Professional Documents
Culture Documents
net/publication/350639103
CITATIONS READS
8 742
6 authors, including:
All content following this page was uploaded by Bohan Liu on 11 August 2022.
Abstract
Purpose – This paper aims to demonstrate the principle and practical applications of hyperspectral object detection, carry out the problem we now
face and the possible solution. Also some challenges in this field are discussed.
Design/methodology/approach – First, the paper summarized the current research status of the hyperspectral techniques. Then, the paper
demonstrated the development of underwater hyperspectral techniques from three major aspects, which are UHI preprocess, unmixing and
applications. Finally, the paper presents a conclusion of applications of hyperspectral imaging and future research directions.
Findings – Various methods and scenarios for underwater object detection with hyperspectral imaging are compared, which include preprocessing,
unmixing and classification. A summary is made to demonstrate the application scope and results of different methods, which may play an
important role in the application of underwater hyperspectral object detection in the future.
Originality/value – This paper introduced several methods of hyperspectral image process, give out the conclusion of the advantages and
disadvantages of each method, then demonstrated the challenges we face and the possible way to deal with them.
Keywords Machine intelligence, Image processing
Paper type Literature review
This work was supported by the National Key Research and Development
The current issue and full text archive of this journal is available on Emerald Program of China under Grant 2018YFC0831000.
Insight at: https://www.emerald.com/insight/0260-2288.htm Contributions: Zhixin Wang and Peng Xu contributed equally to this work
and they are the co-first authors.
Figure 1 Diagram of hyperspectral image processing that transforming non-uniform noise into a Gaussian distribution
necessitated a principal component noise reduction method based
on an improved cumulative variance percentage. According to the
characteristics of high correlation between the hyperspectral spatial
domain and spectral domain data, Gao (1993) proposed a new
residual adjusted local standard deviation hyperspectral image
noise evaluation method. The basis of the abovementioned
hyperspectral noise reduction estimation method is that the
hyperspectral image and its noise are distributed independently,
and some studies have investigated the correlation between the
noise and the hyperspectral image to estimate the noise reduction
achieved with this method.
In addition to the two methods that we discussed above,
there are also some novel methods that have been proposed
recently. For example, Fekri-Ershad et al. (2018) proposed a
method for texture images, which is based on real word
spelling correction and LBP (local binary pattern). First, the
2. Hyperspectral images preprocessing pixels most similar to noisy desired pixels in terms of
textural features are generated using LBP; next, this method
The hyperspectral image preprocessing method is an
uses a two-gram algorithm to select the best of all
important component of hyperspectral image processing.
candidates. The results show that this approach can obtain a
An effective preprocessing method can reduce the adverse
high level of accuracy, a low blurring effect and a low
effects caused by irrelevant information and may provide a
computation cost.
more reliable data source for the subsequent analysis of
hyperspectral images.
2.2 Dimension reduction of hyperspectral images
The most notable feature of hyperspectral remote sensing is
2.1 Noise reduction of hyperspectral images
that its image data contain information from hundreds of
In an actual application scenario, due to poor heat dissipation
spectral bands. The approximate continuous characteristic
in the enclosed environment, noise may be present in the
spectral lines of different ground features make the spectral
captured image. As the resolution of the hyperspectral camera
characteristics more obvious, but at the same time, the
is generally not high, the noise may cause the target to
similarity and duplication of information cause data
be masked. Therefore, it is necessary to reduce the noise of the processing to be more difficult. High-dimensional data
hyperspectral image. We need to analyze and evaluate the processing has always been a major problem in the
image, filter the noise and keep the image feature information application of spectral remote sensing in real scenes.
while suppressing the noise. In recent years, there have been Improving point classes and detection accuracy, reducing
two commonly used noise reduction methods: spatial noise data volume, increasing efficiency and reducing dimensional
reduction (Othman and Qian, 2006) and spectral domain noise processing are particularly important for hyperspectral data.
reduction (Hangai et al., 2009). There are two methods for hyperspectral data dimensionality
The spatial noise reduction method is comparable to two- reduction, namely, the feature selection dimensionality
dimensional image noise reduction methods, which reduction method and the dimensionality reduction method
primarily include such methods as the geological average based on feature extraction.
method (Curran and Dungan, 1989), the local mean Spectral feature selection, also known as band selection, selects
method (Gao, 1993) and the optimization method based a portion of the bands from all the bands of the hyperspectral
on Gaussian waveform extraction. The spatial noise image data whose information can represent the main features of
reduction method can only be used to reduce the noise of a the original hyperspectral image to the maximum extent. Band
single image because the relationships among the image set selection is divided into wrapper-type and filter-type selection
are not considered. Thus, the spatial noise reduction methods according to the different objective functions. Figure 3
method is not useful for hyperspectral images. Thus, while shows the ground measurement of soil electrical conductivity
using the spatial noise reduction method, we also need to using regression [Figure 3(a)], instance-based [Figure 3(b)] and
perform spectral domain noise reduction. Figure 2 shows regression trees [Figure 3(c)]. The average result of seven
the images at 470.93 nm from the airborne visible infrared randomly generated band sets is shown in Figure 3(d), which is
imaging spectrometer (AVIRIS) data cube and the noisy generated as a baseline of comparison with Figures 3(a)–3(c).
data cube, as well as their corresponding spectral derivative This result shows that the best unsupervised outcome results in a
images. The average noise level is not visible in the noisy smaller prediction error.
signal in Figure 2(c), but the noise is clearly manifested in Wrapper-type feature selection uses classification accuracy as
the derivative domain in Figure 2(d). an indicator to judge the quality of the selected band. Through
Minimal noise separation is considered to be a common method wrapper-type feature selection, one can obtain a better feature
of noise reduction in the spectral domain. Huang et al. (2015) subset than the filter method, but it requires considerable
noted that this method can improve the signal-to-noise ratio by calculation time. Compared with the wrapper-type band
reducing uncorrelated noise in hyperspectral data. The problem selection, the filter type uses a different objective function and
Underwater object detection Sensor Review
Zhixin Wang et al.
