Abstract

This study investigates the potential of machine learning-based statistical downscaling to generate high-resolution (1 km) urban climate projections over Nicosia, Cyprus. A dynamical and a hybrid dynamical–statistical framework was applied to CMIP6 data under the SSP2-4.5 scenario to assess mid-twenty-first-century summer thermal conditions, focusing on daily maximum (T2max) and minimum (T2min) 2 m air temperature. The WRF model was used to dynamically downscale ERA5 reanalysis data to produce reference datasets for training three statistical models of increasing complexity: Multiple Linear Regression (MLR), Artificial Neural Networks (ANN), and Convolutional Neural Networks (CNN). All models successfully reproduced the spatial temperature patterns simulated by WRF, with the CNN achieving the best performance and improved spatial consistency. Evaluation against observations confirmed that the statistical models capture daily maximum and minimum temperature variability with accuracy comparable to the dynamical model. Both the WRF model and the trained statistical models were subsequently applied to CMIP6-driven predictors to produce high-resolution projections. Results indicate pronounced summer warming over Nicosia by mid-century (1.9 °C for T2max and 1.1 °C for T2min), with strong spatial heterogeneity and intensified heat hotspots in compact urban areas (LCZ 3). These findings highlight the value of integrating dynamical and machine learning approaches for efficient urban-scale climate assessments.

IPC Classification

Keywords