Abstract

Flood risks are escalating under climate change, necessitating advanced methods to improve runoff prediction and multivariate flood-risk assessment. In this study, a physics–XGBoost-based Bayesian model averaging with bivariate copulas (PXB-BVC) framework was developed by integrating the Soil and Water Assessment Tool (SWAT), the Hydrologiska Byråns Vattenbalansavdelning (HBV) model, Extreme Gradient Boosting (XGBoost), Bayesian model averaging (BMA), and bivariate copulas. Spatially detailed underlying surface parameters including 30 m land-use data derived from the 2000 China land-use remote sensing monitoring data were pre-processed and reclassified using ArcGIS to support spatially explicit hydrological simulation. The framework was applied to the Xiangxi River Basin (XXRB), China, under four general circulation models and three shared socioeconomic pathways. PXB-BVC improved daily runoff simulation by combining process-based hydrological information with nonlinear machine learning correction, achieving Nash–Sutcliffe efficiency (NSE) values of 0.95 during calibration and 0.89 during validation. Future runoff generally increased from the near-term to the late-century period, with stronger changes under SSP585 and Sen slopes reaching up to 0.46 m3 s−1 yr−1, although the magnitude and significance of trends varied among GCMs. The dependence structures among flood peak, flood volume, and flood duration showed non-stationary behavior under future climate forcing, with Kendall’s tau for peak–volume pairs mostly ranging from 0.6 to 0.8. The revised bivariate return-period analysis further indicates that inferred flood-risk changes depend on the joint risk definition. Under SSP245 and ACCESS-ESM1–5, OR-type joint return periods show that representative near-future 50-year events may become more frequent in 2061–2100, whereas AND-type return periods show weaker and less uniform changes among flood-characteristic pairs. Conditional probability analysis also indicates enhanced compound risk under high-emission conditions: given an extreme peak flow, the probability of accompanying high flood volume increases from 0.23 to 0.56, while the probability of prolonged duration increases from 0.18 to 0.45. These results demonstrate that the PXB-BVC framework can support non-stationary multivariate flood-risk assessment and provide useful information for climate-resilient water-resource management and infrastructure planning.

IPC Classification

Keywords