Reproduce “TSGAN: An Optical-to-SAR Dual Conditional GAN for Optical based SAR Temporal Shifting” on OpenResearch

About this paper

Moien Rangzan, Sara Attarchi, Richard Gloaguen, Seyed Kazem Alavipanah

This research introduces TSGAN, a dual conditional Generative Adversarial Network that reformulates Optical-to-SAR (Opt2SAR) image translation as a "SAR Temporal Shifting" problem. The model takes a prior SAR image, current optical data, and an optical change map to generate a current SAR image, which addresses the ill-posed nature of Opt2SAR and reduces the generation of unrealistic details.

Key insights

• Reformulating Opt2SAR as a temporal modification task, guided by an existing SAR image and optical change information, is effective in mitigating the ill-posed nature of the problem and significantly reducing the "fiction" phenomenon.
• The combination of a change-weighted loss function and an explicit optical change map input is crucial for training robust models that accurately modify changed regions while preserving high fidelity in unchanged areas.
• Specific architectural elements, such as a 1:1 convolution for SAR input and attention mechanisms (Squeeze-and-Excitation, Global-Local Attention Module), improve the network's ability to selectively transfer and synthesize information across modalities and time.

Summary by alphaXiv.