Using deep learning to identify (urban) form and function in satellite imagery - the case of Great Britain


Martin Fleischmann
Dani Arribas-Bel

How we arrange “stuff” in cities matters…

Source: A map of every building in America (New York Times)

… it matters a lot

Spatial Signatures

A characterisation of space based on form and function designed to understand urban environments

A characterisation of space based on form and function designed to understand urban environments

A characterisation of space based on form and function designed to understand urban environments

A characterisation of space based on form and function designed to understand urban environments

16 signature types, 3 groups

Countryside (3 types)

Periphery (4 types)

Urban (9 types)

The issue

Data

Form
  • OS OpenMap
  • OS OpenRoads
Function
  • (Business) Census
  • OpenStreetMap
  • Geolytix
  • Listed buildings
  • CDRC
  • CORINE / Sentinel 2
  • VIIRS

Possible solution?

Sentinel 2

What do we want to do?

train a neural network
understand the role of geography

Exploration

Image classification

Neural network architecture

Chip size effect

80x80m: 13760 chips, which is 74 % of maximum

160x160m: 2718 chips within, which is 57 % of maximum

320x320m: 423 chips within, which is 35 % of maximum

640x640m: 38 chips within, which is 13 % of maximum

Overall accuracy 42.8%

Wild countryside

Urbanity

Co-location

Multi-output regression

320x320m, chips capturing the proportion

Overall accuracy 43.5%

Wild countryside

Urbanity

Probability modelling

*the accuracy is based on a different sample than in previous cases (WIP)

A way forward

  1. Deploy probability modelling on GB
  1. Image segmentation
  1. Alternative NN architecture including additional context in a single model

Conclusions

  1. Relationship between signatures and satellite data is fuzzy
  1. Chip size needs to balance information and relation to input geometry
  1. Co-location information needs to be embedded in the model

Using deep learning to identify (urban) form and function in satellite imagery - the case of Great Britain

Martin Fleischmann
Dani Arribas-Bel
[@martinfleis]
[@darribas]
m.fleischmann@liverpool.ac.uk
d.arribas-bel@liverpool.ac.uk