Hyperspectral Images Unmixing Based on Abundance Constrained Multi-Layer KNMF

Due to the low spatial resolution of sensors, hyperspectral images contain mixed pixels. The purpose unmixing is decompose pixels into a series endmembers and abundance fractions. In order improve performance nonlinear algorithm for images, method, i.e., constrained multi-layer kernel non-negative matrix factorization (AC-MLKNMF), presented. Firstly, MLKNMF presented iteratively structure, then AC-MLKNMF based on by adding sparseness constraint total variation regularization characterize piecewise smooth structure maps according distribution characteristics actual ground-objects. Experimental results synthetic real datasets show that proposed can accuracy compared with single-layer KNMF, it also superior factorization, KNMF without pure pixels, sparse NMF, MLKNMF.