Approximate Computing of Remotely Sensed Data: SVM Hyperspectral Image Classification as a Case Study

被引：42

作者：

Wu, Yuanfeng ^{[1
]}

Yang, Xinghua ^{[2
]}

Plaza, Antonio ^{[3
]}

Qiao, Fei ^{[2
]}

Gao, Lianru ^{[1
]}

Zhang, Bing ^{[1
]}

Cui, Yabo ^{[1
]}

机构：

[1] Chinese Acad Sci, Inst Remote Sensing & Digital Earth, Key Lab Digital Earth Sci, Beijing 100094, Peoples R China

[2] Tsinghua Univ, Inst Circuits & Syst, Dept Elect Engn, Tsinghua Natl Lab Informat Sci & Technol, Beijing 100084, Peoples R China

[3] Univ Extremadura, Dept Technol Comp & Commun, Hyperspectral Comp Lab, Caceres 10071, Spain

来源：

IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING | 2016年 / 9卷 / 12期

基金：

中国国家自然科学基金;

关键词：

Approximate computing; energy saving and low power consumption; high-performance computing; hyperspectral remote sensing; onboard processing; ANOMALY DETECTION; IMPLEMENTATION; ARCHITECTURE; PERFORMANCE; EXTRACTION; ACCURACY; DESIGN;

D O I：

10.1109/JSTARS.2016.2539282

中图分类号：

TM [电工技术]; TN [电子技术、通信技术];

学科分类号：

0808 ; 0809 ;

摘要：

Onboard processing systems are becoming very important in remote sensing data processing. However, a main problem with specialized hardware architectures used for onboard processing is their high power consumption, which limits their exploitation in earth observation missions. In this paper, a novel strategy for approximate computing is proposed for reducing energy consumption in remotely sensed onboard processing tasks. As a case study, the implementation of support vector machine (SVM) hyperspectral image classification is considered by using the proposed approximate computing framework. Experimental results show that the proposed approximate computing scheme achieves up to 70% power savings in the kernel accumulation computation procedure with negligible degradation of classification accuracy as compared to the traditional ripple carry adder (RCA) precise computation. This is an important achievement to meet the restrictions of onboard processing scenarios.

引用

页码：5806 / 5818

页数：13

共 29 条

[1] Kernel-based methods for hyperspectral image classification [J].