Dimensionality Reduction on Hyperspectral Data for Solids Analysis

Dimensionality Reduction on Hyperspectral Data for

Solids Analysis

Annalisse BoothUtah State University

Electrical and Computer Engineering DepartmentResearch Experience for Undergraduates 2009

Hyperspectral Imaging: An Overview

Source: http://www.yellowstoneresearch.org

• Records information across electromagnetic spectrum

• Spectral band correlates to certain range of wavelength

• Bands combined to form cube

• Hundreds to thousands of bands per cube

• 258 bands in current data

January 11, 2008 17:41:25, wavelength 46

Solids Hyperspectral Data

• 3 months data

• Camera on tripod, but shaken

• Cleaned up by Mckay

• Turned into video, RGB approximations

• Wrote other applicable codes

Gathering Tools for Analysis

An example of a Locally Linear Embedding (LLE)

• Multidimensional Scaling (MDS)• Principle Component Analysis (PCA)• Locally Linear Embedding (LLE)• Isomap (weighted geodesic distances)• Maximum Variance Unfolding (MVU)

Comparing Techniques

Source: Boundary Constrained Manifold Unfolding. Bo, Hongbin, Wenan. 2008.

Comparing Techniques

Source: Boundary Constrained Manifold Unfolding. Bo, Hongbin, Wenan. 2008.

Comparing Techniques

Source: Boundary Constrained Manifold Unfolding. Bo, Hongbin, Wenan. 2008.

Work Still Uncompleted

• Write program to choose pixels from each substance through time

• Compare pixels of each substance to self and other substances

• Analysis in Isomap for preliminary results

• Write code for Riemmanian Manifold Learning (RML)

• Execute code on data

• Write code for Boundary Constrained Manifold Unfolding

• Execute new code, compare