Harmonizing Synthetic and Real Remote Sensing Imagery for Object Detection via Multiscale Adversarial Training (MSAT) Framework

Recent advances in generative modeling have enabled synthetic data generation for remote sensing; however, object detectors trained on synthetic imagery suffer from synthetic–real domain gaps that limit generalization. To address this challenge, we propose Multiscale Adversarial Training (MSAT), a unified framework that combines a SinGAN-style multiscale generator for harmonizing synthetic objects with real satellite imagery and a Detector framework that performs multiscale object detection while providing realism-aware feedback to progressively improve synthetic image quality. Experiments on NWPU VHR-10, DOTA-v2.0, AID, and PatternNet demonstrate that MSAT outperforms YOLOv10m, RTMDet, DINO, Swin-T + Faster R-CNN, and DETR under identical evaluation settings. MSAT achieves relative mAP@0.5 improvements of up to +12.9% on real datasets and +17.9% under hybrid training. Across settings, MSAT exhibits large relative mAP@0.5 improvements as compared to benchmark baseline models when transitioning from real to hybrid data, +6.50% to +12.37%, demonstrating strong robustness under domain shift. Ablation studies confirm contribution of each component, and synthetic image quality evaluation demonstrates that MSAT reduces FID and LPIPS scores across all four hybrid datasets compared with the vanilla SinGAN baseline, indicating improved alignment with real image distributions. These results demonstrate that realism-aware harmonization and detection-guided adversarial learning effectively reduce synthetic–real domain discrepancy and improve remote sensing object detection. Code is publicly available at https://github.com/Faryalaurooj/ Multi-Scale-Adversarial-Training-MSAT-framework.