Abstract
Accurate mapping of low-income areas (LIAs) is vital for urban planning and providing essential services like water and sanitation. However, traditional mapping is hindered by data scarcity and spatial heterogeneity across different cities. We present a Point-to-Region (PTR) co-learning framework that leverages high-resolution satellite imagery and internet connectivity data. By contextualizing local "points" within broader "regional" signatures, the model improves generalization to unseen cities. Our approach utilizes attention mechanisms and a dynamic "Region-Gate" to distinguish between universal and city-specific features. Experimental results across three Kenyan cities demonstrate significant performance gains over standard baselines, particularly in cross-city generalization.
Introduction
Providing safe water and sanitation (WASH) is a critical challenge in fast-growing African cities. Effective intervention requires fine-scale maps of low-income areas (LIAs), but these are difficult to produce due to two major hurdles: data scarcity and spatial heterogeneity. Ground-truth labels typically require expensive household surveys, and the visual "signature" of poverty varies significantly between regions—for instance, green space may signify wealth in arid climates but neglect in tropical ones.
To address these challenges, this study utilizes multi-source public data, including Sentinel-2 multi-spectral imagery (10m resolution) and auto-tracked internet connectivity data. Ground truth was established through extensive field surveys in Kisumu, Malindi, and Nakuru, Kenya.
Methods
The proposed Point-to-Region (PTR) Co-learning framework shifts the focus from classifying isolated 10m satellite patches ("points") to evaluating them alongside a representative sample of the entire city ("region").
The architecture includes several key innovations:
- Attention-based Regional Aggregation: An attention mechanism filters out non-informative features, such as clouds or water bodies, ensuring the regional signature remains clean.
- Region-Gate: This component dynamically decides whether to classify a point based on universal patterns (e.g., swimming pools as wealth indicators) or city-specific variant patterns.
- Asynchronous Update Strategy: To manage the computational load of city-wide training, point-level and region-level features are updated at different frequencies.
- Time-series Augmentation: The model treats different seasonal captures as separate "regions," increasing robustness to environmental changes like wet and dry seasons.
Results
The framework was evaluated using a "leave-one-city-out" approach. The PTR models (PTR-g1 and PTR-g2) outperformed standard baselines in 12 out of 15 experimental scenarios. Notably, in Kisumu, the PTR model achieved an F1-score of 0.506 using a U-Net backbone, compared to 0.360 for the base model. The framework proved to be architecture-agnostic, improving performance across various backbones like PSPNet and DeepLabV3+. Sensitivity analysis further confirmed that the attention mechanism effectively maintains accuracy even in the presence of significant "noise" like open water.
Conclusion
While the model showed strong generalization, performance was lower in Malindi due to its unique internet signatures and high cloud interference. Future work will explore additional auxiliary features and incorporate spatial fairness to prevent algorithmic bias. This co-learning approach offers a scalable path for high-resolution poverty mapping in data-constrained environments, with potential applications extending to environmental monitoring and broader urban planning.
Summary updated