Comparing ML Methods for Downscaling Near-Surface Air Temperature over the Eastern Mediterranean

Abstract

Near-surface air temperature (Ta) is a key variable in global climate studies. Global climate models such as ERA5 and CMIP6 predict various parameters at coarse spatial resolution (>9 km). As a result, local phenomena such as the urban heat islands are not reflected in the model’s outputs. In this study, we address this limitation by downscaling the resolution of ERA5 (9 km) and CMIP6 (27 km) Ta to 1 km, employing two different machine learning algorithms (XGBoost and Deep Learning). Our models leverage a diverse set of features, including data from satellites (land surface temperature and normalized difference vegetation index), from ERA5 and CMIP6 climate models (e.g., solar and thermal radiation, wind), and from digital elevation models to develop accurate machine learning prediction models. These models were rigorously validated against observations from 98 meteorological stations in the East Mediterranean (Israel) using a standard cross-validation technique as well as a leave-one-group-out on the station ID evaluation methodology to avoid overfitting and dependence on geographic location. We demonstrate the sensitivity of the downscaled Ta to local land cover and topography, which is missing in the climate models. Our results demonstrate impressive accuracy with the Deep Learning-based models, obtaining Root Mean Squared Error (RMSE) values of 0.98 °C (ERA5) and 1.86 °C (CMIP6) for daily Ta and 2.20 °C (ERA5) for hourly Ta. Additionally, we explore the impact of the various input features and offer an extended application for future climate predictions. Finally, we propose an enhanced evaluation framework, which addresses the problem of model overfitting. This work provides practical tools and insights for building and evaluating Ta downscaling models. The code and data are publicly shared online.