Optimizing the Efficiency of Water Pollution Tracing Based on Three-Dimensional Fluorescence Spectra Extracted from Characteristic Excitation Wavelengths
Traditional excitation-emission matrix (EEM) fluorescence spectroscopy suffers from prolonged scanning times, data redundancy, high instrument cost, and bulkiness, hindering rapid on-site water pollution source tracing. This study proposes a novel classification method combining fixed characteristic excitation wavelength scanning with support vector machine (SVM) to enhance efficiency. A total of 180 EEM samples were collected from six pollution source categories: chemical fiber dyeing and finishing, wool textile dyeing and finishing, leather processing, metal surface processing, papermaking, and domestic sewage. Parallel factor analysis (PARAFAC) identified characteristic fluorescent components and excitation wavelengths. Correlation analysis and feature importance analysis further reduced these to seven characteristic excitation wavelengths. SVM and random forest (RF) models were constructed using both the reduced and original EEM datasets. Results demonstrated that models based on the seven characteristic excitation wavelengths maintained high recognition accuracy while significantly improving efficiency. The SVM model achieved the best performance, with runtime reduced from 243.05 s to 34.56 s (an 86% decrease) and recognition accuracy reaching 94.4%. Precision, recall, and F1-score metrics confirmed the robust performance of SVM with characteristic wavelengths, particularly for metal surface processing wastewater. This study provides an efficient and reliable method for rapid water pollution tracing by simplifying EEM scanning and integrating SVM, offering high application value. Future work will optimize feature selection strategies and explore additional sample categories and model combinations to broaden applicability.