Research Article

Download as pdf or txt
Download as pdf or txt
You are on page 1of 13

Hindawi

Computational Intelligence and Neuroscience


Volume 2020, Article ID 8886932, 13 pages
https://doi.org/10.1155/2020/8886932

Research Article
Adoption of Machine Learning in Intelligent Terrain
Classification of Hyperspectral Remote Sensing Images

Yanyi Li,1,2 Jian Wang ,1 Tong Gao,3 Qiwen Sun,1 Liguo Zhang,4 and Mingxiu Tang1
1
College of Geomatics, Shandong University of Science and Technology, Qingdao 266500, China
2
Chinese Society of Optimization, Overall Planning and Economic Mathematics, Beijing 100089, China
3
College of Ocean Science and Engineering, Shandong University of Science and Technology, Qingdao 266500, China
4
Shandong Provincial Institute of Land Surveying and Mapping, Jinan 250001, China

Correspondence should be addressed to Jian Wang; 201701020822@sdust.edu.cn

Received 11 May 2020; Revised 15 July 2020; Accepted 10 August 2020; Published 1 September 2020

Academic Editor: Raşit Köker

Copyright © 2020 Yanyi Li et al. ,is is an open access article distributed under the Creative Commons Attribution License, which
permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

To overcome the difficulty of automating and intelligently classifying the ground features in remote-sensing hyperspectral images,
machine learning methods are gradually introduced into the process of remote-sensing imaging. First, the PaviaU, Botswana, and
Cuprite hyperspectral datasets are selected as research subjects in this study, and the objective is to process remote-sensing
hyperspectral images via machine learning to realize the automatic and intelligent classification of features. ,en, the basic
principles of the support vector machine (SVM) and extreme learning machine (ELM) classification algorithms are introduced,
and they are applied to the datasets. Next, by adjusting the parameter estimates using a restricted Boltzmann machine (RBM), a
new terrain classification model of hyperspectral images that is based on a deep belief network (DBN) is constructed. Next, the
SVM, ELM, and DBN classification algorithms for hyperspectral image terrain classification are analysed and compared in terms
of accuracy and consistency. ,e results demonstrate that the average detection accuracies of ELM on the three datasets are
89.54%, 96.14%, and 96.28%, and the Kappa coefficient values are 0.832, 0.963, and 0.924; the average detection accuracies of SVM
are 88.90%, 92.11%, and 91.68%, and the Kappa coefficient values are 0.768, 0.913, and 0.944; the average detection accuracies of
the DBN classification model are 92.36%, 97.31%, and 98.84%, and the Kappa coefficient values are 0.883, 0.944, and 0.972. ,e
results also demonstrate that the classification accuracy of the DBN algorithm exceeds those of the previous two methods because
it fully utilizes the spatial and spectral information of hyperspectral remote-sensing images. In summary, the DBN algorithm that
is proposed in this study has high application value in object classification for remote-sensing hyperspectral images.

1. Introduction classification of remote-sensing hyperspectral images has


important theoretical value and practical significance.
Remote-sensing hyperspectral technology is a comprehen- However, a hyperspectral image has high resolution and
sive new technology. Remote-sensing hyperspectral images large data volume; hence, hyperspectral data should be
can effectively retain the spatial and spectral information of detected using a more detailed method than those that are
ground objects. Object detection has important application applied to traditional multispectral images. Traditional
value in remote sensing, and the analysis of terrain changes machine learning methods, such as SVM, are commonly
can provide timely information regarding changes in large- used in the classification of hyperspectral remote-sensing
scale ground objects on the Earth surface [1, 2]. Remote- images [4]. Traditional machine learning methods often
sensing hyperspectral technology has been widely applied in require model training on a large amount of data, and the
agriculture, geology, and ecology [3]. Mastering the surface data that are used for the training must have similar dis-
object information is of substantial significance for im- tribution characteristics; hence, it is difficult to obtain
proving the surrounding environment. ,erefore, the training sample data for some machine learning methods
8483, 2020, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2020/8886932, Wiley Online Library on [21/07/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
2 Computational Intelligence and Neuroscience

