A High Resolution Urban and Rural Settlement Map of Africa Using Deep Learning and Satellite Imagery

Technical Summary: A High-Resolution Urban and Rural Settlement Map of Africa

1. Problem Statement

Accurate mapping of urban and rural settlements is essential for sustainable development, spatial planning, and monitoring socio-economic disparities. However, existing global datasets (such as GHSL-SMOD, GHS Degree of Urbanisation, and GRUMP) suffer from several critical limitations, particularly when applied to the African continent:

• Coarse Spatial Resolution: Most existing products operate at a ~1 km resolution, which fails to capture small, informal, or dispersed settlements common in Africa.
• Methodological Inconsistency: Many datasets rely on rule-based classifications (e.g., population density thresholds) or administrative boundaries, which do not reflect actual land-use patterns or rapid urbanization.
• Temporal Lag: Existing maps are often static or provided at long, discrete intervals, making it difficult to track rapid peri-urban transitions.
• Sensitivity to Census Errors: Reliance on census data can propagate inaccuracies in regions with poorly maintained official statistics.

2. Methodology

The authors propose a deep learning-based framework to generate a High-Resolution Urban-Rural (HUR) map at 10 m resolution covering Africa from 2016 to 2022.

• Deep Learning Architecture: The study employs a DeepLabV3 semantic segmentation model with a ResNet-50 backbone. This architecture uses Atrous Spatial Pyramid Pooling (ASPP) to capture multi-scale contextual information, allowing the model to recognize irregular settlement morphologies (e.g., distinguishing a small village from surrounding cropland).
• Multi-Source Data Integration:
  • Input Data: Landsat-8 multi-spectral imagery (30 m) and VIIRS nighttime lights (463 m) are used to provide spectral and luminosity features.
  • Target Data (Label Augmentation): To create high-quality training labels, the authors combined ESRI Land Use Land Cover (LULC) (10 m) with GHS-SMOD (1 km). They used the SMOD data to "subdivide" the ESRI "Built-up" class into specific "Urban" and "Rural" subclasses, creating a refined three-class target: Urban, Rural, and Non-Human Settlement (NonHS).
• Training Strategy:
  • Handling Imbalance: To prevent the model from being biased toward dominant land cover classes (like forests or crops), they implemented inverse probability weighting in the cross-entropy loss function.
  • Spatial Validation: To avoid data leakage due to spatial autocorrelation, the authors used a country-wise partitioning approach for five-fold cross-validation.
• Inference & Smoothing: To prevent "edge effects" (artifacts at the borders of image tiles), a sliding-window approach with a cropping-based overlap strategy was used, retaining only the central, high-context portion of each predicted tile.

3. Key Contributions

• High-Resolution Dataset: The creation of the first continent-wide, 10 m resolution urban-rural map for Africa (2016–2022).
• Context-Aware Classification: Moving beyond rule-based population thresholds to a data-driven approach that considers the spatial relationship between settlements and their environment.
• Robust Evaluation Framework: The use of a multiple-imputation framework to validate the map against the Demographic and Health Surveys (DHS). This accounts for the intentional spatial displacement (privacy protection) used in DHS data, providing a statistically rigorous assessment of accuracy.

4. Results

The HUR dataset was evaluated against the independent DHS dataset and compared to the existing GHS-SMOD product:

• Superior Accuracy: At the continental scale, the HUR maps achieved an overall accuracy of ~65–66% and a Kappa coefficient of ~0.47–0.48, significantly outperforming SMOD (which had ~56% accuracy and ~0.37 Kappa).
• Improved Precision/Recall Balance: While SMOD tended to overestimate urban extents (high recall, low precision), the HUR model provided a much more balanced classification, better delineating the boundaries of urban and rural areas.
• Granularity: The 10 m resolution allowed for the detection of small settlements and complex peri-urban zones that were previously "smoothed out" or merged into larger classes in 1 km resolution products.

5. Significance and Applications

• Sustainable Development (SDGs): The map directly supports SDG 11 (Sustainable Cities and Communities) and indirectly supports SDG 1 (No Poverty) and SDG 3 (Good Health and Well-Being) by providing the spatial granularity needed to target infrastructure, healthcare, and sanitation services.
• Policy and Research: It provides a harmonized, spatially consistent "spatial prior" for researchers studying migration, poverty, and health, allowing them to classify survey clusters (like DHS) into urban or rural categories without relying on inconsistent national definitions.
• Climate Resilience: High-resolution data enables better monitoring of land-use changes and the environmental impacts of urban expansion on rural sustainability.