IEEE Geoscience and Remote Sensing Magazine - June 2013 - 10

in hyperspectral image classification. Section V addresses
hyperspectral target detection. Section VI reviews the main
problems and methods in model inversion and estimation
of physical parameters, and finally Section VII outlines several strategies to accelerate the hyperspectral image computations using different hardware architectures.
II. Data FUsIon
In this section, we will discuss hyperspectral processing
techniques (image in-image out), that fuse spatial and
spectral information from one or multiple hyperspectral
observations, or a combination of hyperspectral images
and other image sources. We will refer to this processing as
data fusion. In Fig. 4, a schematic overview of the different
strategies is given.
A. RestoRAtion
Signal processing techniques can be applied to restore or
improve the signal-to-noise ratio (SNR) and/or the spatial
resolution. In the case of gray-scale images, many denoising and deconvolution techniques were developed to restore
SNR and spatial resolution. It is clear that a band-by-band
treatment of the restoration problem in HSIs would not
benefit from the high spectral redundancy. The traditional
image restoration techniques are extended to account for
this spectral redundancy. In this way, hyperspectral image
denoising techniques were recently developed by, e.g.,
employing spatial-spectral information [17] or employing
tensor decompositions and multilinear algebra [18]. In [19],
restoration of hyperspectral images was proposed based on
anisotropic diffusion filtering. Remark that all the above
mentioned methods preserve the original spatial and spectral sampling and thus do not improve the spatial resolution.

Restoration

B. spectRAl DAtA Fusion
Here, we discuss the fusion of spectral bands of an HSI, in
this way removing high spectral redundancy. Since the high
number of bands causes dimensionality problems, a dimensionality reduction of the hyperspectral vectors can highly
facilitate the analysis afterwards. The goal is to obtain an
image of reduced number of bands while trying to preserve
the most useful spectral information as possible. The simplest way is to select a few of the available bands, but it is
clear that better performance can be obtained when bands
are fused together. Traditionally, methods based on PCA
are applied that decorrelate bands. In many occasions, the
dimensionality reduction is applied for an improved classification afterwards. This topic is treated in Section IV
(IV.A.1 and IV.A.2).
A specific application of spectral data fusion is the visualization of HSIs. A user may need to visualize hyperspectral image data for exploration purposes, e.g., for generating
ground reference data. However, an HSI contains far more
image bands than can be displayed on a standard tristimulus display. By fusion of the spectral bands, an image of limited number of bands can be generated, e.g., a panchromatic
image or an RGB image; how to fuse preserving as much
information as possible is an issue. In [20], hyperspectral
images are linearly projected onto color matching basis
functions specifically designed as RGB primaries of a standard tristimulus display. A spatio-spectral approach allows
to retain spatial details as well, and often generally generates
high-contrast images. Spatio-spectral methods that were
developed use e.g. wavelet transforms to fuse multiresolution information of the image bands [21], Markov Random
Fields that model the spatial relationship between neighboring pixels [22] or constrained optimization to enforce
spatial smoothness [23]. In Fig. 5,
four different color visualizations
of an AVIRIS hyperspectral image
of 224 spectral bands are shown,
obtained by PCA and the methods of
[20]-[22] respectively.

Spatial-Spectral Data Fusion
(Single Frame Superresolution)

Spectral Data Fusion

Multisource Data Fusion

Spatial Data Fusion (Multiframe Superresolution)
FIGURE 4. A schematic overview of the five different hyperspectral data fusion methodologies.

c. spAtiAl DAtA Fusion (MultiFRAMe supeRResolution)
The term (geometric) superresolution (SR) refers to the enhancement
of the spatial resolution of imaging sensors by inferring information at the subpixel level. Subpixel
image information is for instance
available as subpixel shifts of multiple low-resolution observations
(multiframe SR). In practice, the
images are subsampled by dividing each pixel into m # m subpixels and interpolating the pixel gray
levels. Then, corresponding areas
between the multiple observations

ieee Geoscience and remote sensinG maGazine

june 2013

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - June 2013