Robust Estimation of Mahalanobis Distances in Hyperspectral Images
Abstract
This dissertation develops new estimation methods that fit Johnson distributions and generalized Pareto distributions to hyperspectral Mahalanobis distances. The Johnson distribution fit is optimized using a new method which monitors the second derivative behavior of exceedance probability to mitigate potential outlier effects. This univariate distribution is then used to derive an elliptically contoured multivariate density model for the pixel data. The generalized Pareto distribution models are optimized by a new two-pass method that estimates the tail-index parameter. This method minimizes the mean squared fitting error by correcting parameter values using data distance information from an initial pass. A unique method for estimating the posterior density of the tail-index parameter for generalized Pareto models is also developed. Both the Johnson and Pareto distribution models are shown to reduce fitting error and to increase computational efficiency compared to previous models.
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 01, 2006
- Accession Number
- ADA462686
Entities
People
- Eduardo C. Meidunas
Organizations
- Air Force Institute of Technology