A global synthesis and machine learning prediction of biodegradable dissolved organic carbon in freshwater systems using optical indices = 광학 지표를 활용한 담수 시스템 내 생분해성 용존유기탄소의 전 지구적 종합 분석 및 머신러닝 예측

다국어 초록 (Multilingual Abstract)

Biodegradable dissolved organic carbon （BDOC） is the most reactive fraction of dissolved organic matter （DOM） and a major driver of carbon dioxide （CO2） emissions from inland waters, yet its large-scale assessment is constrained by incubation-based measurements. This study conducted a global meta-analysis integrating harmonized BDOC observations, optical DOM indices, and machine-learning （ML） modeling to enable scalable BDOC prediction across freshwater systems. A total of 1,063 paired observations were compiled from 17 studies across eight countries, with cross-study comparability improved through correction of variable incubation periods using first-order degradation kinetics, and standardization of UV–visible absorbance wavelengths. The harmonized dataset revealed pronounced hydrological and climatic controls on BDOC dynamics. Biodegradation rate constants were consistently higher in rivers than in lakes, with maximum values observed in temperate rivers during the wet season. Seasonal analyses showed a decoupling between bulk dissolved organic carbon （DOC） and BDOC, with dry conditions favoring DOC accumulation and aromaticity, while wet conditions enhanced inputs of bio-labile DOM. Across climatic gradients, BDOC exhibited greater sensitivity to environmental variability than bulk DOC. Using DOC concentration, absorbance at 254 nm, and humification index （HIX） as predictors, nonlinear ML models outperformed linear regression. Extreme gradient boosting （XGBoost） achieved the highest predictive
performance （R2 〉 0.6）, with stronger accuracy for riverine than lacustrine systems. Overall, this study demonstrates that harmonized optical datasets combined with ML modeling provide a robust and scalable alternative to incubation-based BDOC measurements, advancing quantitative assessment of freshwater carbon biodegradability and inland-water carbon cycling.

목차 (Table of Contents)

• Ⅰ. Introduction 13
• 1.1 Importance and challenges of BDOC in freshwater systems 14
• 1.2 Research background 18
• 1.2.1 Climatic distribution of BDOC dataset 18
• 1.2.2 Seasonal distribution of BDOC dataset 24
• 1.2.3 Waterbody type distribution of BDOC dataset 30
• 1.3 Research objectives 36
• Ⅱ. Materials and Methods 39
• 2.1 Keyword analysis 40
• 2.2 Data collection 43
• 2.3 Input variables adjustments 45
• 2.3.1 Adjustment of DOM incubation periods from different studies 45
• 2.3.2 Adjustment for UV-based indices 48
• 2.4 Input variables used for BDOC prediction 50
• 2.5 Multiple Linear Regression 54
• 2.6 Machine Learning algorithms 55
• 2.6.1 Deep Neural Network (DNN) 57
• 2.6.2 Extreme Gradient Boosting (XGBoost) 60
• 2.6.3 Model performance evaluation 64
• 2.6.4 Shapley Additive exPlanations (SHAP) 65
• 2.7 Data preprocessing for ML 66
• 2.8 Statistical analysis 67
• Ⅲ. Results and Discussion 68
• 3.1 Comparison of DOM biodegradation rates across water bodies and climatic zones 69
• 3.2 Climatic and seasonal distributions of the meta-analysis data 72
• 3.2.1 Climatic distributions of BDOC-related parameters and BDOC concentrations 72
• 3.2.2 Seasonal distributions of BDOC-related parameters and BDOC concentrations 76
• 3.3 Bivariate relationships between BDOC and predictor variables 79
• 3.4 Comparison of predictive model performance for BDOC estimation 82
• 3.5 Hydrological influences on BDOC prediction 84
• 3.5.1 Effects of waterbody type on model performance 84
• 3.5.2 Model interpretation using SHAP analysis 86
• 3.6 Environmental implications and future studies 88
• Ⅳ. Conclusions 92
• References 95