Artificial Intelligence Assisted Mid-Infrared Laser Spectroscopy In Situ Detection of Petroleum in Soils

• Published in: Applied sciences 2020-02, Vol.10 (4), p.1319
• Main Authors: Galán-Freyle, Nataly J., Ospina-Castro, María L., Medina-González, Alberto R., Villarreal-González, Reynaldo, Hernández-Rivera, Samuel P., Pacheco-Londoño, Leonardo C.
• Format: Article
• Language: English
• Subjects: Artificial intelligence; artificial intelligence (ai); Brown soils; chemometrics; Discriminant analysis; Environmental impact; Hydrocarbons; Infrared lasers; Infrared spectroscopy; Laser spectroscopy; Lasers; mid-infrared (mir) laser spectroscopy; multivariate analysis; Oil spills; Pattern recognition; Petroleum; Pollutants; Pollution detection; Potassium; Principal components analysis; Quantum cascade lasers; quantum cascade lasers (qcls); Remote sensing; Sandy soils; Sediment pollution; Selectivity; soil; Soil contamination; Soil mixtures; Soil pollution; Soil types; Spectroscopy; Statistical analysis; Statistical methods; Support vector machines
• ISSN: 2076-3417; 2076-3417
• DOI: 10.3390/app10041319

A simple, remote-sensed method of detection of traces of petroleum in soil combining artificial intelligence (AI) with mid-infrared (MIR) laser spectroscopy is presented. A portable MIR quantum cascade laser (QCL) was used as an excitation source, making the technique amenable to field applications. The MIR spectral region is more informative and useful than the near IR region for the detection of pollutants in soil. Remote sensing, coupled with a support vector machine (SVM) algorithm, was used to accurately identify the presence/absence of traces of petroleum in soil mixtures. Chemometrics tools such as principal component analysis (PCA), partial least square-discriminant analysis (PLS-DA), and SVM demonstrated the effectiveness of rapidly differentiating between different soil types and detecting the presence of petroleum traces in different soil matrices such as sea sand, red soil, and brown soil. Comparisons between results of PLS-DA and SVM were based on sensitivity, selectivity, and areas under receiver-operator curves (ROC). An innovative statistical analysis method of calculating limits of detection (LOD) and limits of decision (LD) from fits of the probability of detection was developed. Results for QCL/PLS-DA models achieved LOD and LD of 0.2% and 0.01% for petroleum/soil, respectively. The superior performance of QCL/SVM models improved these values to 0.04% and 0.003%, respectively, providing better identification probability of soils contaminated with petroleum.