Evaluation of spectral unmixing using nonnegative matrix factorization on stationary hyperspectral sensor data of specifically prepared rock and mineral mixtures
Hyperspectral sensors are used to identify materials via spectroscopic analysis. Often, the measured spectra consist of mixed materials and depending on the problem, the mixture ratio and the pure material spectra are wanted. In this paper, linear spectral unmixing is performed using the Nonnegative Matrix Factorization to analyze its correlation to ground truth data. The results are compared to Nonnegative Least Squares unmixing using manually selected endmembers from the image. Additionally, the effect of different endmember extraction algorithms and abundance initialization methods for NMF are investigated. To test the validity of the method, several checkerboard patterns of different ground minerals/rocks with predefined mixtures were prepared. It was shown that good initialization is beneficial in terms of approximation error and correlation to ground truth.
展开▼