Change Detection in Road Networks During Conflicts: A Deep Learning Framework with Aerial Photography

Abstract

Monitoring transformations in road infrastructure is critical for disaster response and recovery in conflict regions. However, manual analysis of road damage is time-consuming and limited in scale, while the speed of changes requires near real-time monitoring. This paper presents a deep learning framework using Hybrid Spectral Networks (HybridSN) for change detection in road networks during conflicts, applied to the city of Old Mosul, Iraq. The HybridSN model achieved high accuracy in road detection, with recall of 0.990 in 2014 and 0.982 in 2022. Change detection analysis revealed a substantial reduction of 13,249.83 m in total road length from 20,186.93 m in 2014 to 6,937.10 m in 2022, indicating widespread damage. The quantitative results demonstrate the capabilities of the proposed approach in assessing road network changes through conflict.