[5, 6]. Deep learning is often applied in multiple fields. ,is hyperspectral images, and the fusion of multiscale structural
method can be used in the recognition of objects, behav- features insensitive to image noise was conducted by kernel
iours, and images, for example. Studies have shown that principal component analysis. ,e results present that the
deep learning algorithms can be used in the feature ex- method also has high robustness in the extraction of image
traction of remote-sensing image blocks [7]. However, structure features with intense noise [16]. Mishra et al.
relatively few applications of deep learning algorithms in the proposed a two-dimensional nonsecondary sampling
classification of hyperspectral remote-sensing images have wavelet transform method and applied it to the noise re-
been demonstrated. duction of hyperspectral images. ,e results illustrate that
even images with continuous noise of high wavelength can
2. Literature Review achieve automatic noise reduction [17]. Machine learning
can remove the noise effectively in hyperspectral images, but
2.1. Application of Machine Learning in Image Classification. the object detection in hyperspectral images has important
Machine learning technology can learn through training data, significance for the application of remote-sensing technol-
then finds the development trend of data from the diversified ogy. Zhou et al. proposed a method for hyperspectral image
database, and realizes the automatic processing of data classification that is based on the compact and discrimi-
analysis [8]. Machine learning has achieved excellent results native stacked autoencoder framework. After applying it to
in the processing of nonlinear data, such as image, text, and data classification, it is found that the method can effectively
voice, while deep learning technology in machine learning has classify ground objects in hyperspectral images [18]. Hang
a stronger advantage in image recognition. Garcia-Floriano et al. proposed a discrimination model, which is based on
et al. proposed a method for classification and recognition of recurrent neural networks, and applied it to the discrimi-
medical images that were based on support vector machines, nation of hyperspectral image learning features. ,e results
and the results presented that the method could be effectively illustrate that the model can extract spectral-spatial image
used in the diagnosis and classification of diseases [9]. features [19].
Sudharshan et al. conducted a classification of breast tumor Machine learning algorithms can realize high classifi-
biopsy images based on deep learning and found that this cation accuracy in image classification and recognition, but
method has high classification accuracy and does not require relatively few studies have been conducted on the classifi-
image labeling [10]. Now machine learning method is widely cation of ground objects in hyperspectral images. ,erefore,
used in medical image recognition, but it is also studied in a classification model that is based on SVM, ELM, DBN, and
hyperspectral image processing. Li et al. constructed a clas- the spectrum-space characteristics of remote-sensing
sification model of hyperspectral images that was based on hyperspectral images is proposed. It is applied to three
deep learning method to solve the shortcomings of traditional hyperspectral datasets, namely, PaviaU, Botswana, and
machine learning method [11]. Lv and Han proposed a Cuprite, and its accuracy is compared with those of various
method that was based on the multiple reduced kernel ex- classification models in the terrain classification of hyper-
treme learning machine, applied it to the efficient classifi- spectral image features. ,is study aims at providing a
cation of hyperspectral images, verified it with PaviaU and theoretical basis for increasing the efficiency of object rec-
other databases, and found that the model has a high clas- ognition in remote-sensing hyperspectral images for real-
sification effect [12]. Murphy and Maggioni proposed an izing intelligent object recognition.
unsupervised learning method for hyperspectral image
clustering based on spatial regularized random walk, which 3. Methodology
was found to have lower computational complexity by
marking [13]. At present, more experts and scholars have 3.1. Remote-Sensing Hyperspectral Image Segmentation Based
studied the application of machine learning and deep learning on Spectral-Spatial Characteristics. Different ground objects
methods in hyperspectral image classification, but there is no show different spectral characteristics and spatial distribu-
precision assessment of object classification with different tion characteristics; hence, it is necessary to identify and
recognition algorithms. judge image categories according to the information char-
acteristics and spatial distribution characteristics of terrain
spectral images [20]. Assuming that hyperspectral image
2.2. Processing of Remote-Sensing Hyperspectral Image. data ϖM×N×L constitute a cube (where M, N, and L represent
Remote-sensing hyperspectral images have been widely used the length, width, and band, respectively, of the data), the
in military, medical, and agricultural monitoring fields. In classification performance of the data depends mainly on the
the process of acquisition and transmission of hyperspectral image category, the dimension of the spectral data, the
images, they will be affected by illumination, atmosphere, number of samples that are used during training, the
and radiation; hence, there will be a lot of noise in hyper- classifier, and the classification method. Since the classifi-
spectral images, which will affect the credibility of image cation of hyperspectral data is similar to metaclassification, it
data and bring inconvenience to subsequent processing and can be followed from the whole variable space. A class of
analysis [14, 15]. ,erefore, much research is focused on the mean vectors is used to represent the coordinates in the
feature extraction of noise in hyperspectral images. Duan eigenspace. ,e data are classified by using a classification
et al. proposed a multiscale total variation method, which function to divide the region. ,e classification process of
was applied to the extraction of structural features in hyperspectral images is illustrated in Figure 1.
8483, 2020, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2020/8886932, Wiley Online Library on [21/07/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Computational Intelligence and Neuroscience 3

Image acquisition Classification accuracy Input

Preprocessing Classification results


Remote sensing image
Supplementary
Selection of samples Classification processing
information

Feature recognition Feature extraction Spectrum Dimensionality reduction

Figure 1: Classification process of remote-sensing hyperspectral Classification graph Segmentation graph


images.

