The current protocols in fire debris analysis rely on ignitable liquid pattern recognition and the identification of target compounds. These practices allow fire debris analysts to determine whether a sample contains or is absent of ignitable liquid residue and to classify that type of ignitable liquid based upon subjective thresholds. A statistical approach using computationally mixed datasets was devised in this research to generate an objective approach to the classification of fire debris samples and to determine error rates. Quadratic, linear, and partial least squares linear discriminant analysis (QDA, LDA, and PLS-DA) models were developed and evaluated to determine the effects of different population distributions and the overall performance on known and unknown ground-truth fire debris samples. The evaluation of the method performance was determined by generating receiver operator characteristic (ROC) curves and calculating an area under the curve (AUC) based on log-likelihood ratios (LLR) scores. The Substrate Database was also expanded within this work to aid fire debris analysts in casework and to provide a more relevant population when generating statistical models. The most optimal population distribution was determined to consist of a uniform population of equal contribution of substrate and ignitable liquids, with each ignitable liquid class represented. The QDA model performed the best when evaluating the cross validation datasets, calculating an AUC of 0.975 ± 0.005. The PLS-DA model calculated the highest AUC for a limited validation 16 known ground-truth dataset (0.991 AUC) in comparison to the other models. All models were evaluated by determining an analyst's threshold for large scale burn data of unknown ground-truth. All models determined a conservative threshold (LLR = 1) as a cutoff score by the informed analyst.
Identifer | oai:union.ndltd.org:ucf.edu/oai:stars.library.ucf.edu:etd-7839 |
Date | 01 January 2019 |
Creators | Allen, Alyssa |
Publisher | University of Central Florida |
Source Sets | University of Central Florida |
Language | English |
Detected Language | English |
Type | text |
Format | application/pdf |
Source | Electronic Theses and Dissertations |
Page generated in 0.0016 seconds