Algorithm Development for Hyperspectral Anomaly Detection

Loading...
Thumbnail Image

Files

umi-umd-5691.pdf (3.69 MB)
No. of downloads: 1596

Publication or External Link

Date

2008-08-12

Citation

DRUM DOI

Abstract

This dissertation proposes and evaluates a novel anomaly detection algorithm suite for ground-to-ground, or air-to-ground, applications requiring automatic target detection using hyperspectral (HS) data. Targets are manmade objects in natural background clutter under unknown illumination and atmospheric conditions. The use of statistical models herein is purely for motivation of particular formulas for calculating anomaly output surfaces. In particular, formulas from semiparametrics are utilized to obtain novel forms for output surfaces, and alternative scoring algorithms are proposed to calculate output surfaces that are comparable to those of semiparametrics. Evaluation uses both simulated data and real HS data from a joint data collection effort between the Army Research Laboratory and the Army Armament Research Development & Engineering Center. A data transformation method is presented for use by the two-sample data structure univariate semiparametric and nonparametric scoring algorithms, such that, the two-sample data are mapped from their original multivariate space to an univariate domain, where the statistical power of the univariate scoring algorithms is shown to be improved relative to existing multivariate scoring algorithms testing the same two-sample data. An exhaustive simulation experimental study is conducted to assess the performance of different HS anomaly detection techniques, where the null and alternative hypotheses are completely specified, including all parameters, using multivariate normal and mixtures of multivariate normal distributions. Finally, for ground-to-ground anomaly detection applications, where the unknown scales of targets add to the problem complexity, a novel global anomaly detection algorithm suite is introduced, featuring autonomous partial random sampling (PRS) of the data cube. The PRS method is proposed to automatically sample the unknown background clutter in the test HS imagery, and by repeating multiple times this process, one can achieve a desirably low cumulative probability of taking target samples by chance and using them as background samples. This probability is modeled by the binomial distribution family, where the only target related parameter--the proportion of target pixels potentially covering the imagery--is shown to be robust. PRS requires a suitable scoring algorithm to compare samples, although applying PRS with the new two-step univariate detectors is shown to outperform existing multivariate detectors.

Notes

Rights