S2O-SARUNet: SAR-to-optical self-attention Res-UNet for grayscale images colorization
Abstract Synthetic Aperture Radar (SAR) imagery provides all weather, day and night Earth observation but lacks the spectral richness of optical imagery, which limits interpretability in many applications. This work presents an enhanced SAR-to-Optical (S2O) translation model architecture based on a conditional Generative Adversarial Network (cGAN) with a Residual U-Net generator augmented self-attention, which we will call S2O-SARUNeUsing the SEN1–2 dataset, comprising globally distributed, multi-season SAR–optical pairs, and integrating VGG-based perceptual loss, cosine-annealed learning rates, and early stopping to maximize structural and spectral normalization in the discriminator on GPU hardware (RTX A2000, 4 GB).dB, with a peak SSIM of 0.961 and best case PSNR of 27.5 dB, with LPIPS of 0.302, FID of 114.67, SAM of 8.60°, and ERGAS of 36.59. The proposed configuration outperformed the evaluated baseline U-Net, transformer-augmented, and diffusion-refined variants in terms of SSIM and PSNR. However, comparisons with literature-reported methods revealed metric-dependent trade-offs, and uniform superiority across all metrics is not claimed. The results demonstrate the feasibility of the proposed local–global–adversarial configuration for SAR-to-optical translation under the reported resource-constrained setting.
Authors
- Fahima A. Maghraby (ORCID: https://orcid.org/0000-0002-2547-5673)
- Mohamed A. Aborizka (ORCID: https://orcid.org/0000-0003-1154-6407)
- Ahmed Mohamed Abosrea Abdelaziz
Publication Details
- Journal
- Scientific Reports
- Published
- 2026-11-15
- DOI
- https://doi.org/10.1038/s41598-026-73580-x
- Primary Topic
- Generative Adversarial Networks and Image Synthesis
- Type
- article
- Field-Weighted Citation Impact
- 0.00