As illustrated in Figure 1, the classification process of Spectral-spatial classification


remote-sensing hyperspectral images can be divided into the
following steps: (I) Image acquisition: the data from remote-
Output
sensing hyperspectral image databases being mainly used,
namely, the University of Pavia (PaviaU) dataset, Botswana Figure 2: Spectral-spatial classification of remote-sensing hyper-
dataset, and Cuprite dataset; (II) image preprocessing, such spectral images.
as geometric calibration and atmospheric correction; (III)
sample selection; (IV) recognition of features in an image;
Commonly used kernel functions include polynomial
(V) feature extraction in image; (VI) classical disposal: in
functions [(xi , xj ) + a]q , Sigmoid function tanh[v(xi ,
this study, SVM, ELM, and the deep learning algorithm
xj ) + c], redial basis kernel function exp(− (|x − xi |2 )/σ 2 ),
being used to classify hyperspectral images; (VII) classifi-
B-spline kernel function B2N+1 [v|x − xi |], and Fourier
cation results; and (VIII) classification accuracy. ,e com-
function ((sin(N + (1/2))(xi − xj ))/sin(1/2)(xi − xj )).
monly used classification accuracy evaluation methods
In order to address the problem of generalization,
include overall classification accuracy, average classification
penalty coefficient and relaxation factor are introduced to
accuracy, and Kappa value.
correct the SVM classification results.
During the pretreatment of remote-sensing hyper-
ELM is widely used to solve various nonlinear problems
spectral images, a method that is based on watershed and
due to its specific characteristics. Based on the ELM
spatial regularization is mainly used to segment images. ,e
structure, ELM is mainly a feedforward neural network with
spectral-spatial model classification framework is illustrated
a single hidden layer composed of an input layer, hidden
in Figure 2.
layer, and output layer [22]. Assuming the random sample
size is N, then the following equation can be obtained:
3.2. Brief Introduction to SVM and ELM. When using SVM
⎨ Xi � 􏼂xi1 , xi2 , . . . , xin 􏼃T ∈ Rn ,

to solve nonlinear problems, it is necessary to select a (4)
⎩ T
suitable kernel function and to map the samples in a low- ti � 􏼂ti1 , ti2 , . . . , tim 􏼃 ∈ Rm .
dimensional space to a special space in a high-dimensional
space. ,e optimal solution of the hyperplane is calculated in ,en, the expression of neural network with a single
this space [21]. When nonlinear problem is solved by SVM, hidden layer is as follows:
the expression of the nonlinear mapping is as follows: L
x ⟶ φ(x). (1) oj � 􏽘 βi g􏼐Wi Xj + bi 􏼑, j � 1, 2, . . . , N. (5)
i�1
,e above equation can be converted into the following
Among them, g(x) is an activation function; β is the
equation:
output weight of each component of the hidden layer; bi is
n
1 n the bias of each component of the ith hidden layer; and Wi is
Q(α) � 􏽘 αi + 􏽘 α α y y φ xi 􏼁φ􏼐xj 􏼑. (2) the input weight of each component of the ith hidden layer.
i�1
2 i,j�1 i j i j

Among them, φ(xi )φ(xj ) is the inner computing.


Nonlinear mapping can be used to solve nonlinear 3.3. Image Classification Based on DBN. In the case of a great
problems, but it increases the difficulty. ,erefore, instead of number of samples, unsupervised learning method gradually
kernel computation in a particular space, the space function becomes an operational approach to machine learning. RBM
K can be input: is an unsupervised mapping learning method, which in-
cludes the input layer and hidden layer, and the connection
K􏼐xi , xj 􏼑 � φ xi 􏼁φ􏼐xj 􏼑. (3) between them is a full connection [23]. ,ere is a connection
weight between any two nodes in RBM. If the number of the
Among them, K(xi , xj ) is the kernel function. hidden layer nodes in RBM is N and the number of input
8483, 2020, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2020/8886932, Wiley Online Library on [21/07/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
4 Computational Intelligence and Neuroscience

layer nodes is M, then the probability of activation of the ...


hidden layer node nj is as follows: Output layer

⎝b + 􏽘 W m ⎞ ⎠ ...
p􏼐nj | m􏼑 � σ ⎛ j i,j i , i � 1, 2, . . . , M, j � 1, 2, . . . , N. Hidden layer 4
i RBM 4

(6)
Hidden layer 3 ...
Among them, σ is the activation function. ,en, the
probability of the hidden layer, the input layer, and node mi, RBM 3
which are activated, is as follows:
Hidden layer 2 ...

⎧ N

⎪ p(n | m) � 􏽑 p􏼐nj | m􏼑, RBM 2



⎪ j�1



⎪ Hidden layer 1 ...

