Research On Automatic Target Detection For Hyperspectral Imagery

Hyperspectral imagery has the unique characteristic of acquiring spectral andspatial information simultaneously, which brings the hyperspectral detectionadvantages when dealing with target detection problem under complex conditions.Hyperspectral target detection has become a charming research field in ATR in recent'years. Our research of hyperspectral target detection mainly focus on several keyissues, including how to describe the statistical characteristic of hyperspectral imagerydata more precisely, how to utilize the spatial information to enhance the detection,how to suppress clutter while keep the target signal and how to explore the implicitinformation besides the spectral and spatial information in hyperspectral data. Themain contributions of this dissertation are as follows:1. The likelihood ratio test in hyperspectral target detection is simplified tomerely testing background likelihood by virtue of maximum entropy criterion. Thesample number of spatial low. probability small target is scanty, which makes itdifficult to constraint the moments of target signal. Thus, the unknown statistics oftarget signal can be obtained in terms of maximum entropy criterion while introducingno unwarranted information. In this case, the hyperspectral target detection is reducedto merely test the single background likelihood.2. An anomaly target detector for hyperspectral imagery is constructed usingsamples-depending multimode model (SDMM) and single likelihood test. In the casethat the whole hyperspectral data is viewed as samples of background ensemble due tothe low probability of target sample, SDMM is used to estimate the density of thebackground, which is then combined with the single likelihood test to construct thedetector. Experiment results demonstrate the effectiveness of our detector to detect theanomaly target in hyperspectral imagery.3. A two-step detector, which is consists of orthogonal subspace projection andlikelihood ratio test, is presented in order to detect subpixel target in hyperspectralimagery. The detector synthesizes advantages of orthogonal subspace projectionmethod and likelihood ratio test, therefore enhancing.the detection performance.Effectiveness of our hyperspectral detector to detect the subpixel target is testified inexperiment results.4. The high-order moment of quadratic of multivariate random variable is used to separate the structured background adaptively, which let the residual data be anapproximation of spatial whitened Gaussian random process. Under this condition, thetarget shape subspace is constructed using a prior target shape information and thenthe 2-dimensional shape information is transformed to high-dimensional spectralfeature which is to be matched. Consequently, a shape-spectral target detector isderived. Experiment results show that our detector is effective to detect targets withdifferent shape and spectral signature.5. A figure combining probability density, curve estimated from detection valueand detection value of target is constructed to evaluate the performance of small targetdetector for hyperspectral imagery.6. In order to utilize the spatial scale information in the large target detectionproblem for hyperspectral data, the high-dimensional multiscale autoregressive modelis defined. Based on it, a detector that is used for detectingtarge target is constructed,utilizing Markov property of the high-dimensional markov random process, theequality between joint distribution of various multiscale observations and itsregression noise and multivariate t distribution statistics of the regression noise.Experiment results show that our detector is effective to detect large target inhyperspectral imagery.