This talk…

… is not (directly) about migration, climate change, or economic development

… is about using different data (📡 🌎) to study socio-economic systems

… uses cities as a useful illustration, but “the sky is the limit”

Cities and …

Cities and …

Why satellite data?

• Low (marginal) cost
• Near real-time
• Ripe for disruption

i.e., this is our “20-year-big-thing”…

But…

• Unstructured
• Big
• Only measure what can be “seen”

Why satellite data?

British Signatures

Source: Sentinel-2 cloudless

Wild countryside (320x320m)

Urbanity (320x320m)

Predicted class (320x320m)

Experiment’s dimensions

Chip size

[74%]

[57%]

[35%]

[13%]

(Spatial) data augmentation

Model architecture

EfficientNetB4 Image Classification Multi-Output Regression

Evaluation

Metrics Standard κ, accuracy, F1 Spatial Joint Counts

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)

Wrap up

Wrap up

Wrap up

• We need to measure cities more, more frequently
• We can thanks to the 📡 🌎 + 💻 + 🤖 combo
• How is still W.I.P., but with a stress on the P

Appendix

Building Spatial Signatures

Enclosed Tessellation	Embedding form & function	Spatial Signatures
+

Meaningful spatial units

Hierarchy

Sub-classifications