Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Understanding the effects of Urban Green Spaces and Building Characteristics on Land Surface Temperature using multi-source geodata and Deep Learning

Wienker, Frederik LU (2025) In Student thesis series INES NGEM01 20251
Dept of Physical Geography and Ecosystem Science
Abstract
Urban areas are increasingly facing heat stress which poses a significant health risk to growing populations. Land Surface Temperature (LST), as a strong indicator for urban temperature analysis, is substantially influenced by the composition of a city. Urban Green Spaces (UGS) and buildings are especially important in determining LST and therefore are strong factors in identifying potential urban heat island effects (UHI). This study investigates how UGS and building characteristics influence LST, using the example of Berlin as study area. The study’s approach also aims to set up a Deep Learning (DL)-based workflow testing UNet and DeeplabV3+ models’ capacities in segmenting UGS from Sentinel-2 imagery in 10m resolution. UNet showed... (More)
Urban areas are increasingly facing heat stress which poses a significant health risk to growing populations. Land Surface Temperature (LST), as a strong indicator for urban temperature analysis, is substantially influenced by the composition of a city. Urban Green Spaces (UGS) and buildings are especially important in determining LST and therefore are strong factors in identifying potential urban heat island effects (UHI). This study investigates how UGS and building characteristics influence LST, using the example of Berlin as study area. The study’s approach also aims to set up a Deep Learning (DL)-based workflow testing UNet and DeeplabV3+ models’ capacities in segmenting UGS from Sentinel-2 imagery in 10m resolution. UNet showed superior segmentation results with F1-Score of 0.83 and an Intersection over Union (IoU) of 0.71. Both models gave promising results in identifying larger UGS like urban parks but encountered some segmentation deficits in identifying smaller UGS patches. To explore the relationship between LST and identified urban features, a Random Forests (RF) approach not only assessed how well these features can explain LST. It also analysed and determined each feature’s importance on the LST prediction. By integrating both building and UGS-related features, a high predictive accuracy with R² of 0.8 was achieved. Larger UGS were also found to have the greatest potential in cooling its proximate environment. Building characteristics, on the other hand, showed less substantial LST correlations. The proposed framework supports an open-data-driven planning approach in segmenting UGS and analysing important urban LST associations, important for climate-adaptation planning. (Less)
Popular Abstract
The urban heat island effect, where temperatures in cities get higher compared to their rural surroundings, has a significant impact on its citizens. Higher temperatures in increasingly dense areas can increase discomfort, energy use and health risks. The most important indicators for urban temperatures are Green Spaces and Buildings. So, how do we assess these effects effectively?
This project intends to contribute to this general problem both methodologically and practically, focusing on a specific case study. The first objective is to use state-of-the-art Deep Learning models to identify Urban Green Spaces from freely-available 10m resolution satellite imagery. Secondly, the aim is to assess the relationship between Land Surface... (More)
The urban heat island effect, where temperatures in cities get higher compared to their rural surroundings, has a significant impact on its citizens. Higher temperatures in increasingly dense areas can increase discomfort, energy use and health risks. The most important indicators for urban temperatures are Green Spaces and Buildings. So, how do we assess these effects effectively?
This project intends to contribute to this general problem both methodologically and practically, focusing on a specific case study. The first objective is to use state-of-the-art Deep Learning models to identify Urban Green Spaces from freely-available 10m resolution satellite imagery. Secondly, the aim is to assess the relationship between Land Surface Temperature and different urban characteristics, representing Green Spaces and Built-up areas.
To explore the first research objective, two Deep Learning models were trained to automatically extract Green Spaces from Sentinel-2 imagery. This binary classification was conducted over the main city of Berlin. The models’ performance were evaluated using ESA’s Land Cover Classification data as validation. After tuning the model’s performance, the best output map covering urban green spaces was selected. Secondly, building footprints and heights from Berlin were collected and a downscaled Land Surface Temperature Dataset was derived from August 2024. An explainable machine learning technique was used to analyse the relationship between urban features and local temperatures. The analysis focused not only on pixel-based values but also accounted for spatial autocorrelation by including neighbourhood-based values.
The pre-trained deep learning segmentation models performed similarly well in terms of metrics, with accuracy, precision and F1 scores all above 0.8. Visually, however, the UNet model produced clearer green space boundaries, while DeeplabV3+ tended to over-smooth and generalise. The temperature analysis showed that larger urban green spaces have a noticeable cooling effect. Areas of 1 hectare or more were on average 2°C cooler than similar built-up areas. Building features such as height and density showed weaker and more inconsistent effects. While taller buildings (over 25 m) had a slight cooling effect, buildings overall were associated with higher surface temperatures, but with less impact than expected. (Less)
Please use this url to cite or link to this publication:
author
Wienker, Frederik LU
supervisor
organization
course
NGEM01 20251
year
type
H2 - Master's Degree (Two Years)
subject
keywords
Physical Geography and Ecosystem analysis, Urban Green Spaces (UGS), Land Surface Temperature (LST), Remote Sensing (RS), Deep Learning Segmentation
publication/series
Student thesis series INES
report number
714
language
English
id
9204052
date added to LUP
2025-06-23 13:21:39
date last changed
2025-06-23 13:21:39
@misc{9204052,
  abstract     = {{Urban areas are increasingly facing heat stress which poses a significant health risk to growing populations. Land Surface Temperature (LST), as a strong indicator for urban temperature analysis, is substantially influenced by the composition of a city. Urban Green Spaces (UGS) and buildings are especially important in determining LST and therefore are strong factors in identifying potential urban heat island effects (UHI). This study investigates how UGS and building characteristics influence LST, using the example of Berlin as study area. The study’s approach also aims to set up a Deep Learning (DL)-based workflow testing UNet and DeeplabV3+ models’ capacities in segmenting UGS from Sentinel-2 imagery in 10m resolution. UNet showed superior segmentation results with F1-Score of 0.83 and an Intersection over Union (IoU) of 0.71. Both models gave promising results in identifying larger UGS like urban parks but encountered some segmentation deficits in identifying smaller UGS patches. To explore the relationship between LST and identified urban features, a Random Forests (RF) approach not only assessed how well these features can explain LST. It also analysed and determined each feature’s importance on the LST prediction. By integrating both building and UGS-related features, a high predictive accuracy with R² of 0.8 was achieved. Larger UGS were also found to have the greatest potential in cooling its proximate environment. Building characteristics, on the other hand, showed less substantial LST correlations. The proposed framework supports an open-data-driven planning approach in segmenting UGS and analysing important urban LST associations, important for climate-adaptation planning.}},
  author       = {{Wienker, Frederik}},
  language     = {{eng}},
  note         = {{Student Paper}},
  series       = {{Student thesis series INES}},
  title        = {{Understanding the effects of Urban Green Spaces and Building Characteristics on Land Surface Temperature using multi-source geodata and Deep Learning}},
  year         = {{2025}},
}