Work

Phy-OSShip-MS Dataset: A Paired Optical–SAR Ship Imagery Generation Method via Hierarchical Physics-Guided Style Transfer

Year: 2026

Type: article

Abstract: High-quality spatiotemporally paired optical and Synthetic Aperture Radar (SAR) datasets are fundamental for cross-modal ship detection tasks. However, due to distinct imaging mechanisms, difficulties in acquiring spatiotemporal consistency, and data privacy concerns, such public datasets are extremely scarce. Current cross-modal SAR image synthesis methods often struggle to balance physical consistency and perceptual fidelity, resulting in images with scattering distributions that violate imaging mechanisms and speckle statistics that deviate from real data. To address these challenges, this article proposes a Hierarchical Physics-Guided Style Transfer framework for Optical-to-SAR image generation. In this architecture, style transfer serves as an underlying generative backbone to ensure perceptual fidelity, while a novel “Radiometric-Geometric-Fusion” three-level physical guidance strategy is introduced to explicitly constrain the generation process. First, a scene-universal category-level radiometric prior is constructed to establish physical intensity references for different categories. Second, phase congruency features are introduced to compensate for the deficiency of radiometric priors in fine structural depiction, precisely preserving the critical geometric details of strong scattering targets like ships. Finally, a synergistic fusion mechanism based on physical attribute intersection is established to apply directional gain to strong scattering targets by coupling material attributes with geometric saliency, thereby ensuring intensity-level physical consistency. Experimental results demonstrate that the proposed method outperforms existing image-to-image translation models in terms of texture details and scattering characteristics. Consequently, we establish a Physics-guided Optical-SAR Ship Imagery dataset with Multi-Source attributes (Phy-OSShip-MS), providing target-level synthetically paired samples as an effective generative data augmentation strategy to reliably alleviate the data scarcity issue in cross-modal ship detection. Downstream evaluations show that our dataset significantly improves cross-domain generalization, boosting Average Precision (AP50) on unseen domains by over 56%. The constructed dataset will be made publicly available athttps://github.com/lilili0612/Phy-OSShip-MS-Dataset.

Source: IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing

DOI: 10.1109/jstars.2026.3699023