⎝c + 􏽘 W n ⎞
p m i | n􏼁 � σ ⎛ ⎠ (7)

⎪ j i,j j , RBM 1

⎪ i

⎪ Input layer ...


⎪ M
⎪ p(m | n) � 􏽑 p m | n􏼁.

⎩ i
i�1 Figure 3: ,e basic structure of DBN.

RBM training process is mainly divided into the fol-


lowing steps: (I) the data are input into the input layer, and of RBM1 are obtained, and RBM1 is used as the visible layer
the probability that the hidden layer and the input layer are to train RBM2 via the same approach. (III) Similarly, all
activated is calculated by using equation (7). (II) After RBMs in DBN are obtained, and the initial parameter value
obtaining the distribution of each node in the hidden layer, of DBN is obtained after completion. ,en, the network
Gibbs sampling method is used to extract the sample nj in parameters are optimized. (IV) ,e contrastive wake-sleep
the hidden layer. (III) ,e sample nj is used to reconstruct algorithm is used to optimize and generate DBN, and the BP
the input layer, and equation (7) is used to calculate the algorithm is used to optimize and discriminate DBN. (V)
probability of the input layer being activated. (IV) After When the parameters are optimized by the BP algorithm, if
obtaining the different conditions of the reconstructed input the error between the actual value and the expected value of
layer nodes, Gibbs sampling method is used to extract mj the output does not satisfy the requirements, back-
from the input layer samples. (V) After the reverse calcu- propagation is conducted. ,e stochastic gradient descent
lation, the activation probability and distribution probability method is used to correct the reverse parameters. When the
of the hidden layer are obtained again. (VI) w + λ(p number of iterations reaches the maximum and the target
(nj | mi )mi − p(nj′ | m′i)) ⟶ W is used to update the net- data have been obtained, the training is complete.
work weight, where λ represents the learning rate. ,e basic framework of DBN-based terrain classification
In this study, a single layer RBM that contains 50, 100, method for remote-sensing hyperspectral images, which is
150, 200, 250, and 300 hidden layer nodes is constructed, and constructed in this study, is illustrated in Figure 4.
the effects of the number of nodes on the spectral recon- It is concluded from Figure 4 that the DBN-based terrain
struction performance and the classification accuracy are classification framework for remote-sensing hyperspectral
compared. ,en, the number of unsupervised iterations is images that is constructed in this study contains two layers of
set as 50, 100, 200, 300, and 400 to evaluate the impact of the DBN, and the outermost layer of DBN is connected with a
number of iterations on the classification accuracy. ,e Softmax classifier. ,e Softmax classification layer optimizes
learning rate in RBM is set as 0.01, 0.05, 0.1, 0.15, 0.3, and the parameters in DBN via the BP method, and it can fa-
0.45, and the performances at these learning rates are cilitate the direct output of the image category label.
compared in terms of the classification accuracy.
,e optimal RBM parameter is selected and DBN is built. 4. Experiments and Results
DBN is composed of a multilayer RBM structure, and the
training method of DBN is layer-by-layer training of RBM ,e total number of samples that are used for model training
[24]. ,e basic structure of the DBN constructed based on in the PaviaU, Botswana, and Cuprite databases is 3000, and
RBM in this study is presented in Figure 3. the number of samples for testing is 1000. When evaluating
As illustrated in Figure 3, the classic DBN contains an the model classification performance, the CPU is Intel
input layer, a hidden layer, and an output layer. ,e i5− 3470, dual-core, and 4GB of memory. ELM classification
structure contains four hidden layers and four RBM is realized in LibELM open interface. ,e SVM classification
structures. In this study, the training methods for DBN are is C++ version. And DBN is MATLAB version. In order to
mainly divided into the following steps: (I) ,e data that better evaluate the effects of different methods on the
must be trained are input into RBM1, and the training of classification of hyperspectral remote-sensing images,
DBN that is based on RBM is conducted using the RBM qualitative and quantitative evaluation methods are selected
training method. (II) After the RBM training, the parameters to evaluate the classification results. ,e quantitative
8483, 2020, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2020/8886932, Wiley Online Library on [21/07/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Computational Intelligence and Neuroscience 5

...
...

...
...

...

...

...
...
...
...

...
...

...

