Raman Peaks Feature-Based Machine Learning for Raman Spectroscopy Quantification of Inorganic Pollutants

This study investigates how Raman peak features influence the accuracy of pollutant concentration quantification in Raman spectroscopy using feature-based machine learning. A Raman peak feature-based approach is applied to data acquired from laboratory Raman equipment for inorganic anion mixtures, with the objective of identifying processing strategies suitable for low-resolution Raman instruments. A dataset comprising nitrate, nitrite, and sulphate dissolved in water at varying concentrations was used to develop a pre-processing pipeline, train and validate models, and perform linear regression using peak- and area-based features extracted from analyte-specific spectral fingerprint regions. Signal downsampling was employed to simulate reduced spectral resolution and evaluate feature robustness under peak broadening conditions. Results indicate that peak-based regression is more sensitive to reduced peak resolution for nitrate and sulphate, whereas area-based regression produces more stable prediction errors across downsampling factors. Conversely, for nitrite, peak-based regression is less affected by downsampling, while area-based regression exhibits degraded performance. Area-based regression achieved mean absolute percentage errors below 7% and 16% for nitrate concentrations above 3833 mg/L and 1916 mg/L, respectively, and generally below 16% for sulphate concentrations above 1916 mg/L. For nitrite, peak-based regression yielded errors within 4%.