Mapping Temporary Slums from Satellite Imagery using a Semi-Supervised Approach

One billion people worldwide are estimated to be living in slums, and documenting analyzing these regions is a challenging task. As compared regular slums; the small, scattered temporary nature of slums makes data collection labeling tedious time-consuming. To tackle this problem detection, we present semi-supervised deep learning segmentation-based approach; with strategy detect initial seed images zero-labeled settings. A small set samples (32 our case) automatically discovered by temporal changes, which manually labeled train segmentation representation module. The module gathers high dimensional image representations, transforms representations into embedding vectors. After that, scoring uses vectors sample from large pool unlabeled generates pseudo-labels for sampled images. These their added training update modules iteratively. analyze effectiveness technique, construct geographically marked dataset slums. This constitutes more than 200 potential slum locations (2.28 square kilometers) found sieving sixty-eight thousand 12 metropolitan cities Pakistan covering 8000 kilometers. Furthermore, proposed method outperforms several competitive semantic baselines on similar setting. code will made publicly available.