pyESDv1.0.1: an open-source Python framework for empirical-statistical downscaling of climate information

Abstract

Abstract. The nature and severity of climate change impacts vary significantly from region to region. Consequently, high-resolution climate information is needed for meaningful impact assessments and the design of mitigation strategies. This demand has led to an increase in the application of empirical-statistical downscaling (ESD) models to general circulation model (GCM) simulations of future climate. In contrast to dynamical downscaling, the perfect prognosis ESD (PP-ESD) approach has several benefits, including low computation costs, the prevention of the propagation of GCM-specific errors, and high compatibility with different GCMs. Despite their advantages, the use of ESD models and the resulting data products is hampered by (1) the lack of accessible and user-friendly downscaling software packages that implement the entire downscaling cycle, (2) difficulties reproducing existing data products and assessing their credibility, and (3) difficulties reconciling different ESD-based predictions for the same region. We address these issues with a new open-source Python PP-ESD modeling framework called pyESD. pyESD implements the entire downscaling cycle, i.e., routines for data preparation, predictor selection and construction, model selection and training, evaluation, utility tools for relevant statistical tests, visualization, and more. The package includes a collection of well-established machine learning algorithms and allows the user to choose a variety of estimators, cross-validation schemes, objective function measures, and hyperparameter optimization in relatively few lines of code. The package is well-documented, highly modular, and flexible. It allows quick and reproducible downscaling of any climate information, such as precipitation, temperature, wind speed, or even short-term glacier length and mass changes. We demonstrate the use and effectiveness of the new PP-ESD framework by generating weather-station-based downscaling products for precipitation and temperature in complex mountainous terrain in southwestern Germany. The application example covers all important steps of the downscaling cycle and different levels of experimental complexity. All scripts and datasets used in the case study are publicly available to (1) ensure the reproducibility and replicability of the modeled results and (2) simplify learning to use the software package.

Comment on egusphere-2024-958

“…The empirical downscaling techniques involve laborious steps that must be carefully addressed to ensure the quality of the local climate series reconstructions, as pointed out in Boateng and Mutz (2023). RASCAL is a Python library that implements the analog method in a clear and simple way.…”

“…It is an object-oriented library with four main blocks or classes: Station, Predictor, Analogs and RSkill. This library is a valuable complement to other empirical downscaling libraries, such as pyESD from Boateng and Mutz (2023), which is based on machine learning downscaling methods and focus on generating monthly time series. RASCAL is based on classical statistical methods, which produce results that are easier to interpret physically, and additionally, it is more focused on daily resolution reconstructions rather than monthly, which allows for the calculation of…”

“…However, selecting and processing the predictor correctly is crucial for achieving accurate local weather reproduction. The selection should be based on our knowledge of atmospheric dynamics and the local climate of the study area, as pointed out by other authors (Boateng and Mutz, 2023). After selecting a predictor variable that is expected to have a strong relationship to the predictand variable, it is necessary to choose a predictor field domain that can identify relevant synoptic patterns for the study area.…”