Rapid and uncertainty-aware source tracing of accidental river pollution using surrogate-assisted Bayesian inversion

Rapid identification of accidental pollutant releases is essential for reducing ecological risks in rivers, but direct inverse simulation is often too slow for emergency use. This study developed a surrogate-assisted Bayesian inversion framework for sudden surface water pollution events in the Dongliao River, northeastern China. A process-based hydrodynamic-water quality model was used to generate a virtual accident database. A re-optimized Jellyfish Search-based kernel extreme learning machine (JS-KELM) was then used as a fast surrogate for the nonlinear source-response relationship. The surrogate was coupled with a Bayesian-PSO inversion scheme to estimate source variables and quantify conditional uncertainty. Across 20 pollutant-section outputs, the JS-KELM achieved an independent test-set R² of 0.9565 and Pearson's r of 0.9781. Comparisons with GPR, SVR, ANN, and JS-BP showed that GPR and SVR also performed strongly, while JS-KELM offered a favorable balance between accuracy, speed, small-sample robustness, and repeated inversion calls. Under multi-section monitoring, the MAP estimates had a mean relative error of 5.12%, and reconstructed concentrations reached R² = 0.9962 and MAPE = 2.73%. Under reduced monitoring, the mean relative error increased to 10.96%, while concentration reconstruction remained accurate, with R² = 0.9965 and MAPE = 4.41%. Posterior distributions further identified weakly constrained variables and plausible source ranges. External validation with Missouri River tracer data supported the framework's transferability. The proposed approach provides a rapid, physically informed, and uncertainty-aware tool for emergency river pollution source tracing.