Support Vector Machines for Unmixing Geological Mixtures

摘要

In this paper, we describe a Support Vector Machine (SVM) based approach for spectral unmixing of intimate geological mixtures. In this approach, we use the spectral distance of a pixel from the bounding hyperplane for a given endmember as a measure of purity of the pixel with respect to that specific endmember. The approach is implemented by first identifying the pure pixels in the image using various algorithms for estimating pixel purity, and then using spectral similarity measures to identify the endmembers. The pure pixels are then used to train a series of SVMs for all endmembers using the "one-against-all" approach. The trained SVMs are then used to process all pixels, and the spectral distance of each pixel from the bounding hyperplane for all endmembers are estimated. The approach is demonstrated using a simulated and real-world hyperspectral data. The results indicate that our approach outperforms linear and bilinear spectral unmixing approaches.