...
Spectral
Image vector Softmax

DBN
Figure 4: Classification framework of hyperspectral remote-sensing image based on DBN.

evaluation indices include the time of classification use, the (8.27%), and bare soil (2.92%). In addition, the spatial
overall accuracy of classification, the average accuracy of resolution of the information is 30 m.
classification, and Kappa coefficient. ,e spectral characteristics of features are compared, and
the reflectance is output once every 5 bands. According to
Figure 8, features of ground objects such as water, nettle
4.1. Data Processing and Spectral Curve Analysis. ,e PaviaU grass, and bare soil in the Botswana remote-sensing image
dataset is a remote-sensing hyperspectral image dataset that data vary substantially in the visible-light shortwave infrared
was collected by the university of Pavia in Italy in 2002 based region, while the spectral curves of Robinia pseudoacacia
on ROSIS sensor, which contains 115 spectral bands with a forest, Robinia shrub, and Robinia pseudoacacia are not
wavelength range of 0.43 ∼ 0.86 μm. ,e size of the dataset is easily distinguished nor are the ground object categories,
610 ∗ 340 pixels, and according to Figure 5, the image data such as Brassica oleifera and Robinia shrub.
contain mainly information on 9 land types: asphalt road ,e Cuprite dataset consists of AVIRIS hyperspectral
(15.50%), grassland (43.60%), sand grain (4.91%), trees image data that were obtained by the United States Geo-
(7.16%), sheet metal (3.14%), bare soil (11.76%), asphalt roof logical Survey in 1995. ,ere are 50 spectral bands in the
(3.11%), floor tile (8.61%), and shadow (2.21%). ,e spatial wavelength range of 1.99 ∼ 2.48 μm in the image data. ,e
resolution of the information is approximately 1.3 m. size of the dataset is 350 ∗ 400 pixels. As shown in Figure 9,
,e spectral characteristics of ground objects are com- there are 8 main types of land information in this dataset:
pared, and the reflectance is output once for every 5 bands. muscovite (8.04%), muscovite + chlorite (11.87%), tuff
According to Figure 6, metal sheets, trees, grassland, and (4.21%), opal (31.97%), dickite (7.60%), kaolinite (22.43%),
sand grain in the remote-sensing image data set of PaviaU alunite (3.39%), feldspar (10.49%), muscovite (8.04%),
show large differences in the reflectance spectra of ground muscovite + chlorite (11.87%), tuff (4.21%), opal (31.97%),
objects in the visible and near-infrared bands. ,e reflec- dickite (7.60%), kaolinite (22.43%), alunite (3.39%), and
tance patterns of the bare soil and sand grain categories are feldspar (10.49%). ,e space rate of the information is ap-
highly similar. Only a small difference is observed in the red- proximately 20 m.
near-infrared band. ,e spectral characteristics of features are compared, and
,e concentration dataset consists of image data of the the reflectance is output once for every 5 bands. According
Botswana delta that were collected in 2001 using a Hyperion to Figure 10, the spectral characteristics of ground objects
EO-1 sensor, which senses 145 spectral bands with a such as opal and alunite in the dataset differ significantly in
wavelength range of 0.4 ∼ 2.5 μm. ,e size of the dataset is the range of the shortwave infrared region, whereas the
1476 ∗ 256 pixels. Figure 7 shows that the dataset consists spectral characteristics of kaolinite, tuff, and other ground
mainly of 14 types of terrain information: water (8.31%), objects are highly similar.
nettle grass (3.09%), flood plain grassland 1(7.74%), flood
plain grassland 2(6.63%), reed (8.27%), riverside (8.27%),
cliff (7.98%), island (6.26%), Robinia pseudoacacia forest 4.2. Influence of the Parameter Settings on the Classification
(9.67%), Robinia shrub (7.65%), Robinia pseudoacacia Accuracy of the DBN Model. In this study, the constructed
(9.38%), Brassica oleifera (5.56%), mixed bean wood DBN model is used for spectral reconstruction of interior
8483, 2020, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2020/8886932, Wiley Online Library on [21/07/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
6 Computational Intelligence and Neuroscience

Original Ground truth

Grassland (18649, 43.60%) Sand grain (2099, 4.91%)


Asphalt (6631, 15.50%) Metal sheets (1345, 3.14%)
Bare soil (5029, 11.76%) Asphalt roof (1330, 3.11%)
Floor tile (3682, 8.61%) Shadows (947, 2.21%)
Trees (3064, 7.16%)
Figure 5: PaviaU data image and real ground objects map.

5000

4000
Reflectivity

3000

2000

1000

0
0 20 40 60 80 100 120
Number of bands

Meadows Asphalt
Soil Bricks
Trees Gravel
Sheets Biteumen
Shadows
Figure 6: ,e spectral curve of each category of ground object in the PaviaU data image.

and boundary points of remote-sensing images. When the Figure 11, the spectral reconstruction performance of DBN
number of hidden layer nodes in DBN model is 200, the model on interior points of ground objects is higher than
image in the PaviaU database is identified. According to that on boundary points of ground objects. ,erefore, the
8483, 2020, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2020/8886932, Wiley Online Library on [21/07/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Computational Intelligence and Neuroscience 7

Original

Ground truth

Robinia pseudoacacia forest (314, 9.67%) Floodplain 1 (251, 7.74%)


Robinia pseudoacacia (305, 9.38%) Robinia shrub (248, 7.65%)
Water (270, 8.31%) Floodplain 2 (215, 6.63%)
Riverside (269, 8.27%) Island (203, 6.26%)
Reed (269, 8.27%) Brassica oleifera (181, 5.56%)
Mixed bean wood (268, 8.27%) Nettle (101, 3.09%)
Cliff (259, 7.98%) Soil (95, 2.92%)
Figure 7: Botswana data images and the real ground object map.

