Urban Grammar

“Learning from Deep Learning”
Lessons from using computer vision to identify (urban) form and function in open data satellite imagery

UCL-QML Turing workshop
Dani Arribas-Bel
@darribas
Martin Fleischmann
@martinfleis

“Previous season…”

This “season”

What

Explore the extent to which neural networks can recognise spatial signatures from satellite imagery

Why

  • Learn about Spatial Signatures (scale, context)
  • Explore the potential of NNs for cities
  • Work towards more frequent Spatial Signatures

Experiments setup

Dimensions to explore

Chip size


[74%]

[57%]

[35%]

[13%]

(Spatial) data augmentation

Model architecture

EfficientNetB4

  • Image Classification
  • Multi-Output Regression

Evaluation

Metrics

 Summarisation

Results

Summarisation summary

  • Extra ML pays off
  • M.O.R. worse in general, better within class
  • Spatial context always improves performance
  • Scale: larger is better, except for spatial patterning
  • Spatial sliding rarely (within-class)

Conclusions

  • Space matters for the spatial signatures
  • There’s value in combining NNs & other ML
  • A bit closer to frequent Spatial Signatures

Urban Grammar

“Learning from Deep Learning”
Lessons from using computer vision to identify (urban) form and function in open data satellite imagery

UCL-QML Turing workshop
Dani Arribas-Bel
@darribas
Martin Fleischmann
@martinfleis