Sub-pixel Target Detection From Hyperspectral Remote Sensing Imagery

Imaging spectrometer can provide images with very high spectral resolution by collecting many spectrally continuous images, named hyperspectral images, at the same time. Different materials are distinguishable by the spectral difference revealed in hyperspectral images. The spatial pattern of the images is also helpful to position the potential target with spectral difference from the background. These characteristics make hyperspectral images suitable for target detection task. However, due to the complex distribution of different ground objects and the limited spatial resolution of the hyperspectral images, a pixel in the hyperspectral images is usually composed of different land objects and the targets usually reside in the sub-pixel scale. Sub-pixle target detection is a difficulty for extracting information from hyperspectral images and has been of great interest to the researchers in target detection domain since the end of last century. This thesis focuses on the following problems in sub-pixel target detection from hyperspectral images:detection with low dimension data, target spectral variety, contamination caused by targets to the background statistics.The main research work and the corresponding contributions of this dissertation are as following:Target detection in hyperspectral images originates from the singal detection theory. So the singal processing theory is summarized. The signals received by the imaging spectrometer come from the array of multiple senors acquiring different radiances from the same ground scene. This is the array signal processing in hyperspectral images. Besides, the three most important methods for constructing the target detectors are detailed:generalized likelihood ratio test, constant false alarm rate detector and matched detector theory.Linear mixture model is the most widely used model for target detection in hyperspectral images. The physical basis, the description and the limitation of the model are summarized. Based on the linear mixture model, the linear spectral unmixing based target detection method is introduced. The fully constrained least squares method is used to solve the problem and an active/passive sets based numerical calculation method is researched on. Experiments show the limitation of the method. Besides, a constrained energy minimization method based on minimum noise fraction transformation (MNF-CEM) is proposed. In this method, minimum noise fraction (MNF) reduces the dimension of the hyperspectral imageries and separates the noise from the hyperspectral images. Then a finite impulse response (FIR) filter is used to detect the sub-pixel targets in the low dimension images. In this way, the computations of ill-conditional matrix inverse and virtual dimension of the hyperspectral imageries are unnecessary. Experiments show that this method can restrain the influence of the noise and is an effective sub-pixel target detection method for hyperspectral imageries.Subspace based methods can solve the spectral variety problem. Two subspace based detection methods are proposed in this paper. Traditional matched subspace detectors rely on the estimated endmembers and the according abundance to model the background and construct a generalized likelihood ratio test based detector. However, these endmembers and abundances have no physical meanings. So we use the abundances from the spectral unmixing so as to introduce the physically meaningful endmembers and abundances into the matched subspace detector. The proposed method is called spectral unmixing based matched subspace detector. The other method is the local OSP (Orthogonal Subspace Projection) method which uses the local neighbor pixels to represent the background information of the pixels under observation instead of the endmembers from the whole image. Experiments show that with physically meaningful endmembers and the according abundances the proposed matched subspace detector models the background better and performs better than its counterpart detector and the other OSP detectors; local OSP output a better result than OSP which proves that the pixel is related more with the local information than the whole set of pixels in the image, so it is more suitable to model the pixel's background with its neighboring pixels.To further investigate the performance with more accurate background information, two hybrid detectors are proposed. Both of them make use of abundances from spectral unmixing, and a background endmembers selection procedure is introduced into the detectors. The aim of the background endmembers selection is to use the correct kinds of background endmembers in the detection, while traditional methods use the same endmembers in the detection of each pixel. A structured background hybrid detector and an unstructured background hybrid detector are developed. First, the endmembers selection procedure is performed to choose the correct kinds of background endmembers in each pixel. Second, spectral unmixing is done with the endmembers selection information to get more accurate abundances. Then, the abundances are used to construct the structured background hybrid detector and the unstructured background hybrid detector and the endmembers selection information is taken into consideration in the structured detector again. Experiments show that the hybrid detectors with endmembers selection procedure perform better than the hybrid ones without endmembers selection procedure. It is concluded that with more accurate information to model the reality of the background ground objects the detection performance would be improved.Current anomaly detection methods are susceptiable to the anomalies in the background statistics. And they either use a local background statistics or a global statistics. A random selection based anomaly detector is proposed in this thesis. The method uses a random selection procedure to get background statistics from the image. Eachtime, not only the local neighboring pixles but also a number of blocks containting the same number of pixels are selected from random positions of the image. In this way, both local and global statistics of the pixel under observation are taken into consideration. The selection procedure is performed several times and each time the detection procedure is done successively with the constructed background statistics. Finally, the detection results are confused. Besides, the real-time version of the random selection based detector is implemented. Experiments show that the separability between targets and background by the proposed random selection based anomaly detector is larger than that by SSRX detector. And the performance of the real-time version detector also shows better detection and less false alarms on the boundaries of different backgrounds than current real-time anomaly detectors..