7000 concluded from Figure 12 that when the number of the


6000
hidden layer nodes in the network is 200, the recognition
accuracy is the highest. ,e recognition accuracies for im-
5000 ages in the PaviaU, Botswana, and Cuprite databases are
91.93%, 97.59%, and 98.73%, respectively. ,erefore, in this
Reflectivity

4000
study, the number of the hidden layer nodes in the DBN
3000 model is set to 200 for subsequent experiments.
2000 ,e effect of the number of unsupervised training it-
erations on the accuracy of DBN model recognition is
1000 evaluated. As shown in Figure 13, when the number of
0 unsupervised training iterations is 100, the recognition
0 20 40 60 80 100 120 140 accuracies of images in the PaviaU, Botswana, and Cuprite
Number of bands databases are the lowest, namely, 90.14%, 92.87%, and
91.37%, respectively. When the number of unsupervised
Robinia pseudoacacia forest Brassica oleifera training iterations is 300, the recognition accuracies of
Robinia pseudoacacia Riverside images in the PaviaU, Botswana, and Cuprite databases are
Water Mixed bean wood the highest, namely, 91.99%, 96.88%, and 98.41%, respec-
Reed Floodplain 1 tively. ,erefore, in this study, the number of unsupervised
Cliff Soil training iterations of the DBN model is set at 300 for
Robinia shrub Island
subsequent experiments.
Floodplain 2 Nettle
,en, the influence of the learning rate on the recog-
Figure 8: ,e spectral curve of each category of ground objects in nition accuracy of the DBN model is evaluated. As presented
the Botswana data image. in Figure 14, when the learning rate is 0.01, the recognition
accuracies of images in the PaviaU, Botswana, and Cuprite
interior feature points of ground objects in remote-sensing databases are the lowest, namely, 89.97%, 94.80%, and
hyperspectral images are selected for the final experiments. 94.75%, respectively. When the learning rate is 0.15, the
,e learning rate is set at 0.01, and the number of un- recognition accuracies of images in the PaviaU, Botswana,
supervised training iterations is 400. ,e effects of the and Cuprite databases are the highest, namely, 91.33%,
number of nodes in hidden layers (50, 100, 150, 200, 250, and 96.69%, and 98.15%, respectively. ,erefore, the learning
300) in the PaviaU, Botswana, and Cuprite databases on the rate of the DBN model is set as 0.15 for subsequent
accuracy of DBN model recognition are compared. It is experiments.
8483, 2020, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2020/8886932, Wiley Online Library on [21/07/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
8 Computational Intelligence and Neuroscience

Original Ground truth

Opal (509, 31.97%) Muscovite (128, 8.04%)


Kaolinite (357, 22.43%) Dickite (121, 7.60%)
Muscovite + chlorite (189, 11.87%) Tuff (67, 4.21%)
Feldspar (167, 10.49%) Alunite (54, 3.39%)
Figure 9: Cuprite data images and the real ground object map.

1.0
450 0.5
400 0.8
0.4
Reflectance_inside

Reflectance_edge
350
0.3 0.6
300
Reflectivity

0.2 0.4
250
200 0.1
0.2
150 0.0
100 0.0
–0.1
50 0 50 100 150 200 250 300
0 10 20 30 40 50 Spectral bland number
Number of bands
Reconstructed_inside
Opal Muscovite Raw_inside
Feldspar Alunite Error_inside
Tuff Muscovite + chlorite Error_edge
Kaolinite Dickite Raw_edge
Reconstructed_edge
Figure 10: ,e spectral curve of each category of ground objects in
the Cuprite data image. Figure 11: Image spectral reconstruction based on DBN model.

