【目的】 针对光学遥感影像薄云去除后易出现影像失真、可用性降低的问题,提出一种融合慢特征分析(SFA)与生成对抗网络(GANs)的薄云去除方法(SFGAN),以提升影像质量,为林业遥感数据分析提供可靠支持。【方法】首先,设计慢变特征模块,通过计算初始影像的云反射率及高维特征慢变化度,将慢变特征向量与随机初始向量融合作为生成网络输入,增强生成器对云层特征的辨识能力;其次,利用云反射率作为鉴别器约束因子,通过对抗博弈迭代优化生成高质量无云影像,提升网络对云层与地物的区分能力。【结果】在公开数据集RICE1和PRSC上的试验表明:SFGAN模型在RICE1数据集上表现较为出色,无云影像的峰值信噪比(PSNR值)为33.740 7,结构相似性(SSIM值)为0.958 2;在PRSC数据集上,PSNR值为24.341 3,SSIM值为0.879 2,即本方法的性能量化指标及视觉质量均优于其他方法。基于Landsat 8影像的泛化实验进一步验证,SFGAN在真实云与模拟云场景中均能有效恢复地物细节,且处理单幅影像仅需0.98 s。【结论】SFGAN通过融合慢特征分析与生成对抗网络,能显著降低薄云对林业光学遥感影像的干扰,从数据源头提升影像可用性与分析准确性。

To address the issue of image distortion and reduced usability caused by thin cloud removal in optical remote sensing images, this study proposes a novel thin cloud removal method: SFGAN, that integrates slow feature analysis (SFA) with generative adversarial networks (GANs), aiming to enhance image quality and provide reliable data support for forestry remote sensing analysis.【Method】First, a slow-varying feature module is designed to calculate cloud reflectance and high-dimensional feature slowness. The slow-varying feature vectors are concatenated with random initial vectors as the generator input, improving cloud feature recognition. Second, cloud reflectance is utilized as a discriminative constraint factor to iteratively optimize the discriminator, thereby generating high-quality cloud-free images through adversarial training.【Result】Experiments on public datasets RICE1 and PRSC demonstrate that the SFGAN outperforms existing methods in both quantitative metrics (e.g., PSNR=33.740 7 and SSIM=0.958 2 on RICE1,PSNR=24.341 3 and SSIM=0.879 2 on PRSC) and visual assessments. Validation using Landsat 8 imagery shows SFGAN achieves superior cloud removal effects in both real and simulated cloud scenarios, with a processing time of 0.98 seconds per image.【Conclusion】The SFGAN framework effectively mitigates thin cloud interference in forestry optical remote sensing images by synergizing SFA and GANs, significantly improving data usability and analytical accuracy at the source level.