Performance evaluation of chemical plume detection and quantification algorithms

Abstract

Hyperspectral Imaging sensors provide a wealth of spatial, and more importantly, spectral information, which can be used in a wide variety of applications, including chemical plume detection. In theory, every gaseous chemical has a unique spectrum, by studying, the effect this spectrum has on electromagnetic radiation from the background of a scene, we are able to perform chemical plume detection and quantification. Analysis of the accuracy, strengths, and weaknesses of detection and quantification algorithms is a difficult task, as one typically lacks ground truth data for physical plume parameters such as location, concentration, and temperature. In order to better understand the performance of our algorithms, we developed a tool which allows us to embed synthetic plume into real background data with control of these parameters. In this thesis, we first develop a radiative transfer model for chemical plumes. Using this model, we then build a suite of detection and quantification algorithms, as well as our plume embedding routine. Finally, using our semi-synthetic data, we study the impact of various physical plume parameters, including concentration and thermal contrast, among others, on the results of our algorithms.