The problem of automating the data analysis of microplastics following a spectroscopic measurement such as focal plane array (FPA)-based micro-Fourier transform infrared (FTIR), Raman, or QCL is gaining ever more attention. Ease of use of the analysis software, reduction of expert time, analysis speed, and accuracy of the result are key for making the overall process scalable and thus allowing nonresearch laboratories to offer microplastics analysis as a service. Over the recent years, the prevailing approach has been to use spectral library search to automatically identify spectra of the sample. Recent studies, however, showed that this approach is rather limited in certain contexts, which led to developments for making library searches more robust but on the other hand also paved the way for introducing more advanced machine learning approaches. This study describes a model-based machine learning approach based on random decision forests for the analysis of large FPA-μFTIR data sets of environmental samples. The model can distinguish between more than 20 different polymer types and is applicable to complex matrices. The performance of the model under these demanding circumstances is shown based on eight different data sets. Further, a Monte Carlo cross validation has been performed to compute error rates such as sensitivity, specificity, and precision.
© 2021 The Authors. Published by American Chemical Society.