Uncertainty analysis for AI-based regional flood frequency analysis methods in Australia

This study undertakes an uncertainty analysis for two regional flood frequency analysis techniques (random forest regression (RFR) and multiple linear regression (MLR)). This uses data from 201 gauged catchments of south-east Australia. To quantify uncertainty, a synthetic dataset comprising 15,000 samples from the original 201 catchments’ data was generated via Monte Carlo simulation. Model performance was evaluated using Monte Carlo Cross-Validation. Key statistical indicators, including Relative Error, Absolute Relative Error, and Median Relative Error (REr) were used to assess the model performance. Results show that RFR consistently outperforms MLR with narrower uncertainty bounds. Median REr for MLR (original data) ranges 40.32 ± 5.09% to 54.94 ± 7.51%, MLR (synthetic data) 38.69 ± 0.64% to 51.79 ± 0.88%, and RFR (synthetic data) 36.81 ± 0.68% to 50.84 ± 0.89%. These results demonstrate that artificial intelligence (AI) based RFR provides lower error rage and more consistent uncertainty estimates compared to the MLR.