IEEE Geoscience and Remote Sensing Magazine - December 2017 - 62

TABLE 6. THE CPU PROCESSING TIME IN SECONDS FOR DIFFERENT RESTORATION APPROACHES USED IN THE EXPERIMENT.
2-D WAVELET

3-D WAVELET

FOSRPDN

LRMR

NAILRMA

SPAWMARS

1.16

1.72

1.09

25.70

4.79

0.62

fully automatic version of an SVD-based sparse regularization with reduction [236] with rank selection. The MATLAB
codes for FOSRPDN and SPAWMARS are available online in
[275] and [276], respectively.
The blue line in Figure 10 indicates the noise levels, and,
therefore, the effectivity of the HSI restoration techniques
can be compared with respect to the noise levels. It can be
seen that 3-D wavelet restoration considerably improves
on conventional 2-D wavelet techniques. Also, FOSRPDN,
which is based on a 2-D wavelet and an FOSRP, outperforms
3-D wavelet restoration, which confirms the importance of
the spectral correlations. Finally, SPAWMARS, which utilizes
a wavelet-based low rank, outperforms the other techniques
used in this experiment. Note that Wavelab Fast (a fast wavelet toolbox), which is provided in [277], was considered for
the implementation of wavelet transforms. A Daubechies
wavelet with two and ten coefficients for spectral and spatial
bases, respectively, in five decomposition levels was used in
all the experiments. The central processing unit (CPU) processing time in seconds for different restoration approaches
used in the experiment confirms that SPAWMARS is the
most efficient method (Table 6). All the experiments in this
section were performed in MATLAB on a computer having an Intel Core i7-4710HQ CPU at 2.5 GHz, with 12 GB
of memory, and a 64-b operating system.
CHALLENGES IN HYPERSPECTRAL IMAGE
DENOISING AND IMAGE RESTORATION
HSI restoration and modeling face several future challenges.
1) The HSI model selection and noise parameter estimation need more attention. For instance, a model selection criterion that is not restricted to the Gaussian noise
model would be very useful for HSI modeling and restoration. The main advantage of a model selection criterion is that it provides an instrument to compare the restoration techniques without using a simulated (noisy)
HSI but rather using the observed HSI itself.
2) It is of interest to investigate the contributions of the various HSI restoration approaches, such as CD, unmixing,
and resolution enhancement, as preprocessing steps for
further HSI analysis.
3) In mixed-noise scenarios, investigating the dominant
noise type within HSIs should be considered.
4) HSI restoration approaches, which are computationally
efficient and parameter free (automatic), can simply be
used as a preprocessing step in real-world applications.
Fast computing techniques may be considered for the
62

swift implementation of HSI restoration approaches in
the future.
CHANGE DETECTION
In this section, we define a CD problem in a pair of bitemporal HSIs (those in the multitemporal domain can be addressed straightforwardly, pair by pair). Let X 1 and X 2 be
two HSIs acquired over the same geographical area at times
t 1 and t 2, respectively. The hyperdimensional difference
image X D, i.e., the SCVs, can be computed by subtracting
the bitemporal images pixel by pixel, i.e.,

X D = X 2 - X 1. (3)

Let X = {~ n, X c} be the set of all classes in X D, where ~ n
is the no-change class and X c = {~ C 1, ~ C 2, ..., ~ C k} be the
set of K possible change classes. The considered binary CD
problem can be formalized to separate the ~ n and X c classes without distinguishing different classes in X c ; the objective of the multiclass CD task is to detect the changed pixels
in X c and identify their classes in " ~ C 1, ~ C 2, ..., ~ C k ,.
Continuous satellite observation resulted in the acquisition of a large number of multitemporal remote-sensing
images. By analyzing these images, a better understanding
of the changes and evolutions on Earth's surface can be
gained. CD is a technique that enables the land cover changes occurring over a geographical area at different observation times to be identified [278]. In past decades, CD has
played important roles in various multitemporal remotesensing applications, such as urban sprawl analysis, disaster
loss evaluation, and forest and environmental monitoring
[279]-[281]. For optical remote-sensing images, CD based
on multispectral images has been intensively investigated
because of the availability of multispectral sensors onboard
the last-generation Earth observation satellites. With the
launch of new-generation Earth observation satellites carrying hyperspectral sensors, there are further opportunities to
implement CD in multitemporal HSIs.
Unlike multispectral images, detailed spectral sampling in HSIs allows the potential detection of more spectral variations, especially those with subtle, spectrally insignificant changes [282]. Compared with the abrupt changes
that present in coarse multispectral images, changes in
HSIs are more sophisticated, implicit, and structurally
complex. One emerging problem within this context is
the detection of small and subtle changes compared
to the large unchanging background in multitemporal
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

DECEMBER 2017

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - December 2017