Feasibility of Conditional Tabular Generative Adversarial Networks for Ecologically Plausible Synthetic River Water-Quality Data: A Statistical and Ecological Similarity Assessment

Reliable biological datasets, especially those integrating biotic indices such as the Saprobic Index, are scarce, limiting machine and deep learning applications in aquatic ecosystem assessments. This study evaluates Conditional Tabular Generative Adversarial Networks (CTGANs) for generating synthetic datasets that combine physico-chemical parameters with a biological index (Saprobic Index) from multiple monitoring stations in the lower Danube River. Beyond univariate distributional agreement, we assess whether ecologically meaningful multivariate relationships are preserved in the synthetic tables. To support this, we propose an ecology-oriented validation workflow that combines distributional tests with correlation structure and clustering diagnostics across stations. Real monitoring datasets were statistically modelled and recreated using CTGANs, then qualitatively assessed for realism. Comparisons between synthetic and real data employed box plots, Wilcoxon rank-sum tests, correlation matrices, and K-means clustering across stations. Stable variables, including pH, total dissolved solids, and chemical oxygen demand, were well replicated, showing no significant distributional differences (p > 0.05). Conversely, dynamic parameters such as dissolved oxygen, total nitrogen, and suspended solids exhibited notable discrepancies (p < 0.05). Correlation analyses indicated that several strong associations present in the observed data (e.g., total nitrogen–nitrate and total nitrogen–electrical conductivity) were substantially weaker in the synthetic dataset. Overall, a CTGAN can reproduce several marginal patterns but may fail to preserve key ecological linkages, which constrains its use in ecological relationship-dependent inference. While promising for exploratory modelling and general trend analysis, synthetic data should be applied cautiously for studies involving seasonally influenced, biologically significant parameters.