Abstract

Standoff detection and identification of trace chemicals in hyperspectral infrared images is an enabling capability in a variety of applications relevant to defense, law enforcement, and intelligence communities. Performance of these methods is impacted by the spectral signature variability due to the presence of contaminants, surface roughness, nonlinear effects, etc. Though multiple classes of algorithms exist for the detection and classification of these signatures, they are limited by the availability of relevant reference datasets.

In this work, we first address the lack of physics-based models that can accurately predict trace chemical spectra. Most available models assume that the chemical takes the form of spherical particles or uniform thin films. A more realistic chemical presentation that could be encountered is that of a non-uniform chemical film that is deposited after evaporation of the solvent which contained the chemical. This research presents an improved signature model for this type of solid film. The proposed model, called sparse transfer matrix (STM), includes a log-normal distribution of film thicknesses and is found to reduce the root-mean-square error between simulated and measured data by about 25% when compared with either the particle or uniform thin film models. When applied to measured data, the sparse transfer matrix model provides a 0.10-0.28 increase in classification accuracy over traditional models.

There remain limitations in the STM model which prevent the predicted spectra from being well-matched to the measured data in some cases. To overcome this, we leverage the field of domain adaptation to translate data from the simulated to the measured data domain. This thesis presents the first one-dimensional (1D) conditional generative adversarial network (GAN) to perform spectrum-to-spectrum translation of reflectance signatures. We apply the 1D conditional GAN to a library of simulated spectra and quantify the improvement with the translated library. The method demonstrates an increase in overall classification accuracy to 0.723 from the accuracy of 0.622 achieved using the STM model when tested on real data. However, the performance improvement is biased towards data included in the GAN training set.

The next phase of the research focuses on learning models that are more robust to different parameter combinations for which we do not have measured data. This part of the research leverages elements from the field of theory-guided data science. Specifically, we develop a physics-guided neural network (PGNN) for predicting chemical reflectance for a set of parameterized inputs that is more accurate than the state-of-the-art physics-based signature model for chemical residues. After training the PGNN, we use it to generate a library of predicted spectra for training a classifier. We compare the classification accuracy when using this PGNN library versus a library generated by the physics-based model. Using the PGNN, the average classification accuracy increases to 0.813 on real chemical reflectance data, including data from chemicals not included in the PGNN training set.

The products of this thesis work include methods for producing realistic trace chemical residue reflectance signatures as well as demonstrations of improved performance in active spectroscopy classification applications. These methods provide great value to a range of scientific communities. The novel STM signature model enables existing spectroscopy sensors and algorithms to perform well on real-world problems where chemical contaminants are non-uniform. The 1D conditional GAN is the first of its kind and can be applied to many other 1D datasets, such as audio and other time-series data. Finally, the application of theory-guided data science to the trace chemical problem not only enhances the quality of results for known targets and backgrounds, but also increases the robustness to new targets.

Publication Date

4-9-2021

Document Type

Dissertation

Student Type

Graduate

Degree Name

Imaging Science (Ph.D.)

Department, Program, or Center

Chester F. Carlson Center for Imaging Science (COS)

Advisor

John Kerekes

Advisor/Committee Member

Charles Bachmann

Advisor/Committee Member

John Fisher

Campus

RIT – Main Campus

Share

COinS