The Uncertainty of Slum Mapping

Kurzfassung

Slums are complex, dynamic urban environments often characterized by poor housing and inadequate infrastructure. Their detection at a global scale remains challenging due to data scarcity and the varied morphological structures of slums. This study introduces an uncertainty-aware approach to slum mapping using machine learning techniques, addressing the limitations of previous binary classification methods by estimating the probability of slum presence. To ensure robust slum probability estimations, multiple uncertainty-aware strategies were implemented: Test-Time Augmentation (TTA): Predictions were stabilized by applying randomized augmentations to each image tile, helping to mitigate model uncertainty. Test-Time Dropout (TTD): A dropout rate of 0.3 was used during inference, capturing both model-based and data-related uncertainties. Overlapping Image Tiles: Each location was analyzed from five overlapping 224×224 pixel image tiles, with predictions averaged to reduce edge effects and noise. Ensemble of CNNs: Four deep learning architectures (ResNet-18, ReXNet-150, EfficientNet-B4, and MobileNetV3) were combined into an ensemble. Their outputs were averaged to generate a final slum probability score, capturing variations in model responses. Using transfer learning, models were initially pre-trained on four reference cities with well-documented slums (Caracas, Mumbai, Nairobi, Rio de Janeiro). The learned features were then adapted to each of the 55 target cities, ensuring robust predictions even in data-scarce environments. The study successfully produced probabilistic slum maps for 55 cities, offering insights into the spatial uncertainty of slum areas. The overall accuracy of the method varied between classes: urban and vegetation areas achieved high F1 scores (~92-95%), while slum identification exhibited high recall (87.19%) but lower precision (47.81%), indicating some misclassification. The study revealed significant inter- and intra-city variability in slum structures, underscoring the need for probabilistic rather than binary classifications. By integrating uncertainty estimation, this research provides the first large-scale dataset with quantified slum probabilities, improving understanding of informal settlements for urban planning and policy-making.