Master Thesis defense by Mari Michelle Knudson

Title: Urban Heat Island Dynamics in a Changing Climate: Insights from Machine Learning

Abstract: 
Using the new DANish regional atmospheric ReAnalysis (DANRA) dataset from the Danish Meteorological Institute, this thesis utilized machine learning (ML) to explore the canopy layer urban heat island (UHI) phenomenon in Copenhagen and its interactions with background weather patterns and urban surface characteristics. First, a baseline climatology for the UHI in Copenhagen was established over 2003-2023 using data from in-situ meteorological stations, finding a maximum value of 8.7°C, with a mean value of 2.4 °C. The daily maximum UHI intensity was found to correlate strongly with wind speed, solar radiation intensity, and the rural diurnal temperature range (DTR). A seasonal cycle was seen for the daily maximum UHI intensity, with the highest mean value in summer and lowest in winter.

A predictive model of the spatially varying daily maximum UHI intensity was then built using a combination of DANRA and station observation data as meteorological input, along with remote sensing based urban surface predictors. After initial testing, the model using LightGBM regression was found to be the best-performing and its hyperparameters were optimized. The resulting model had a performance on a held-out test set of R2 = 0.71 and RMSE = 0.57 °C, with better explanatory power under calm and clear conditions and less under windy and cloudy conditions.

The feature importance tests for the best performing model suggest that the rural DTR, wind speed, and solar radiation intensity were the most important to the model meteorological predic- tors, and local climate zone (LCZ) class, population density, distance to the coast, vegetation fraction, and elevation were the most important geospatial predictors. The model results support the feasibility of the method in predicting the maximum daily UHI intensity over Copenhagen, and future improvements and model applications are suggested.

Supervisors: Jens Hesselbjerg Christensen, Alexander Baklanov, and Maher Sahyoun
Censor: Martin Drews, DTU