Enhancing precipitation intensity estimation using ERA5-land reanalysis with statistical and machine learning approaches
Alireza Abdolmanafi, Bahram Saghafian, Saleh Aminyavari
Abstract
• ERA5-Land exhibits limitations in certain areas, particularly for longer return periods, necessitating bias correction. • The RF method performed better for shorter return periods, while QM was more effective for longer return periods. • Integrating reanalysis data with machine learning techniques can significantly improve precipitation intensity estimates. Intensity-Duration-Frequency (IDF) curves are essential tools in hydrology, illustrating the relationship between precipitation intensity, duration, and frequency. In areas lacking dense meteorological networks, reanalysis and satellite-based data offer practical alternatives for estimating precipitation intensity. This study uses ERA5-Land reanalysis data to estimate 6-hourly precipitation intensities via IDF curves from 1982 to 2023 for 71 meteorological stations across Iran. To improve estimation accuracy, two post-processing techniques Quantile Mapping (QM) and Random Forest (RF) were applied. Evaluation was conducted at two levels: nationally using Root Mean Squared Error (RMSE) and BIAS, and regionally using Mean Absolute Error (MAE) and Kling-Gupta Efficiency (KGE) across five regions. Results show that ERA5-Land data has limitations, particularly for longer return periods, underscoring the need for bias correction. RF performed better for shorter return periods (5 and 50 years), while QM was more effective for longer ones (100 and 200 years). Regionally, QM had greater impact in the Caspian coast and mountainous areas (R2, R3, R4), whereas RF was more effective in arid and semi-arid regions (R1, R5). Spatial analysis indicated that bias correction was most beneficial in areas with high precipitation variability. RF demonstrated better generalization and consistently reduced overall errors more effectively than QM. These findings highlight the importance of region-specific correction methods and suggest that combining reanalysis data with machine learning and statistical techniques can improve flood risk assessments and hydrological planning.