Fluorescence Analysis of River DOM Spectra Using PARAFAC in Combination with a Self-Organizing Map to Distinguish Organic Matter Sources

This study used parallel factor method (PARAFAC), self-organizing map (SOM), and random forest models to study the dissolved organic matter (DOM) sources and characteristics in rivers with varying non-point source inputs. The artificial canal Cihuai New River (CH) and the Gouqu (GQ) which are heavily polluted by agricultural surface sources were selected as the study objects. The PARAFAC model resolved four chemical components. C1 comprises two peaks, C1 (T1) (UVC fulvic acid) and C1 (T2) (humic-like acid). C2 includes two peaks, C2 (T1) (tyrosine-like protein) and C2 (T2) (tryptophan-like protein). C3 has two peaks, C3 (T1) (humic-like) and C3 (T2) (UVA fulvic acid). C4 is identified as humic-like fulvic acid. The SOM model shows that the degree of humification in the GQ, which is influenced by agricultural non-point source pollution, is higher than that of the unaffected CH. The primary source of humic substances in the river is agricultural non-point source pollution. CH is influenced by surrounding human activities and the eutrophication of water bodies, resulting in a higher level of autochthonous characteristics and biological activity in DOM. Random Forest model indicated that the C3 was the most sensitive (R2 = 0.88) to river's changes and therefore it is a good indicator of river's water quality. And NH4+ has a strong driving effect on the water quality of both rivers. Principal Component Analysis (PCA) reveals that the agricultural river DOM (GQ) is mainly composed of humic substances, while the artificial river DOM (CH) is predominantly derived from autochthonous sources. The combination of PARAFAC, SOM, and random forest methods helps overcome the limitations of traditional approaches and provides a scientific basis for the management of river water quality pollution. Differences in DOM spectra of rivers with different surface source intensities.Analyzing three-dimensional fluorescence data using machine learning.Combine PARAFAC with self-organizing maps and random forest models to complement the limitations of traditional analytical methods.