Gain Mechanism Analysis of Optical-to-synthetic Aperture Radar (SAR) Cross-modal Data Augmentation for SAR Object Detection
Photogrammetric Engineering & Remote Sensing · 2026
DOI: 10.14358/pers.26-00107r2 ↗Kurz gefasst
Fachartikel in der Zeitschrift Photogrammetric Engineering & Remote Sensing, 2026. Verfasst von Changlin Xiao und 7 weiteren Autor:innen.
Abstract (Originalsprache)
Optical-to-SAR (O2S) generation is increasingly used to ease the scarcity of annotated synthetic aperture radar (SAR) samples for object detection. Most prior work judges synthetic data by visual metrics and offers little guidance on mixing ratios, detector dependence, or true gain bounds. Using coregistered optical–SAR pairs (QXSLAB_SAROPT), we evaluated a grayscale control and three generators—CycleGAN, Cross-CycleGAN, and E3Diff—with five detectors (YOLO11s, CenterNet, Faster R-CNN, RT-DETR-l, and DINOv3-YOLOv12) under progressive mixing on ship detection and seven-class aircraft recognition. From prior knowledge of generator objectives and detector inductive biases, we hypothesized that pairing is not universal and therefore grouped both sides into simple structural types before the experiments, with falsifiable matching predictions. Through multi-dimensional experiments, we find: (1) grayscale cannot emulate SAR scattering and causes negative transfer; (2) pixel scores (peak signal-to-noise ratio [PSNR]/ structural similarity index measure [SSIM]) are decoupled from mean average precision (mAP) and so selection should be detection-driven; (3) no generator is universally best—pairing is type- and scene-dependent—and complex multi-class aircraft scenes partially falsify it, demonstrating that the a priori hypothesis is predictive in simple scenes but not predictive in complex multi-class scenes; (4) mixing has a useful window rather than “more is better,” with larger gains for lower-baseline detectors (up to +6.1/+11.0 mAP@0.5 on ships/aircraft); (5) gains mainly improve detection/recall (mAP@0.5), with limited tight-localization benefit; and (6) abundant real SAR yields diminishing returns. These results provide hypothesis-driven pairing and dosing guidance for O2S augmentation in SAR detection.
Zitieren
APA
Xiao, C., Yang, W., Ma, G., Wang, X., Niu, Y., Lu, Y., Lu, J., & Chen, W. (2026). Gain Mechanism Analysis of Optical-to-synthetic Aperture Radar (SAR) Cross-modal Data Augmentation for SAR Object Detection. Photogrammetric Engineering & Remote Sensing. https://doi.org/10.14358/pers.26-00107r2
Vancouver
Xiao C, Yang W, Ma G, Wang X, Niu Y, Lu Y, et al. Gain Mechanism Analysis of Optical-to-synthetic Aperture Radar (SAR) Cross-modal Data Augmentation for SAR Object Detection. Photogrammetric Engineering & Remote Sensing. 2026. doi:10.14358/pers.26-00107r2
BibTeX
@article{Xiao2026,
author = {Xiao, Changlin and Yang, Weichen and Ma, Guangdi and Wang, Xiang and Niu, Yiming and Lu, Yilong and Lu, Jiangping and Chen, Wei},
title = {Gain Mechanism Analysis of Optical-to-synthetic Aperture Radar (SAR) Cross-modal Data Augmentation for SAR Object Detection},
journal = {Photogrammetric Engineering & Remote Sensing},
year = {2026},
doi = {10.14358/pers.26-00107r2}
}
EUVIMED