Driving factors and predictive model of dissolved N2O concentrations in a complex aquatic network.

Inland water networks, comprising hydrologically integrated rivers, agricultural ditches, and aquaculture ponds, are significant N2O sources, yet their complexity impedes accurate quantification. Here we developed an integrated framework combining structural equation modeling (SEM), machine learning (ML), and SHapley Additive exPlanations (SHAP) to bridge causal inference with nonlinear predictive modeling in China's Taihu Basin. Our results demonstrate that NO3--N and water temperature (WT) dominate N2O variability, explaining significantly more variance than discrete water body categories. This framework successfully reconciled the dual role of dissolved organic carbon (DOC). SEM identifies DOC as a macroscopic sink driven by the complete denitrification of nitrate to N2 (standardized effect = -0.143), while SHAP reveals its role as a microscopic catalyst that enhances N2O production efficiency per unit of nitrate. Although the ensemble model achieved high accuracy (test R2 = 0.70), the parsimonious model using four routine parameters (NO3--N, DO, NH4+-N, and WT) proved more suitable for regional assessment, demonstrating satisfactory predictive capability (test R2 = 0.54) and successfully reconstructing basin-wide spatiotemporal patterns. This study provides a scalable and transferable methodology for unlocking the driving mechanisms of complex aquatic ecosystems, offering a robust tool for basin-scale N2O estimation and targeted greenhouse gas mitigation.