Combining parallel factor analysis and machine learning for the classification of dissolved organic matter according...

Area of Science:

• Environmental Chemistry
• Biogeochemistry
• Analytical Chemistry

Background:

• Dissolved organic matter (DOM) fluorescence analysis using Parallel Factor (PARAFAC) analysis is crucial for understanding its biogeochemical cycling.
• Existing PARAFAC-based methods lack holistic approaches for distinguishing DOM sources.
• Accurate source identification is vital for comprehensive carbon flow assessments in ecosystems.

Purpose of the Study:

• To develop and evaluate machine learning methods (MLM) for classifying DOM sources based on PARAFAC-analyzed excitation-emission matrices (EEMs).
• To assess the impact of experimental treatments and dataset properties on DOM classification accuracy.
• To establish a robust method for distinguishing between riverine and leached DOM fluorescence signatures.

Main Methods:

• Classified 1029 PARAFAC-analyzed EEMs from leaf leachates, rivers, and standards using four MLMs.
• Evaluated classification performance across various subsets, including whole EEMs, size-fractionated DOM, mixtures, and quenched samples.
• Optimized PARAFAC models from 10 to 12 components to improve residual peak removal and classification accuracy.

Main Results:

• The 12-component PARAFAC model enhanced classification accuracy, particularly for datasets with size-fractionated DOM or diverse sources.
• MLM achieved high accuracy in classifying riverine DOM (up to 87.0%), leachates (up to 92.5%), and distinguishing between them (97.2%).
• Multilayer perceptron and support vector machines demonstrated superior performance among the tested MLMs, with N-way partial least-squares discriminant analysis showing comparable results.

Conclusions:

• Combining PARAFAC with MLM provides an effective strategy for classifying DOM based on fluorescence signatures.
• PARAFAC isolates meaningful fluorescent species, while MLM simultaneously classifies EEMs into distinct categories.
• This integrated approach enhances the accurate accounting of carbon flows by identifying diverse DOM sources and their contributions.