Optimizing Machine Learning-Based Prediction of Terrestrial Dissolved Organic Matter in the Ocean Using Fluorescence...

Area of Science:

• Environmental chemistry
• Oceanography
• Data science

Background:

• Marine dissolved organic matter (DOM) is a complex mixture vital to the global carbon cycle.
• Arctic climate change increases terrestrial organic carbon release into marine systems.
• Accurate assessment of DOM composition is crucial for understanding its sources and fate.

Purpose of the Study:

• To compare machine learning (ML) models for predicting terrestrial DOM using molecular formula data.
• To optimize ML techniques for accuracy and computational efficiency in analyzing LC-FTMS data.
• To identify key molecular features indicative of terrestrial DOM signatures.

Main Methods:

• Comparison of Random Forest (RF), Support Vector Machines, and Generalized Linear Models (GLMs).
• Systematic evaluation of data preprocessing, normalization, and ML techniques.
• Application of feature selection, Shapley values, and permutation importance for analysis.

Main Results:

• Generalized Linear Models (GLMs) with sum normalization achieved the highest accuracy (5.7% NRMSE) and efficiency.
• Random Forest (RF) models were robust but less accurate and computationally intensive.
• Feature selection significantly improved all models, reducing the number of features needed for robust predictions.

Conclusions:

• GLMs provide a scalable and accurate approach for predicting terrestrial DOM from LC-FTMS data.
• ML enhances the analysis of complex marine DOM, aiding in understanding carbon cycling.
• This study provides a blueprint for applying ML to high-resolution mass spectrometry data in marine science.