Graph Neural Networks for Urban Heat Island Forecasting using Satellite data
Development of a GNN-based architecture operating on satellite imagery, urban topology, and meteorological data to predict neighborhood-scale temperature distributions.
Requirements
- M.Sc. in Data Science, Computer Science, Artificial Intelligence, Mathematics, or similar
- Strong knowledge of Python and deep learning frameworks (PyTorch, Lightning)
- Experience with computer vision and time series analysis
- Experience with geospatial data processing and satellite imagery (Optical and thermal satellite data, DEM, land cover)
- Understanding of graph manipulation (NetworkX, OSMnx)
- Understanding of Graph Neural Networks (GCNs, GraphSAGE)
Description
Urban Heat Islands (UHIs) are localized zones within cities that experience significantly higher temperatures than their surroundings due to dense built environments, limited vegetation, and anthropogenic heat. Analyzing the thermal behavior of air at the urban level requires very high resolutions to accurately resolve neighborhood-scale variability. While both Numerical Weather Prediction (NWP) and image-based Deep Learning (DL) can model these environments, they are both often computationally intensive, requiring high memory costs and powerful hardware.
This thesis proposes evolving a current CNN-based temperature mapping approach into a graph-based formulation that explicitly reflects the physical topology of the urban environment. The research focuses on constructing a city graph where intersections and point of interest serve as nodes and road segments serve as edges. Each node in the graph will contain feature vectors derived from remote sensing data, including thermal bands, elevation, land-use fractions and other meteorological indicators.
The model will be trained to predict the daily maximum temperature as the target value for each node location. By shifting from a dense raster to this targeted graph structure, the approach aims to reduce the computational complexity and memory requirements, as inference is performed only on meaningful urban elements rather than every pixel in a grid.
The research will address the following objectives:
- Temperature Data Sourcing: Searching for and integrating historical temperature data from urban weather stations to act as ground-truth nodes.
- Graph Construction & Manipulation: Using OSMnx and NetworkX to build a 3-D urban graph where edges encode spatial distances.
- Multimodal Integration: Incorporating satellite-derived variables and meteorological indicators as node features.
- Model Architecture Development: Implementing GNN architectures, like GCNs or GraphSAGE.
- Visualization: Implementing interpolation techniques to visualize results.