Figure 2 Example of noise reduction, figure from Othman and Qian (2006)
the calculation complexity is low. These objective functions are retains more information but also reduces the calculation time. In
generally based on entropy (Samadzadegan and Partovi, 2010), addition, spectral feature extraction obtains some research results
graph theory (Bajcsy and Groves, 2004), separability (Su and in nonlinear processing. The representative algorithms of
Du, 2006), minimum estimated abundance covariance (Wu nonlinear feature extraction based on the kernel function include
et al., 2012), Fisher discrimination (Chang, 2013) and other kernel linear discriminant analysis (LDA) and kernel PCA.
structures. The filter band selection algorithm is relatively Padmavathi et al. (2010) proposed a method combining KPCA
simple and obtains different results for different applications in and scale-invariant feature transform (SIFT), which is well-
terms of accuracy and stability. In addition, in recent years, the suited for blur, illumination change and rotation of the input
band selection algorithm based on the intelligent search image. However, the theoretical proof of the working principle of
strategy has received extensive attention and successively, the kernel function is inadequate and belongs to implicit
genetic algorithms (Jing and Ning, 2009), clone selection (De mapping. Therefore, further research is needed in the
Castro and Von Zuben, 2000) and ant colony optimization construction of the kernel function and parameter selection. The
(Gao et al., 2016) have been applied to feature selection. nonlinear feature extraction method based on manifold learning
Ghamisi and Benediktsson (2014) proposed a band selection has been a heavily researched topic in recent years. This method
algorithm based on a genetic algorithm and particle swarm embeds the points in the high-dimensional data space with a low-
optimization, which can automatically select effective features. dimensional manifold to achieve the purpose of reducing the data
Spectral feature extraction is the mathematical transformation dimension. Although this method can better reflect the nonlinear
of the original data to achieve a reduction in hyperspectral data structure of the data, it often brings a greater computational
while enhancing features as much as possible. Spectral feature burden.
extraction can be divided into two types, namely, the linear There are also some other methods we can take advantage of
extraction method and the nonlinear extraction method. The to reduce dimensionality. For instance, Ershad and Hashemi
currently widely used feature selection extraction method is the (2011) proposed a new concept called dispelling classes
principal component analysis (PCA) dimensionality reduction gradually, which can be used to increase the separability of
method. Using this method, dimensionality reduction not only classes based on their labels. Wold (1966) proposed sparse
Underwater object detection Sensor Review
Zhixin Wang et al.
Figure 3 Example of band selection of hyperspectral images, figure from Bajcsy and Groves (2004)
generalized partial least squares (PLS) selection (PLS), which method provides us more space to choose the optimal method for
was originally developed to reduce the high-dimensional specific situations.
predictor space based on the construction of latent variables as
linear combinations of original predictors. 3. End-member unmixing of hyperspectral images
The research content of unmixing is focused on the unmixing
2.3 Summary of preprocess of hyperspectral images of hyperspectral images taken by remote sensing. At present,
In this section, we presented two major aspects (Figure 4) of underwater hyperspectral unmixing research has not been
hyperspectral image preprocessing. One aspect is noise reduction, carried out widely due to the lack of underwater detection
and the other is dimension reduction. In each aspect, we described equipment. Therefore, based on the unmixing of remote
several methods that are often used in most scenarios. Subjected to sensing images, this article briefly describes a future feasible
the calculation capability, it is difficult to perform a real-time in situ method for underwater hyperspectral image unmixing.
process; therefore, the data set was processed in the lab. This Spectral unmixing first requires the establishment of a
spectral mixing model. According to the spatial scale of
material mixing and physical distribution, spectral mixing can
Figure 4 Diagram of hyperspectral image preprocess method
be divided into linear mixing and nonlinear mixing modes.
Linear mixing assumes that the photons reaching the sensor
interact with only the end-member. Conversely, when the
target distribution scale is small, the photons interact with
multiple substances, resulting in nonlinear mixing.
The linear mixed model has simple modeling and clear physical
meaning. This model is a widely used mixed pixel decomposition
model and is suitable for applications on larger scales. For fine
spectral analysis or small target detection on a smaller scale, such
as when the distance from the camera is on only a meter scale, a
nonlinear hybrid model is required for unmixing.
where mi represents the ith end-member, h represents the noise Although prototype analysis solves the non-convex problem of
item in the image and si represents the proportion of the end- NMF, how to adjust the relaxation factor remains a problem in
member in the pixel. Under the conditions of equations (2) and the presence of high-mixed pixels. However, the objective
(3), all the pixels in the hyperspectral image are considered and function has convexity, which is suitable for end-member
equation (1) is rewritten as: extraction and helps to address the end-member extraction of
highly mixed pixels. Figure 5 shows an abundance result
X ¼ MS 1 H (4) example of prototype analysis, which demonstrates that better
results can be obtained by using prototype analysis compared to
X is the L P dimension, L is the spectral dimension that is NMF methods. As shown in Figures 5(g)–(i), the multi-end-
hyperspectral image band data, P represents the number of member unmixing method using kernel archetypal analysis
end-members in the image, M is an L k dimension end- (AA) obtains a more practical decomposition result than
member matrix, S is the abundance vector of X It is a k P unmixing based on a single terminal member. The conclusion
matrix and H is an L P matrix. is that the kernel AA can generate pure prototype/end members
Linear model unmixing mainly includes the following within the class, which is more robust to noise and outliers.
methods, namely, non-negative matrix factorization (Rajabi
3.1.3 Sparse unmixing
and Ghassemian, 2014), prototype analysis (Zhao et al., 2015)
The goal of sparse unmixing is to estimate the abundance of the
and sparse unmixing (Arngren et al., 2011).
end-members appearing in the spectral image on the basis of
3.1.1 Non-negative matrix factorization the known end-member spectrum library, provided that the
Given the observation spectrum X, the problem of solving the mixed pixels can be expressed as a linear combination of
end-member matrix M and the abundance matrix S is multiple pure-end-member spectra; therefore, selecting some
transformed into the inverse problem of equation (1). of the existing pure-end-member spectral information to
According to the principle of matrix decomposition, the approximate the mixed-pixel spectrum will inevitably lead to
observed spectral matrix can be disassembled into the product sparse abundance. The basic model of sparse unmixing is as
of the end-member matrix and the abundance matrix and meet follows (Xiaoyan et al., 2013):
the non-negativity requirements; therefore, nonnegative matrix 1
minw Y hW 2F 1 l W 1;1
factorization (NMF) is suitable for spectral unmixing. 2 (7)
However, as the actual number of end-members is considerably s:t: W 0
smaller than the number of bands, there are infinitely many
solutions for non-negative matrix factorization; therefore, the The advantage is that the end-member library is known, which
problem of non-unique solutions needs to be overcome. can reduce the impact of inaccurate end-member extraction on
The advantage of NMF is that it can handle the situation of the abundance matrix. However, as the external environment
high-mixed pixels, but its objective function is non-convex, and cannot be avoided when constructing the end-member, the
the solution is not unique, which easily produces false end end-member library has poor transportability.
members.
