Classification of Substances Combining Standoff Laser Induced Fluorescence and Machine Learning

Abstract

Contaminated objects and areas must be handled carefully depending on the underlying pollution. There are methods which require short distances, others the collection of samples or even direct contact to the hazardous, and some of the established techniques take long to reach a conclusion. A fast standoff method for predicting the potential hazard can be achieved by examining the laser induced fluorescence spectra of the substances of interest. The samples are excited by low-energy laser pulses of two alternating wavelengths. The datasets are measured for almost 50 agents, including fuels, pesticides and bacteria and represent the basis for a subsequent classification procedure. Therefore, the investigated materials are grouped in seven classes depending on their origin and utilization. The majority of the dataset is used in a training phase to create predictive models, which are tested with the remaining signals to qualify the classification. After all, the single spectra of the test set are classifed with an error rate less than 0.1 % in predicting the correct class. With a statement like this frst responders would be able to choose the right preventive measure for a rescue or decontamination procedure.