Robust validation of pattern classification methods for laser-induced breakdown spectroscopy

Abstract

Laser-induced breakdown spectroscopy (LIBS) is an emerging technology that is suitable for a variety of material identification applications. For LIBS to successfully transition from the laboratory into field applications, the sensor must be paired with the appropriate algorithms for accurate and robust processing of the LIBS spectra. In this study we will report on the results of testing classification methods on eight distinct classification tasks using LIBS datasets. Results suggest that standard cross-validation techniques may not accurately estimate generalization performance and a proposed "leave-one-sample-out" approach to experiment design for classifier validation may provide a more robust measure of performance.