3.2 End-member unmixing based on a nonlinear mixing
3.1.2 Prototype analysis spectral model
Although NMF-based spectral unmixing can handle the Assuming that the target is nonlinearly mixed (Liu et al., 2009),
situation of high-mixed pixels to achieve a certain unmixing each pixel vector in the image can be described by equation (8)
effect, its non-negativity causes the method to easily fall into (LI et al., 2011):
local extreme solutions and it is easy to produce false end- y ¼ f ðM; aÞ 1 n (8)
member matrices. The spectral unmixing method of prototype
analysis can be used to address this issue (Zhang and Li, 2016). where M is the spectral vector corresponding to the p end-
Given Y ¼ ðy1 ; y2 ; ; yk Þ 2 RLK is the hyperspectral image members contained in the pixel, a is the abundance of each
under k pixels and L bands, the main target is to find N end-member in pixel point y, n is the additional noise term and
prototype samples, h ¼ ðu 1 ; u 2 ; ; u N Þ 2 RLN , the target f is the end-member matrix M and unknown nonlinear function
function can be described as: of abundance a.
Underwater object detection Sensor Review
Zhixin Wang et al.
Figure 5 Example of the result of unmixing abundance using prototype analysis: (a–c) result of MV-NMF; (d–f) result of AA; (g-i) result of kernel AA
(p = 2)
The nonlinear spectral decomposition of hyperspectral images Licciardi and Del Frate (2011) applied a multilayer perceptron
usually includes two steps, namely, end-member extraction and neural network to combine feature dimensionality reduction and
nonlinear spectral unmixing. First, the number of end- abundance estimation and proposed and evaluated a new method.
members contained in the hyperspectral image is determined; Liu et al. (2009) combined a self-organizing mapping neural
next, the spectral curve of each end-member is identified and network with fuzzy membership to achieve the decomposition of
finally, the percentage of each end-member is determined mixed pixels in remote sensing images.
through different nonlinear unmixing methods. These methods
include analysis methods based on neural networks, analysis 3.2.2 Nonlinear unmixing method based on kernel function
methods based on sum functions and nonlinear analysis The kernel function method is an effective nonlinear data
methods based on manifold learning. analysis method. The data with nonlinear relationships are
mapped to the high-dimensional feature space through
3.2.1 Unmixing method based on artificial neural network nonlinear mapping such that the data present a linear
Neural networks (Carpenter et al., 1999) have functions such as relationship in the feature space and are then processed by the
self-organization, self-learning and high fault tolerance; therefore, linear method.
they have unique advantages in solving complex nonlinear The following nonlinear mapping is considered (Broadwater
problems. Over the past decade, many neural network models and Banerjee, 2009):
have been used for the nonlinear decomposition of mixed pixels
(Broadwater et al., 2007). Atkinson et al. (1997) used the w : x 2 Rn ! w ðxÞ 2 CN (9)
multilayer perceptron method to decompose AVHRR (Advanced
Very High-Resolution Radiometer) data and confirmed that the The data are mapped from the input space R to the feature
method is better than the linear mixed pixel decomposition space N, and the nonlinear relationship is converted into a
method. Carpenter et al. (1991) used the neural network advanced linear relationship.
very high-resolution radiometer (AVHRR) model to identify the Definition of kernel function: For all x, z [ X, X Rn, if the
mixed type of trees in the forest, showing that the ARTMAP function k satisfies:
adaptive resonance theory map (ARTMAP) model is better than
the linear mixed model. Liu and Yang (2013) proposed the kðx; zÞ ¼ w ðxÞ w ðzÞ (10)
adaptive resonance theory mixture MAP (ART-MMAP) model
on the basis of ARTMAP and optimized the ARTMAP model. k is the kernel function and hs²i is the inner product.
Underwater object detection Sensor Review
Zhixin Wang et al.
Kernel function matrix: For the input space vector set Figure 6 Classification results using kernel-based linear spectral
S ¼ fx1 ; x2 ; ; xj g, the kernel function matrix K is defined as mixture analysis methods
an l l matrix and the matrix elements are:
Kij ¼ h w ðxi Þ w ðxj Þi ¼ kðxi ; xj Þ (11)
distance (Licciardi and Del Frate, 2011; Broadwater and Figure 7 SAM classification result of wreck transects
Banerjee, 2011), spectral angle mapper (SAM) and spectral
wave features (Liu et al., 2011). Considering the difference
method of class-based feature extraction and classification
efficiency, classification accuracy can be used to import
decision criteria.
4.1.1.1 Spectral angle mapper. Based on the limited category-
based reference vectors, the spectral angle between unlabeled
samples and reference samples of different categories is
calculated. The following formula is used to obtain the
minimum value to classify unlabeled samples into related
categories:
hx; ri
u ¼ cos1 (14)
jxj jrj expands the difference of between-class means and narrows the
within-class scatter value. The optimization process of v is as
where x is the unlabeled target vector and r is the reference follows:
vector. T
When the samples that are used for classification can be v Sb v
v ¼ arg max b ðv Þ ¼ arg max T (16)
linearly separated, the traditional SAM method works well. Liu v v v Sv v
et al. (2011) proposed a kernel-based SAM method to achieve
nonlinear separability. Both x and r values are mapped into a P
K
where Sb ¼ 1n nk ð m k m Þð m k m ÞT is the between-class
higher-dimensional space F, while x ! w ðxÞ; r ! w ðr Þ; k¼1
w P
K P
R ! F , equation (14) becomes: variance, Sv ¼ 1n m k Þðxi m k ÞT is the within-
i2Ik nk ðxi
k¼1
h w ðxÞ; w ðr Þi class variance, and m k and Ik are the sample mean and class
u ¼ cos1 (15)
j w ðxÞj j w ðr Þj index value, respectively.
Features can be extracted by LDA from vectors. When LDA
Figure 7 is a SAM classification result. As shown in the figure, is used on the image data, the image should first be converted
misclassification can be found in all instances. The two bottles into a unidirectional vector. This conversion may cause some
were successfully classified, and the wood classification failed problems, such as a large dimension of the scatter matrix, a high
with a very modest correct classification. calculation cost and a small sample size. Xu and Zhang (2010)
4.1.1.2 Linear discriminant analysis. LDA is a widely used proposed image-based LDA (IBLDA), which is one of the
method in feature extraction (Weinberger et al., 2010; Gu and typical methods that can be used to extract features effectively
Feng, 2013; Bao et al., 2013). In supervised classification, given and directly from the image.
a set with n labeled samplesfAi ; Bi ; yi gni¼1 , yi are class labels. Compared with the original method, IBLDA defines a
The standard LDA classifier defines Y = v T x to map the complex matrix C = A 1 iB to describe samples, and the
unlabeled spectrum x to another space. The mapped samples original matrix C can be converted into a complex vector Z with
can be linearly separated by the projection matrix v . LDA Z = CX.
Underwater object detection Sensor Review
Zhixin Wang et al.
