IEEE Geoscience and Remote Sensing Magazine - December 2022 - 135

in hyperspectral image classification. With the support of
these professional libraries, the coding difficulty for developers
is greatly reduced. CUDA's good ecosystem and ease
of use support it becoming the first choice for researchers
using GPUs for remote sensing data processing.
OpenCL is a programming framework for heterogeneous
platforms that can support various types of CPUs,
GPUs, and other types of processors. Because of its universality,
many researchers choose OpenCL for the parallel
development of GPUs. Zhu et al. [234] realized change
detection based on SAR image data by using an AMD Radeon
HD 6870 GPU and OpenCL and achieved 63-145×
acceleration. Bilotta et al. [22] applied OpenCL to develop
program coding that was suitable for the parallel processing
of the classic brightness temperature difference ash
cloud detection algorithm of the CPU and GPU. To use the
powerful computing of a GPU and keep the portability of
program code, Peternier et al. [143] implemented a parallel
processing program for the SAR image focuser based on
OpenCL, which can run on either a GPU or a CPU without
any code modification. Bernabe et al. [18] implemented
a hyperspectral image target detection method based
on OpenCL and solved the problem of the cross-platform
portability of parallel algorithms. At present, CUDA and
OpenCL are the two most widely used tools for GPU development;
C++AMP and Open Accelerators can also be
used. Because there are few studies on remote sensing
data processing based on other tools, this article will not
expand on them.
APPLICATION OF CPU/GPU COLLABORATION AND
COMPUTING OPTIMIZATION IN REMOTE SENSING
Deep collaborative computing between a GPU and a CPU
has great potential
Although researchers have achieved higher speed-up
ratios by using GPU acceleration, the acceleration efficiency
varies. To give full play to the computing performance
of the GPU, many researchers have optimized its
parallel computing strategy in memory management, data
exchange, and task scheduling and in combination with
other parallel computing technologies. In view of the limitation
of a large number of loads and the low data transmission
bandwidth of GPU memory and host memory, Liu
et al. [114] adopted data buffering, texture memory stores,
multithreading, and other technologies to optimize data
loading, memory usage, and data transmission. Sun et al.
[166] proposed an improved remote sensing image correction
method based on the rational functional model and
proposed an optimization strategy to maximize device
occupancy and improve memory access efficiency and
instruction throughput when using CPUs and GPUs in
cooperation. Zhang et al. [222] developed a dynamic task
scheduler based on reinforcement learning for lidar data
processing to realize the optimal dynamic and automatic
allocation of computing resources.
Liu et al. [112] put forward a three-level parallel strategy
of job-level, task-level, and image processing-level parallelism
for retrieving aerosol optical depth (AOD) in long time
series. The first two levels are coarse-grained parallelism,
which is realized by multi-DAG, while the image processing
level is fine-grained parallelism, which is realized by
the GPU. Tan et al. [168] used a two-level parallel computing
framework to classify support vector machine (SVM)
hyperspectral images and realized the data processing-level
COCCi COCC1 CCC GPU
...
in complex and heterogeneous big
data processing and analysis. With cooperation between
a CPU and a GPU in the processing of a multitask flow,
the data dependence among subtasks may increase the
data exchange cost between the GPU and CPU, so it is
very important to make rational use of both resources.
Wong et al. [191] optimized the division of labor between
a CPU and a GPU in the geometric correction processing
of lidar data. First, the CPU filters and eliminates invalid
data, then the GPU completes computation-intensive
tasks, such as geometric correction, optimizing the overall
processing speed. Fang et al. [63] used a CPU/GPU
combination for the near-real-time processing of Zi-Yuan
3 (ZY-3) images and proposed a CPU/GPU workload distribution
scheme to fully invoke the computing resources
(Figure 3). This scheme divided the CPU roles into a control
CPU core (CCC) and computing CPU cores (COCCs).
The COCCs are responsible for taking part of the processing
work to reduce the computing pressure of the GPU,
the CCC is responsible for managing the COCCs and the
GPU (controlling data exchange, task initiation, and so
on), and the three types of processors complete the data
processing tasks together.
DECEMBER 2022 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
RRC
...
MTFC
GC
COCCs Work
Idle
GPU Works
GPU Data Transfer and
Kernel Launch
FIGURE 3. The CPU/GPU collaborative strategy in near-real-time
processing of ZY-3 nadir images [63]. COCC: computing CPU core;
CCC: control CPU core. RRC: relative radiometric correction; MTFC:
modulation transfer function compensation; and GC: geo-correction.
135
Time Line
IEEE Geoscience and Remote Sensing Magazine - December 2022

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