Journal of the Indian Society of Remote Sensing (December 2018) 46(12):2069–2079

https://doi.org/10.1007/s12524-018-0878-8(0123456789().,-volV)(0123456789().,-volV)

RESEARCH ARTICLE

Quaternion-Based Sparse Model for Pan-Sharpening of IRS Satellite

Images
D. Synthiya Vinothini1 • B. Sathya Bama1

Received: 11 July 2017 / Accepted: 22 October 2018 / Published online: 26 October 2018
 Indian Society of Remote Sensing 2018

Abstract
This paper considers the pan-sharpening problem of the IRS satellite images from the perspective of vector sparse
representation model using quaternion matrix analysis. It selects the sparse basis in quaternion space, which uniformly
transforms the color channels into an orthogonal color space. Moreover, the proposed quaternion model for pan-sharpening
is more efficient than the conventional sparse model as the hyper-complex representation of color channels conserves the
interrelationship among the chromatic channels. This paper also proposes a quaternion forward–backward pursuit algo-
rithm that preserves the inherent chromatic structures in terms of spatial and spectral details during the vector recon-
struction. The experimental result validates the efficacy of the proposed quaternion model and shows its potential as a
powerful pan-sharpening tool for IRS data even for cloudy multispectral data.

Keywords Image fusion  Indian remote sensing satellite  Multispectral image  Panchromatic image  Pan-sharpening 
Quaternions  Remote sensing  Sparse representation

Introduction The Indian Earth Observation (EO) satellite, Indian

Remote Sensing IRS-P5 Cartosat-2, provides panchromatic
There is a thriving need for fine spectral and high spatial image at a higher spatial resolution (0.65 m) than that of
resolution in remote sensing applications which involves the multispectral image provided by the linear imaging
accurate description of morphological features, geometrical self-scanning sensor (LISS-IV) of ResourceSat-2 (5.8 m).
shapes and structures, better distinction between categories Pan-sharpening is the task of merging both these data to
and faithful interpretation capabilities. But unfortunately overcome the limitations of information obtained from the
due to the trade-off in spectral and spatial resolution of the individual sensors and can achieve a better understanding
sensors, the remote sensing platforms provide coarse of the observed scene (Joshi et al. 2006).
multispectral images and fine panchromatic image which is The vital role in the pan-sharpening approach is to uti-
the reflectance data representative of a range of wave- lize the relationship between the low-resolution multi-
lengths. With the advancement of sensors in the Indian spectral (LRMS) image and the high-resolution
remote sensing scenario, the reflectance data are available panchromatic (HRP) image (Jiang et al. 2014). Recently,
at finer resolution. This development leads to the emer- this image fusion has been a topic of interest for many
gence of pan-sharpening scheme which merges comple- researchers and can be categorized into component sub-
mentary information from different sensors. stitution (CS), ARSIS concept based, restoration based,
compressed sensing (CoS). The CS approach is based on
transformation of resampled LRMS into another space
& D. Synthiya Vinothini where one of its components which is closely related to the
synthiya@tce.edu
HRP will be substituted by the HRP itself and then trans-
B. Sathya Bama formed back to its original space (Choi et al. 2011). The
sbece@tce.edu
widely used component substitution method includes
1
Department of Electronics and Communication Engineering, intensity–hue–saturation (IHS) (Tu et al. 2001), principal
Thiagarajar College of Engineering, Madurai, component analysis (PCA) (Chavez et al. 1991) and Gram–
Tamil Nadu 625015, India

123
2070 Journal of the Indian Society of Remote Sensing (December 2018) 46(12):2069–2079