The difference between-class mean values is expanded by the Kasapoglu and Ersoy (2007) proposed a method using
optimization process of IBLDA, while within-class scatter boundary vector detection in which more attention is given to
values are reduced, as shown in equation (17): the effective division of the feature space to obtain a higher
T accuracy of rare categories. Axelsson et al. (2013) proposed a
X Gb X method in which explosive substances were used in
X ¼ arg max b ðX Þ ¼ arg max (17)
X X X T Gv X hyperspectral images. These types of classification methods are
based on the hyperspectral image function and can be classified
L
P H accurately and quickly.
where Gb ¼ L1 Cl C Cl C is the between-class
l¼1 4.1.2 Classifier models based on spectral features
P
L P H
r j
scatter matrix, Gv ¼ rL
1
j¼1 Cl Cl Clj Cl is the An increasing number of classifier models have been
l¼1 successfully used in hyperspectral image classification, such as
within-class scatter matrix, C is the mean of all training statistical Gaussian (Yao et al., 2010) and Bayesian models
samples, C l denotes the mean of all the training samples of the (Mohamed and Farag, 2005), neural network models (Hou,
lth category and Clj denotes the jth training pattern of the lth 2011), support vector machines (Melgani and Bruzzone, 2004)
class. (SVM), kernel-based methods and random forests (Ham et al.,
The X that maximizes the objective function should be the 2005). The operation of random forest is to build a large
eigenvector corresponding to the largest eigenvalue. We choose number of decision trees during training and then output the
[X1, X2, . . ., Xt] as the initial value. class output mode of a single tree. All these models have
4.1.1.3 Principal component analysis and matrix-based princi- addressed the classification procedure from different angles
pal component analysis. As an effective statistical method for and have good performance.
reducing dimensions, orthogonal transformation is used by 4.1.2.1 Gaussian process for classification. (Yao et al., 2010),
PCA using orthogonal transformation to map the original discussed the Gaussian process method. Given that different
vector to the new orthogonal vector space (Plaza et al., 2009). categories of variables interact independently, the traditional
The image should be converted into vectors before we use PCA Gaussian process can be extended to infinite dimensions by this
to extract the feature from vectors, which is similar to LDA. In theory. Therefore, an assumption can be made that any finite
this case, Yang et al. (2004) proposed two-dimensional PCA subset of samples follows a Gaussian distribution.
(2DPCA), which is a simple image projection technology. Given a set D = {xi, yi}, i = 1,2,. . ., n, where yi is the label.
Predicting the label of the new test set based on category x by
Traditional 2DPCA is described as follows:
Given a training set X = [A1, A2,. . ., An], image A can be calculating the output probability p (y jD, x ) is the goal. It is
mapped into an m-dimensional vector Y by using an n- assumed that the probability is only related to the value
function f:
dimensional X. The discrimination index of X can be measured
by the total divergence of the projected feature vector. As a pðyi jfi Þ ¼ Uðyi fi Þ (21)
result, the trace of the covariance matrix can obtain the
discrimination index of X, as shown below: where U is Gaussian cumulative distribution function (CDF).
h i Thus:
T
fðX Þ ¼ tr ðSX Þ ¼ X T E ðA EAÞ ð A EAÞ X (18) ðz
1 x2
UðzÞ ¼ pffiffiffiffiffiffiffi e 2 dx (22)
We can change equation (18) into: 1 2p
Figure 8 Diagram of neural network model maximization but also error minimization. In d-dimensional
space, equation (25) describes the optimal hyperplane:
g ðxÞ ¼ v T x 1 b (25)
Figure 10 (A) Pseudocolor map of pipeline environments, (B) unsupervised classification result of (A)
4.2 Underwater biological object detection L1 and later machine learning was used to correlate health
Hyperspectral image analysis applied to underwater sampling conditions with hyperspectral reflection. The results show that
has been used to research marine life (Table 2). In some hyperspectral imaging has the potential for marine life health
studies, the pigment composition and absorption monitoring and environmental degradation assessment.
characteristics show relations to the reflection spectra.
Therefore, to verify whether hyperspectral imaging can be used 4.3 Seafloor properties detection
as a marine species identification and classification tool based With the development of society, minerals on land are gradually
on spectral characteristics, pigments were extracted from four consumed. Due to the increasing demand for mineral
species of benthic animals (spoonworm (Bonellia viridis), resources, there is currently a need to locate new deposits.
sponges (Isodictya palmata), Hymedesmia paupertas and There are additional mineral resources in the oceans for
Hymedesima sp.) and analyzed by liquid chromatography, humans to exploit, not limited to manganese nodules, cobalt-
liquid chromatography-mass spectrometry and nuclear rich crusts and polymetallic sulfides.
magnetic resonance. The results show a negative correlation Manganese nodules consist of iron oxides and manganese
between the absorption spectrum and hyperspectral reflection. oxides, which are often exposed or half-buried on the seafloor
Therefore, a spectral database can be established based on a 4,000 m–6,000 m (Margolis and Burns, 1976) below the
specific optical fingerprint for the automatic detection and surface. Cobalt-rich crusts contain abundant strategic cobalt,
classification of pigment organisms in the ocean. which is found 800 m–2,500 m (Hein et al., 2000) beneath the
The spectral characteristics of different types of coral algae surface. Polymetallic sulfide contains cuprum, lead, aurum,
were analyzed by Mogstad and Johnsen (2017) using a similar argentum and other metallic resources. This mineral is
method (Figure 11), and he identified coral algae pigments, primarily detected on the mid-ocean ridges, rift valley and
such as R-phycoerythrin and chlorophyll, by hydrothermal vents with a depth of approximately 1,500 m–
spectrophotometry, high-performance liquid chromatography 3,700 m (Herzig and Hannington, 1995) beneath the surface.
and hyperspectral imaging techniques. The results show that However, exploration for minerals in the ocean is highly difficult
due to the similarity of the reflection spectra, it is difficult to due to mining costs and the deep-sea environment. Such
distinguish among different types of coral algae. However, coral methods as in situ sampling, transient electromagnetics, seismic,
algae can still be distinguished as a whole from other objects by underwater photography sonar surveying and spectroscopy are
supervised classification in shallow water investigations. the most commonly used mineral exploration methods
Some research progress and efforts have also been made in (Malehmir and Bellefleur, 2009; Gehrmann et al., 2019; Wong
health monitoring and ecological assessment using et al., 2020; Schoening et al., 2017; White et al., 2006).
hyperspectral images. Letnes et al. (2019) exposed a cold-water In recent years, several research studies conducted at sea
coral sample to a solution of the toxic compound 2- have explored the distribution of seafloor deposits through
methylnaphthalene at concentrations of 0 mg L1 to 3.5 mg underwater hyperspectral imaging (Table 3).
