Enhancing ERA5 precipitation with improved predictor selection for regional climate change assessment

Abstract

Accurate regional precipitation projections are critical for effective climate impact assessment and adaptation planning. This study presents a novel methodology for enhancing ERA5 reanalysis precipitation data through optimized predictor selection and statistical downscaling using the Multivariate Adaptive Regression Splines (MARS) algorithm. Four distinct predictor selection scenarios: a full 26-variable model, a reduced 14-variable model based on correlation and physical relevance, a compact 6-variable model emphasizing simplicity, and a station-specific model derived from All Possible Regression (APR), were used along with the MARS algorithm. Predictor variables were selected through traditional correlation analyses (Pearson and Spearman), the APR-based approach, and performance-based evaluation using MARS. The resulting downscaled models were evaluated using different performance metrics, including Kling-Gupta Efficiency (KGE), Nash-Sutcliffe Efficiency (NSE), normalized Root Mean Square Error (nRMSE), and the coefficient of determination (R-2). The Western Black Sea Basin in T & uuml;rkiye, with monthly precipitation data from 32 meteorological stations (1979-2023), was selected as an application to apply the newly proposed dual-stage approach. Results demonstrated that all MARS-enhanced models significantly outperformed the raw ERA5 data, particularly in inland regions where ERA5 performance was initially poor. The APR-based model emerged as the top performer across most stations, while the 6-variable model provided a strong balance between accuracy and simplicity. While the nRMSE initially reached around 77% at some stations, it was significantly reduced to 24.6%, 29%, 26.4%, and 25.1% under the 26-variable, 14-variable, 6-variable, and APR scenarios. The KGE nearly doubled, reaching approximately 0.7-0.9 across all scenarios, confirming the substantial improvement applied to the ERA5 precipitation data. This approach, integrating correlation-based and predictive performance-driven variable selection, proved effective in refining regional precipitation projections. The methodology can be adapted to other regions or climate variables, offering a replicable framework for improving the usability of reanalysis data in hydrological and climate impact studies.