4.3. Comparison of Image Classification Results Based on the ,e effect of ELM, SVM, and DBN on the feature
@ree Classification Algorithms. ,e results of SVM, ELM, classification in the Botswana data image is compared. It can
and DBN in the classification of PaviaU images are evalu- be found from Table 2 that the classification time of ELM is
ated. Table 1 illustrates that the classification time of ELM the shortest (34.55 s) and the SVM classification time is the
model is the shortest (32.17 s) while the classification time of longest (330.91 s). After comparing the classification accu-
SVM is the longest (605.36 s). ,e DBN model has the racy, the overall classification accuracy and average classi-
highest overall classification accuracy and average classifi- fication accuracy of DBN are the highest (98.17%, 97.31%).
cation accuracy (90.54%, 92.36%) while the SVM model has ,e Kappa coefficients of ELM, SVM, and DBN models are
the lowest overall classification accuracy and average clas- 0.963, 0.913, and 0.944, respectively. As shown in Figure 16,
sification accuracy (86.17%, 88.90%). ,e Kappa coefficients the accuracy of DBN model and ELM model in the clas-
of ELM, SVM, and DBN are 0.832, 0.768, and 0.883, re- sification of image features is significantly higher than that of
spectively. It can be concluded from Figure 15 that SVM, SVM model.
ELM, and DBN can effectively complete the classification of ,e effect of ELM, SVM, and DBN on the terrain
ground objects in PaviaU data images, but DBN has higher classification in Cuprite data image is compared. It can be
classification accuracy. concluded from Table 3 that the classification time of ELM is
8483, 2020, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2020/8886932, Wiley Online Library on [21/07/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Computational Intelligence and Neuroscience 9

92.5 99 100
90.0 96 96
87.5 93
92
85.0 90

Accuracy (%)
Accuracy (%)

Accuracy (%)
87 88
82.5
84 84
80.0
81
77.5 80
78
75.0 76
75
72.5
72 72
70.0
0 100 200 300 400 0 100 200 300 400 0 100 200 300 400
PaviaU Botswana Cuprite

Nodes: 50 Nodes: 200 Nodes: 50 Nodes: 200 Nodes: 50 Nodes: 200


Nodes: 100 Nodes: 250 Nodes: 100 Nodes: 250 Nodes: 100 Nodes: 250
Nodes: 150 Nodes: 300 Nodes: 150 Nodes: 300 Nodes: 150 Nodes: 300
Figure 12: ,e influence of the number of hidden layer nodes on the recognition accuracy.

92.5 97 99

96 98
92.0 97
95
96
Accuracy (%)

Accuracy (%)

Accuracy (%)
91.5
94 95

91.0 93 94
93
92
90.5 92
91 91
90.0 90 90
n = 50

n = 100

n = 200

n = 300

n = 400

n = 50

n = 100

n = 200

n = 300

n = 400
n = 50

n = 100

n = 200

n = 300

n = 400

PaviaU Botswana Cuprite

Figure 13: ,e influence of unsupervised training on recognition accuracy.

92
98
96
90 96
94

88 92 94
Accuracy (%)
Accuracy (%)
Accuracy (%)

90 92
86
88 90
84
86 88
82 86
84

80 82 84
0 100 200 300 400 0 100 200 300 400 0 100 200 300 400
PaviaU Botswana Cuprite

Learning Learning Learning Learning Learning Learning


rate: 0.01 rate: 0.15 rate: 0.01 rate: 0.15 rate: 0.01 rate: 0.15
Learning Learning Learning Learning Learning Learning
rate: 0.05 rate: 0.30 rate: 0.05 rate: 0.30 rate: 0.05 rate: 0.30
Learning Learning Learning Learning Learning Learning
rate: 0.10 rate: 0.45 rate: 0.10 rate: 0.45 rate: 0.10 rate: 0.45
Figure 14: ,e influence of learning rate on recognition accuracy.
8483, 2020, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2020/8886932, Wiley Online Library on [21/07/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
10 Computational Intelligence and Neuroscience

Table 1: ,e results of classification evaluation indices of PaviaU data.


Indices ELM SVM DBN
Classification time (s) 32.17 605.36 336.19
Overall classification accuracy (%) 86.32 86.17 90.54
Average classification accuracy (%) 89.54 88.90 92.36
Kappa coefficient 0.832 0.768 0.883

Original SVM ELM DBN


Figure 15: ,e classification results of PaviaU data.

Table 2: ,e results of classification and evaluation indices of Table 3: ,e results of classification and evaluation indices of
Botswana data. Cuprite data.
Indices ELM SVM DBN Indices ELM SVM DBN
Classification time (s) 34.55 330.91 127.64 Classification time (s) 24.83 35.22 25.39
Overall classification accuracy (%) 96.39 93.26 98.17 Overall classification accuracy (%) 97.74 95.27 99.06
Average classification accuracy (%) 96.14 92.11 97.31 Average classification accuracy (%) 96.28 91.68 98.84
Kappa coefficient 0.963 0.913 0.944 Kappa coefficient 0.924 0.944 0.972