Schmidt (GS) process (Laben and Brower 2000). The high- The goal of this paper is to take advantage of the
pass filtering (HPF) (Chavez et al. 1991) and multi-reso- interesting properties of quaternion model in the context of
lution analysis (MRA) (Amolins et al. 2007; Rao et al. image pan-sharpening. This paper proposes a pan-sharp-
2016) methods try to implement the concept of ARSIS. It ening scheme using the quaternion-based sparse model.
includes A trous Wavelet Transform (AWT) (Garzelli et al. The key contribution of this work is that
2004) or Laplacian Pyramid or Content-Based Decision
• The chromatic channels are represented as quaternion
(CBD) (Aiazzi et al. 2003), in which the high-frequency
matrix, which increases the correlation among the color
information extracted from the HRP image is injected into
channels.
the LRMS image that lacks the spatial information.
• Dictionary construction for quaternion-based sparse
The restoration-based model is inspired by the single-
model, which does not require dictionary training;
image super-resolution technique (Jiang et al. 2014), where
hence, it is practical and time-efficient.
LRMS and HRP are assumed to be the degraded obser-
• Quaternion forward–backward pursuit (QFBP) algo-
vations of the high-resolution multispectral (HRMS) image
rithm for estimation of quaternion sparse vector which
(Li and Leung 2009; Zhang et al. 2012). The CoS tech-
is an iterative forward–backward scheme, which
nique shows impressive results in the pan-sharpening
requires no prior knowledge of the sparsity and has
application (Li and Yang 2011). This solves the image
an advantage of removing misidentified atoms from the
reconstruction as an inverse problem using sparse repre-
support.
sentation of data which has been successful in the field
• Quaternion pan-sharpening of IRS satellite data using
recently. The topic became influential after the effective
the proposed sparse model to reconstruct information
contribution of Li and Yang (2011) that presumed to have a
efficiently.
sparse representation in a dictionary randomly sampled
from a database of unavailable HR multispectral images The rest of the paper is organized as follows. In ‘‘Pan-
acquired by compatible sensors. The studies have docu- sharpening with Quaternion-Based Sparse Model’’ section,
mented several interesting dictionary construction the quaternion-based dictionary construction and quater-
methodologies which play a significant role in determining nion sparse vector estimation are briefly introduced and the
the fusion quality. A more practical method was used by proposed quaternion pan-sharpening scheme is explained.
Jiang et al. (2012) where dictionaries are constructed from Experimental results and the discussion on the qualitative
a set of database available. and quantitative comparisons with the conventional algo-
Another motivating effort was obtaining the dictionary rithms are provided in ‘‘Experimental Results and Dis-
from the data readily available at hand, considering only cussion’’ section. Finally, in ‘‘Conclusion’’ section, the
the corresponding scene under observance. Typically, the work is summarized and concluded.
dictionary is constructed from the HRP (Jiang et al. 2014;
Zhu and Bamler 2013; Guo et al. 2014), both the HRP and
the original LRMS (Li et al. 2013), the synthetic HRMS Pan-Sharpening with Quaternion-Based
images (Cheng et al. 2014) obtained from the conventional Sparse Model
pan-sharpening schemes, or the details of the HRP and
LRMS (Vicinanza et al. 2015) to reconstruct the details of Motivated by the quaternion representation of the sparse
HRMS efficiently. The results discussed in the above- model (Xu et al. 2015; Yu et al. 2014), this work proposes
mentioned algorithms are inspiring, yet the constructed a novel idea of pan-sharpening which utilizes the super-
dictionary lacks correlation among the chromatic channels. resolution capability of the quaternion matrix representa-
The state-of-the-art pan-sharpening methods based on tion. The proposed pan-sharpening scheme with quater-
sparse representation rely on independent spectral sensor, nion-based sparse model (PSQS) comprises three major
resulting in significant loss of spectral information, as they phases: (1) quaternion dictionary construction, (2) quater-
do not consider the correlation among the spectral chan- nion sparse vector estimation and (3) quaternion super-
nels. However, hyper-complex numbers, in particular the resolution reconstruction of the HRMS image. The concept
quaternion model, allows handling of the chromatic chan- diagram of PSQS is shown in Fig. 1. Given the LRMS
nels as a single entity, preserving the correlation among (x) image and the HRP (Y) image, the HRMS (X) image is
them (Yu et al. 2014). The super-resolution capability and reconstructed.
robustness of the quaternion sparse reconstruction tech-
nique give higher spatial and spectral resolution with less Quaternion Dictionary Construction
spectral distortion compared with other existing methods
(Xu et al. 2015). Conventionally, in the compressed sensing pan-sharpening
algorithms dictionary construction is based on either

