ESA CLIM4Cities - Urban Thermal Signal Downscaling (UTS-D) model

The Urban Thermal Signal Downscaling (UTS-D) machine-learning model downscales near-surface air temperature fields for the Danish functional urban areas of Copenhagen, Aarhus, Aalborg, and Odense. The Random Forest model takes the Danish regional atmospheric reanalysis (DANRA) and geospatial predictors associated with urban heat island processes as input:

• Relative Humidity (DANRA)
• Air Temperature at 2m (DANRA)
• Air Temperature at 2m, from previous 3h (DANRA)
• Day of the Year
• Hour of the day
• Month
• Wind Velocity (DANRA)
• Maximum vertical angle to terrain toward the dominant wind direction
• Distance to coastline
• Fraction of area covered by impervious surface
• Fraction of area covered by tree canopy
• Bowen ratio associated with Local Climate Zone classification
• Surface altitude

The model is trained to predict the bias between the DANRA Air temperature at 2m and the air temperature data from crowdsourced citizen weather station observations (Netatmo), allowing for the generation of high-resolution temperature fields at approximately 0.002 x 0.002º resolution from the original 2.5 × 2.5 km reanalysis grid. The resulting air temperature maps allows for neighbourhood-scale analysis of urban thermal patterns and supports applications in urban climate research, heat-risk assessment, and local climate adaptation planning.

The model development was conduced within the CLIM4cities project funded by the European Space Agency (ESA Contract No. 4000143628/24/I-DT, AI Trustworthy Applications for Climate).