A Study Of Hyperspectral Image Unmixing Based On Low-rank Representation

As an emerging technology, hyperspectral imaging is developed in the field of remote sensing in the early 1980’s and hyperspectral data has the characters of multi-band and high spectral resolution. Hyperspectral remote sensing has a good application prospect in the thematic map and relief map of drawing and update, environmental monitoring and disaster evaluation, detection and identification of rock ore, fine classification of crops and vegetation, and other fields. However, the widely existence of mixed pixels has become an outstanding problem in the hyperspectral remote sensing application.Therefore, how to effectively solve the problem of mixed pixels, has become one of the challenges faced by the hyperspectral image processing technology. Thus, this paper focuses on the hyperspectral image unmixing algorithm and the author’s major contributions are outlined as follows:1. A new abundance estimation algorithm via prior structured low-rank representation is proposed. The algorithm uses spectral library as endmembers. Compared with the size of the spectral library, the number of endmembers in a given scene is usually much smaller. Due to this fact, the abundance matrix is low-rank. Considering the high mutual coherence of the library signatures, we proposed the prior structured low-rank representation model which contains a dictionary pruning strategy to modify the original low-rank representation model. The proposed algorithm can not only better capture the spatial structure of abundance matrix but also mitigate the aforementioned effect. Thus the proposed algorithm improves the accuracy of unmixing.2. A new abundance estimation algorithm via spatial consistency based low-rank representation is proposed. This algorithm includes spatial information on the low-rank representation model by means of the spatial consistency regularizer. This regularizer accounts for spatial homogeneity: it is very likely that two neighboring pixels have similar fractional abundances for the same endmembers. And it also accounts for the boundary problem. Due to using both spectral information and spatial contextual information, the algorithm can achieve a better unmixing result.3. A new abundance estimation algorithm via endmembers consistency based low-rank representation is proposed. Compared with the number of endmembers in the whole image, the number of endmembers in part of the scene is usually smaller. Due to this fact, the algorithm uses sparse representation method to evaluate the contribution of different endmembers of the mixed pixels in different regions and includes it on the low-rank representation model by means of the endmembers consistency regularizer.Experimental results demonstrate the efficacy of the proposed algorithm.