LAUSR

Abstract The significant issues in remote sensing image fusion are enhancing the spatial details and preserving the essential spectral information. The classical pan-sharpening methods often incur spectral distortion and still striving to produce the fused images with prominent spatial and spectral attributes. Motivated by the desirable results of sparse representation (SR) theory, a novel pan-sharpening method is developed based on SR of high frequency (HF) components over a multi-scale learned dictionary (MSLD). MSLD technique acquires the capability of extracting the intrinsic characteristics of images, wherein, it possess the features of both multi-scale representation and learned dictionaries. In this paper, the dictionaries are adaptively learned from HF sub-images derived from the two versions of panchromatic image, realized at different spatial resolutions. A fast and computationally efficient algorithm is used for dictionary learning. The notion of SR together with patch recurrence over different scales is incorporated to estimate the high frequency details. The fused image is reconstructed by injecting the band specific spatial details into the up-sampled multi-spectral images. The performance of the proposed method is appraised with the datasets from different satellite sensors namely, QuickBird, IKONOS, WorldView-2 and Pleiades. The observations inferred from visual perception and quality indices analysis manifest the efficiency of proposed method over several well-known methods for the datasets considered at reduced-scale and full-scale resolutions. Further, the quantitative analysis of obtained performance measures confirms the efficacy of the proposed method for the reduced-scale and full-scale data sets. Especially, at a reduced-scale, proposed method yields an optimal value of Correlation coefficient, Structural similarity and Q4. In a comparative sense, usage of the proposed method at full-scale results in 4% and 2.56% improvement in the Spatial distortion index for QuickBird and WorldView-2 data respectively contrary to the best reported outcome obtained from Sparse Representation of injected details (SR-D) scheme. Invariably, for full-scale data, the QNR attains its optimal value.