Figure 11 (a) Unclassified transect with training sites (32 pixels within positioning and ROV attitude transformations. However, by
each circle). (b) Classification results of spectral angle mapper (SAM), (c) using the median spectrum method and fine geographic
classification results of minimum distance, (d) classification results of reference to create the reference spectrum, the negative effects
binary encoding, (e) classification results of spectral information can be effectively eliminated. Due to the lack of changes in
divergence (SID), (f) classification results of parallelepiped actual lighting conditions and inherent optical properties
(IOPs) of seawater close to the seafloor, the actual result is
pseudoreflectance data. Two supervised classification methods
are used; as a result, the SVM exhibits better performance than
the SAM method in detecting manganese nodules. Figure 12
shows the color image, the corrected pseudocolor image and
the classification results. Based on the classification results, the
coverage of manganese nodules in the survey area was
calculated. As part of the sediment was buried with many
manganese nodules, the coverage rate was estimated to be
between 1% and 9%.
To date, few reports on the use of hyperspectral images for
deep sea mineral exploration have been presented at
conferences and in journals. The research is still in the
exploratory stage, and more in-depth studies are needed.
Considering the advances that we have already made, this
technology should receive more attention. For actual deep-sea
mining, cost is a very important factor; therefore, the quality
and approximate reserves of the deposit must be found.
However, there are still some challenges to be solved in
hyperspectral imaging for mineral exploration. For example,
the ultraviolet and infrared bands are almost unavailable due to
the IOPs of seawater, but many minerals lack visible and
qualitative features in the visible band. More sensitive sensor
spectrum analysis methods (such as spectrum decomposition)
might represent a means of solving these problems.
Furthermore, different types and contents of optical fingerprint
databases of deposits should be established for accurate
detection and identification.
Hyperspectral images are used by Johnsen et al. (2013) to 4.4 Summary of applications
obtain optical fingerprints of manganese nodules, corals and In this section, we present three potential applications of
other samples, leading to the finding that manganese nodules underwater hyperspectral imaging (Figure 13). Through the
can be identified by their reflectance spectrum features. study of classification, object detection and other methods, we
According to different classification methods, various objects of found that it is still difficult to detect large areas of underwater
financial interest (such as multicomponent substrates targets only by hyperspectral analysis.
manganese nodules and rusted pieces of iron) could be Although there are some applications for underwater
identified by underwater hyperspectral images. hyperspectral imaging, its use is still relatively scarce. This
The first hyperspectral imaging survey of deep-sea imaging is rarely performed because while valuable objects, such
manganese nodules was conducted by Dumke et al. (2018a, as minerals, are located in the deep sea, detection and mining
2018b). The hyperspectral data were obtained by hyperspectral costs are high. However, it is conceivable that with the gradual
imaging through remotely operated vehicle (ROV) deployed in development of hyperspectral image processing technology and
the manganese nodule mining area of the Peru Basin in the its further integration with other technologies, such as acoustic
Southeast Pacific. Underwater hyperspectral images are technology, hyperspectral images will gradually become a
influenced by ROV movements, the accuracy of underwater powerful tool for human understanding of the underwater world.
Figure 12 (a) Underwater hyperspectral imaging radiance data in RGB (R: 645 nm, G: 571 nm, B: 473 nm), (b) adjusted pseudoreflectance data in
pseudocolor map, (c) SVM classification result of (b)
Figure 13 Potential applications of underwater hyperspectral imaging hyperspectral images is limited, resulting in some differences
in the application of hyperspectral imaging technology
underwater and in the air. In addition, due to the military and
economic strategic value of underwater hyperspectral
imaging technology, the sharing of technology in this field has
been greatly restricted.
Underwater hyperspectral imaging may provide a useful
method for observing and understanding the underwater
world. Although there are few reports on underwater
hyperspectral imaging technology, there are a number of
systems using this technology that are already in commercial
use. In the foreseeable future, underwater hyperspectral
5. Conclusion imaging may gradually become a more mature method for
Compared with traditional images, hyperspectral images can observing the underwater world.
provide optical images obtained in hundreds of bands. This
article introduces the use of hyperspectral in the underwater
environment from the three aspects of preprocessing
References
hyperspectral images, spectral unmixing and object detection Arngren, M., Schmidt, M.N. and Larsen, J. (2011),
and subsequently compares various existing methods. With “Unmixing of hyperspectral images using Bayesian non-
hundreds of spectral images available in different bands, negative matrix factorization with volume prior”, Journal
spectral and spatial details are contained in the hyperspectral of Signal Processing Systems, Vol. 65 No. 3, pp. 479-496.
image cubes, which provide researchers with a better Atkinson, P.M., Cutler, M. and Lewis, H. (1997), “Mapping
understanding of image information. New methods have been sub-pixel proportional land cover with AVHRR imagery”,
developed that include not only classification but also object International Journal of Remote Sensing, Vol. 18 No. 4,
detection to provide a more specific interpretation of image pp. 917-935.
information, and most of them have shown good performance. Axelsson, M., Friman, O., Johansson, I., Nordberg, M. and
However, the following challenges still exist. Östmark, H. (2013), “Detection and classification of explosive
In actual underwater applications, due to the complexity of substances in multi-spectral image sequences using linear
the underwater environment and the stability of the subspace matching”, 2013 IEEE International Conference on
underwater vehicle, the acquisition of underwater Acoustics, Speech and Signal Processing, IEEE, pp. 3492-3496.
Underwater object detection Sensor Review
Zhixin Wang et al.
Bajcsy, P. and Groves, P. (2004), “Methodology for “First hyperspectral imaging survey of the deep seafloor:
hyperspectral band selection”, Photogrammetric Engineering high-resolution mapping of manganese nodules”, Remote
& Remote Sensing, Vol. 70 No. 7, pp. 793-802. Sensing of Environment, Vol. 209, pp. 19-30.
Bao, J., Chi, M. and Benediktsson, J.A. (2013), “Spectral Dumke, I., Purser, A., Marcon, Y., Nornes, S.M., Johnsen, G.,
derivative features for classification of hyperspectral remote Ludvigsen, M. and Søreide, F. (2018b), “Underwater
sensing images: experimental evaluation”, IEEE Journal of hyperspectral imaging as an in situ taxonomic tool for deep-
Selected Topics in Applied Earth Observations and Remote sea megafauna”, Scientific Reports, Vol. 8 No. 1, pp. 1-11.
Sensing, Vol. 6 No. 2, pp. 594-601. Ershad, S.F. and Hashemi, S. (2011), “To increase quality of
Bazi, Y. and Melgani, F. (2006), “Toward an optimal SVM feature reduction approaches based on processing input
classification system for hyperspectral remote sensing datasets”, 2011 IEEE 3rd International Conference on
images”, IEEE Transactions on Geoscience and Remote Sensing, Communication Software and Networks, IEEE, pp. 367-371.
