Quantum cascade laser imaging (LDIR) and machine learning for the identification of environmentally exposed...

Monitoring of microplastics in environmental samples is relevant to the scientific world, as well as to environmental agencies and water authorities, in particular considering increasing efforts to decrease emissions and the growing concern of governments and the public. Therefore, rapid accurate detection and identification of microplastics including polymers, despite their degradation in the environment, is crucial. The degradation has a significant impact on the infrared spectra of the microplastics and can impede the identification process. This work presents a novel approach to addressing the problem of identification of weathered microplastics. A quantum cascade laser (LDIR) was used to record the infrared spectra of various polymeric particles (81,291 individual particles). Using a combination of pristine and weathered particles, two supervised machine learning (ML) models, namely Subspace k-Nearest Neighbor (Sub-kNN) and Boosted Decision Tree (BDT), were trained to recognize the spectrum characteristics of labeled particles and then used to identify unlabeled samples, with an identification accuracy of 89.7% and 77.1% using 10-fold cross validation. About 90% of the samples could be identified via the Sub-kNN or BDT models. Subsequently, a non-supervised ML model, namely, Density-based Spatial Clustering of Applications with Noise (DBSCAN), was used to cluster samples which could not be labeled from the supervised ML model. This enabled the detection of additional subgroups of microplastics. Manual labelling can then be carried out on a selection of spectra per group (e.g., centroids of each cluster), hence accelerating the identification process and allowing to add new labeled samples to the initial supervised ML. Although expert efforts are still needed, the proposed method greatly lowers labeling efforts by using the combined supervised and unsupervised learning models. In the future, the use of deep neural networks could further boost the implementation of these kinds of approaches for polymer and microplastic identification in environmental settings.