**6. Conclusions**

In this paper, we proposed the use of a state of the art ADGAN, Skip-GANomaly, for unsupervised post-disaster building damage detection, using only imagery of undamaged buildings from the pre-epoch phase. The main advantage of this method is that it can be developed in the pre-event stage and deployed in the post-event stage, thus aiding disaster response and recovery. Special attention was given to the transferability of this method to other geographic regions or other typologies of damage. Additionally, several Earth observation platforms were considered, since they o ffer di fferent advantages for data variety and data availability. Specifically, we investigated (1) the applicability of ADGANs to detect post-disaster building damage from di fferent remote sensing platforms, (2) the sensitivity of this method against di fferent types of pre-processing or data selections, and (3) the generalizability of this method over di fferent typologies of damage or locations.

In line with earlier findings, we found that the satellite-based models were sensitive against the removal of objects that contained a high visual variety: vegetation and shadows. Removing these objects resulted in an increase in performance compared to the baseline. Additionally, in order to reduce the visual variety in the original images, experiments were conducted with varying image patch sizes. No clear di fference in performance of di fferent patch sizes was observed. UAV-based models yielded high performance when detecting earthquake-induced damage. Contrary to satellite-based models, UAV-based models trained on smaller patches obtained higher scores.

UAV imagery contained small GSDs and showed damage in high detail. Therefore, models based on UAV-imagery transferred well to other locations, which is in line with earlier findings. Models based on satellite-imagery did not transfer well to other locations. The results made it evident that image characteristics (patch size and GSD), and the characteristics of the disaster induced damage (large-scale and small-scale), play a role in the ability of satellite-based models to transfer to other locations.

Compared to supervised approaches, the obtained results are good achievements, especially considering the truly unsupervised and single-epoch nature of the proposed method. Moreover, the limited time needed for training in the pre-event stage and for inference in the post-event stage (see Section 3.4) make this method automatic and fast, which is essential for its practical application in post-disaster scenarios.

**Author Contributions:** Conceptualization, S.T. and F.N.; Analysis, S.T.; Methodology, S.T. and F.N.; Supervision, F.N., N.K. and G.V.; Original draft, S.T. and F.N.; Review and editing, S.T., F.N., N.K. and G.V. All authors have read and agreed to the published version of the manuscript.

**Funding:** Financial support has been provided by the Innovation and Networks Executive Agency (INEA) under the powers delegated by the European Commission through the Horizon 2020 program "PANOPTIS–Development of a decision support system for increasing the resilience of transportation infrastructure based on combined use of terrestrial and airborne sensors and advanced modelling tools", Grant Agreement number 769129.

**Conflicts of Interest:** The authors declare no conflict of interest.
