The abiologically and biologically driving effects on organic matter in marginal seas revealed by deep learning-assisted model analysis.

The biogeochemical processes of organic matter exhibit notable variability and unpredictability in marginal seas. In this study, the abiologically and biologically driving effects on particulate organic matter (POM) and dissolved organic matter (DOM) were investigated in the Yellow Sea and Bohai Sea of China, by introducing the cutting-edge network inference tool of deep learning. The concentration of particulate organic carbon (POC) was determined to characterize the status of POM, and the fractions and fluorescent properties of DOM were identified through 3D excitation-emission-matrix spectra (3D-EEM) combined parallel factor analysis (PARAFAC). The results indicated that the distribution of POM and DOM exhibited regional disparity across the studied sea regions. POM demonstrated greater heterogeneity in the South Yellow Sea (p < 0.05), and in contrast, all three fluorescent components of DOM displayed a higher degree of heterogeneity in the Bohai Sea (p < 0.05). To delve into the drivers of the discrepancy, artificial neural network (ANN) models were constructed, incorporating 15 extra abiotic and biotic parameters. Under optimal parameter setting, ANNs achieved a maximum Pearson correlation coefficient (PCC) of 0.87 and a minimum Root Mean Squared Error (RMSE) of 0.23. The model identified turbidity and temperature as the most influential factors, accounting for the variation in the heterogeneity of POM and DOM across different sea regions, respectively. Additionally, the result highlighted the significant role of pico-size photosynthetic organisms among biological predictors, which may suggest their pivotal, yet often underappreciated, role in blue carbon cycles. In conclusion, this research introduces advanced deep-learning modeling techniques, providing novel insights into the biogeochemical processes of organic matter in marginal seas.