Deep Learning Architectures for Automated Infrastructure Damage Detection from Post-Disaster Satellite Imagery in Nigerian Conflict and Flood Zones

Rapid, automated detection of infrastructure damage from post-disaster satellite imagery is a critical humanitarian intelligence capability, yet deep learning-based damage detection methods developed primarily on dataset from Syria, Turkey, and the United States have demonstrated limited transferability to the distinct building morphologies, material types, and image contexts of West African disaster environments. This dissertation develops and evaluates deep learning architectures specifically optimised for automated infrastructure damage detection from post-disaster satellite imagery in Nigerian conflict and flood zones, making an original methodological contribution to humanitarian remote sensing. A novel annotated satellite image dataset of 48,000 labelled building instances across six disaster events in North East and North Central Nigeria is constructed from commercial satellite imagery acquired pre- and post-disaster, constituting the first Nigerian-specific building damage annotation corpus. Three deep learning architectures are developed and compared: a transfer-learned ResNet-50 baseline, a modified dual-encoder U-Net with cross-temporal attention, and a novel Multi-Scale Damage Signature Network designed to resolve mud-brick and concrete building damage patterns simultaneously. Domain adaptation experiments assess the contribution of Nigerian-specific training data to cross-domain performance. Results demonstrate that the proposed Multi-Scale Damage Signature Network outperforms the transfer-learned baseline by 18.4 F1 points on the Nigerian test set. The dissertation contributes a new architectural paradigm for damage detection in low-income country built environments and publishes the annotated dataset under open access. Keywords: deep learning, damage detection, satellite imagery, disaster, Nigeria.