High-quality cloud-free optical image generation using multi-temporal SAR and contaminated optical data

This paper addresses the challenge of missing optical image generation, i.e., generating unavailable optical images, for instance, due to cloud cover. We propose CRGenNet, an image generation network that incorporates innovative modules, for instance the DownUp Block and Fusion Attention to enhance accuracy. Furthermore, we constructed the TCSEN12 dataset, which includes multi-temporal SAR and optical observations in a flood-affected region, offering a benchmark for evaluating image generation methods under challenging conditions, characterized by cloud-contaminated auxiliary imagery and substantial temporal change. Experimental results validate the effectiveness, achieving superior visual effects and quantitative evaluation results that exceed those produced by comparison methods. On the TCSEN12 dataset, it achieved a Peak Signal-to-Noise Ratio (PSNR) of 26.978 and Structural Similarity Index Measure (SSIM) of 0.648, as compared to 26.225 and 0.622, using MTS2ONet, while on the public DTSEN1-2 dataset, it achieved a PSNR of 88.9263 and SSIM of 0.9999, as compared to 36.0435 PSNR and 0.9896 SSIM, respectively, for MTS2ONet. The study concludes that CRGenNet provides a robust method for integrating multi-modal data with complex temporal dynamics, while TCSEN12 establishes a more realistic benchmark by focusing on future-scene prediction under challenging cloud and post-disaster conditions.