Rapid urban expansion in low- and middle-income countries has driven the growth of informal settlements in major cities such as Lahore, Karachi, and Mumbai. However, large-scale mapping of these areas is hampered by annotation scarcity and data quality issues, including high spectral ambiguity between formal and informal structures and significant annotation noise, according to a research paper published on arXiv. To address this, a team of researchers led by Muhammad Taha Mukhtar and colleagues has introduced SLUM-i, a semi-supervised segmentation framework designed to improve informal settlement mapping.
The Benchmark Dataset
The researchers constructed a benchmark dataset for Lahore from scratch, and derived companion datasets for Karachi and Mumbai from verified administrative boundaries, totaling approximately 900 km² of urban area. This collection is supplemented by four cities from prior literature across Sub-Saharan Africa and Latin America, with comprehensive data quality assessments provided for each city, the study reports.
| City | Country | Source |
|---|---|---|
| Lahore | Pakistan | Constructed from scratch |
| Karachi | Pakistan | Derived from verified administrative boundaries |
| Mumbai | India | Derived from verified administrative boundaries |
| Four additional cities | Sub-Saharan Africa and Latin America | From prior literature |
The SLUM-i Framework
SLUM-i is a semi-supervised segmentation framework that mitigates the class imbalance and distribution mismatch inherent in standard semi-supervised learning pipelines, the authors explain. It integrates two key components:
- Class-Aware Adaptive Thresholding: A mechanism that dynamically adjusts confidence thresholds to prevent minority class suppression.
- DINOv2-based unlabeled pool filter: Removes out-of-distribution tiles prior to training to reduce covariate shift.
Both components are architecture-agnostic and add no inference overhead, according to the paper.
Performance Gains
Extensive experiments across seven cities spanning three continents, repeated over five random seeds, demonstrated gains of up to +5.9 pp mIoU over state-of-the-art semi-supervised baselines. The study notes that both components contribute to this improvement without increasing computational burden during inference.
Gains of up to +5.9 pp mIoU over state-of-the-art semi-supervised baselines, with both components being architecture-agnostic and adding no inference overhead.
Implications for Urban Mapping
The research directly addresses critical challenges in mapping informal settlements, which are often poorly documented yet vital for urban planning and resource allocation. By providing a benchmark dataset and an effective semi-supervised framework, SLUM-i enables more accurate and scalable mapping in regions where labeled data is scarce. The methodology's architecture-agnostic design allows it to be integrated into existing computer vision pipelines, making it practical for deployment by urban planners and development organizations.
For enterprise technology leaders, the approach demonstrates how advanced AI techniques like semi-supervised learning and vision transformers (DINOv2) can overcome data quality and annotation bottlenecks. Such capabilities are relevant for applications beyond informal settlements, including infrastructure monitoring and logistics planning in complex urban environments.