123
Journal of the Indian Society of Remote Sensing (December 2018) 46(12):2069–2079 2071

Fig. 1 Concept diagram of the proposed pan-sharpening scheme with quaternion-based sparse model (PSQS)

monochromatic model (Zhu and Bamler 2013; Guo et al. moments of the HRP image with those of each LRMS band
2014), or concatenation model (Jiang et al. 2012, 2014; Li improves the performances of the algorithm (Vivone et al.
et al. 2013; Cheng et al. 2014; Vicinanza et al. 2015). This 2015).
section proposes a quaternion-based sparse model for dic- ðY  lY ÞrI
tionary construction. The RGB color space can be repre- Yeq ¼ þ lI ð2Þ
rY
sented in the pure quaternion form (Xu et al. 2015; Yu
et al. 2014). Hence, the dictionary matrix can be repre- X~ ¼ x~ þ ðYeq  IÞ : ð3Þ
sented as D_ ¼ Dp þ Dr i þ Dg j þ Db k where D_ 2Hs x n is
the quaternion dictionary with its components
Quaternion Sparse Vector Estimation
Dp ; Dr ; Dg ; Db 2Rp x n , where the scalar part is the dic-
tionary constructed (Zhu and Bamler 2013) from the The sparse-based signal recovery has an advantage of
panchromatic image and the imaginary part consists of the adapting themselves to the data, while the super-resolution
three chromatic channel dictionaries constructed (Zhu and capability, correlation and the robustness of the quater-
Bamler 2013) independently. Similarly, the quaternion nionic sparse reconstruction are anticipated to increase the
image vector and its corresponding quaternion sparse resolution in terms of both spatial and spectral dimension
coefficient are represented by x_ ¼ xp þ xr i þ xg j þ xb k and with reduced distortion.
a_ ¼ ap þ ar i þ ag j þ ab k, respectively. Now, the quater- In pan-sharpening applications, the sparse coefficients
nion-based sparse model is represented as are estimated predominantly by basis pursuit (BP), K-sin-
min jjajj
_ 0; s:t: x_ ¼ D_ a;
_ ð1Þ gular value decomposition (K-SVD) (Cheng et al. 2014;
a_
Jiang et al. 2012), which is not computationally efficient,
where a_ 2 Hn is the quaternion sparse vector coefficient of least absolute shrinkage and selection operator (LASSO)
the input data x_ 2Hs ,with its components ap ; ar ; ag ; ab 2 Rn (Guo et al. 2014) and orthogonal matching pursuit (OMP)
and xp ; xr ; xg ; xb 2Rs . Since the equivalent LR and HR (Cheng et al. 2014; Li et al. 2013; Vicinanza et al. 2015),
multispectral patches share the same quaternion sparse which is fast. OMP iteratively enlarges the support esti-
vector coefficient over the corresponding quaternionic mate, i.e., once an incorrect atom enters the support, it
dictionaries D_ h and D_ l , the dictionary pair is constructed cannot be removed leading to data misrepresentation and
directly from the corresponding panchromatic and multi- thus the signal recovery will be distorted. This can be
spectral channels and does not require training. The dic- recovered by forward–backward pursuit (FBP) algorithm
tionary D_ l ¼ Dy þ Dxr i þ Dxg j þ Dxb k is constructed using (Karahanoglu and Erdogan 2013) which includes a second
stage known as backward step to prune the support esti-
the down-sampled version of HRP and the available LRMS
mate in each iteration. This paper proposes a quaternion
data, while D_ h ¼ DY þ DX~r i þ DX~g j þ DX~b k utilizes the
forward–backward pursuit (QFBP) which solves the
HRP and the matched image X. ~ The intensity component, I, quaternion sparse estimation problem satisfying the fol-
of the LRMS equalizes the HRP data according to Eq. (2). lowing equation:
In general, preprocessing by matching the two first

123
2072 Journal of the Indian Society of Remote Sensing (December 2018) 46(12):2069–2079

a_ ¼ arg min _ 22 ;
jjx_  D_ ajj s:t: jjajj
_ 0  k; ð4Þ The QFBP algorithm estimates the quaternionic sparse
a_ coefficient vector a_ of the quaternion measurement vector x_
where k is the sparsity factor denoting the maximum over the quaternion dictionary D. _ This two-stage iterative
number of nonzero coefficients to represent a signal. The algorithm expands the support estimate S_ in the forward
pseudo-code of QFBP is given in Algorithm 1. step by the factor g which is the forward step size. This is

