融合Sentinel-2数据的高分五号高光谱数据降尺度

Downscaling GF-5 hyperspectral images by fusing with Sentinel-2 images

高分五号(GF-5)是中国首颗实现对大气和陆地综合观测的全谱段高光谱卫星,其搭载的可见短波红外高光谱相机AHSI提供的遥感数据拥有极高的光谱分辨率。然而,AHSI数据的空间分辨率为30 m,较低的空间分辨率限制了应用场景。为实现GF-5数据的降尺度,本文通过融合10 m哨兵二号(Sentinel-2)多光谱数据,生成10 m GF-5高光谱数据。在方法上,针对现有先进的信息损失引导的图像融合方法ILGIF(Information Loss-Guided Image Fusion)在高光谱图像降尺度中计算时间成本高的问题,本文提出了其快速版本FILGIF(Fast ILGIF)。另一方面,在降尺度过程中,本文考虑并估计了30 m GF-5高光谱数据和10 m Sentinel-2数据之间的尺度转换点扩散函数PSF(Point Spread Function),提高融合数据质量。实验结果验证了融合Sentinel-2数据用于GF-5高光谱数据降尺度的可行性。同时,结果表明:在获得与ILGIF相当精度的前提下,FILGIF大幅提高了运行效率;尺度转换PSF对降尺度过程有着重要影响,其准确的估计有助于获得更高精度的降尺度结果。

GF-5 is the first hyperspectral satellite in China that can acquire comprehensive observations of the atmosphere and land surface.The Advanced Hyper Spectral Imager(AHSI)onboard GF-5 is a sensor that can acquire data covering visible near-infrared(VNIR)and short-wave infrared(SWIR)wavelengths with a very fine spectral resolution(i.e.,5 nm for VNIR and 10 nm for SWIR).However,the spatial resolution of GF-5 AHSI data(i.e.,30 m)is relatively coarse for the extraction of land cover information in several cases,such as small-sized buildings and roads in urban areas.To produce GF-5 data with fine spatial and spectral resolutions,in this paper,GF-5 hyperspectral images were downscaled to 10 m by spatial-spectral image fusion with 10 m Sentinel-2 multispectral images.To deal with the large computational burden of the advanced Information Loss Guided Image Fusion(ILGIF)method and the ubiquitous effect of the Point Spread Function(PSF),this paper also introduced a fast and accurate method for downscaling GF-5 data.A fast ILGIF(FILGIF)method was proposed.In this method,the original GF-5 hyperspectral data were transformed to a new feature space via principal component analysis(PCA),and the ILGIF-based spatial-spectral image fusion was implemented for the first few principal components.The fused components coupled with the remaining ones were transformed back to the original space to produce the 10 m downscaled results.The scale transformation optimal PSF between 10 m Sentinel-2 and 30 m GF-5 data was estimated adaptively for each band of GF-5 to enhance downscaling.Experimental results show that by fusing with the 10 m Sentinel-2 data,the 30 m GF-5 hyperspectral data can be downscaled effectively to 10 m.The FILGIF and ILGLF methods obtain greater accuracy than the area-to-point regression kriging(ATPRK)and approximate ATPRK(AATPRK)methods.Moreover,the computational cost of FILGIF is 30 times lower than that of ILGIF,and the accuracy of the downscaling results can be improved by considering the PSF effect adaptively for each band.Sentinel-2 images are suitable for downscaling GF-5 hyperspectral images.The proposed FILGIF method can achieve a comparable accuracy compared with ILGIF while significantly reducing computational costs.Highly accurate downscaling results are obtained when the PSF effect is considered appropriately.