Unsupervised deep clustering of high-resolution satellite imagery reveals phenotypes of urban development in Sub-Saharan Africa

Abby Metzler, Ricky Nathvani, Viktoriia Sharmanska, Wenjia Bai, Simon Moulds, Nkechi S. Owoo, Iris Ekua Mensimah Fynn, Emily Muller, Esaie Dufitimana, Ghafi Kondi Akara, George Owusu, Samuel Agyei‐Mensah, Majid EzzatiThe Science of The Total Environment 2025 · DOI

TL;DR

This paper introduces a novel hierarchical deep learning approach to analyze high-resolution satellite imagery for characterizing urban development in Sub-Saharan Africa. By using unsupervised clustering, the method identifies distinct "phenotypes" of urban environments, offering a scalable and data-driven way to monitor rapidly changing cities where traditional data is scarce.

What Problem Does This Paper Address?

Urbanization in Sub-Saharan Africa is rapid and complex, leading to diverse urban landscapes. Effectively monitoring and planning for this growth is challenging due to a lack of timely, granular data on urban built and natural environments. Traditional methods often rely on supervised learning, requiring extensive labeled data that is difficult and costly to obtain in these regions. This paper addresses the need for a method that can consistently and coherently analyze multidimensional urban features across multiple cities using readily available satellite imagery, without requiring pre-labeled data.

Methods

The core of the methodology is a hierarchical two-stage unsupervised deep learning approach using the DeepCluster framework.

  1. Joint Feature Extraction: A Convolutional Neural Network (CNN), specifically VGG-16 pre-trained on ImageNet, is used to learn image representations (features) from satellite image tiles across all four cities simultaneously. This step leverages commonalities in urban structures.
  2. City-Specific Clustering: The learned representations are then clustered independently for each city using a k-means algorithm. This two-stage process balances comparability across cities with the ability to capture city-specific nuances. The authors compared this hierarchical approach (Approach C) with a joint one-stage model (Approach A) and separate one-stage models for each city (Approach B), finding Approach C to be the most efficient and interpretable.

Datasets

  • Satellite Imagery: High-resolution (0.3 meters per pixel) RGB satellite images from Maxar Technologies for four major Sub-Saharan African cities: Accra (Ghana), Dakar (Senegal), Dar es Salaam (Tanzania), and Kigali (Rwanda). Images were pre-processed into 256x256 pixel tiles.
  • External Data for Interpretation: Datasets on urban characteristics including building area, building count, building orientation, NDVI (Normalized Difference Vegetation Index), population density, distance to roads, and average building size. These were used after clustering for interpretation and validation, not for training the unsupervised model.

Key Insights and Ideas

  • Unsupervised Learning for Urban Phenotyping: The study demonstrates that unsupervised deep learning can effectively identify meaningful "phenotypes" of urban development directly from satellite imagery, bypassing the need for extensive manual labeling.
  • Hierarchical Approach for Cross-City Comparability: The proposed hierarchical two-stage clustering method successfully balances the need for consistent feature extraction across cities with the ability to capture unique urban characteristics within each city.
  • Visual Signatures of Urban Development: The research highlights how distinct visual patterns in satellite imagery correspond to specific urban development characteristics, such as building density, road networks, and vegetation types, even in data-scarce regions.
  • Colonial Legacies in Urban Form: Similarities in urban patterns among coastal cities (Accra, Dakar, Dar es Salaam) are linked to shared colonial urban planning influences, manifesting in grid-like road structures and building orientations.

Key Results

  • The hierarchical approach (Approach C) achieved the lowest Median Absolute Deviation (MAD) scores across cities, indicating more homogeneous clusters and better overall performance compared to joint or separate one-stage models.
  • The clustering identified several super-groups of urban environments: Vegetation, Agriculture, Empty land, Water, Densely populated areas, and Other mixed environments.
  • Accra, Dakar, and Dar es Salaam showed analogous patterns, with dense settlements (20-54% of urban area), peri-urban intermix (21-44%), natural vegetation (9-13%), and agricultural land (8-15%).
  • Kigali exhibited unique characteristics, including a sparser urban core (23%) and significant wildland-urban intermix (19% of vegetation).
  • Water was a distinct cluster in Accra (2% of area), while Empty land was identified in Accra and Dakar (8-10%).
  • Predictive modeling (XGBoost with SHAP values) confirmed that building information (area, count, orientation) and NDVI were key predictors for cluster membership across cities, with variations in importance for specific clusters.

Limitations and Future Work

  • Data Availability and Cost: High-resolution satellite imagery can be expensive, and obtaining it for many cities or for longitudinal studies remains a challenge.
  • Temporal Resolution: The study used composite images from a specific time period, making it sensitive to seasonal variations, particularly for vegetation. Future work could incorporate multi-temporal imagery to capture intra-annual changes and seasonal transformations.
  • External Data Mismatch: Some external datasets used for interpretation had coarser resolutions than the satellite imagery, potentially obscuring fine-scale relationships.
  • Limited External Variables: The number of external variables used for interpretation was limited. Incorporating more detailed data on building height, agricultural land, and other features could improve interpretation.
  • Future Directions: The authors suggest extending the approach to more cities, conducting longitudinal assessments to monitor urban change, and exploring different hierarchical clustering configurations. They also highlight the potential for near real-time analysis with improved computational resources.

Summary updated

Abstract

Sub-Saharan Africa and other developing regions have urbanized extensively, leading to complex urban features with varying presence and types of roads, buildings and vegetation. We use a novel hierarchical deep learning framework and high-resolution satellite images to characterize multidimensional urban environments in multiple cities. Application of the model to images from Accra, Dakar, and Dar es Salaam identified areas with analogous patterns of building density, roads and vegetation. These included dense settlements within the metropolitan boundary (20-54% of urban area), peri-urban intermix of natural and built environment (21-44%), natural vegetation (9-13%) and agricultural land (8-15%). Kigali, with its mountainous geography and post-colonial expansion, exhibited unique urban characteristics including a sparser urban core (23%) and significant wildland-urban intermix (19% of vegetation). Other notable clusters were water (2% of area of Accra) and empty land (8-10% of Accra and Dakar). Our results demonstrate that unlabeled satellite images with unsupervised deep learning can be used for consistent and coherent near-real-time urban monitoring, particularly in regions where traditional data are scarce.
BibTeX
@inproceedings{metzler2025unsupervised,
  author = {Abby Metzler and Ricky Nathvani and Viktoriia Sharmanska and Wenjia Bai and Simon Moulds and Nkechi S. Owoo and Iris Ekua Mensimah Fynn and Emily Muller and Esaie Dufitimana and Ghafi Kondi Akara and George Owusu and Samuel Agyei‐Mensah and Majid Ezzati},
  title = {{Unsupervised deep clustering of high-resolution satellite imagery reveals phenotypes of urban development in Sub-Saharan Africa}},
  booktitle = {The Science of The Total Environment},
  year = {2025},
  doi = {10.1016/j.scitotenv.2025.179739},
}
Figure