A sparse Gaussian sigmoid basis function approximation of hyperspectral data for detection of solids

• Published in: Statistical analysis and data mining Vol. 12; no. 6; pp. 489 - 495
• Main Authors: Lanker, Cory, Smith, Milton O.
• Format: Journal Article
• Language: English
• Published: Hoboken Wiley Subscription Services, Inc., A Wiley Company 01.12.2019; Wiley Subscription Services, Inc; Wiley Blackwell (John Wiley & Sons)
• Subjects: basis function decomposition; Basis functions; classification; curve fitting; feature extraction; LASSO regression; minerals; Reflectance curves; Spectra; spectral characterization; target detection
• ISSN: 1932-1864; 1932-1872
• DOI: 10.1002/sam.11433

We define a new characterization of emissivity and reflectance curves for compositional exploitation of hyperspectral data. Our method decomposes each spectrum into a sparse set of Gaussian sigmoid components using penalized regression. Detection is based on the combination of Gaussian sigmoid components unique to a target material. Focusing on the presence of spectral upslopes and downslopes rather than spectral correlations makes detection more robust to both target variation and spectral variability from atmosphere and background encountered during the collection process. We present simulation studies that demonstrate the potential to reduce false positive rates without compromising sensitivity. Characterization of long‐wave infrared (LWIR) experimental data validates our method using minerals of different particle sizes, measurement angles, and collection conditions.