Global Urban Metabolism Clustering Framework
2020 - 2024
UMERA
SIERA
Project Team
PI: Prof. John E. Fernandez
Lead Researcher: Dr. Norhan Bayomi
Project Collaborators
Bangkok, Thailand; Ben Guerir, Morocco.
Supported By
The funders of the project: Bangkok Bank, Environmental Solutions Initiative, University Mohammed VI Polytechnic Research Program
A three-stage machine learning pipeline, combining multivariate outlier detection, variational autoencoder encoding, and agglomerative clustering with evolutionary hyperparameter optimization, that produces metabolism-based urban typologies for over 11,000 cities worldwide, grouping cities by functional resource-use similarity rather than geography or income.
Conventional urban classifications rely on geography, income brackets, or population size, categories that obscure the enormous variation in how cities actually consume water, energy, and food. Two cities on opposite sides of the globe may share strikingly similar metabolic profiles, while neighbouring cities within the same country can differ dramatically. The Global Urban Metabolism Clustering Framework (GUM-CF) was developed by the Urban Metabolism Group to address this gap by producing data-driven typologies grounded in measured and predicted per-capita resource flows.
The framework operates through four sequential stages. Stage 1 - Outlier Detection: A multivariate outlier detector (MMVBACON) separates extreme-consumption cities, typically large metropolises whose population and area follow Zipf’s Law, into a distinct outlier group, preventing them from distorting the clustering of the majority of cities. Stage 2 - Deep Clustering Encoding: The non-outlier data is passed through a Variational Autoencoder (VAE) with k-means integrated directly into the latent space (a deep clustering architecture). Rather than treating encoding and clustering as separate operations, the VAE jointly learns a low-dimensional representation and cluster-aware structure, producing embeddings that are both denoised and aligned with cluster boundaries. The VAE is trained to convergence using KL-Divergence and Mean Squared Error objectives with the Adam optimizer.
Stage 3 - Agglomerative Clustering: Agglomerative Clustering is applied to the deep-clustering embeddings to assign cities to final typological groups, capturing hierarchical relationships between urban profiles that flat partitioning methods miss. Stage 4 - Evolutionary Optimization: CMA-ES (Covariance Matrix Adaptation Evolution Strategy) jointly optimizes the hyperparameters of both the VAE architecture and the clustering step, maximizing internal validation metrics without manual tuning. The outlier group is then reunited with the main clusters to yield the final N+1 typology.
