Long-range human detection in SAR: The C2A cluster-based synthetic approach

Abstract

Limited training datasets hamper the application of remote Unmanned Aerial Vehicles (UAVs) for hydrological disaster search. However, current datasets are not able to model the visual complexity and long-range views typically found in flooded areas. To overcome this drawback, we propose a synthetic data pipeline using the Combination to Application (C2A) cluster-based method. The method explicitly uses object localization, clustering instead of uniform positioning. Synthetic human instances are superimposed on authentic aerial shots to create natural groupings. Experiments on YOLOv11 show that the detection performance of small objects is significantly improved when trained on densely clustered data rather than randomly distributed. Experimental evaluations using state-of-the-art datasets demonstrate that our method achieves competitive performance. The proposed model outperforms dedicated maritime detectors and those developed for terrestrial terrains.