IntroductionRapid detection of war-induced damage is vital for humanitarian response and post-war reconstruction, yet large-scale, well-annotated benchmarks for war-damage change detection are still lacking.MethodsWe present a large-scale, high-resolution bi-temporal optical remote sensing benchmark for war-damage change detection, named WD-CD. WD-CD contains 9,687 bi-temporal image pairs of 512 × 512 pixels collected from five conflict-affected theaters across Europe, Africa, and Asia, covering a total area of 1,055.83 km2. It provides annotations for 16 target categories and up to six change states. We evaluated representative deep learning models, performed cross-dataset comparisons, and conducted intercontinental transfer experiments.ResultsThe benchmark evaluations demonstrated consistent and competitive performance while revealing the challenges of damage-state discrimination. The intercontinental transfer experiments showed that building geometry generalized well across regions, whereas damage morphology and background landscapes were region-specific.DiscussionWD-CD addresses the lack of large-scale, fine-grained benchmarks for war-damage change detection and provides a foundation for evaluating model performance and cross-regional generalization. To support open science and reproducibility, the dataset will be made available upon request for non-commercial academic research.