123
Journal of the Indian Society of Remote Sensing (December 2018) 46(12):2069–2079 2073

done by finding g atoms in the dictionary, which are

maximally correlated with the residue n. _ Then, the
quaternion orthogonal projection of the vector onto the
support is estimated by (8). The superscript ‘i’ denotes the
quaternionic pseudo-inverse operation (Xu et al. 2015; Yu
et al. 2014). The b atoms that contribute minimum in the
projection are then removed in the backward step, where b Fig. 2 Spectral range of IRS data
is the backward step size for pruning the support. The a_ is
now estimated by (11). The residue is then updated by (12), been performed on the data from the sensors: first, the full-
and the iteration terminates if the energy of the residue is and reduced-scale validation on the cloud-free multispec-
less than the allowable parameter e or if the maximum tral and panchromatic data, and then the full-scale valida-
sparsity j is obtained or if the maximum number of iter- tion of the clouded multispectral and panchromatic data. A
ations qmax is achieved. major requirement for pan-sharpening application is that
the overhead times of the satellite sensors should be as
Quaternion Super-resolution Reconstruction close as possible. In our case, the IRS dataset is from
Madurai region, India, which was obtained by the sensors
Each quaternionic LRMS patch (x_n ) has to be represented in February 2015.
sparsely over the constructed LR dictionary. The funda- The comparison is made with the following popular
mental principle behind this work is that the corresponding methods: additive a trous wavelet transform with unitary
HRMS patch can also be represented by the same sparse injection model (ATWT) (Garzelli et al. 2004), additive
vector over a similarly constructed HR dictionary. For each wavelet luminance proportional (AWLP) (Ranchin and
LRMS patch x_n , a quaternion sparse vector a_ is estimated Wald 2000), Brovey transform (Gillespie et al. 1987),
by QFBP algorithm. band-dependent spatial detail (BDSD) with local parameter
Now, the sparse vector coefficient of each atom is estimation (Garzelli et al. 2008), Gram–Schmidt adaptive
multiplied with the dictionary D_ h using quaternion analysis (GSA) (Aiazzi et al. 2007), partial replacement adaptive
to generate a HRMS patch using (14), which is tiled back component substitution (PRACS), intensity–hue–saturation
to reconstruct the desired HRMS data. model (IHS) (Tu et al. 2001), sparse fusion of image
(SparseFI) (Zhu and Bamler 2013) and online coupled
X_ n ¼ D_ h  a_ n ; ð14Þ dictionary learning (OCDL) (Guo et al. 2014). The results
where the symbol  denotes quaternion multiplication. The are assessed visually and qualitatively at both full- and
proposed quaternion-based sparse model for pan-sharpen- reduced-scale resolution and are reported and discussed in
ing explores the efficiency of the quaternion analysis of the the forthcoming section.
sparse vector for pan-sharpening application. Though it
produces encouraging results, it is not anticipated to Quality Assessment of Cloud-Free Data in Full
dominate all the conventional methods in its performance. Resolution

This is the more practical method for validating the effi-

Experimental Results and Discussion ciency of the pan-sharpening algorithms, as it performs
image fusion mechanism directly on the available data at
This section analyzes the performance of the proposed hand. For the sake of visual evaluation, the results of the
method at both reduced and full scale and compares with aforementioned conventional algorithms and the proposed
that of the recently developed algorithms in the field. The methods are exhibited in Fig. 3. The pan-sharpened results
experimental simulation is conducted on IRS data. The of a small region are displayed in Fig. 4 for an enlarged
fusion is performed between the data from two IRS satellite view and better visual comparison.
sensors: Cartosat-2 with a spatial resolution of 0.65 m, The most popular quality metrics used for this validation
while its spectral range covers the interval of 0.5–0.85 lm is QNR (quality with no reference) index, which is the
and linear imaging self-scanning sensor (LISS-IV) MX product of spectral (Dk ) and spatial distortion (Ds), spatial
mode of ResourceSat-2 with a spatial resolution of 5.8 m. correlation coefficient (SCC) between the HRP image and
It is composed of three spectral bands: green the pan-sharpened product and finally the time (in seconds)
(0.52–0.59 lm), red (0.62–0.68 lm) and near-infrared taken to execute the algorithm. These values are reported in
(NIR) (0.77–0.86 lm). Spectral ranges of the IRS data are Table 1.
shown in Fig. 2. Three sets of experimental validation have Based on the analysis of the reported results, the pro-
posed pan-sharpening scheme of quaternion-based sparse

