Exploitation of Intra-Spectral Band Correlation for Rapid Feature Selection, and Target Identification in Hyperspectral Imagery

摘要

This research extends the work produced by Capt. Robert Johnson for detecting target pixels within hyperspectral imagery (HSI). The methodology replaces Principle Components Analysis for dimensionality reduction with a clustering algorithm which seeks to associate spectral rather than spatial dimensions. By seeking similar spectral dimensions, the assumption of no a priori knowledge of the relationship between clustered members can be eliminated and clusters are formed by seeking high correlated adjacent spectral bands. Following dimensionality reduction Independent Components Analysis (ICA) is used to perform feature extraction. Kurtosis and Potential Target Fraction are added to Maximum Component Score and Potential Target Signal to Noise Ratio as mechanisms for discriminating between target and non-target maps. A new methodology exploiting Johnson's Maximum Distance Secant Line method replaces the first zero bin method for identifying the breakpoint between signal and noise. A parameter known as Left Partial Kurtosis is defined and applied to determine when target pixels are likely to be found in the left tail of each signal histogram. A variable control over the number of iterations of Adaptive Iterative Noise filtering is introduced. Results of this modified algorithm are compared to those of Johnson's AutoGAD 2007.