A machine learning approach to the identification of chemical substances from lidar measurements

In the last decades, the application of LiDAR/DIAL measurements to remote sensing and atmospheric physics has been consolidated from both the experimental and the interpretation point of view. The laser and optic technologies involved have become very sophisticated and the quality of the results have reflects this fact. These techniques are therefore seriously considered also for defence applications, for example for the survey of large areas to detect the release of chemical agents. On the other hand, for a reliable deployment of these techniques in real life applications, robust data analysis tools are required, an aspect to which not enough consideration is typically accorded during the design phase of the instrumentation. In this paper, it is shown how the absorption signals generated by various chemical substances can be processed to maximise the success rate of their identification. The developed classification methods are based on state of the art classification trees. The quality of the proposed technique is well supported by simulations based on the HITRAN database. Significant efforts have been devoted to the issue of providing an estimate of the robustness against noise of the classification provided by the machine learning tools.