Date of Award

Summer 8-2012

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Geography and Geology

Committee Chair

Dr. Carl Reese

Committee Chair Department

Geography and Geology

Committee Member 2

Dr. Bandana Kar

Committee Member 2 Department

Geography and Geology

Committee Member 3

Dr. George Raber

Committee Member 3 Department

Geography and Geology

Committee Member 4

Dr. David Cochran Jr.

Committee Member 4 Department

Geography and Geology

Committee Member 5

Dr. Gregory Carter

Committee Member 5 Department

Geography and Geology

Abstract

This dissertation uses post storm imagery processed using wavelet transforms to investigate the capability of wavelet transform-based methods to classify post storm damage of residential areas. Five level Haar, Meyer, Symlets, and Coiflets wavelet transform decompositions of the post storm imagery are inputs to damage classification models of post hurricane and tornado damage. Hurricanes Ike, Rita, Katrina, and Ivan are examined as are the 2011 Joplin and Tuscaloosa tornadoes.

Wavelet transform-based classification methods yielded varying classification accuracies for the four hurricanes examined, ranging from 67 percent to 89 percent classification accuracy for classification models informed by samples from the storms classified. Classification accuracies fall when the samples being classified are from a hurricane not informing the classification model, from 17 percent for Rita classified with an Ike-based model, 41 percent for Rita classified with an Ike-Katrina-based model, to 69 percent for Rita classified with an Ike-Katrina-Ivan-based model.

The variability within and poor classification accuracy of these models can be attributed to the large variations in the four hurricane events studied and the significant differences in impacted land cover for each of these storms.

Classification accuracies improved when these variations were limited via examination of residential areas impacted by 2011 Joplin and Tuscaloosa tornadoes. Damage classification models required as few as nineteen to as many as fifty nine wavelet coefficients to explain the variability in the hurricane storm data samples, and included all four wavelet functions studied. A similar analysis of the tornado damaged areas resulted in a damage classification model with only six wavelet coefficients, four Meyer-based, one Symlets-based and one Haar-based. Classification accuracies ranged from 96 percent for samples included in the model formation to 85 percent for samples not included in the model formation.

The damage classification accuracies found for tornado storms suggests this model is suitable for operational implementation. The damage classification accuracies found for the hurricane storms suggests further investigation into methods that will reduce the variability attributable to land cover and storm variability.

Share

COinS