Real-time Target Detection Algorithms For Hyperspectral Imagery

With the development of imaging spectroscopy and the improvement of spectral resolution,hyperspectral imaging has received more and more attention.Hyperspectral imagery could distinguish subtle difference of ground objects and detect targets that can hardly or even not be detected by multi-spectral imagery,with the help of abundant spectral information.Anomaly detection(AD)has been a major task in practical use since it can uncover targets without any priori knowledge.In applications of anomaly detection real time processing is particularly important and crucial.This is because many anomalies,such as moving targets,may not stay too long and the duration of their presence is very short.Most importantly,they may show up suddenly and instantly,then disappear quickly afterwards.Therefore,for an algorithm to be able to detect these targets in a timely fashion,the process must be real time.On the other hand,hyperspectral imagery is giving more and more detailed spectral information as the technique develops,but at the same time,the ever-expanding hyperspectral dataset has also brought tremendous pressure and challenges to data storage,transmission and further processing.Real-time processing could reduce these pressures and decrease information loss caused by data compression and transmission,and also enhance detection and classification efficiency.This dissertation considers how to realize real-time anomaly detection for hyperspectral imagery using modern signal processing techniques and detection theory,and has a strong focus on the speed and efficiency characteristics.The main innovation contributions and important research findings of this dissertation are as follows.Firstly,despite the fact that there are many real-time processing algorithms have been proposed for anomaly detection in the literature,technically speaking they are not true realtime processing algorithms but actually fast computational algorithms which can process data without much time delay.A true real-time detector should meet the following requirements: 1)a real time processing algorithm must be carried out “causally” in the sense that the data samples used for data processing should be only those up to the data sample vector currently being processed.In other words,no future data sample vectors after the current data sample vector can be allowed for data processing;2)a true real-time processing algorithm must produce its output at the same time as an input comes in.It should be near real-time with theassumption that the processing time is negligible.However,form a practical appoint of view such a time delay is determined by a specific application.For example,in surveillance and reconnaissance applications,finding moving targets is imminent for decision-making and the responding time must be very short.In this case,very little time delay should be allowed.As another example,for applications of crop monitoring in agriculture,the time to respond can be minutes or hours.Far more interestingly,for environmental changes in ecology,the allowable time delay can be longer even to days and months.So,once an algorithm can meet a required time constraint,it can be considered as a real-time processing algorithm.Secondly,the commonly used RX detector is actually the Mahalanobis distance between the background and pixel to be detected,which needs to calculate the mean and covariance statistics of the entire set of data sample vectors,which certainly cannot be implemented in real time in a causal manner.To resolve this issue,a new real-time algorithm is proposed for hyperspectral anomaly detection.It is an innovational process derived from the concept of Kalman filtering by updating needed processing information only provided the currently being processed pixel via an iterative equation.The recursive equation is derived using Woodbury's identity.In this case,this algorithm could largely reduce the processing time and decrease needed storage space without reducing the detection effect.Thirdly,it is crucial that sometimes the global detectors are not ideal for small targets and local anomalies.So the global real-time detector proposed above is not suitable for this kind of anomalies and a local detector with real-time capability is needed.A sliding causal matrix window is firstly presented.But it is time consuming due to the need of repeated counting of local mean and covariance matrix while the sliding window moves.A sliding causal array window with recursive characteristic is then proposed to solve this issue.By the derivation of recursive equation,this new algorithm could realize the local detection in a realtime and causal manner which is supposed to be the local real-time causal anomaly detection algorithm.Then,in real-time processing procedure,background suppression is very important because in real world applications with no availability of groundtruth,its effectiveness is generally performed by visual inspection.In this case,the detection performance depends on not only gray level intensity,but also contrast between background and detected targets.Thispaper proposes an adaptive background smoothing method for a better background suppression,and proposes ARTCRXD and ARTCCEM algorithms for both anomaly detection and target detection with target of interest.Finally,the above algorithms are based on a pixel-by-pixel processing procedure which is suitable for Whiskbroom Imaging Spectrometer.However,hyperspectral data from Framing Imaging Spectrometer is no longer obtained pixel-by-pixel but band-by-band.In order to solve this problem,a progressive band processing method is proposed for Framing Imaging Spectrometer,PBCEM and PBRXD are proposed for both anomaly detection and target detection with target of interest,by which the detector could get the progressive results with the data acquisition procedure.