Leveraging neural networks for high-resolution urban heat assessments and adaptation modelling

Machine Learning techniques and neural network usage have experienced a sharp increase in applications in the last years in the domain of climate modelling and operational weather forecasting. With respect to high-resolution climate over cities however, their application is currently still limited, mainly attributed to a lack of accurate and long-term model simulation available to train these AI models.

The UrbClim model is an urban boundary layer model which offers fast and accurate long-term assessments of urban climate for any city in the world at spatial resolutions up to 100m. In this work we applied the model on 9 urban areas within the Basque Country for a period of 2001-2020. From this data, heat-related health indicators were calculated. Using this large source of data, we created a neural network model that allowed to expand the 100m resolution indicator outputs to the full Basque Country.

Since the neural network setup uses different climate and surface characteristics as prediction variables, the model furthermore allows to assess the impact and effectiveness of changes in the urban surface parameters on the climate within the city. Based on this, the impact of several nature-based solutions, e.g. greening in the city, unsealing,… can be assessed quantitatively at unprecedented execution times.

This setup has been integrated into an interactive web application, allowing policy makers, health practitioners and urban planners not only to assess vulnerable areas within the Basque Country, but also to simulate hypothetical adaptation scenarios and quantify in a matter of minutes the impact of different nature-based solutions on the climate and their potential positive benefits of reducing e.g. dangerous heat levels.

This work is executed within the Adaptation Modelling Framework for Destination Earth.