Vol. 44 No. 11, pp. 3374-3385. Fauvel, M., Tarabalka, Y., Benediktsson, J.A., Chanussot, J.
Broadwater, J. and Banerjee, A. (2009), “A comparison of and Tilton, J.C. (2012), “Advances in spectral-spatial
kernel functions for intimate mixture models”, 2009 First classification of hyperspectral images”, Proceedings of the
Workshop on Hyperspectral Image and Signal Processing: IEEE, Vol. 101 No. 3, pp. 652-675.
Evolution in Remote Sensing, IEEE, pp. 1-4. Fekri-Ershad, S., Fakhrahmad, S. and Tajeripour, F. (2018),
Broadwater, J. and Banerjee, A. (2010), “A generalized kernel “Impulse noise reduction for texture images using real word
for areal and intimate mixtures”, 2010 2nd Workshop on spelling correction algorithm and local binary patterns”,
Hyperspectral Image and Signal Processing: Evolution in Remote International Arab Journal of Information Technology, Vol. 15
Sensing, IEEE, pp. 1-4. No. 6, pp. 1024-1030.
Broadwater, J. and Banerjee, A. (2011), “Mapping intimate Foglini, F., Angeletti, L., Bracchi, V., Chimienti, G., Grande,
mixtures using an adaptive kernel-based technique”, 2011 V., Hansen, I.M., Meroni, A.N., Marchese, F., Mercorella,
3rd Workshop on Hyperspectral Image and Signal Processing: A. and Prampolini, M. (2018), “Underwater hyperspectral
Evolution in Remote Sensing (WHISPERS), IEEE, pp. 1-4. imaging for seafloor and benthic habitat mapping”, 2018
Broadwater, J., Chellappa, R., Banerjee, A. and Burlina, P.
IEEE International Workshop on Metrology for the Sea;
(2007), “Kernel fully constrained least squares abundance
Learning to Measure Sea Health Parameters (MetroSea),
estimates”, 2007 IEEE international geoscience and remote
IEEE, pp. 201-205.
sensing symposium, IEEE, pp. 4041-4044.
Foglini, F., Grande, V., Marchese, F., Bracchi, V.A.,
Carpenter, G.A., Gopal, S., Macomber, S., Martens, S. and
Prampolini, M., Angeletti, L., Castellan, G., Chimienti, G.,
Woodcock, C.E. (1999), “A neural network method for
Hansen, I.M. and Gudmundsen, M. (2019), “Application of
mixture estimation for vegetation mapping”, Remote Sensing
hyperspectral imaging to underwater habitat mapping,
of Environment, Vol. 70 No. 2, pp. 138-152.
Southern Adriatic sea”, Sensors, Vol. 19 No. 10, p. 2261
Carpenter, G.A., Grossberg, S. and Reynolds, J.H. (1991),
Gao, B. (1993), “An operational method for estimating signal
“ARTMAP: supervised real-time learning and classification
to noise ratios from data acquired with imaging
of nonstationary data by a self-organizing neural network”,
spectrometers”, Remote Sensing of Environment, Vol. 43
Neural Networks, Vol. 4 No. 5, pp. 565-588.
Chang, C. (2013), Hyperspectral Data Processing: algorithm No. 1, pp. 23-33.
Design and Analysis, John Wiley & Sons. Gao, L., Zhuang, L., Wu, Y., Sun, X. and Zhang, B. (2016),
Charalambous, C.D. and Logothetis, A. (2000), “Maximum “A quantitative and comparative analysis of different
likelihood parameter estimation from incomplete data via the preprocessing implementations of DPSO: a robust
sensitivity equations: the continuous-time case”, IEEE endmember extraction algorithm”, Soft Computing, Vol. 20
Transactions on Automatic Control, Vol. 45 No. 5, pp. 928-934. No. 12, pp. 4669-4683.
Cochrane, S., Ekehaug, S., Pettersen, R., Refit, E.C., Hansen, Gehrmann, R., North, L.J., Graber, S., Szitkar, F., Petersen,
I.M. and Aas, L. (2019), “Detection of deposited drill S., Minshull, T.A. and Murton, B.J. (2019), “Marine
cuttings on the sea floor-a comparison between underwater mineral exploration with controlled source electromagnetics
hyperspectral imagery and the human eye”, Marine Pollution at the TAG hydrothermal field, 26° N mid-atlantic ridge”,
Bulletin, Vol. 145, pp. 67-80. Geophysical Research Letters, Vol. 46 No. 11, pp. 5808-5816.
Curran, P.J. and Dungan, J.L. (1989), “Estimation of signal- Ghamisi, P. and Benediktsson, J.A. (2014), “Feature selection
to-noise: a new procedure applied to AVIRIS data”, IEEE based on hybridization of genetic algorithm and particle
Transactions on Geoscience and Remote Sensing, Vol. 27 No. 5, swarm optimization”, IEEE Geoscience and Remote Sensing
pp. 620-628. Letters, Vol. 12 No. 2, pp. 309-313.
De Castro, L.N. and Von Zuben, F.J. (2000), “The clonal Gu, Y. and Feng, K. (2013), “Optimized laplacian SVM with
selection algorithm with engineering applications”, distance metric learning for hyperspectral image
Proceedings of GECCO, pp. 36-39. classification”, IEEE Journal of Selected Topics in Applied
Ding, S. and Chen, L. (2009), “Classification of hyperspectral Earth Observations and Remote Sensing, Vol. 6 No. 3,
remote sensing images with support vector machines and pp. 1109-1117.
particle swarm optimization”, 2009 International Conference on Ham, J., Chen, Y., Crawford, M.M. and Ghosh, J. (2005),
Information Engineering and Computer Science, IEEE, pp. 1-5. “Investigation of the random Forest framework for
Dumke, I., Nornes, S.M., Purser, A., Marcon, Y., Ludvigsen, classification of hyperspectral data”, IEEE Transactions on
M., Ellefmo, S.L., Johnsen, G. and Søreide, F. (2018a), Geoscience and Remote Sensing, Vol. 43 No. 3, pp. 492-501.
Underwater object detection Sensor Review
Zhixin Wang et al.
Hangai, M., Yamamoto, M., Sakamoto, A. and Yoshimura, N. Liu, K., Wong, E., Du, E.Y., Chen, C.C. and Chang, C.
(2009), “Ultrahigh-resolution versus speckle noise- (2011), “Kernel-based linear spectral mixture analysis”,
reduction in spectral-domain optical coherence IEEE Geoscience and Remote Sensing Letters, Vol. 9 No. 1,
tomography”, Optics Express, Vol. 17 No. 5, pp. 4221-4235. pp. 129-133.
