Application of cGAN algorithm to divide a parcel of land (given a road network) into segmented housing lots

1. Typology of Subdivisions and
Parcel Lot Division With Image-
to-Image Translation Machine
Learning Algorithms
By: Matthew To and Guo Yu Wei

2. Research Questions
 How to simplify the approach to divide various
subdivisions into typologies?
 Is an "image-to-image translation with conditional
adversarial nets" a viable machine learning
algorithm for subdivision planning?

3. Subdivision Classification – Means to an
End?
What should subdivisions be classified for?
Identification for
planning design
Simplifying the
pix2pix algorithm

4. Subdivision Classification – Means to an
End?
What should subdivisions be classified for?
Identification for
planning design
Simplifying the
pix2pix algorithm
Same goal!

5. Subdivision Classification Proposed
Techniques
 Manual classification
 Individual judgement by principles
 Planning design resources
 Classification by statistics of attributes
 Density of housing via average parcel size
 Road to subdivision ratio
 Machine learning?
 Do we classify both target and input data?

6. Subdivision Classification
 Manual classification
 Few training datasets (30, planned 100)
 Shape analysis easier with humans than computers
 Selective modification
 Automated classification may be used once number
of datasets far exceeds human capabilities or
reliable method of classification is found
Originally from Report for the President's Conference on
Home Building and Home Ownership, 1932 by Thomas
Adams and Walter Baumgarten
Retrieved from Street Standards and the Shaping of Suburbia
by Southworth and Ben-Joseph

7. Parcel Lot Division with pix2pix
 pix2pix = conditional generative adversarial network (cCAN)
 TensorFlow implementation
 1. Takes a collection of image pairs
 2. Creates a model
 3. Uses model to produce output
 Why pix2pix?
 Designed to be general-purposed
 No other accessible algorithm currently

8. Conditional Generative Adversarial
Network (cGAN)?
 Conditional = takes user input to create output
 Generative = creates estimated outputs that improves with more training
 Adversarial = creates two neural networks (generator and discriminator) to
compete for a better output
 (Artificial) Neural Network = machine learning system which takes input,
processes the data within “hidden layers”, and returns an output
Images retrieved from https://affinelayer.com/pix2pix/, by Hesse

9. Conditional Generative Adversarial
Network (cGAN)?
 Conditional = takes user input to create output
 Generative = creates estimated outputs that improves with more training
 Adversarial = creates two neural networks (generator and discriminator) to
compete for a better output
 (Artificial) Neural Network = machine learning system which takes input,
processes the data in “hidden layers”, and returns an output
Images modified from https://affinelayer.com/pix2pix/, by Hesse

10. Methodology
 Training and target data retrieval and clean-up
 Export shapefiles into images
 Categorize training data
 Train and test pix2pix model
 Perform image segmentation on output

11. Methodology
 Training and target data retrieval and clean-up
 Export shapefiles into images
 Categorize training data*
 Train and test pix2pix model
 Perform image segmentation on output
* Only one result will have training data categorized

12. Training Data
 Study area:
 Independent of source, however data is from Kitchener-Waterloo area
 Rules:
 When training with borders, roads must be one colour, lots (pre and post
development) must be another
 When training with four colours, no lot with same colour can intersect
 All lots in training data must have a road bordering
 No redundancies (adds too much weight)

13. Results!

14. Experiment 1: Three Datasets
Input:
Target:
Output:

15. Experiment 1: Analysis
 It tried.
 Was unable to represent borders
 Patterns did not adapt to input polygon

16. Experiment 2: Larger Borders
Input:
Target:
Output:

17. Experiment 2: Larger Borders
 Borders are better drawn compared to Experiment 1
 Borders do not connect well
 Shapes are unstructured
 Lots generated in-land without a bordering road
 Rule where every lot must border a road was not
enforced

18. Experiment 3: 29 Datasets

19. Experiment 3: Outputs

20. Experiment 3: Outputs

21. Experiment 3: Rules Broken!

22. Experiment 3: Analysis
 Lines are shaky, but structured
 Lots are very small
 Lots with no backdoor neighbours are somewhat accurately represented
 Large parcel areas are poorly segmented with weird lot generation in the
middle
 Areas with backdoor neighbours (between two roads) are inconsistent
 Overall a major improvement from Experiment 1 and 2
 Better cleanup may have changed result of data

23. Experiment 4: Categorized Datasets
 Categorized based on road surrounding subdivision and dense neighbourhoods
 11 Training Datasets

24. Experiment 4: Outputs (native)

25. Experiment 4: Outputs

26. Experiment 3 (left) vs. 4 (right)

27. Experiment 4: Analysis
 Lines are shaky
 Better than Experiment 3 at parcel division based on its own specialization
 Parcels without backdoor neighbours are represented fairly well, not as good
as Experiment 3
 Consistent lot sizes
 Bubbly housing lots similar to Experiment 2

28. Four Colour Theorem
 Any arrangement of polygons on a 2D surface only requires four colours
without any of the colours intersecting

29. Experiment 5: Four (Five) Colour
Theorem
 Selected nine random parcels
 Assigned red to roads, four common colours for lots in training data

30. Experiment 5: Outputs

31. Experiment 5: Analysis
 Colour separation is fairly discrete (easy image
segmentation)
 Output similar to experiment 2 and 4; bubbly output
 Lots without backdoor neighbours are somewhat well
divided
 No coherent structure

32. Limitations
 About 30 training datasets at the most
 Many trained models use tens of thousands
 Even less for categorized datasets (Experiment 4)
 Algorithm is extremely resource intensive (512x512 is soft-cap limit)
 16 minutes to train 29 images with a GTX 1060 6GB
 Various technical variables and terminology makes it difficult for non-
scientists or statisticians to understand in-depth
 Epoch (Iterations)
 Layers/Generator and Discriminator Filters
 Gradients
 Steps

33. Overall Results Analysis
 None of the experiments produce data close to target
data
 Experiment 3 hints more training data produces better
shaped lots
 Experiment 4 hints that categorization maintains
consistency (against target data)
 Application of four colour theorem (Experiment 5) is
indeterminant

34. Next Steps
 Multiple types of subdivision classification
(at least one category with 50+ datasets)
 Train at least 100 subdivisions (aggregate)
 Play with pix2pix internal variables (max
epochs, layers, steps)

35. Thanks for listening!
Questions or suggestions?