A Neural Network-based downscaling method for near surface temperature in urban areas

Julia Garcia Cristobal, Jean Wurtz, Antoine Doury, Blanka Balogh, Valéry Masson, Olivier Mestre

Urban Climate · 01 August 2026 · Météo-France (incl. CNRM)

In one sentence. UNet with Sobel-penalty loss downscales AROME and ARPEGE to 300 m urban air temperature over Paris, with mean bias of about -0.3 to -0.5 °C against stations.

Abstract

Urban weather forecasts at hectometric resolution are essential for effective warning systems and sustainable urban planning, especially in the context of climate change and rapid urban expansion. While numerical weather prediction (NWP) systems can simulate the Urban Heat Island (UHI) effect, their computational cost limits their operational use. This study presents a convolutional neural network (CNN) based statistical downscaling (SD) method to emulate 300 m resolution 2 m air temperature over the Paris region, using coarse resolution inputs from AROME (2.5 km) and ARPEGE (10 km) operational models. This method leverages a UNet architecture, optimized with a Sobel operator penalty loss preserving fine-scale temperature gradients, and includes static and dynamic predictors. The approach is first validated in a super-resolution framework, using degraded high-resolution Meso-NH simulations (1.2 km to 9.6 km) to assess the model’s ability to recover fine-scale details. This demonstrates the potential of the methodology, achieving a Root Mean Squared Error (RMSE) of 0.37 °C when reconstructing 300 m temperature fields. The method is further evaluated using operational AROME and ARPEGE forecasts, validating against high-resolution Meso-NH simulations and in-situ observations from Météo-France’s stations. When compared to observations, the downscaled AROME SD model achieves − 0.33 ° C mean bias over the urban station, and the ARPEGE SD, whose inputs do not include any urban parameterization, a mean bias of − 0.46 ° C . This study demonstrates the efficiency of the SD in recovering hectometric-scale patterns, including UHI intensity and spatial heterogeneity, while offering a computationally efficient alternative to traditional NWP systems.

Labels

Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.