Hyperspectral mixed pixel decomposition policy merging simple linear iterative clustering

Abstract: The decomposition efficiency of mixed pixels is an important subject in the research of remote sensing application. Current mature extraction algorithms of the endmember are pure pixel index (PPI), N-FINDR, vertex component analysis (VCA), sequential maximum angle convex cone (SMACC), alternating volume maximization (AVMAX), minimum volume enclosing simplex (MVES), etc. Extracting pure spectra from all pixels of an image by those algorithms has relatively slow processing and low accuracy. Therefore, in this paper, we introduced an integration of simple linear iterative clustering (SLIC) super pixel segmentation of hyperspectral mixed pixel decomposition algorithm. The number of pixels in a hyperspectral image was one of the main reasons that affected pixel unmixing (especially endmember extraction). Super pixel segmentation techniques can compose adjacent pixels with similar characteristics into image blocks, and retain useful information for further image processing, thus significantly reducing the number of pixels involved in endmember extraction and solving the problems of low processing and low accuracy in an effective way. In this paper, endmember extraction method was modified. We used Simple Linear Iterative Clustering (SLIC) algorithm to segment hyperspectral image pixels into super pixel sets, and conducted experimental comparisons on impact of different dimension reduction methods such as principal components analysis (PCA), minimum noise fraction (MNF), correspondence between different RGB (six kinds), different color spaces (RGB, LAB), different data formats（JPG, BIN）and different algorithms' parameters K on hyperspectral image super pixel segmentation results. Furthermore, the impact of SLIC super pixel segmentation results on two typical endmember extraction algorithms was analyzed. The results showed that, with the increase of K value, the decomposition of mixed pixels gradually increased. The reason for such change was that as the K value increased, the number of pixels increased, and the size of the image was reduced and was close to a normal pixel. The JPG (lossy compression) data format of time was shorter than BIN (lossless compression) data format because a lossy compression can reduce the image size occupied in memory or disk space. The root-mean-square error (RMSE) of SLIC+MVES was slightly higher than that of MVES, and lower than that of AVMAX, with the time much less than actual MVES. In other words, relative to MVES, SLIC+MVES greatly increased processing speed, at the expense of a small amount of sacrifice in accuracy. AVMAX consumed little time itself because it did not need to solve linear programming problems repeatedly. So after adding super pixel segmentation, the time shortened for AVMAX was not significant, but the time for SLIC computing had increased. When increasing K, the computation time increased gradually, and the RMSE was flat or deceasing. When K was large enough, the results of SLIC+MVES was approximate to that of MVES. The difference in the RGB color space and Lab color space was not obvious. SLIC algorithm used Lab color space because it did not rely on the advantages of the device, but also had its own advantages, i.e., a wide color gamut. In most cases, dimensionality reduction by MNF was better than that of PCA. As to the MNF dimensionality reduction method, the data in JPG format, with Lab color space were more favorable for unmixing results. In addition, the value of SLIC parameters from 5 to10 was more appropriate. The SLIC super pixel segmentation algorithm was simple and improved the efficiency of unmixing, with a great practical value.