123
2074 Journal of the Indian Society of Remote Sensing (December 2018) 46(12):2069–2079

Fig. 3 Visual analysis of full-scale resolution assessment of IRS dataset a original LRMS from LISS-IV sensor, b original HRP from Cartosat-2,
c ATWT, d AWLP, e Brovey, f BDSD, g GSA, h PRACS, i IHS, j SparseFI, k OCDL, l PSQS (proposed method)

model (PSQS) achieves the best performance with good representation, the proposed PSQS gives lower Dk value
visual appearance and better quality metrics for most of the which indicates the decrease in the spectral distortion while
indexes with a reduced spatial distortion and a comparably maintaining higher spatial correlation.
better reduction in spectral distortion. ‘ The experiment is conducted on a i7 processor system
From the visual analysis, IHS, GSA, PRACS and the with MATLAB 14. The run time is mainly concentrated on
proposed method show better spatial features. These the compressed sensing approach, where OCDL is an
methods also have higher SCC values indicating higher iterative method and required longer time, while PSQS can
spatial correlation with respect to the panchromatic image, give better result quicker than OCDL.
but for IHS the spectral distortion is high. The proposed
PSQS method has a higher QNR index with lesser spatial Quality Assessment of Cloud-Free Data
and spectral distortion, while SparseFI and OCDL methods in Reduced Resolution
have better spatial correlation with good SCC value, but
greater spectral distortion with increased Dk value. This This section is devoted to describe the reduced resolution
spectral distortion is due to the lack of correlation among assessment of the pan-sharpening algorithms, where the
the spectral channels while constructing the dictionary. By available multispectral image is low-pass-filtered and
increasing the spectral channel correlation using quaternion degraded. The experiments are then conducted on the

123
Journal of the Indian Society of Remote Sensing (December 2018) 46(12):2069–2079 2075

Fig. 4 Visual analysis of full-scale resolution assessment of IRS dataset, a small region from a original LRMS, b original HRP, c ATWT,
d AWLP, e Brovey, f BDSD, g GSA, h PRACS, i IHS, j SparseFI, k OCDL, l PSQS (proposed method)

Table 1 Quantitative evaluation results of the experiments conducted Table 2 Quantitative evaluation results of the experiments conducted
at full scale at reduced scale
Dk Ds QNRI SCC Time (s) UIQI SAM ERGAS

ATWT 0.4391 0.3831 0.3460 0.8984 12.5300 ATWT 0.6964 3.1289 3.8011
AWLP 0.4089 0.3823 0.3652 0.8808 12.9174 AWLP 0.6922 3.2112 4.0740
Brovey 0.0885 0.0142 0.9755 0.7733 0.2706 Brovey 0.7285 1.9264 3.5620
BDSD 0.4554 0.7367 0.1434 0.1179 1.8079 BDSD 0.5427 4.8759 4.5304
GSA 0.0393 0.0119 0.9493 0.9099 5.9728 GSA 0.5320 5.7819 5.2505
PRACS 0.0678 0.1352 0.8062 0.9432 7.1877 PRACS 0.6833 1.8254 3.0987
IHS 0.5359 0.9417 0.0271 0.9870 0.1179 IHS 0.2227 10.7961 19.6370
SparseFI 0.5080 0.7458 0.1251 0.7625 2.1192 SparseFI 0.3451 9.9366 6.1576
OCDL 0.2758 0.5188 0.3485 0.7951 5.1352 OCDL 0.4708 7.3512 4.2419
Proposed method 0.0058 0.0102 0.9898 0.9843 2.1297 Proposed method 0.7319 1.7453 3.4551
Ideal value 0 0 1 1 Low Ideal value 1 0 0

