Unpaired optical-to-SAR image translation with coordinate attention and differentiable histogram loss
Wenbo Yu, Jiamu Li, Tian Tian, Feng Zhou · International Journal of Remote Sensing · 2026
The high acquisition cost of synthetic aperture radar (SAR) imagery has been a persistent obstacle to advanced deep learning researches. Recently, image translation techniques have emerged as promising solutions for augmenting SAR datasets by translating readily available optical images into SAR-like representations. However, the substantial stylistic differences between optical and SAR images pose significant challenges in accurately extracting optical image semantics and replicating SAR image styles, especially when co-registered data is unavailable. To address this challenge, we propose an unpaired optical-to-SAR image translation (O2SIT) method, named extract-and-transform generative adversarial network (ET-GAN). First, we introduce cascaded coordinate attention (CA) bottleneck blocks that enhance the positional information of feature maps, thereby precisely extracting optical image semantics. Second, to better capture SAR style characteristics, we employ histograms as auxiliary supervision by constructing a differentiable histogram using kernel density estimation and global average pooling. On this basis, the squared earth mover distance is adopted as an additional loss to guide the generator in producing synthetic images with pixel distributions similar to real SAR images. Experimental results on SEN12, WHU-SEN-City, and GaoFen aircraft detection (GF-AD) dataset demonstrate that ET-GAN achieves competitive SAR image generation performance compared to other state-of-the-art methods, with PSNR of 17.11 on SEN12 and FID of 168.16 on GF-AD. Transfer learning results demonstrate that the images generated by ET-GAN can bring about 3% accuracy improvement to SAR aircraft detection.