Abstract

In hyperspectral imaging, identification and detection tasks are often done by comparing the spectrum of an unknown material to labeled spectra in a spectral library. However, it is intractable to populate a spectral library with measurements of all possible materials in all possible conditions. We propose modeling the weights resulting from matrix decomposition as functions of material condition(s) to generate synthetic spectra at unmeasured conditions to augment spectral libraries. Matrix decompositions considered here include principal component analysis (PCA), non-negative matrix factorization (NMF), and independent component analysis (ICA).We demonstrate this method by generating (1) synthetic long-wave infrared (LWIR) emissivity spectra of calcite conditioned by grain size and (2) synthetic emissivity spectra of mixtures of olivine and nontronite in the visible to short-wave infrared (VNIR-SWIR), conditioned by mixture fraction and grain size. For the calcite spectra, modeling the weights of two components derived from PCA produced reconstructions with an average sample-wise RMSE of 0.0055-for comparison, the RMS difference between the spectra of the largest and smallest grain size samples was 0.25, and the physical range of emissivity is 0 to 1. Using NMF and four components, the olivine-nontronite mixture spectra were reconstructed with an average sample-wise RMSE of 0.0054, outperforming the linear mixture model (average RMSE of 0.21)-for comparison, the RMS difference the between pure nontronite and olivine spectra was 0.31. Because both the weight and the spectral shapes embedded in the components can be readily examined, this method offers an interpretable way of predicting the spectrum of a material at an unmeasured condition.

Degree

College and Department

Ira A. Fulton College of Engineering; Civil and Construction Engineering

Rights

https://lib.byu.edu/about/copyright/