degraded data, such that the original multispectral data will discussion as of full resolution holds good for that of the
serve as a reference for evaluating the performance of the reduced resolution. However, a more detailed analysis can
algorithms. The performance of the proposed method is be performed within the algorithms, the increased value in
compared with the aforementioned well-established algo- SAM for most of the methods including SparseFI indicates
rithms, and their corresponding quality is assessed using spectral distortion, which is reduced in PSQS as the cor-
the quality metrics spectral angle mapper (SAM) (Yuhas relation is introduced between the spectral channel, and in
et al. 1992) that measures the spectral distortion between PRACS the image is sharpened by substituting the sum of
the fused and the multispectral image, the global index for the weighted HRP and kth band of the LRMS image
pan-sharpening Erreur Relative Globale Adimensionnelle instead of replacing it by the intensity component, which
de Synthese (ERGAS) (Wald 2002) and universal image decreases the spectral distortion. In fact, the proposed
quality index (UIQI) or Q index (Wang 2002). Quantita- PSQS outperforms the conventional algorithm in most of
tively evaluated results are summarized in Table 2. the performance index. This is because of the fact that the
Next, the performance of the algorithms is discussed HR dictionary is constructed directly from both the HRP
based on the results of the quantitative measures. It is and LRMS images, which is used as a reference for vali-
evident that the reduced-scale resolution validation per- dating the quality of the pan-sharpened product.
forms in accordance with that of the full resolution. Similar

123
2076 Journal of the Indian Society of Remote Sensing (December 2018) 46(12):2069–2079

Table 3 Quantitative evaluation results for cloud removal application Quality Assessment of Cloud Removal
PSNR RMSE SSIM SAM Application

ATWT 28.6150 9.5780 0.8797 4.1628 Though few algorithms perform equivalently better, the
AWLP 28.6178 9.5475 0.8640 4.2548 proposed algorithm outperforms for pan-sharpening and
Brovey 26.6735 8.9779 0.8736 4.9614 cloud removal of cloudy multispectral data. Few of the
BDSD 25.3126 11.9292 0.2349 5.8457 pan-sharpening algorithms have found to be useful in the
GSA 25.7628 13.3404 0.8733 6.7748 cloud removal application, particularly the component
PRACS 28.6160 9.5668 0.6241 2.9231 substitution methods where the panchromatic data replace
IHS 40.9228 2.5728 0.9890 12.6984 the clouded multispectral information. But the advantage
SparseFI 28.3549 9.7534 0.8427 10.8573 of our approach is that a satisfactory cloud-free pan-
OCDL 28.3612 9.3826 0.8652 9.2851 sharpened multispectral data can be obtained even though
Proposed method 31.8756 6.7773 0.9822 2.6935 the amount of cloud cover is large. The accuracy of the
Ideal value High Low 1 0 experimental results is assessed by the well-known metrics:
peak signal-to-noise ratio (PSNR), root-mean-square error

Fig. 5 Visual analysis of full-scale resolution assessment of IRS e Brovey, f BDSD, g GSA, h PRACS, i IHS, j SparseFI, k OCDL,
dataset for cloud removal application a original LRMS from LISS-IV l PSQS (proposed method)
sensor, b original HRP from Cartosat-2, c ATWT, d AWLP,

123
Journal of the Indian Society of Remote Sensing (December 2018) 46(12):2069–2079 2077

Fig. 6 Visual analysis of full-scale resolution assessment of IRS dataset for cloud removal application, a small cloudless region from a original
LRMS, b original HRP, c ATWT d AWLP, e Brovey, f BDSD, g GS, h PRACS, i IHS, j SparseFI, k OCDL, l PSQS (proposed method)

Fig. 7 Visual analysis of full-scale resolution assessment of IRS dataset for cloud removal application, a small region from a original LRMS,
b original HRP, c ATWT, d AWLP, e Brovey, f BDSD, g GS, h PRACS, i IHS, j SparseFI, k OCDL, l PSQS (proposed method)

