GSD-Conditioned Generalization of UAV-Trained Oriented Vehicle Detectors Toward Satellite-Relevant Spatial Sampling

Abstract

This study examines how train–test ground sampling distance (GSD) mismatch affects UAV-based oriented vehicle detection under controlled source-scene and network-input conditions. Five datasets at 10, 15, 20, 25, and 30 cm/pixel were generated by direct downsampling from the same 1509 UAV source frames while preserving vehicle annotations and data-partition assignments. The GSD conversion procedure was empirically evaluated using 25 image measurements, yielding an overall mean absolute error of 0.180 cm/pixel, root mean square error of 0.225 cm/pixel, and mean absolute relative error of 0.891%. YOLO11m-OBB models were trained under two augmentation settings and three random seeds. The best validation checkpoint from each of 30 training runs was evaluated across all five test GSDs, producing 150 primary cross-GSD evaluations. Mean matched and mismatched AP50−95 were 0.7749 and 0.4915 under controlled augmentation and 0.8166 and 0.7346 under default augmentation, respectively. Across the five equally weighted test GSDs, 15 cm/pixel training achieved the highest mean AP50−95 reaching 0.6249 ± 0.0026 and 0.8049 ± 0.0018 under controlled and default augmentation. For 15-to-30 cm/pixel transfer, recall at a confidence of 0.25 increased from 0.5967 to 0.8407 with default augmentation. Nevertheless, recall remained only 0.0495 for vehicles with an OBB short side below 2 pixels under matched 30 cm/pixel evaluation. These results show that cross-GSD performance depends on augmentation, relative sampling, and object pixel footprint. The findings are limited to the UAV-resampling conditions evaluated here and do not establish a universal optimal GSD or direct transfer to actual satellite imagery.