Simulation and evaluation of local daily temperature and precipitation series derived by stochastic downscaling of ERA5 reanalysis

2025-07-02

AI Summary

ERA5 reanalysis data exhibit limited representativeness at local scales—particularly for precipitation—hindering high-resolution climate impact assessments. To address this, we propose a stochastic downscaling method that integrates in situ station observations with ERA5 data within a nonlinear regression framework. Our approach innovatively couples generalized additive models (GAMs), regression splines, and ARMA time-series modeling to jointly enhance spatial localization and temporal dynamism while preserving physical consistency. Validated across over 4,000 European stations at daily resolution over a 60-year period, the downscaled temperature and precipitation series significantly outperform raw ERA5: local biases are reduced by more than 35%, and the representation of extreme precipitation events is markedly improved. This method establishes a high-accuracy, broadly applicable downscaling paradigm for regional hydrological modeling and climate impact studies.

Technology Category

• Machine Learning: Ensemble Methods
• Reasoning under Uncertainty: Stochastic Optimization
• Knowledge Representation and Reasoning: Geometric, Spatial, and Temporal Reasoning

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Graph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphs
• Web Mining and Content Analysis: Robustness and generalizability of Web mining methods

Abstract

Reanalysis products such as the ERA5 reanalysis are commonly used as proxies for observed atmospheric conditions. These products are convenient to use due to their global coverage, the large number of available atmospheric variables and the physical consistency between these variables, as well as their relatively high spatial and temporal resolutions. However, despite the continuous improvements in accuracy and increasing spatial and temporal resolutions of reanalysis products, they may not always capture local atmospheric conditions, especially for highly localised variables such as precipitation. This paper proposes a computationally efficient stochastic downscaling of ERA5 temperature and precipitation. The method combines information from ERA5 and surface observations from nearby stations in a non-linear regression framework that combines generalised additive models (GAMs) with regression splines and auto-regressive moving average (ARMA) models to produce realistic time series of local daily temperature and precipitation. Using a wide range of evaluation criteria that address different properties of the data, the proposed framework is shown to improve the representation of local temperature and precipitation compared to ERA5 at over 4000 locations in Europe over a period of 60 years.

Problem

• Improving local temperature and precipitation accuracy from ERA5
• Downscaling ERA5 data using GAMs and ARMA models
• Enhancing representation of local weather over 4000 European sites

Innovation

• Stochastic downscaling of ERA5 temperature and precipitation
• Combines GAMs, regression splines, and ARMA models
• Improves local temperature and precipitation representation