High-resolution multi-sensor fusion for urban temperature downscaling: integrating physical covariates and stochastic extremes for health impact assessment

Under the impact of ongoing climate change, urban environments exhibit warming rates significantly exceeding those of surrounding rural areas. Assessing the resulting impacts on human health requires high-resolution data that current Earth Observation (EO) products struggle to provide. We address the inherent trade-off between temporal frequency and spatial detail by integrating a multisensor data suite, comprising SEVIRI (high temporal), MODIS (daily), and Landsat LST (high spatial) EO data, weather reanalysis and dense urban station network data, to train a machine learning-based downscaling framework. 

Our methodology generates city-wide temperature maps at 100m resolution every hour. To capture the complex physical drivers of the urban canopy layer, the model incorporates a diverse array of covariates, including land cover, spectral indices, sky view factor, and building morphology. The model also accounts for energy balance variables such as anthropogenic heat emissions and utilizes a precipitation proxy to simulate the effects of evaporative cooling on surface temperatures. 

We validate this framework across the city of Newcastle (UK), utilizing the high-density Urban Observatory sensor network to train and benchmark model accuracy against ground-truth data. We further evaluate the model’s transferability through case studies in the cities of London (high-density metropolitan landscape) and Southampton (coastal-urban interactions), representing environments with sparse ground networks. 

A key aspect is the generation of design heatwaves and cold snaps. Rather than relying solely on historical records, we employ stochastic modeling to produce extreme weather series that allow for the high-resolution quantification of the full spectrum of thermal risk, including unprecedented events. These outputs serve as the primary forcing data for indoor thermal simulations and population exposure models within the ETHOS project. Ultimately, this framework provides clinicians and local authorities with the precise spatial risk information needed to protect vulnerable populations during thermal extremes.