECEMBER 2022 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
for satellite image simulation. On the Tianhe 1 A supercomputer,
16,800 CPU cores were used to achieve near-linear
acceleration efficiency.
Assessing the scale and dynamics of an emergency or
disaster requires real-time analysis of global nighttime multispectral
satellite images. Zhizhin et al. [228] developed
a high-performance data processing system, including a
MapReduce computing model, workflow, job queue, virtual
computing node, and cluster file system, by using the
supercomputer of the Kurchatov Research Institute, in Moscow,
Russia, and completed the real-time detection of rapid
changes of night light brightness of the Suomi National PolarOrbiting
Partnership satellite, which had a 700-m spatial resolution.
Crop type classification plays an important role in
crop insurance, financial market forecasting, and so on. Cai
et al. [27] used ROGER and Blue Waters to classify crop types
based on time series Landsat data and deep learning. ROGER
has a large storage space, which is used for the rapid preprocessing
of large amounts of spatial data, and Blue Waters has
stronger computing power, which is used for intensive training
and testing and establishing classification models. Kurte
et al. [91] used 5,426 nodes of the Titan supercomputer to
map human settlements, with a coverage area of 685,675 km2,
based on WorldView-3 satellite remote sensing images with a
resolution of 0.31-m. For a dataset with a size of 13 TB, the
processing time was only 6 h.
Compared with cloud computing and GPU research,
there are relatively few supercomputers in the field of remote
sensing. The main reason is the high construction cost,
maintenance cost, and energy consumption of supercomputers,
and there are some problems, such as a lack of convenience
and a high threshold. At present, supercomputers are
mostly used in high-tech fields and cutting-edge technology
research, such as military, medical, energy, aerospace, astronomy,
physics, and other fields. With an increased number
of supercomputers and a lower construction threshold,
the application of supercomputers in the field of remote
sensing big data needs to be further explored.
MESSAGE PASSING CALCULATION
Many application programming interfaces (APIs) and implementation
libraries that support serial computer process
communication are called message passing models. The message
passing interface (MPI) is the most important parallel
programming tool based on message passing, which has
the advantages of good portability, powerful functionality,
high efficiency, and many free, efficient, and practical
implementation versions. Almost all parallel computer
manufacturers support it. It has become the de facto parallel
programming standard in message passing [34]. The
MPI has been continuously updated and expanded since it
appeared in the 1990s. With hundreds of commands, the
MPI can cope with most situations in parallel computing,
and now it is the main model of HPC. Task execution and
data exchange are two core problems in parallel computing
programming using the MPI. Reasonable task allocation,
127
IEEE Geoscience and Remote Sensing Magazine - December 2022

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - December 2022
