Research On The MVC-NMF Improved Algorithm For Endmembers Extraction Based On Hyperspectral Image

Hyperspectral remote sensing can obtain rich spectrum information, which can promote the development of feature recognition. Unfortunately, due to the impact of the spatial resolution of the sensor and the complexity of natural features, mixed pixels exist in the hyperspectral image, which can affect the accuracy of object recognition. The key of mixed pixels unmixing is endmember extraction, so this paper mainly studies the endmember extraction algorithm.In this paper, endmember extraction algorithms based on linear mixture model is summing up. And the comparison of the most frequently used endmember extraction algorithms are compared. Three kinds of endmember extraction algorithms and their advantages and disadvantages are introduced in the paper. The MVC-NMF algorithm is able to extract endmembers with relatively high precision, and at the same time, it can get the abundance of the endmembers. But the antinoise ability of MVC-NMF algorithm is weak, the algorithm need be improved.In order to improve the precision and the anti-noise ability, the MVC-NMF algorithm is still needed to improve the accuracy of the endmembers. MVC-NMF algorithm combines the least squares analysis and geometric convex. It combines the least squares analysis and geometric convex, and uses the interactive projection gradient algorithm to minimize the matrix of the objective function, at the same time, it can get the endmembers and abundance of the endmembers. The precision of the extractedendmembers is higher than that of the VCA algorithm and ATGPalgorithm. To improve the accuracy of the endmembers, MVC-NMF algorithm must consider the spatial arrangement of pixels and the sparse characteristic of abundance matrix. So the idea for improving precision of the extracting endmembers is based on searching spatial information in this paper, with separating and amplifying sparse decomposition by using Lagrange method. The method can obtain more effective the spectrum of endmembers in mixed pixels unmixing.In this paper, both simulated data and real hyperspectral image data are used for the experimental analysis of the endmembers extraction. Experiments with different signal to noise ratio(30db,40 db and 50db), different window sizes(3 × 3, 5 × 5 and 7×7)and different number of endmembers(6, 7, 8, 9 and10)were carried out respectively based on simulated data, experiments with different window sizes(3×3, 5 ×5 and 7×7)was carried out respectively based on real hyperspectral image data. And using six kinds of accuracy evaluation methods(SAD, SID, SRMSE, AAD, AID and ARMSE) to assess the accuracy of the endmembers extracting. The results of experimental show that:(1) the antinoise ability of improved MVC-NMF algorithm is better than that of the three algorithms;(2) the improved algorithm is more accurate than the MVC-NMF algorithm in the same experimental conditions;(3) the running time of the improved MVC-NMF algorithm is similar to the running time of the MVC-NMF algorithm. These three advantages prove that the MVC-NMF algorithm can obtain better endmember s than the MVC-NMF algorithm.