We return from another record-breaking summer of heatwaves and extreme weather. Yet as days shorten and evenings cool, the topic of heat and urban development will remain a pressing concern for both science and planning. We illustrate how pre-trained deep learning models and open data can make it faster and less resource-intensive to predict the cooling impact of design interventions in urbanising regions.
Rising Urban Heat: A Pressing Challenge for Planners
Despite enormous scientific progress, overwhelming public attention and the pressing urgency of the issue, practical solutions to reduce local climate impacts are far from being implemented at scale. Current climate impact modelling sits at two extremes: high-resolution climate models that require proprietary software and expert knowledge, or more practical plug-and-play open-source models that generate more aggregate values on a neighbourhood scale.
As a result, current planning practice often relies on very accurate, but one-off, climate maps provided by government agencies. These maps are used to identify problem areas and propose general strategies to counteract, for example, local heat island effects. At present, only the most prestigious projects might commission dedicated climate modelling – and even then, this is the exception rather than a prerequisite. This reflects a broader science–action gap: while we know the fundamentals and can demonstrate them scientifically, what is missing are established processes that connect concrete design decisions with their microclimatic effects.
AI-Tools Make Microclimate Impacts Visible – and Actionable!
However, an increasingly used alternative are so-called emulator models. Instead of running full-scale physical simulations, these models are trained to reproduce the results of climate models based on fewer and simpler inputs. This makes it possible to estimate, for example, how a change in building height or new green space affects local temperature – without advanced climate expertise.
To briefly take a dive into the technical side: our approach uses a deep convolutional network pre-trained to generate maps of PET (Physiologically Equivalent Temperature) in Celsius for the Canton of Zürich. The model learns from openly available data: elevation models, RGB orthophotos, and building height maps. Even with this reduced input, the emulator reaches a test accuracy of around 97%.
This approach opens up the possibility of testing design scenarios quickly, cheaply, and without the bottleneck of specialised software, bringing climate-sensitive planning closer to everyday practice.

Dr. Michal Switalski is a postdoctoral researcher at the Planning of Landscapes and Urban Systems (PLUS, IRL) research group. He works on integrating diverse disciplines – ranging from climate and economics to place and place-making – into spatial and landscape planning, by developing measurement methods that draw on both traditional machine learning and deep learning.