SARLCNet: supervised generative adversarial SAR layover compensation framework

Considering the layover phenomenon in SAR orthoimage, the existing layover compensation (LC) methods suffer from insufficient data and the difficulty of layover area detection. In this study, a supervised SAR LC method based on domain adaptation was proposed. Considering the lack of multi-view SAR image data and the difficulty of accurate detection of layover areas, a high-precision SAR layover extraction method was proposed, and then the SAR LC dataset (SAR-LCD) was constructed. Furthermore, a supervised generative adversarial network is used to construct the SARLCNet, which converts optical data into simulated SAR images. In this method, the optical content and SAR style are coupled, and SAR mask supervision information is added into the framework, which effectively improves the performance of LC in SAR images. The experimental results with the SAR-LCD show that the proposed method can achieve efficient and accurate SAR LC tasks. Extensive experiments on single and regional SAR images have shown that SARLCNet achieves a compensation rate at 100% while maintaining structural similarity to the original SAR image and well radiometric consistency, which verifies the effectiveness of our approach. The SARLCNet method can be regarded as a reliable way to compensate for SAR layover.