融合多光谱影像的高光谱影像厚云去除方法

云遮挡对高光谱影像的应用造成了不可忽视的影响。现有云去除方法通常利用时域近邻的同源影像提供辅助信息。然而,高光谱影像(如GF-5和EO-1高光谱影像)较低的时间分辨率导致同源辅助影像中可能存在较大的地物覆盖变化。时间分辨率更高的多光谱影像(如Landsat 8 OLI影像)能提供时间上更接近于高光谱云影像的辅助信息,从而减少地物覆被变化带来的影响。为应对高光谱和多光谱波段之间差异较大的问题,本文基于空谱随机森林(spatial-spectral-based random forest,SSRF)方法,提出一种利用多光谱影像(Landsat 8 OLI影像)对高光谱影像进行厚云去除的方法,将其简记为SSRF_M。SSRF_M较强的非线性拟合能力使其能够综合利用多光谱影像所有波段的有效数据对各个高光谱波段进行重建。本文使用GF-5和EO-1高光谱影像进行模拟云去除试验,视觉和定量评价结果均表明,与利用时间间隔更长的同源辅助影像的方法相比,本文方法能获得更高精度的云下信息重建结果。

Thick cloud removal of hyperspectral images by fusing with multispectral images

The cloud contamination issue poses a significant obstacle to the application of hyperspectral images. Existing cloud removal methods usually use temporally close images from the same sensors as cloudy images to provide auxiliary information. Unfortunately, the coarse temporal resolution of hyperspectral images (GF-5 and EO-1 hyperspectral images) may result in great land cover changes. The finer temporal resolution of multispectral images (Landsat 8 OLI images) allows to provide auxiliary information temporally closer to the hyperspectral cloudy images, thus reducing the effect uncertainty caused by land cover changes. To deal with the large spectral differences between auxiliary multispectral bands and cloud-contaminated hyperspectral bands, this paper applied the spatial-spectral-based random forest (SSRF) method to use multispectral images (Landsat 8 OLI images) for cloud removal of hyperspectral images, namely, the SSRF_M method. Benefiting from the strong nonlinear fitting ability, the proposed SSRF_M method can use simultaneously the effective information from multiple bands of the auxiliary multispectral image for cloud removal of each hyperspectral band. In this paper, the GF-5 and EO-1 hyperspectral images were used for cloud simulation experiments. The visual and quantitative evaluation results show that compared with the strategy using homologous auxiliary images, the proposed SSRF_M method can reconstruct the information under clouds more accurately.