(RMSE), structural similarity (SSIM) and SAM. Table 3 effectively reconstruct the multispectral data with less
records the values of the accuracy assessment metrics. spectral distortion.
Though IHS seems to remove clouds with better PSNR The main limitation in cloud removal is the availability
and SSIM value, it has poor SAM value which leads to of multi-temporal multispectral images. But this proposed
high spectral distortion, since IHS is just mere replacement approach has been successful in reconstructing a cloud-free
of the intensity component by panchromatic data. Com- multispectral image with the availability of a panchromatic
pared to SparseFI and OCDL, the RMSE value is less for data of the same scene taken at same period of time.
PSQS as the reconstruction error is minimized by the
QFBP algorithm which has an advantage of improving its
support in each iteration by pruning the least contributing Conclusion
atom. This preserves the inherent structures in terms of
spatial and spectral details during the super-resolution This paper proposes a new pan-sharpening scheme based
reconstruction. It is evident from the reported data and on quaternion representation of sparse vectors and the
visual analysis of Figs. 5, 6 and 7 that the proposed method constructed dictionary. The novelty of the work lies in
can preferably remove the clouds and its shadow and increasing the correlation among the spectral channels in

123
2078 Journal of the Indian Society of Remote Sensing (December 2018) 46(12):2069–2079

