Abstract
Extreme heat has become one of the deadliest natural disasters that humans have to face. Especially for residents living in cities, the urban heat island effect makes the consequences of hightemperatures even more severe. In order to study future temperatures at the micro level of a
city, it is necessary to model the effect of the city’s properties (land use, population, etc.) on the
local temperature.
However, most models that can make long-term predictions are low-resolution outputs, and
only by downscaling their results can they be applied to the study of specific cities. The most
commonly used physical models for downscaling are accurate but often at the cost of high computational costs and long computational times. Meanwhile, with the rise of machine learning
and deep learning research in recent years, using statistical models to supplement or even replace
physical models has gradually become a problem worth studying.
Therefore, this thesis attempts to develop machine leaning models and deep leaning models
with different architectures to emulate UrbClim, a numerical urban model that has been successfully applied in hundreds of European cities. In particular, the 2m temperature variable is
the focus of our attention.
We used variables that are the same as the UrbClim input data, including ERA5 reanalysis
data and geospatial data, and built two models, MLP and U-Net, for performance comparison.
The final results show that the U-Net model can better replicate the UHI effect and generalize
to unseen data from unseen city and time periods, and has the potential to be further studied
in the urban climate modelling task
| Date of Award | 10 Jun 2025 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Lesley De Cruz (Promotor), Sara Top (Co-promotor) & Andrei Covaci (Advisor) |
Cite this
- Standard