Bayesian Nonparametric Methods for Material Identification from Large Remotely Sensed Hyperspectral Space-Time Data Sets

• Activity Number: 166
• Type: Topic Contributed
• Date/Time: Monday, July 30, 2012 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics and the Environment
• Author(s): Candace Berrett*+
• Companies: Brigham Young University
• Keywords: signal processing ; spatial dependence ; remote sensing ; image analysis

Abstract

Hyperspectral images are large 3-D data cubes containing observations over a wide range of spectral wavelengths at each pixel of the image. Ideally, the observed spectrum at each pixel would exactly match the true spectral signal of the material being captured; however, the observed spectrum is a mixture of many interacting signals and measurement error. The goal is to use this large, noisy and messy data to determine spectral and emissivity fingerprints for each material. One mechanism for doing this is to first reduce the amount of data by clustering the pixels by similar spectra, and then determine the underlying signal within each cluster by accounting for environmental and measurement noise. The flexibility and feasibility of Bayesian nonparametric methods make them an ideal tool for doing this. Combining this stochastic process with the physical process of Plank's Law, and allowing for spatial and temporal dependence, we use a large dataset containing 30 different materials, 258 wavelengths for each pixel, observed across a period of 3 months, to cluster and identify posterior distributions of spectra associated with each material.