Hyperspectral surface reflectance reconstruction based on non-negative matrix factorization and multispectral results

Different from the principal component analysis (PCA), non-negative matrix factorization (NMF) can provide more direct interpretation owning to the non-subtractive combinations of non-negative basis vectors, and many practical problems also require non-negative basis vectors rather than the orthogonal vectors with alternating positive and negative. In this study, we develop a hyperspectral surface reflectance reconstruction method based on NMF and multispectral results in several wavelength bands. In order to test our spectral reconstruction method, the spectral datasets of typical surface types are extracted from the spectral library of John Hopkins University (JHU), which include the soil, vegetation, manmade materials, sedimentary fine and coarse rock. The prior surface reflectance or emissivity results are selected from only four wavelength bands (2.13, 3.75, 3.96, 4.05 μm) from shortwave infrared to Mid-infrared, which can be easily obtained from the surface product of Moderate-resolution Imaging Spectroradiometer (MODIS). Based on the JHU spectral dataset and NMF, the hyperspectral surface reflectance in the spectral range of 2-5μm with the step of 25 nm can be reconstructed consistently. In addition, the hyperspectral reconstruction effects by NMF are quantitatively investigated, in which the root mean square error and the mean absolute error is about 0.016 and 0.01, respectively.