Performance Evaluation of Machine Learning and Traditional Statistical Approaches in Bias Correction of CMIP6 Precipitation Data

Document Type : Research Paper

Authors

• mohammad fallah kalaki ¹
• Asghar Azizian ²
• Hadi Ramezani Etedali ³

¹ PHD Student in Water Resources Engineering, Water Science and Engineering, IKIU University, Qazvin, Iran

² Associate Professor, Water Science and Engineering Department, Imam Khomeini International University (IKIU), Qazvin, Iran

³ Professor, Water Science and Engineering Dept., IKIU University, Qazvin, Iran

Abstract

Keywords

• Climate change
• Bias correction
• Machine learning
• Poldokhtar watershed

Main Subjects

• ...

Introduction

Increasing greenhouse gas concentrations and climate change significantly affect water resources and hydrological processes, while extreme events such as floods and droughts impose substantial economic and social impacts (Almazroui et al., 2020; Kim et al., 2021; Aryal et al., 2019; UNISDR, 2009; Jongman et al., 2012; Hirabayashi et al., 2021; Houshmand Kouchi et al., 2019). General Circulation Models (GCMs) are the primary tools for predicting climate change, but their coarse spatial resolution requires bias correction for local and regional studies (Sachindra et al., 2018; Taylor et al., 2012). The CMIP6 project, providing SSP and RCP scenarios, enables higher-accuracy simulations of future climate changes (Eyring et al., 2016; O’Neill et al., 2016). Traditional bias correction methods, such as Linear Scaling (LS) and Quantile Mapping (QM), effectively reduce systematic errors but have limitations in capturing nonlinear relationships and temporal dependencies (Shiru and Park, 2020; Jaiswal et al., 2022; Heshmati et al., 2025). Recently, machine learning algorithms, including XGBoost and LSTM, have shown improved performance in modeling complex precipitation patterns and providing more accurate bias corrections (Li et al., 2023; Tanimu et al., 2024). This study evaluates the performance of both traditional and machine learning-based bias correction methods on CMIP6 precipitation outputs under SSP scenarios in the Poldokhtar watershed, aiming to reduce uncertainty and improve hydrological predictions and water resources management.

Method

This study focuses on the Poldokhtar watershed, a sub-basin of the Karkheh River in western Iran, covering an area of 2,073 km², with the Kashkan River as its main stream. The data used include daily projections from General Circulation Models (GCMs) in 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 within the watershed were used to correct biases in the GCM outputs. Three bias correction methods Linear Scaling (LS), Quantile Mapping (QM), and the machine learning algorithm XGBoost were evaluated. Additionally, a Long Short-Term Memory (LSTM) recurrent neural network was applied to capture long-term temporal dependencies and achieve more accurate precipitation adjustments. Model performance before and after bias correction was assessed using statistical metrics including MAE, RMSE, Pearson correlation coefficient, NSH, and KGE. Based on these metrics, GCMs were ranked at each station, and the cumulative ranks across the three stations were used to select the best-performing model for the watershed. This approach enables more accurate precipitation predictions and supports the analysis of climate change impacts on water resources.

Results

In this study, the performance of 23 CMIP6 climate models in simulating precipitation at three stations Poldokhtar, Afrineh, and Doab Veisian was evaluated. Results indicated that the raw model outputs exhibited considerable bias, with high RMSE and MAE values and low correlation and NSE, performing even worse than the simple observational mean. After applying bias correction methods, the linear scaling (LS) approach only slightly reduced errors and provided limited improvements in performance metrics, while the quantile mapping (QM) method generally worsened model performance across most indices. The XGBoost machine learning algorithm significantly reduced systematic errors and improved RMSE, MAE, NSE, and KGE, leading to the identification of selected models such as CanESM6, MIROC6, and GISS-E2-1-G. The LSTM approach also reduced mean errors but failed to reproduce variance and temporal correlation structures. Future precipitation projections using XGBoost-corrected CanESM6 indicated increases in winter and spring precipitation and notable monthly seasonal shifts, particularly under more severe SSP370 and SSP585 scenarios, while summer precipitation remained largely unchanged. These findings highlight that combining climate models with advanced machine learning techniques substantially enhances the reconstruction of historical precipitation and the assessment of climate change impacts.

Conclusions

In this study, the performance of 23 CMIP6 climate models in simulating precipitation at three stations—Poldokhtar, Afrineh, and Doab Veisian—was evaluated. Raw model outputs exhibited significant errors and low correlation. The LS method provided only slight improvements, while QM performed worse. Machine learning approaches, particularly XGBoost, substantially enhanced model accuracy by reducing MAE and RMSE and improving NSE and KGE. The XGBoost-corrected CanESM6 model showed the best performance and was used for future scenario projections. Predictions indicate increased precipitation in spring and autumn and slight decreases in winter, highlighting important applications for water resources management and climate change forecasting.

Author Contributions

“Conceptualization, M.F.K. and A.A.; methodology, M.F.K and A.A.; software, M.F.K.; validation, M.F.K. and A.A.; formal analysis, M.F.K., A.A. and H.R.E.; investigation, M.F.K., A.A. and H.R.E.; resources, X.X.; data curation, M.F.K and A.A.; writing original draft preparation, M.F.K.; writing review and editing, M.F.K and A.A.; visualization, M.F.K and A.A..; supervision, A.A. and H.R.E.; project administration, A.A.; funding acquisition, A.A. All authors have read and agreed to the published version of the manuscript.”

Data Availability Statement

Data available on request from the authors.

Acknowledgements

We sincerely thank all individuals who contributed to the completion of this research.

Ethical considerations

The authors avoided data fabrication, falsification, plagiarism, and misconduct.

Conflict of interest

The author declares no conflict of interest.