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

Articles in Press

Document Type : Research Paper

Authors

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

10.22059/ijswr.2025.402248.670007

Abstract

Global climate models (GCMs) are key tools for analyzing and predicting future climate trends. However, these models often exhibit systematic errors (bias) in simulating climatic parameters. In this study, the impact of traditional statistical bias correction methods Linear Scaling (LS) and Quantile Mapping (QM) and machine learning–based methods Extreme Gradient Boosting (XGBoost) and Long Short-Term Memory neural networks (LSTM) was evaluated to improve the performance of CMIP6 models from the NEX-GDDP dataset in predicting precipitation for the historical period (1961–2014) and the future period (2025–2100) in the Poldokhtar watershed, an area highly sensitive to hydrological variability and frequent destructive floods. The results indicated that raw climate model outputs exhibit significant bias and are unsuitable for direct hydrological applications. The LS method moderately reduced errors and improved performance indices, while QM increased RMSE (up to 10.3 mm) and decreased NSE (down to –2.5). Among machine learning methods, XGBoost achieved the highest accuracy with increases in r (up to 0.67), NSE (up to 0.44), and KGE (by more than 0.4), whereas LSTM effectively corrected systematic errors but was limited in reproducing variability and temporal correlation. These findings provide a valuable foundation for future-oriented climate change analyses and water resource management in the Poldokhtar basin.

Keywords

Main Subjects

Iranian Journal of Soil and Water Research

Articles in Press, Accepted Manuscript
Available Online from 15 December 2025

Files

Share

How to cite

Statistics