Hein, J.R., Koschinsky, A., Bau, M., Manheim, F.T., Kang, J. Liu, L., Wang, B. and Zhang, L. (2009), “Decomposition of
and Roberts, L. (2000), “Cobalt-rich ferromanganese crusts mixed pixels based on bayesian self-organizing map and
in the pacific”, Handbook of Marine Mineral Deposits, Vol. 18, gaussian mixture model”, Pattern Recognition Letters, Vol. 30
pp. 239-273. No. 9, pp. 820-826.
Herzig, P.M. and Hannington, M.D. (1995), “Polymetallic Liu, X. and Yang, C. (2013), “A kernel spectral angle mapper
massive sulfides at the modern seafloor a review”, Ore algorithm for remote sensing image classification”, 2013 6th
Geology Reviews, Vol. 10 No. 2, pp. 95-115. International Congress on Image and Signal Processing (CISP),
Hou, X. (2011), “Research of model of quantum learning IEEE, pp. 814-818.
vector quantization neural network”, Proceedings of 2011 Maes, W.H. and Steppe, K. (2019), “Perspectives for remote
International Conference on Electronic & Mechanical Engineering sensing with unmanned aerial vehicles in precision
and Information Technology, IEEE, pp. 3893-3896. agriculture”, Trends in Plant Science, Vol. 24 No. 2,
Huang, W., Xiao, L., Liu, H. and Wei, Z. (2015), pp. 152-164.
“Hyperspectral imagery super-resolution by compressive Malehmir, A. and Bellefleur, G. (2009), “3D seismic reflection
sensing inspired dictionary learning and spatial-spectral imaging of volcanic-hosted massive sulfide deposits: insights
regularization”, Sensors, Vol. 15 No. 1, pp. 2041-2058. from reprocessing halfmile lake data, New Brunswick,
Jiang, J., Zhang, J., Yang, G., Zhang, D. and Zhang, L. (2010), Canada”, Geophysics, Vol. 74 No. 6, pp. B209-B219.
“Application of back propagation neural network in the Manfreda, S., Mccabe, M.F., Miller, P.E., Lucas, R., Pajuelo
classification of high resolution remote sensing image: take Madrigal, V., Mallinis, G., Ben Dor, E., Helman, D., Estes,
remote sensing image of Beijing for instance”, in 2010 18th L. and Ciraolo, G. (2018), “On the use of unmanned aerial
International Conference on Geoinformatics, IEEE, pp. 1-6.
systems for environmental monitoring”, Remote Sensing,
Jing, F. and Ning, S. (2009), “Research of band selection based
Vol. 10 No. 4, p. 641
on improved genetic algorithm and support vector machine
Margolis, S.V. and Burns, R.G. (1976), “Pacific deep-sea
to select”, Journal of Wuhan University of Technology, No. 18,
manganese nodules: their distribution, composition, and
p. 31.
origin”, Annual Review of Earth and Planetary Sciences, Vol. 4
Johnsen, G., Volent, Z., Dierssen, H., Pettersen, R., Van
No. 1, pp. 229-263.
Ardelan, M., Søreide, F., Fearns, P., Ludvigsen, M. and
Melgani, F. and Bruzzone, L. (2004), “Classification of
Moline, M. (2013), “Underwater hyperspectral imagery to
hyperspectral remote sensing images with support vector
create biogeochemical maps of seafloor properties”, Subsea
machines”, IEEE Transactions on Geoscience and Remote
Optics and Imaging, Elsevier, pp. 508-540e.
Sensing, Vol. 42 No. 8, pp. 1778-1790.
Johnsen, G., Ludvigsen, M., Sørensen, A. and Aas, L.M.S.
Mogstad, A.A. and Johnsen, G. (2017), “Spectral
(2016), “The use of underwater hyperspectral imaging
deployed on remotely operated vehicles-methods and characteristics of coralline algae: a multi-instrumental
applications”, IFAC-PapersOnLine, Vol. 49 No. 23, approach, with emphasis on underwater hyperspectral
pp. 476-481. imaging”, Applied Optics, Vol. 56 No. 36, pp. 9957-9975.
Kasapoglu, N.G. and Ersoy, O.K. (2007), “Border vector Mohamed, R.M. and Farag, A.A. (2005), “Advanced
detection and adaptation for classification of multispectral algorithms for Bayesian classification in high dimensional
and hyperspectral remote sensing images”, IEEE spaces with applications in hyperspectral image
Transactions on Geoscience and Remote Sensing, Vol. 45 segmentation”, IEEE International Conference on Image
No. 12, pp. 3880-3893. Processing 2005, IEEE, pp. II-646.
Kettig, R.L. and Landgrebe, D.A. (1976), “Classification of Ødegård, Ø., Mogstad, A.A., Johnsen, G., Sørensen, A.J. and
multispectral image data by extraction and classification of Ludvigsen, M. (2018), “Underwater hyperspectral imaging:
homogeneous objects”, IEEE Transactions on Geoscience a new tool for marine archaeology”, Applied Optics, Vol. 57
Electronics, Vol. 14 No. 1, pp. 19-26. No. 12, pp. 3214-3223.
Letnes, P.A., Hansen, I.M., Aas, L.M.S., Eide, I., Pettersen, R., Othman, H. and Qian, S. (2006), “Noise reduction of
Tassara, L., Receveur, J., Le Floch, S., Guyomarch, J. and hyperspectral imagery using hybrid spatial-spectral
Camus, L. (2019), “Underwater hyperspectral classification of derivative-domain wavelet shrinkage”, IEEE Transactions on
deep sea corals exposed to 2-methylnaphthalene”, PloS One, Geoscience and Remote Sensing, Vol. 44 No. 2, pp. 397-408.
Vol. 14 No. 2, p. e0209960. Padmavathi, G., Muthukumar, M. and Thakur, S.K. (2010),
Li, X., Wu, X. and Zhao, L. (2011), “Unsupervised nonlinear “Kernel principal component analysis feature detection and
decomposing method of hyperspectral imagery”, Journal of classification for underwater images”, 2010 3rd International
Zhejiang University (Engineering Science), Vol. 45 No. 4, Congress on Image and Signal Processing, pp. 983-988.
pp. 607-613. Palsson, B., Sigurdsson, J., Sveinsson, J.R. and Ulfarsson, M.
Licciardi, G.A. and DEL Frate, F. (2011), “Pixel unmixing in O. (2018), “Hyperspectral unmixing using a neural network
hyperspectral data by means of neural networks”, IEEE autoencoder”, IEEE Access, Vol. 625, pp. 646-656.
Transactions on Geoscience and Remote Sensing, Vol. 49 Pettersen, R., Johnsen, G., Bruheim, P. and Andreassen, T.
