Validation-Aware Building-Exposure Screening along the Lagos–Ibadan Standard Gauge Railway Using Open Building Footprints, Sentinel-2 Imagery, and Spatially Blocked Machine Learning

Abstract

Railway-corridor monitoring requires spatially explicit building-location evidence that is also transparent about data uncertainty. This study develops a validation-aware screening workflow for the 159.989 km Lagos–Ibadan Standard Gauge Railway in Nigeria. A continuous analytical centreline was used to construct three nested screening tiers at 22.86, 200 and 500 m on each side. The 22.86 m tier is an analytical screening distance derived from the 75-foot statutory reference in the Nigerian Railway Corporation Act, not a parcel-specific cadastral boundary. Google Open Buildings version 3 footprints were independently checked against 2,346 manually interpreted buildings across 45 stratified high-resolution imagery tiles. At confidence 0.65, object-level precision, recall and F1 score were 0.491, 0.836 and 0.618. The mapped inventories were 1,202, 50,715 and 139,395 buildings, while validation-informed aggregate sensitivity estimates were 773, 36,761 and 101,261. Area-weighted WorldPop screening exposure was approximately 18,435, 249,397 and 677,202 people, with allocation sensitivity reported separately. Land-cover composition differed significantly among mutually exclusive corridor zones. Under five leakage-controlled spatial folds, a custom adapter-tuned vision-transformer classifier achieved the strongest out-of-fold performance, with F1 score 0.794, classification Jaccard index 0.658 and area under the receiver operating characteristic curve 0.866. The workflow supports inspection prioritisation and uncertainty-aware corridor management, but not cadastral adjudication, enforcement, compensation or annual building-change claims without field verification.