Research On Band Selection For Hyperspectral Image And CUDA Parallel Implementation

By means of imaging spectral technique, hyperspectral remote sensing can obtainrich spectral information formed by surface reflection. It provides the possibility offeature recognition for pixel level or even subpixel level. However, abundant spectralinformation also means massive data and high dimension, which bring big challengeto the hyperspectral image processing. Band selection can not only keep the originalband physical meaning and narrow the data source but also improve the speed ofhyperspectral image processing. But the number of bands to choose must bedetermined before band selection, which can be solved by virtual dimension. Analysisof most band selection methods can discover that band selection process has thecharacteristic of band data parallel computing, which is very suitable for parallelcomputing with GPU. This paper mainly focuses on the research of parallelimplementation of band selection. The specific contents are as follows:Firstly, elaborating the background and significance of hyperspectral image bandselection and its research status at home and abroad. Then unsupervised bandselection methods are analyzed comprehensively.Secondly, introducing the basic concepts of CUDA, the CUDA memory model,CUDA memory optimization and instruction optimization, which can explain thathow to use CUDA to write some high performance algorithm on the GPU and lay atheoretical foundation for further discussion on parallel algorithm based on CUDA.Thirdly, introduce three band selection methods: band selection method based onthe maximum information divergence, band selection method based on high ordermoments and band selection method based on variance and correlation coefficient.Then two band selection methods are proposed: band selection method based oninformation entropy and information divergence and band selection method based onSNR estimation using wavelet transform. According by some experiment results, theinfluences are analyzed on hyperspectral image classfication and anomaly detectionwith various band selection methods.Fourthly, introduce the concept of virtual dimension and its implementationalgorithm HFC. Then parallel HFC algorithm based on CUDA is designed andimplemented. The experiment results shows that the parallel HFC algorithm based onCUDA has significantly acceleration in comparison with the traditional HFC algorithm.Finally, parallel versions based on CUDA for hyperspectral band selection bySNR estimation using wavelet transform and hyperspectral band selection byinformation entropy and information divergence which are proposed in this paper aredeveloped and implemented. The experimental results show that parallel algorithmsbased on CUDA all obtain acceleration to varying degrees when compared with thetraditional implementation, which provides great help for hyperspectral applicationswhich have the features of real time, large amount of data and high complexity ofcomputation.