Unmixing Approach for Hyperspectral Data
Resolution Enhancement Using High Resolution
Multispectral Image with Unknown Spectral
Response Function
Mohamed Amine Bendoumi
Shaanxi Provincial Key Laboratory of Information
Acquisition and Processing (IAP)
School of Electronics and Information
Northwestern Polytechnical University
Xi’an 710129, China
Mingyi He
Shaanxi Provincial Key Laboratory of Information
Acquisition and Processing (IAP)
School of Electronics and Information
Northwestern Polytechnical University
Xi’an 710129, China
Abstract—A novel algorithm based on a Spectral Mixture
Analysis (SMA) technique for enhancing the spatial resolution
of the hyperspectral (HS) image using high spatial-resolution
multispectral (MS) image is described. The proposed algorithm
deals with practical remote sensing situation, where the spectral
relationship between the observed high spatial-resolution MS
image and the estimated high spatial-resolution HS image is
unknown. The high-resolution hyperspectral image is recon-
structed based on the high-spectral information of low spatial-
resolution hyperspectral image represented by endmembers and
high-spatial information of high spatial-resolution multispectral
image represented by abundances. As a result, a SMA diagram
is developed, in which the unmixing process is performed on the
MS and HS images sequentially. The spatial spread transform
matrix of the sensor observation model is used to produce the
matched abundances of the low spatial-resolution HS image, in
order to unmix the later. Finally, we utilize a real HYDICE
data experiments to show the effectiveness of the proposed fusion
algorithm.
I. INTRODUCTION
The development of the remote sensing technology in-
creases continuously the number of the spectral bands from
multispectral (MS) to hyperspectral (HS) data. The Hyperspec-
tral image with finer spectral resolution is useful to identify
the different materials in the scene. Due to the sensor design,
the spatial resolution of the HS image is often lower than that
of MS image. In many applications, images containing both
high spectral and spatial resolution are required. A popular
technique which can overcome this problem is the image
fusion, which tends to merge the spatial and spectral features
in one single image. Different algorithms for the fusion of
MS and HS image have been developed during the past
decades. In which the wavelet based methods [1]-[2] are of
great interesting. However, the results strongly depend on
the spatial and spectral re-sampling methods. Using a high
resolution auxiliary sensor, a maximum a posteriori (MAP)
estimation method was proposed to estimate high spatial HS
image based on the observed HS and MS images [3]-[6]. The
stochastic mixing model (SMM) is utilized to estimate the
underlying spectral characteristics, and then a cost function
is designed to produce an optimized estimation of the high
resolution HS image. Working in the wavelet domain and
using the same estimation framework, the MAP estimation
method showed noise- resistant property [7]. However, the
result for the spatial resolution enhancement is limited to
several principal components (PCs) of the HS image. An-
other approach based on linear mixing model called coupled
nonnegative matrix factorization (CNMF) [8] was recently
proposed for the fusion of multispectral and hyperspectral
images, in which both low spatial resolution HS image and
high spatial resolution MS image are alternately unmixed by
NMF, and the high resolution hyperspectral image is produced
by the endmembers spectra of low resolution HS image and
the abundance maps of high resolution MS image. Overall,
the CNMF method has shown better results compared to the
MAP-SMM method. However, the determination of sensor
properties, such as the spectral response function and the
PSF is needed in the CNMF method. In contrast, an implicit
form of the maximum a posteriori estimation was derived,
where it doesnt require explicit information about the spectral
response function. Therefore, in this paper, a novel algorithm to
enhance the spatial resolution of the hyperspectral image using
a high resolution multispectral image based on linear spectral
unmixing algorithm is proposed, in which the knowledge of
the spectral relationship between the observed high spatial-
resolution MS image and the estimated high spatial-resolution
HS image is not required. In the proposed method, the high
spatial-resolution HS image is synthesized by integrating the
high spectral information contained in the endmembers of low
spatial-resolution HS image and the high spatial information
contained in the abundances of the high spatial-resolution MS
image. Only the spatial spread transform matrix is used to
match the unmixing outputs. Finally, the spatial and spec-
tral performances of the proposed algorithm are verified and
evaluated by comparing it with the CNMF and MAP-SMM
algorithms.
511
978-1-4673-6322-8/13/$31.00
c
2013 IEEE