Research On Spatial-Spectral Anomaly Detection Algorithm Of Hyperspectral Imagery Based On GPU

Hyperspectral remote sensing effectively combines imaging and spectrum technology to reflect the two-dimensional spatial distribution of ground objects and obtain continuous information in hundreds or even thousands of narrow bands.This method of imaging provides more subtle information about the original feature composition for the method of object attribute analysis,enabling object detection or indentification with higher confidence.Target detection based on the need for prior information can be divided into spectral matching detection and anomaly detection,in which anomaly detection without supervision information,has a wider range of applications.The anomaly detection algorithm determines the attributes of the pixel under text by analyzing the degree of different between the pixel under test and the backdround statistical characteristics or background posterior information.Recently,with the increase of the spatial resolution of remote sensing images,the spatial information in hyperspectral data can not be ignored.Combining the spatial and spectral information rationally,using the spatial correlation and similarity of neighborhood information to reduce the abnormal interference in the background,and enhance the separability of anomaly targets and backgrounds and effectively improve the accuracy of anomaly detection;Hyperspectral images have a huge amount of data and bring more computation complexity when the spatial-spectral information has been fully considered.Therefore,how to ensure the original data information is not lost,digging potential information in depth and maintaining high detection efficiency become the key and difficult of detecting hyperspectral image anomalies.Based on the above problems,this paper mainly studies the anomaly detection algorithm of hyperspectral image with spatial-spectral and implementation on GPU,makes full use of the potential spatial information of data and combines with the spectral information to improve the detection accuracy of the algorithm.Then,the parallel system architecture based on GPU/CUDA is designed to take advantage of the powerful general computing power and high storage bandwidth of the GPU platform to realize the efficient algorithm.The main contents of the paper are as follows:First,the classical KRX algorithm makes full use of the non-linear information between bands and obtain better detection performation.However,the increase of the resolution of hyperspectral image makes the background information easily be contaminated.Excessive anomalous interference will lead to the degradation of backdround kernel matrix and affect detection performance,and abundant non-linear kernel functions and inverse covariance matrices,resulting in the algorithm time-consuming.Then,this paper proposes a weighted spetial-spectral KRX algorithm and implementation on GPU.Through reconstruction of pixel under text and combination kernel mapping,effectively reduce the abnormal interference in the background and improve the sparability of the background and anomaly pixel.According to the characteristics of the algorithm,the parallel processing method based on GPU/CUDA model is designed,improving the efficiency of the algorithm.Secondly,aiming at the problem that the really feature can not satisfy the hypothesis of Gaussian distribution in classical statistical algorithms,this paper proposes a spatial-spectral kernel spectral angle mapping algorithm and implementation on GPU.The algorithm doesn't need to assume a background model,and improve performance in spectral angle mapping algorithms.Through using the nonlinear kernel mapping function,the spectral angle distance in the neighborhood are summed up,and the final results are corrected by the extended morphology to remove the noise interference and reduce the false alarm rate.Finally,through using the logical control ability of the CPU and the powerful parallel processing capability of the GPU,the parallel processing model is designed.Through optimizing thread mapping and reducing interaction overhead,the execution efficiency is improved.Finally,due to the interference of noise and abnormal point,the posterior information acquired inaccuracy in the spectral similarity metric algorithm.Then,this paper proposes a Hausdorff distance anomaly detection algorithm based on virtual background spectrum and implementation on GPU.According to the characteristics of the information around pixel under test,an idealized virtual background spectrum is constructed,and combined with an improved Hausdorff distance measure to detect abnormalities.The algorithm effectively reduces the sensitivity of the isolated noise point and reduces the effect of background noise and the abnormal point.In order to reduce the pressure on mass data processing,a parallel system architecture based on GPU/CUDA model is designed.GPU parallel computing ability is used to get better acceleration.