Hyperspectral Linear Subspace Estimation For Anomaly Detection

The technology of anomaly detection from hyperspectral imagery increasingly becomes a focus in hyperspectral image analysis, since an anomaly detector enables one to detect anomalous targets without any prior knowledge about target signatures. Hyperspectral subspace estimation is one of the approaches that accounts for the phenomenon of performance and efficiency decrease caused by huge data volumes in hyperspectral images. For application of anomaly detection, representing hyperspectral data in a low- dimensional linear subspace can yield gains in computational time and complexity, and improve the detection performance by reducing mixture of spectrums correlation between adjacent bands simultaneously.Dimension reduction by linear transformation is one type of subspace estimation method which is most widely used for analysis of hyperspectral imagery. Few of the presented subspace estimation techniques explicitly aims to preserve anomaly vectors with different structure from background subspace, and it makes individual data-vector miss-representations which induces incorrect detection results in the identified subspace. Thus we deal with the problem of preserving the anomaly vectors with the reduced dimensional subspace which can represent background subspace structure using stochastic subspace estimation paradigm. The main work of this paper can be summarized as:(1) We analyze the PCA-based subspace estimation method under the linear mixture model of hyperspectral imagery, and discuss the influence of subspace dimensionality on performance of two statistical anomaly detectors RX and PW-LRT. Then we propose a method for calculating threshold by kernel density estimation which can diminish the detection error caused by assuming statistical model.(2) We prove the two subspace estimation method allowing for nose, MNF and HySime, is optimal F-norm of data misrepresentations. Since the limitation of F norm for representing anomalies, we introduce the MOCA and its 2,?l-norm optimal projection idea. Then a modified m-HySime method is proposed for subspace estimation with robustness to noise and preserving anomalies. The experiments results validate the effectiveness of proposed method and it improves performance of anomaly detectors.(3) Based on analysis of High-order statistics(HOS) non-orthogonal projection, A novel algorithm labeled as Hybrid Indexes for Iterative Projection(HIIP) is proposed to obtain the low-rank subspace. This new method takes the non-orthogonal projections as complements of the orthogonal directions obtained by PCA into consideration. The results of experiments on real PHI data prove that it is suitable for dimensionality reduction on hyperspectral images with small target and obtains more complete signal subspace compared with single projection index.