High-resolution rural poverty mapping in Pakistan with ensemble deep learning

Summary

Peer-reviewed study presenting an ensemble deep learning approach using three CNN models (ResNet-50, ResNet-50V2, ResNet-101) trained on Sentinel-2, VIIRS, and accessibility data to predict chronic poverty at 1 km² resolution across rural Sindh. Validated through hold-out testing (71% recall, 67% precision), 6-fold spatial cross-validation, and ground-truthing with ~7,000 households in Ghotki district (2022). Trained on 1.67 million anonymised household SPS records from SUCCESS and PPRP programmes. Ethics approval: LUMS IRB. Data and scripts publicly available on Figshare.