Sparsity Based Target Detection For Hyperspectral Remote Sensing Imagery

With the characteristic of the high spectral resolution, hyperspectral images (HSI) contain hundreds and even thousands of spectral bands, and convey abundant spectral information, which can distinguish subtle spectral differences even between the similar materials, providing unique advantage for target detection. However, based on the application and practice, although there are many target detection methods in HSI, several problems are still not well solved, including the background estimation problem in anomaly detection, the spectral representation model problem in sub-pixel target detection, and the correlation management problem in the full-pixel target detection.This paper aims to investigate the characteristic of the HSI, focuses on the sparsity characteristic, including the spectral sparseness and the sparse number of the targets. Based on these, this paper proposed several sparsity-based target detection methods for hyperspectral remote sensing images. In detail, for the anomaly target detection without prior target spectral information, the low-rank and sparse matrix decomposition technique was employed for the background estimation problem via combing the background low-rank characteristic and the sparse number of the anomaly targets. Then, with prior target spectral information, there are two kinds of target detection problems based on the target size. For the sub-pixel target detection, the binary hypothesis test was applied to construct the sparse representation model for the sub-pixel target by combining the spectral sparseness and the sparse number of the targets. And for the full-pixel target detection, the multi-task learning technique was introduced to deal with the band correlation problem in HSI through combining the spectral sparseness and the correlation between single bands. For the above three kinds of target detection problem, several hyperspectral remote sensing images were accordingly conducted to validate the effectiveness of the proposed methods.At first, the general design process of target detection in the hyperspectral remote sensing image was introduced, as well as the evaluation criteria. Then, we accordingly introduced typical methods for the anomaly target detection, the sub-pixel target detection and the full-pixel target detection.This paper proposed a low-rank and sparse matrix decomposition-based mahalanobis distance method for hyperspectral anomaly detection (LSMAD). This approach:(1) takes full advantage of the low-rank and sparse matrix decomposition (LRaSMD) technique to set the background apart from the anomalies; (2) explores the low-rank prior knowledge of the background to compute the background statistics, which can simultaneously alleviate the anomaly contamination and the inverse covariance matrix problem; and (3) applies the Mahalanobis distance differences to detect the probable anomalies. Extensive experiments were carried out on three hyperspectral images, and it was found that LSMAD method shows a better detection performance than the current state-of-the-art hyperspectral anomaly detection methods. Moreover, compared with the sparse component-based method, LSMAD method is less sensitive to the sparse parameter.Then, a new sparse representation based binary hypothesis model for hyperspectral sub-pixel target detection (SRBBHD) was proposed. The proposed approach:(1) proposes a sparse representation model based on the the binary hypotheses, where the test sample can be sparsely represented by the training samples from the background-only dictionary under the null hypothesis and the training samples from the union dictionary under the alternative hypothesis; (2) select few samples form the background-only dictionary or union dictionary to sparse represent the test sample, which lead to a competition between the background-only subspace and the target and background union subspace; (3) constructs the detector by comparing the reconstruction residuals under different hypotheses. Extensive experiments were carried out on two hyperspectral images, which reveal that the SRBBHD method shows an outstanding detection performance. What's more, the SRBBHD can improve the spectral representation performance compared to the traditional sparsity model.Finally, this paper proposed a joint sparse representation and multi-task learning method for the full-pixel target detection (JSR-MTL). This approach (1) takes full advantage of the hyperspectral image's similarity via the band cross-grouping strategy to construct multiple detection tasks; (2) based on the spectral sparseness, combines the MTL and the sparse representation technique to learn an elaborate spectral representation model; (3) applies the total reconstruction errors differences accumulated over all the tasks to detect the probable targets. Extensive experiments were carried out on three hyperspectral images, and it was found that JSR-MTL can employ the spectral similarity within the HSI to achieve a better detection performance than the current state-of-the-art hyperspectral target detection methods.This paper indicated that, based on the sparsity of the hyperspectral remote sensing image, these proposed target detection methods can specially overcome the problems within the existing detection methods, and improve the detection performance for the anomaly target without prior target spectral information, the sub-pixel target and the full-pixel target with priori target spectral information.