To forecast DON concentrations in marginal seas of China by machine learning models

Dissolved organic nitrogen (DON) is a critical yet poorly constrained component of the marine nitrogen cycle. Its quantification is notoriously challenging, as it is conventionally determined by subtracting measured inorganic nitrogen species from total dissolved nitrogen (TDN), a process that compounds analytical errors from multiple-steps of measurements. This methodological limitation results in highly uncertainties in DON dataset, hindering accurate assessments of nutrient dynamics, nitrogen loadings and ecosystem productivity. To resolve this, we developed and evaluated a suite of machine learning (ML) models to forecast DON concentrations based on easily obtainable hydrographic and biogeochemical parameters dataset of approximately 1050 samples, encompassing 25 variables collected across spring, summer and winter in the Yellow and East China Seas. We trained and systematically examined five ML algorithms based on three variable panels, full panel (P25), core panel (P15), and minimal panel (P08). The stacked generalization model demonstrated the best performance, achieving an R² of 0.62 and a Root Mean Square Error (RMSE) of 1.06 μmol/L with P25. Notably, even on 8 key parameters (P08), the model retained robust predictive power (R² = 0.58, RMSE = 1.11 μmol/L), establishing an optimal balance between accuracy and operational feasibility. Feature importance analyses identified salinity, dissolved oxygen, latitude, and inorganic nitrogen concentrations as the most impacting predictors, aligning with these physical and biogeochemical mechanisms governing DON distribution. Our study provides a useful, mechanistically interpretable tool to predict DON distribution pattern based on simple environmental parameters. Promisingly, this study offers a transformative approach for DON data quality control and evaluation in regional-scale marine monitoring programs.