SLUM-i: Semi-supervised Learning for Urban Mapping of Informal Settlements and Data Quality Benchmarking
The episode reviews SLUM-i, a semi-supervised learning method for urban mapping of informal settlements using satellite imagery. The team developed two fixes—a DINOv2 filter and Class-Aware Adaptive Thresholding—to overcome class imbalance and covariate shift in pseudo-labeling. Results showed substantial performance gains across seven cities, proving the method is most valuable where data quality is low.