GPU-based Acceleration of the Hyperspectral Band Selection by SNR
Estimation Using Wavelet Transform
Junpeng Zheng
a
, Liaoying Zhao
a
, Xiaorun Li
b
, Xin Zhou
b
, Jing Li
b
a
Institute of Computer Application Technology, Hangzhou Dianzi University, Hangzhou 310018,
China;
b
College of Electrical Engineering, Zhejiang University, Hangzhou 310027, China
ABSTRACT
Band selection provides performance improvement in hyperspectral applications such as target detection, spectral
unmixing and classification. Signal-to-noise ratio estimation (SNRE) as a method can be adjusted for different specific
applications. SNRE is usually used to remove some low SNR bands from original hyperspectral data in a preprocessing
stage and then other band selection methods are applied for the remaining high SNR bands of hyperspectral data to make
the operations more efficient. In this paper, we take advantage of SNRE to select the bands which contain the largest
amount of information. The wavelet transform is first used to realize the signal-noise separation and get the noise standard
deviation of each band, and then the SNRs of all bands are calculated orderly. Considering some time-consuming
operations in SNRE algorithm which can’t satisfy some real time applications are very suitable for high performance
computing(HPC) in parallel, we design a new massively parallel algorithm to accelerate the SNR estimation algorithm on
graphics processing units(GPUs) using the compute device unified architecture(CUDA) language. In addition, the
implementation of our GPU-based SNRE algorithm has extremely explored the possible parallelism in the C code and been
debugged carefully to verify its correctness and efficiency. Experiments are conducted on two sets of real hyperspectral
images and considerable acceleration is obtained.
Key words: graphics processing unit (GPU), compute device unified architecture (CUDA), SNRE, hyperspectral imaging,
band selection, wavelet transform
I. INTRODUCTION
Hyperspectral imagery is a three-dimensional imagery generated by imaging spectrometer simultaneously to the same
surface scenery at dozens even hundred bands, which has abundance of spectral information to detect the materials and
greatly enhances the ability of target detection[1-3]. However huge information also brings many problems in practice,
such as large storage space, long processing time. World-wide researchers all report the difficulties regarding its intrinsic
characteristics of the data complexities. In consideration of the high spectral resolution, spectral correlation among bands is
expected to be very high and the band-to-band spectral information may be overlapped or shared in some aspects
[4].Accordingly, getting rid of the redundant information, creating new hyperspectral space by selecting optimal bands and
lowering the dimensionality of data without losing crucial information becomes more and more important.
According to the information availability, band selection techniques can be divided into two categories: supervised
methods[5-7] and unsupervised methods[8-10].Supervised methods usually select the optimal band set by calculating the
statistical distance between the known category sample regions in the band combination to realize the purpose of more
accurate classification. Compared with supervised band selection techniques, unsupervised methods can achieve automatic
band selection completely with no priori information about objects or classes. Therefore, it is more meaningful to develop
reliable unsupervised band selection methods.
Based on the fact that the power and noise of the image can be separated after wavelet transformation to estimate the
noise variance, a Signal-to-noise ratio estimation method based on wavelet transformation was proposed in [10]. Using
SNRE, we can either remove the bands with low SNR or select the bands with high SNR according to the requirement
details of data processing, and furthermore, the experimental results in [10] indicated that SNRE can be combined with
*121050057@hdu.edu.cn; phone +8618968044859
Proc. of SPIE Vol. 9263 92630C-1