Remote sensing image fusion based on sparse representation has achieved dramatic results. However

classical sparse representation faces two problems. First

the computational complexity of dictionary training is extremely high. Second

the classical sparse representation does not consider the similarity between image patches

which causes performance degradation in solving sparse coefficients. A new structural group sparse representation (SGSR)

which has been successfully applied to image restoration and super-resolution image reconstruction

can effectively solve the two problems. To improve the accuracy and speed of sparse representation of remote sensing image fusion

this study presents a method of remote sensing image fusion based on SGSR. Considering that the spectral range of panchromatic image does not exactly override all the bands of a multispectral image

a luminance component derived from the adaptive weight coefficient of each band image is initially defined to reduce the spectral distortion. Then

the Euclidean distance between the gray scales of image patches is calculated. The most similar structure patches are selected to constitute the structural groups for luminance component image and panchromatic image

and these structural groups are regarded as the basic unit of dictionary learning and sparse representation. The adaptive dictionary learning method by singular value decomposition obtains the adaptive dictionary for every structural group. Furthermore

group sparse coefficient of the luminance component image and panchromatic image are calculated via group sparse representation algorithm. Finally

using the absolute maximum fusion rule

new sparse coefficients can be obtained by part of panchromatic image sparse coefficient substitution

and the high spatial resolution intensity image is reconstructed using the panchromatic image group dictionary and the new sparse coefficients. The high-resolution multispectral image is obtained via the general component substitution framework. SGSR is evaluated by the degraded and nondegraded image fusion experiments on Gaofeng-1

Quickbird

and Worldview images. In the degraded remote sensing image fusion experiment

five quantitative evaluation indices (CC

RMSE

ERGAS

UIQI

and SAM) are used to compare SGSR with the traditional methods (GIHS

PCA

and HPF)

AWLP

and classical sparse representation (SR) method. Results show that five quantitative indices of SGSR are better compared with the other methods. In the nondegraded remote sensing image fusion experiment

psnr

grad

entropy

Dλ

Ds

and QNR are used to compare these image fusion methods. SGSR generally obtained better results

except that spectral information preservation performance of SGSR in the green plant region is slightly worse than those of AWLP. By comparing the speed of training dictionary

SGSR is approximately 10 times faster than SR. SGSR can improve spatial resolution and preserve spectral information in a remote sensing image and has better performance than the traditional methods

AWLP

and SR. Moreover

by comparing SGSR with SR

SGSR significantly reduces the computational complexity. Therefore

SGSR is suitable for multispectral image and panchromatic image fusion of all types of remote sensing images and may be applied in this field.