Monitoring access to piped water and sanitation infrastructure in Africa at disaggregated scales using satellite imagery and self-supervised learning

Tracking access to safely managed drinking water and sanitation (SDG 6) is hindered in many regions by sparse, costly, or delayed survey and census data. We propose a remote-sensing framework that pairs household survey labels with Sentinel-2 satellite imagery and self-supervised Vision Transformer embeddings to learn representations of the built environment without dense task-specific labels. We evaluate a $k$-NN classifier on frozen embeddings to predict (i) piped water access and (ii) sanitation/sewage access at survey locations, and use high-resolution population maps to produce population-weighted national estimates. On held-out validation splits, our best configurations reach 84.19\% accuracy for piped water and 87.17\% accuracy for sanitation.