A fresh benchmark and a semi-supervised segmentation pipeline aim to improve mapping of informal settlements.
The authors built a ground‑truth dataset covering roughly 900 km² of Lahore, plus derived data for Karachi and Mumbai, and added four earlier African and Latin‑American cities. They also introduced a Class‑Aware Adaptive Thresholding step that tweaks confidence cut‑offs for minority classes, and a DINOv2‑based filter that culls out‑of‑distribution tiles before training. Tested on seven cities across three continents, the system achieved up to +5.9 percentage‑points mean IoU over leading semi‑supervised baselines, with no extra cost at inference.
The work matters because existing remote‑sensing maps of slums suffer from noisy labels and spectral overlap with formal housing, limiting aid and planning efforts. By tackling class imbalance and distribution shift, the approach offers a more reliable way to monitor rapid urban growth in data‑starved regions.
Still, the gains are modest and hinge on high‑quality administrative boundaries—a resource not universally available. The method’s architecture‑agnostic claim will be tested as more cities adopt it.