Urban Heat Island Downscaling

Resolving urban heat at street level.

Ongoing Geospatial
  • Python
  • PyTorch
  • Google Earth Engine
  • GCP
  • Landsat
  • Sentinel-2
  • Deep Learning
  • Remote Sensing

Problem

Cities need fine-grained LST maps to evaluate cooling interventions, but Landsat thermal data is too coarse (100m) for street-level analysis. No open high-resolution thermal baseline exists for Berlin.

Solution

Deep learning model downscales Landsat LST to Sentinel-2 resolution (10m) using spectral and structural features. Causal analysis links local adaptation measures (trees, sealed surfaces, water bodies) to temperature differences across the city.

Result

In progress. Target: reproducible Berlin UHI map at 10m resolution + quantified effect estimates per adaptation category.

Lessons Learned

  • To be filled after completion.

Deep Dive

Data stack: Landsat 8/9 thermal band, Sentinel-2 optical bands, Geoportal Berlin (ISU5), DWD climate stations. Infrastructure: GCP (Vertex AI, BigQuery), Google Earth Engine for preprocessing.