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

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. © 2023 National Remote Sensing Bulletin