dictionary construction using quaternion representation and ‘‘chromaticity’’ transformation techniques. Remote Sensing of
analysis for pan-sharpening application. The work also Environment, 22(3), 343–365.
Guo, M., Zhang, H., Li, J., Zhang, L., & Shen, H. (2014). An online
proposes a quaternion forward–backward pursuit algorithm coupled dictionary learning approach for remote sensing image
for estimating quaternion sparse coefficient. This algorithm fusion. IEEE Journal of Selected Topics in Applied Earth
has an advantage over OMP in improving the support Observations and Remote Sensing, 7(4), 1284–1294.
estimate in each iteration by removing the misrepresented Jiang, C., Zhang, H., Shen, H., & Zhang, L. (2012). A practical
compressed sensing-based pan-sharpening method. IEEE Geo-
atoms, thereby improving the reconstruction efficiency. science and Remote Sensing Letters, 9(4), 629–633.
The PSQS method explores the competence of the Jiang, C., Zhang, H., Shen, H., & Zhang, L. (2014). Two-step sparse
quaternion analysis for pan-sharpening approach, which coding for the pan-sharpening of remote sensing images. IEEE
preserves the interrelationship between the chromatic Journal of Selected Topics in Applied Earth Observations and
Remote Sensing, 7(5), 1792–1805.
channels by transforming them to the orthogonal color Joshi, M. V., Bruzzone, L., & Chaudhuri, S. (2006). A model-based
space. The efficacy of the proposed PSQS method is vali- approach to multiresolution fusion in remotely sensed images.
dated both at full- and reduced-scale resolution of IRS data. IEEE Transactions on Geoscience and Remote Sensing, 44(9),
The algorithm has also an advantage of removing the cloud 2549–2562.
Karahanoglu, N. B., & Erdogan, H. (2013). Compressed sensing
and its shadow from the pan-sharpened data. The superi- signal recovery via forward–backward pursuit. Digital Signal
ority of the approach over the well-established algorithms Processing, 23(5), 1539–1548.
in the field is reported and demonstrated by quantitative Laben, C. A., & Brower, B. V. (2000). U.S. Patent No. 6,011,875.
measurement metrics, which proves its potential as a Washington, DC: U.S. Patent and Trademark Office.
Li, S., & Yang, B. (2011). A new pan-sharpening method using a
powerful pan-sharpening tool for remotely sensed images. compressed sensing technique. IEEE Transactions on Geo-
science and Remote Sensing, 49(2), 738–746.
Li, S., Yin, H., & Fang, L. (2013). Remote sensing image fusion via
sparse representations over learned dictionaries. IEEE Transac-
References tions on Geoscience and Remote Sensing, 51(9), 4779–4789.
Li, Z., & Leung, H. (2009). Fusion of multispectral and panchromatic
images using a restoration-based method. IEEE Transactions on
Aiazzi, B., Alparone, L., Baronti, S., Garzelli, A., & Selva, M. (2003). Geoscience and Remote Sensing, 47(5), 1482–1491.
An MTF-based spectral distortion minimizing model for pan- Ranchin, T., & Wald, L. (2000). Fusion of high spatial and spectral
sharpening of very high resolution multispectral images of urban resolution images: The ARSIS concept and its implementation.
areas. In 2nd GRSS/ISPRS joint workshop on remote sensing and Photogrammetric Engineering and Remote Sensing, 66(1),
data fusion over urban areas, 2003 (pp. 90–94). IEEE. 49–61.
Aiazzi, B., Baronti, S., & Selva, M. (2007). Improving component Rao, C. V., Rao, J. M., Kumar, A. S., Jain, D. S., & Dadhwal, V. K.
substitution pansharpening through multivariate regression of (2016). High spatial and spectral details retention fusion and
MS ? Pan data. IEEE Transactions on Geoscience and Remote evaluation. Journal of the Indian Society of Remote Sensing,
Sensing, 45(10), 3230–3239. 44(2), 167–175.
Amolins, K., Zhang, Y., & Dare, P. (2007). Wavelet based image Tu, T. M., Su, S. C., Shyu, H. C., & Huang, P. S. (2001). A new look
fusion techniques—An introduction, review and comparison. at IHS-like image fusion methods. Information Fusion, 2(3),
ISPRS Journal of Photogrammetry and Remote Sensing, 62(4), 177–186.
249–263. Vicinanza, M. R., Restaino, R., Vivone, G., Dalla Mura, M., &
Chavez, P., Sides, S. C., & Anderson, J. A. (1991). Comparison of Chanussot, J. (2015). A pansharpening method based on the
three different methods to merge multiresolution and multispec- sparse representation of injected details. IEEE Geoscience and
tral data- Landsat TM and SPOT panchromatic. Photogrammet- Remote Sensing Letters, 12(1), 180–184.
ric Engineering and Remote Sensing, 57(3), 295–303. Vivone, G., Alparone, L., Chanussot, J., Dalla Mura, M., Garzelli, A.,
Cheng, M., Wang, C., & Li, J. (2014). Sparse representation based Licciardi, G. A., et al. (2015). A critical comparison among
pansharpening using trained dictionary. IEEE Geoscience and pansharpening algorithms. IEEE Transactions on Geoscience
Remote Sensing Letters, 11(1), 293–297. and Remote Sensing, 53(5), 2565–2586.
Choi, J., Yu, K., & Kim, Y. (2011). A new adaptive component- Wald, L. (2002). Data fusion: Definitions and architectures: Fusion
substitution-based satellite image fusion by using partial of images of different spatial resolutions. Paris: Presses des
replacement. IEEE Transactions on Geoscience and Remote MINES.
Sensing, 49(1), 295–309. Wang, Z., & Bovik, A. C. (2002). A universal quality index. IEEE
Garzelli, A., Nencini, F., Alparone, L., Aiazzi, B., & Baronti, S. Signal Processing Letters, 20, 1–4.
(2004). Pan-sharpening of multispectral images: a critical review Xu, Y., Yu, L., Xu, H., Zhang, H., & Nguyen, T. (2015). Vector
and comparison. In Geoscience and remote sensing symposium, sparse representation of color image using quaternion matrix
2004. IGARSS’04. Proceedings, 2004 IEEE International (Vol. analysis. IEEE Transactions on Image Processing, 24(4),
1). IEEE. 1315–1329.
Garzelli, A., Nencini, F., & Capobianco, L. (2008). Optimal MMSE Yu, M., Xu, Y., & Sun, P. (2014, May). Single color image super-
pan sharpening of very high resolution multispectral images. resolution using quaternion-based sparse representation. In 2014
IEEE Transactions on Geoscience and Remote Sensing, 46(1), IEEE International Conference on Acoustics, Speech and Signal
228–236. Processing (ICASSP) (pp. 5804–5808). IEEE.
Gillespie, A. R., Kahle, A. B., & Walker, R. E. (1987). Color Yuhas, R. H., Goetz, A. F., & Boardman, J. W. (1992). Discrimi-
enhancement of highly correlated images. II. Channel ratio and nation among semi-arid landscape endmembers using the
spectral angle mapper (SAM) algorithm.

123
Journal of the Indian Society of Remote Sensing (December 2018) 46(12):2069–2079 2079

Zhang, L., Shen, H., Gong, W., & Zhang, H. (2012). Zhu, X. X., & Bamler, R. (2013). A sparse image fusion algorithm
Adjustable model-based fusion method for multispectral and with application to pan-sharpening. IEEE Transactions on
panchromatic images. IEEE Transactions on Systems, Man, and Geoscience and Remote Sensing, 51(5), 2827–2836.
Cybernetics, Part B (Cybernetics), 42(6), 1693–1704.

123