<![CDATA[Flood monitoring in tropical regions is challenged by persistent cloud cover, which restricts optical imagery, while consistently available Synthetic Aperture Radar (SAR) imagery presents difficulties in visual interpretation. This study employs CycleGAN to translate SAR into optical-like imagery in an unpaired-data scenario within flood-affected areas of Sumatra. Seven training configurations were evaluated, including the default setting, an asymmetric learning rate scheme, a combination of spectral normalization and geometric augmentation, isolated ablations of these components, and variations in the cycle-consistency coefficient. The dataset consisted of 482 Sentinel-1 SAR patches and 446 Sentinel-2 optical patches for training, alongside 276 SAR images for testing, all acquired via Google Earth Engine during the November 2025 flood. The evaluation utilized four metrics: the Structural Similarity Index (SSIM), Learned Perceptual Image Patch Similarity (LPIPS), Fréchet Inception Distance (FID), and the convergence epoch as the primary model-selection criterion. SSIM and LPIPS were calculated for cyclic reconstruction to address the lack of paired optical references. The configuration integrating spectral normalization, a reduced discriminator learning rate, and geometric augmentation achieved the fastest convergence (epoch 43) and the highest performance across all metrics (SSIM = 0.939, LPIPS = 0.025, and FID = 143.00).]]>
Copyrights © 2026