No. 11, pp. 4163-4172. (2014), “Development of hyperspectral imaging as a bio-
Underwater object detection Sensor Review
Zhixin Wang et al.
optical taxonomic tool for pigmented marine organisms”, hydrothermal and cold seep systems (gorda ridge and hydrate
Organisms Diversity & Evolution, Vol. 14 No. 2, pp. 237-246. ridge)”, Geochemistry, Geophysics, Geosystems, Vol. 7 No. 5.
Plaza, A., Benediktsson, J.A., Boardman, J.W., Brazile, J., Wold, H. (1966), “Estimation of principal components and
Bruzzone, L., Camps-Valls, G., Chanussot, J., Fauvel, M., related models by iterative least squares”, Multivariate
Gamba, P. and Gualtieri, A. (2009), “Recent advances in Analysis, pp. 391-420.
techniques for hyperspectral image processing”, Remote Wong, L.J., Kalyan, B., Chitre, M. and Vishnu, H. (2020),
Sensing of Environment, Vol. 113, pp. S110-S122. “Acoustic assessment of polymetallic nodule abundance
Qiu, Q., Wu, X., Liu, Z., Tang, B., Zhao, Y., Wu, X., Zhu, H. using sidescan sonar and altimeter”, IEEE Journal of Oceanic
and Xin, Y. (2017), “Survey of supervised classification Engineering, Vol. 46 No. 1.
techniques for hyperspectral images”, Sensor Review, Vol. 37 Wu, C., Chen, H. and Chang, C. (2012), “Real-time N-finder
No. 3. processing algorithms for hyperspectral imagery”, Journal of
Rajabi, R. and Ghassemian, H. (2014), “Spectral unmixing of Real-Time Image Processing, Vol. 7 No. 2, pp. 105-129.
hyperspectral imagery using multilayer NMF”, IEEE Xiaoyan, T., Kun, G. and Guoqiang, N. (2013), “Advances in
Geoscience and Remote Sensing Letters, Vol. 12 No. 1, pp. 38-42. nonlinear spectral unmixing of hyperspectral images”,
Samadzadegan, F. and Partovi, T. (2010), “Feature selection Remote Sensing Technology and Application, Vol. 28 No. 4,
based on ant colony algorithm for hyperspectral remote pp. 731-738.
sensing images”, 2010 2nd Workshop on Hyperspectral Xu, Y. and Zhang, D. (2010), “Represent and fuse bimodal
Image and Signal Processing: Evolution in Remote Sensing, biometric images at the feature level: complex-matrix-based
IEEE, pp. 1-4. fusion scheme”, Optical Engineering, Vol. 49 No. 3,
Schoening, T., Jones, D.O. and Greinert, J. (2017), p. 037002.
“Compact-morphology-based poly-metallic nodule Yang, J., Zhang, D., Frangi, A.F. and Yang, J. (2004), “Two-
delineation”, Scientific Reports, Vol. 7 No. 1, pp. 1-12. dimensional PCA: a new approach to appearance-based face
Sture, Ø., Ludvigsen, M. and Aas, L.M.S. (2017), representation and recognition”, IEEE Transactions on Pattern
“Autonomous underwater vehicles as a platform for Analysis and Machine Intelligence, Vol. 26 No. 1, pp. 131-137.
Yao, F., Qian, Y., Hu, Z. and Li, J. (2010), “A novel
underwater hyperspectral imaging”, OCEANS 2017-
hyperspectral remote sensing images classification using
Aberdeen, IEEE, pp. 1-8.
gaussian processes with conditional random fields”, 2010
Su, H. and Du, P. (2006), “Study on feature selection and
IEEE International Conference on Intelligent Systems and
extraction of hyperspectral data”, Remote Sensing Technology
Knowledge Engineering, IEEE, pp. 197-202.
and Application, Vol. 21 No. 4, pp. 288-293.
Yu, J. (2010), “Texture image segmentation based on gaussian
Van der Meer, F.D., Van der Werff, H.M., Van Ruitenbeek, F.
mixture models and gray level co-occurrence matrix”, 2010
J., Hecker, C.A., Bakker, W.H., Noomen, M.F., Van Der
Third International Symposium on Information Science and
Meijde, M., Carranza, E.J.M., De Smeth, J.B. and Woldai,
Engineering, IEEE, pp. 149-152.
T. (2012), “Multi-and hyperspectral geologic remote
Yuan, J., Zhang, Y. and Gao, F. (2018), “An overview on linear
sensing: a review”, International Journal of Applied Earth
hyperspectral unmixing”, J. Infrared Millim. Waves, Vol. 37
Observation and Geoinformation, Vol. 14 No. 1, pp. 112-128. No. 5, pp. 553-571.
Wang, Q., Zhang, J., Chen, J. and Zhang, Y. (2011), “An Zhang, L. and Li, J. (2016), “Development and prospect of
improved spectral reflectance and derivative feature fusion sparse representation-based hyperspectral image processing
for hyperspectral image classification”, 2011 IEEE and analysis”, J. Remote Sens, Vol. 20 No. 5, pp. 1091-1101.
International Geoscience and Remote Sensing Symposium, Zhao, G., Jia, X. and Zhao, C. (2015), “Multiple endmembers
IEEE, pp. 1696-1699. based unmixing using archetypal analysis”, 2015 IEEE
Weinberger, K.Q., Sha, F. and Saul, L.K. (2010), “Convex International Geoscience and Remote Sensing Symposium
optimizations for distance metric learning and pattern (IGARSS), IEEE, pp. 5039-5042.
classification [applications corner]”, IEEE Signal Processing Zhou, R., Fan, Q. and Wei, M. (2011), “Solving for
Magazine, Vol. 27 No. 3, pp. 146-158. multimodal function with high dimensions base on
Werdell, P.J., Mckinna, L.I., Boss, E., Ackleson, S.G., Craig, hopfield neural network and immune algorithm”,
S.E., Gregg, W.W., Lee, Z., Maritorena, S., Roesler, C.S. Proceedings of 2011 International Conference on
and Rousseaux, C.S. (2018), “An overview of approaches Electronic & Mechanical Engineering and Information
and challenges for retrieving marine inherent optical Technology, IEEE, pp. 3905-3908.
properties from ocean color remote sensing”, Progress in
Oceanography, Vol. 160, pp. 186-212. Corresponding author
White, S.N., Dunk, R.M., Peltzer, E.T., Freeman, J.J. and Zhi Liu and Zhaojun Liu can be contacted at: liuzhi@sdu.
Brewer, P.G. (2006), “In situ raman analyses of deep-sea edu.cn; zhaojunliu@sdu.edu.cn
For instructions on how to order reprints of this article, please visit our website:
www.emeraldgrouppublishing.com/licensing/reprints.htm
Or contact us for further details: permissions@emeraldinsight.com