Remote identification of microorganisms on various substrates using LIBS and machine learning integration

Abstract

To address the health risks associated with microbial contamination, the development of non-contact remote monitoring technology offers a safe, efficient, and accurate means of detection while minimizing the risk of infection. This study examines the use of remote laser-induced breakdown spectroscopy (LIBS) combined with machine learning algorithms to identify and classify microorganisms on various metal and non-metal substrates. The remote LIBS detection system developed in this study integrates a single-pulse nanosecond laser with a zoomable Cassegrain telescope, significantly enhancing its ability to capture spectral signals from samples located 5 m away. Spectral analysis of ten microbial species identified dual spectral lines of elements such as sodium, potassium, and calcium as the primary features. Classification was effectively achieved by using only these selected features in combination with machine learning algorithms. In the classification phase, four machine learning algorithms, Principal Component Analysis combined with k-Nearest Neighbors (PCA-KNN), Partial Least Squares Discriminant Analysis (PLS-DA), Support Vector Machine (SVM), and Random Forest (RF), were applied to categorize the microbial spectra. Evaluation results indicated that the RF algorithm achieved the highest classification accuracy, reaching 91.0 %. To further enhance the RF model’ s performance, several variable models were introduced, including linear, doublet-line ratio, cross ratio, and hybrid models. SHAP value analysis was used to quantify the contributions of key spectral features to model predictions. The results demonstrated that the hybrid model exhibited the best classification performance, improving RF accuracy to 96.5 %, with sodium and calcium doublet-line ratios being key factors in enhancing classification robustness. Additionally, the best model was used to validate bacterial classification on the surfaces of soil and paper substrates, achieving classification accuracies of 92.5 % and 96 %, respectively. This innovative method leverages the non-contact, remote detection capabilities of LIBS technology combined with the precise classification power of machine learning, enabling efficient microbial species identification. This study confirms the effectiveness of remote LIBS technology for non-invasive, rapid microbial detection and illustrates that optimizing spectral feature selection and integrating appropriate machine learning models significantly enhances the accuracy and robustness of LIBS-based microbial classification. These findings highlight the potential of remote LIBS technology in environmental monitoring and public health safety, while also suggesting new strategies for applications in biological weapon defense and infectious disease control.