Anomaly Detection In Hyperspectral Remote Sensing Imagery:From Robust Background Modeling And Machine Learning Aspects

Hyperspectral remote sensing image data analysis technique has become one of the important research issues in the domain of remote sensing science.With the benefit of imaging spectrometer technique which was raised up in 1980s,hyperspectral remote sensing imagery is capable of obtaining abundant and discriminative spectral information of the land-covers or objects on the earth surface,and providing potent benefit for recognizing or detecting target on the ground surface.Hyperspectral anomaly detection technique,which is proceeded in an unsupervised way,has the ability of detecting the prominent spectrally unique target just through the analysis of hyperspectral image inherent data structure,with no need of the prior target spectral curve.This technique has important research significance and value in many practical application domains,such as rare mineral survey,defense investigation,agricultural monitoring,etc.For detecting the anomaly targets,background(non-anomaly target)suppression is uaually the most important issue for hyperspectral anomaly detection.However,there are many difficult problems in effective background suppression for hyperspectral anomaly detection:1)First,hyperspectral imagery usually has complex background data distribution in the hyperspectral data space,background distribution estimation in simple manners may face difficulties in separating background and anomaly targets.Meanwhile,current hyperspectral anomaly detection methods usually inject anomaly target information in the background distribution step,which will derive inaccurate background distribution estimation result;2)Second,there usually exist phenomena of spectral similarity and nonlinear spectral mixing in hyperspectral imagery.Spectral similarity phenomenon will result in similar spectral characteristics in some spectral band ranges between anomaly targets and some kinds of background land-covers.Meanwhile,nonlinear spectral mixing phenomenon will give rise to complex background features in hyperspectral imagery.These two phenomena both bring challenge for effective hyperspectral background feature learning;3)Moreover,high data redundancy in hyperspectral imagery will affect the diversity decision between background and anomaly information for hyperspectral anomaly detection.The two common used manners to suppress the background information in hyperspectral anomaly detection,which are background distribution estimation and background feature learning,will result in inaccurate background estimation or fail in learning the representative background features by the reason of high data redundancy in hyperspectral imagery.In order to overcome the aforementioned problems with accurately and adequately excavating the background information in hyperspectral imagery,robust background modeling in hyperspectral imagery for hyperspectral anomaly detection,which is combined with a various of effective methods in machine learning domain,is discussed in this dissertation.The main innovation points in this dissertation are presented as follows:(1)Due to the problems that hyperspectral background data distribution in the hyperspectral data space is complex,and the background estimation step is usually contaminated by anomaly target information,this dissertation propose two robust background distribution estimation methods for hyperspectral anomaly detection:robust nonlinear iterative anomaly detector and robust background regression anomaly detector.These two methods both apply the kernel machine learning method to analysis the nonlinear characteristics in hyperspectral imagery for enhancing the separability between hyperspectral background and anomalies.These two methods respectively adopt iterative estimation and data regression manners to establish the robust background estimation process;(2)Due to the problems that there usually exist spectral similarity and nonlinear spectral mixing phenomena in hyperspectral imagery,which will bring challenge for learning effective background features in hyperspectral anomaly detection,this dissertation propose two hyperspectral robust background feature learning methods for anomaly detection.These two methods respectively apply two machine learning methods:slow feature analysis and manifold learning to hyperspectral background feature learning,to effectively suppress the background features and extrude the anomaly targets in the hyperspectral imagery at the same time.These two machine learning methods both adopt nonlinear feature learning manners to exploit the effective feature space which could accurately represent the background land-cover characteristics;(3)Due to the problems that the two common used manners to suppress the background information in hyperspectral anomaly detection,which are background distribution estimation and background feature learning,will result in inaccurate background estimation or fail in learning the representative background features by the reason of high data redundancy in hyperspectral imagery,with the perspective of data compressed sensing and elimination of the data redundancy,in this dissertation,sparse learning method,which is an effective machine learning method,,is adopted to obtain a data redundancy elimination representation for hyperspectral imagery.From this,the sparse dictionary,which could indirectly represent the information in the hyperspectral imagery,will be generated.Based on this,a hyperspectral anomaly detection method which indirectly analysis the background and anomaly information content in hyperspectral imagery with the sparse dictionary is proposed.Meanwhile,two sparse dictionary enhancement processing for hyperspectral anomaly detection are proposed to enhance the diversity between background and anomaly information in the hyperspectral sparse dictionary.