Research On Target Detection Method For Hyperspectral Remote Sensing Image

Hyperspectral image(HSI) obtained by remote sensing systems has provided significant information about the spectral characteristics of the materials in the scene. Typically, a hyperspectral spectrometer provides hundreds of narrow contiguous bands(about 10 nm wide) over a wide range of the electromagnetic spectrum. Different materials are usually assumed to be spectrally separable as they reflect electromagnetic energy differently at specific wavelengths. This property enables discrimination of materials based on the reflectance spectrum obtained by HSI. HSI has found many applications in various fields such as military, agriculture, and mineralogy. One of the most important applications of HSI is target detection(TD), which can be viewed as a binary classifier with the aim of labeling pixels as target or background based on their spectral characteristics.There are two kinds of methods for TD of HSI: the methods based on probability statistics and subspace model and the methods based on sparse representation model. The first kind of method needs the target spectra as prior information. But when the target spectra known as priori knowledge are inaccurate, the performance degrades significantly, which is a common problem with this kind of method. The sparse representation(SR) based method is proposed in recent years. Researchers should first construct a meaningful dictionary, and then solve the SR problem, finally they compare the residuals between the target and background reconstruction results to get the detection outputs. How to construct the dictionary and how to use the sparse coefficients to get the outputs are the main concerns in this field. This thesis mainly contains three aspects:(1) An improved method used for target detection in hyperspectral imagery is proposed. This method can raise target spectra accuracy, so the performance of the target detection methods can be improved. By using the target spectra gotten from the lab as references, the proposed method extracts independent components, which are the closest to the references, from the hyperspectral imagery by means of independent component analysis with references(ICA-R). Then, these independent components are used as target spectra in the following supervised target detection methods. Experimental results on both simulated and real hyperspectral data demonstrate that the proposed method can get more accurate target spectra, which obtains much better performance of target detection.(2) An improved target detection method based on SR is proposed. The spatial-spectral dictionary, which contains both spatial and spectral information, is the first time to be used in hyperspectral target detection. This method also overcomes the problems in the dictionary construction, dictionary partition and the residual comparison steps in the classical SR method. Different to the traditional SR model, this method does not compare the residuals between the target and background reconstruction results. The spatial-spectral dictionary learning method is used to get both the dictionary and the sparse coefficients. Then the sparse coefficients are used as new signatures for each pixel in the SVM classifier.(3) An improved method based on the endmember extraction to construct the dictionary is proposed. This method can get pure background dictionary and target dictionary, which contributes to a more accurate residual comparison. This method uses the endmember extraction algorithm to extract the endmembers from a HSI. The endmembers are used as atoms in the background dictionary. The target spectra known as prior knowledge are used as atoms in the target dictionary. Finally the traditional SR model or the proposed new SR model is used for target detection.