Fast Anomaly Detection Algorithms Based On Causal Systems For Hyperspectral Imagery

Hyperspectral remote sensing imagery,which is characterized by an integration of image and spectrum,has become one of the most significant research subjects in image understanding and interpretation.As its important research branch,hyperspectral anomaly detection can make full use of spectral information and dig potential statistics included in background and targets,thereby having capability to find out those ground targets that are hardly detected by traditional single-spectral and multi-spectral imagery.Recently,with the development of hyperspectral anomaly detection in practical applications and urgent solving of big data problem as it contains,fast detection based on causal systems makes sense.On one hand,traditional hyperspectral anomaly detection outputs all results at a time after collecting and processing the whole hyperspectral data.This noncausal style can be with big limitations in practical applications.For example,moving chariots and spreading fire cannot be detected in real time by using usual processing ways.On the other hand,hyperspectral data collected are vast.In contrast with limited data storage in airborne and spaceborne imaging,heavy data brings much pressure on the storage.And the time delay caused by data transmission is also considerable.Real-time processing requires causality and timeliness.This dissertation focuses on formulating causal systems and improving anomaly detection efficiency and accuracy in these systems,which are presented as follows:1.The push-broom type is a major style in current hyperspectral imaging devices,that is,collecting data line-by-line.According to this push-broom type,two kinds of causal systems including global and local processing can be designed.Based on both causal systems,Progressive Line Processing of Global Causal Fast RX Anomaly Detection(LCF-GRXD)and Progressive Line Processing of Local Causal Fast RX Anomaly Detection(LCF-LRXD)are proposed.Both meet the requirement of causality.But more importantly,they are strictly derived by the covariance-matrix RXD without any information loss,and at the same time the Woodbury matrix identity is employed to reduce computational complexity in terms of the background data estimate mean,the covariance matrix and its inversion.Compared with cross track,the multiple parallel detection tracks are designed to simultaneously process all pixels of a data line in the new local real-time RXD since a line of data are collected at a time,therefore,the proposed algorithm has a good portability to FPGA,GPU and so on.Experimental results by real hyperspectral data show that LCF-GRXD and LCF-LRXD have several times speedup when remain basic detection accuracy2.The causal fast framework based on the RX detection algorithm only considers the first-order and second-order information,accordingly its detection accuracy is relatively low though it is more efficient.In contrast,kernel RX(KRX)anomaly detection algorithm projects data into a feature space,in which high-order information between bands can be employed,thereby leading to better detection accuracy.Inspired by KRX,a progressive pixel causal fast local KRX detector(LCF-KRXD)is proposed.Firstly,a local causal sliding array window is used to model the background statistics,ensuring the causality of the progressive pixel detection system.After that,using the matrix inversion lemma and the Woodbury matrix identity,both the current inverse covariance matrix and estimated mean can be recursively derived from previous moment without extensive repetitive calculation.Experimental results by real hyperspectral data show that LCF-KRXD has at least 49 times speedup when remain basic detection accuracy.3.Kernel collaborative representative detector(KCRD)implements the detection based on the theory that a background pixel can be represented by its neighboring data but an anomaly cannot.In contrast with the LCF-KRXD,KCRD has better detection accuracy,even though it's computational complexity is relatively higher.Accordingly,fast local KCRD algorithm based on a progressive pixel causal system(LCF-KCRD).According to the progressive pixel causal system,the kernel covariance matrix can be recurrsively represented via a recursive thought,which avoids its repeated calculation.The regulation matrix can be optimized if the kernel calculation between pixels are replaced by the complished anomaly values,in this case,the detection accuracy and efficiency are both improved.The inverse operations of matrix can be done through Cholesky decomposition,which further improve the detection efficiency.Experimental results by real hyperspectral data show that LCF-KCRD not only has a robust detection result,but also improve its detection efficiency compared with KCRD.