Multi-Class Classification of Urban Regeneration Using a Siamese Network: An Analysis with Real-World Data from Portland, Oregon Yang Yang Ph.D. Candidate Department of City and Regional Planning UNC Chapel Hill
Department of City and Regional Planning, UNC-Chapel Hill
Global Urban Regeneration Trend
Urban regeneration is a broad and flexible concept, commonly used to describe the physical and social transformation of a significant portion of a city or a cluster of properties \(Lehmann\, 2019\).
Urban regeneration can take various forms, including:
smart infill housing,
development of new infrastructure
adaptive reuse of historic buildings for workplaces, education, tourism, and cultural initiatives
Etc
Department of City and Regional Planning, UNC-Chapel Hill
Detect and Classify Urban Regeneration with AI
Department of City and Regional Planning, UNC-Chapel Hill
The Application of Siamese Network in Urban Regeneration Study
Originally designed for comparing fingerprint, signatures, and faces.
Shows great performance in previous study
Haiping Yang, 2023
Department of City and Regional Planning, UNC-Chapel Hill
Limitation of Current Work
- Current Challenge:
- Reliance on High-Quality Benchmark Datasets. Does it work well with real-world data - Unstable clarity, little urban change, lots of tree cover
- Predominance of Binary Classification
- Urban regeneration should be defined by parcel but not by building
Department of City and Regional Planning, UNC-Chapel Hill
A Real-World Application Experiment
- Manually labelled 2000 sample residential parcels from Portland, Oregon with four types of urban changes between 2010 and 2020:
- No Change
- New Development
- Redevelopment
- Demolition
- Parcel images are taken from the city orthoimagery program
- Choose Portland because:
- Different types of regeneration
- Gradual Urban Regeneration
- Different levels of Tree Coverage
Department of City and Regional Planning, UNC-Chapel Hill
The Application of Siamese Network in Urban Regeneration Study
- Two Identical CNN Encoders to extract features from Aerial Imagery from previous year and later year
- A Full Connected Decoder to Compare Features and Make Classification Decision
- Optimize by a Cross Entropy Loss Function
Department of City and Regional Planning, UNC-Chapel Hill
Test if Attention Mechanism Helps to Improve performance
Many previous studies suggested to use attention mechanism \(Lee et al\.\, 2021; J\. Li et al\.\, 2022\)
We modified the U-Net model to include a similarity attention mechanism for high-level and low-level features \(U\-Net was the best among the three backbone\)
Department of City and Regional Planning, UNC-Chapel Hill
Model Performance
| Backbone | fold 1 | fold 2 | fold 3 | fold 4 | fold 5 |
|---|---|---|---|---|---|
| ResNet | 85.5 | 85.50 | 81.75 | 84.25 | 84.00 |
| Yolo11x | 80.00 | 80.75 | 80.25 | 82.25 | 82.75 |
| U-Net | 84.25 | 85.75 | 85.50 | 85.75 | 85.50 |
| U-Net + Attention | 86.25 | 84.75 | 86.00 | 84.25 | 86.75 |
U-Net Outperform the other two backbones, but adding attention mechanism did not significantly improve accuracy
Department of City and Regional Planning, UNC-Chapel Hill
Confusion Matrix
Siamese network can achieve really good accuracy for most classes except
Redevelopment vs No Change
Department of City and Regional Planning, UNC-Chapel Hill
Things that are hard for Computer to “See”
Department of City and Regional Planning, UNC-Chapel Hill
Advantages of Siamese Network Comparing to Traditional Methods
Siamese Network has a great potential for Real-World Planning Practice, especially for urban planners operating under budget constraints.
Only require satellite imagery, which is widely accessible
More accurate than using building permit records or medium resolution remote sensing data
Light weight and fast
Low cost \(A Big Problem for the Application of LLM\)
Department of City and Regional Planning, UNC-Chapel Hill
Future Work
Transfer learning potential
Local light weight multimodal LLMs \(such as llama 3\.2\-vision\)
Department of City and Regional Planning, UNC-Chapel Hill
Thank you!
Department of City and Regional Planning, UNC-Chapel Hill
Back-Up Slides
Label Standard
| Has a building in 2010 | Has a new construction between 2010 and 2020 | Has a demolition between 2010 and 2020 | Labeled Urban Change Class |
|---|---|---|---|
| Yes or No | No | No | No Change |
| No | Yes | Yes or No | New Development |
| Yes | Yes | Yes or No | Redevelopment |
| Yes | No | Yes | Demolition |
Department of City and Regional Planning, UNC-Chapel Hill
Cross Entropy Loss Function
Department of City and Regional Planning, UNC-Chapel Hill
Average Training Loss Curve
Validation Loss Curve – Yolov11x shows fluctuations
Department of City and Regional Planning, UNC-Chapel Hill
