Document Type : Research Paper
Climate change is one of the most important environmental challenges of the 21st century, significantly affecting hydrological processes, water resources, and extreme events such as floods and droughts. Increasing greenhouse gas emissions have altered precipitation patterns and intensified hydrological extremes worldwide. Floods are among the most destructive natural hazards, causing major economic, social, and environmental damages. CMIP6 climate models under SSP scenarios provide valuable future climate projections, but their outputs often contain biases that require correction for local-scale applications. Recent advances in machine learning have improved bias correction and reduced uncertainties in climate data. Therefore, this study investigates the impact of climate change on flood frequency in the Poldokhtar Basin using CMIP6 projections, machine learning-based bias correction, and hydrological modeling.
The study was conducted in the Poldokhtar Basin, a major sub-basin of the Karkheh River in western Iran. Daily precipitation data from 23 CMIP6 GCMs were obtained from the NEX-GDDP-CMIP6 dataset for both historical and future periods under SSP126, SSP245, SSP370, and SSP585 scenarios. Observed precipitation data from three meteorological stations were used to evaluate and bias-correct the climate model outputs using the XGBoost algorithm. Model performance was assessed using MAE, RMSE, Pearson correlation coefficient, NSE, and KGE. A Multi-Model Ensemble (MME) based on XGBoost was then applied to reduce uncertainty and improve precipitation estimates. Daily streamflow was simulated using the hybrid ANFIS-PSO model with precipitation and lagged precipitation as inputs. Flood frequency analysis was conducted using annual maximum discharge series, and flood magnitudes for 2, 5, 10, 25, 50, and 100 year return periods were estimated using EasyFit under both baseline and future climate conditions.
The results showed that XGBoost significantly improved CMIP6 precipitation simulations by reducing systematic errors. MAE decreased from about 2.1–2.4 mm to 1.1–1.5 mm, and RMSE dropped from 6.2–7 mm to 4–4.8 mm. Correlation coefficients increased to 0.5–0.65, while NSE and KGE also showed notable improvements. The MME-XGBoost model further enhanced performance, with correlation coefficients reaching 0.78–0.84 and NSE values of 0.58–0.70. The ANFIS-PSO model effectively simulated daily streamflow, explaining about 84% of discharge variability during training with an NRMSE of 0.35. Flood frequency analysis indicated stronger climate change impacts at higher return periods. Under SSP370 and SSP585, extreme flood magnitudes increased significantly, with 50-year and 100-year floods rising by about 50% and 68% by 2076–2100 compared to the baseline. Floods also remained concentrated in late winter and early spring, with March showing the highest frequency.
This study shows that integrating machine learning with climate projections improves the assessment of future hydrological extremes. The XGBoost model effectively reduced biases in CMIP6 precipitation data, while the MME-XGBoost approach decreased uncertainties across climate models. The ANFIS-PSO model also provided reliable streamflow simulations under future scenarios. Results indicate a significant increase in extreme flood magnitudes under future climate conditions, especially in high-emission scenarios. Although seasonal flood patterns remain relatively stable, the intensity of extreme floods is expected to rise notably. These findings emphasize the need to update hydraulic design standards, flood management strategies, and climate adaptation policies in the Poldokhtar Basin and similar regions.
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.
All authors contributed equally to the conceptualization of the article and writing of the original and subsequent drafts.
There is no use of any type of artificial intelligence-based technology in this paper.
Data available on request from the authors.
The authors would like to thank anonymous referees for their constructive comments.
The authors have refrained from data fabrication, falsification, plagiarism, and any other form of research misconduct. Ethical principles were fully observed throughout the conduct and publication of this scientific research, and all authors confirm their compliance with these principles.
The authors declare no conflict of interest.
