Neural Networks Based on Synthesized Training Data for the Automated Detection of Chemical Plumes in Passive Infrared Multispectral Images

Abstract

Automated detection of volatile organic compounds in the atmosphere can be achieved by applying pattern recognition analysis to passive infrared (IR) multispectral remote sensing data. However, obtaining analyte-active training data through field experiments is time-consuming and expensive. To address this issue, methodology has been developed for simulating radiance profiles acquired using a multispectral IR line-scanner mounted in a downward-looking position on a fixed-wing aircraft. The simulation strategy used Planck's radiation law and a radiometric model along with the laboratory spectrum of the target compound to compute the upwelling IR background radiance with the presence of the analyte within the instrumental field-of-view. By combining the simulated analyte-active radiances and field-collected analyte-inactive radiances, a synthetic training dataset was constructed. A backpropagation neural network was employed to build classifiers with the synthetic training dataset. Employing methanol as the target compound, the performance of the classifiers was evaluated with field-collected data from airborne surveys at two test fields.