The Research Of Hyperspectral Anomaly Detection Algorithm Based On Background Data Optimized

With the development of aerospace technology and information processing technology,Hyperspectral remote sensing technology has been booming since 1980 s,and has become a attractive topic in the field of remote sensing.Hyperspectral image is a new class of remote sensing image with the property of image and spectrum,containing hundreds of bands,high spectral resolution,and can distinguish the spectral difference with small features,which make it widely used in military,agriculture,geological exploration and other fields.the anomaly target detection is one of hot issues in the field of hyperspectral image information processing field because of no prior knowledge of the target spectral signature is utilized or assumed.Based on analyzing the characteristics of hyperspectral image and the classical algorithm,the thesis suppress the interference that potential anomalies to the estimation of mean and covariance matrix of background by weighting and background optimization,reduce the false alarm rate and improve the detection accuracy.The thesis mainly work as follows:Firstly,the thesis introduces the imaging mechanism and the characteristics of the hyperspectral image,analyses the theory of RX algorithm which frequently used in anomaly detection,then the machine learning and kernel function methods are briefly described,and research the kernel RX anomaly detection algorithm based on kernel mapping,which are used in later research work.Secondly,a wheighted anomaly detection algorithm for hyperspectral image is proposed.Firstly calculate the weight of each pixel in the background,select the pure background,and then each pixel in the new background using their own weight to construct the weighted kernel RX anomaly detection operator for anomaly detection.Two methods are proposed for the selection of weight.The one uses Spectral angle cosine as spectral angle matching criterion to determine the distance between pixels,then k-means clustering is performed on the background pixels of the image to obtain the cluster centers,the spectral angle cosine of background pixel and cluster center is used as the weight.The other obtains background by principal component analysis,projects every pixel into background orthogonal subspace to give every pixel a proper weight.The experiments show that the proposed algorithm has a strong suppression ability to potential outliers and improve the detection accuracy.Finally,a new algorithm for anomaly detection based on local density estimation is proposed.Firstly,calculate the k-distance neighborhood of each pixel in the background by the kernel spectral angle cosine,and then the local outlier factor of each background pixel is obtained by using the points in the neighborhood and compared with the threshold to remove the outlies from the background.Finally,the image is detected by RX anomaly detection algorithm with the statistical information of the refined background.To shorten the running time of the algorithm,calculate the kurtosis in the background before the detection,if kurtosis is greater than the threshold,run the local density estimation anomaly detection algorithm,otherwise,run the RX anomaly detection.The experiments show that the proposed algorithm can effectively separate the background and outlier,provide a guarantee for the estimation of background characteristics,the kurtosis judgment,the efficiency of the algorithm is obviously improved without affecting the detection probability.