5. Discussion
Original ,e spectral feature of a ground object is its electro-
magnetic radiation, which includes reflection, and the
band characteristic is determined by measuring the visible
SVM or invisible light absorption. Ground objects differ in
terms of reflectivity, and reflectivity is often used for
analysis. After analysing the spectral characteristics of
objects in each dataset, it is found that the spectral
ELM morphologies of bare soil and sand grains in the PaviaU
dataset are highly similar, and only a small difference is
observed in the red-near-infrared band. Ground object
DBN categories such as Robinia pseudoacacia forest, Robinia
shrub, and Robinia pseudoacacia in the Botswana dataset
are affected by factors such as mixed pixels; hence, the
Figure 16: ,e classification results of Botswana data. spectral curves of these ground object categories are
difficult to distinguish [25]. In addition, plants such as
the shortest (24.83 s) and the classification time of SVM is Brassica oleifera and Robinia shrubs also exhibit symbiosis
the longest (35.22 s). After comparing the classification in the concentrations of the Botswana data, which can lead
accuracy, the overall classification accuracy and average to similar spectral curves of these ground objects. ,e
classification accuracy of DBN are the highest (99.06%, spectral characteristics of ground object categories such as
98.84%). ,e Kappa coefficients of ELM, SVM, and DBN Kaolinite and Tuff in the Cuprite dataset are highly similar
models are 0.924, 0.944, and 0.972, respectively. As shown in [26]. Analysis of the spectral characteristics of various
Figure 17, the classification accuracy of image features of types of objects is of substantial significance for increasing
DBN model is obviously better than that of SVM model and the classification accuracy and evaluating the classifica-
ELM model. tion performance of a ground object classification model.
8483, 2020, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2020/8886932, Wiley Online Library on [21/07/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Computational Intelligence and Neuroscience 11

Original SVM ELM DBN


Figure 17: ,e classification results of Cuprite data.

DBN is a probabilistic generation model, and it is recognition and classification in hyperspectral images, which
composed of multiple RBM layers. ,e DBN model is widely is consistent with the findings of Maggu et al. that the image
used in image recognition, and it has produced excellent classification model that is based on DBN has high ro-
results. Based on the DBN framework, Samadi et al. pro- bustness [34]. Previous studies on the classification and
posed a method for change detection in SAR images, and it recognition of remote-sensing hyperspectral images focus
was demonstrated that the method realizes high accuracy mainly on the spectral dimension characteristics of image
and detection performance [27]. Ahmad et al. proposed an elements [35]. However, due to the complexity and the
algorithm for the automatic segmentation of liver CT image presence of mixed pixels in natural images, it is not sufficient
features that is based on DBN, and they found that the to analyse the spectral characteristics of pixels. ,erefore, the
accuracy of this method was up to 94.80% [28]. ,erefore, in spectral characteristics and spatial characteristics of ground
this study, DBN is used to evaluate the spectral recon- objects are analysed in the study. ,e study aims at in-
struction of interior and boundary points of terrain images. creasing the classification accuracies of various types of
,e results demonstrate that the errors of spectral recon- ground objects in remote-sensing hyperspectral images.
struction of terrain images based on interior points are Understanding the natural variations of ground objects is of
significantly lower than those based on boundary points; substantial significance. In the future, machine learning
hence, in terrain classification, the spectral reconstruction algorithms can be further investigated from various aspects,
performance of the classification network that selects the such as their loss function curves, to increase the accuracy
internal equinox of the image is higher than that of the and performance in ground object classification of remote-
network that selects the boundary points, which may be why sensing hyperspectral images.
there is much spectral confusion at the boundary points [29].
,e classification time of a model is affected by many factors, 6. Conclusions
such as the tools that are used in the calculation, the
complexity of the model, and the quality of the data [30]. In To study the performance of machine learning on terrain
this study, it is found that when ELM, SVM, and DBN recognition and classification of remote-sensing hyper-
models are used for hyperspectral image classification, the spectral images, an image classification model that is based
ELM model has the shortest classification time. However, on DBN is constructed. It is applied to the classification of
the classification time of the DBN model that is proposed in real hyperspectral image data, and its classification perfor-
this study is between those of the ELM model and the SVM mance is compared with those of SVM and ELM models.
model. ,is is because the DBN model that is constructed in ,e results are as follows:
this study contains 4 layers of RBMs; hence, the complexity (i) Spectral curves that differ in terms of the types of
of this model is high [31]. Subsequently, the Kappa coeffi- ground object information have higher similarity,
cient is used to compare the accuracies of classification and which increases the difficulty of classification of
identification of the models. ,e closer the Kappa coefficient large datasets and affects the accuracy of classifi-
is to 1, the higher the consistency of classification [32]. In cation of different types of ground objects.
this study, it is found that the Kappa coefficients of the DBN-
(ii) Based on the spectral characteristics and spatial
based hyperspectral image feature classification model in
characteristics of ground objects, the ground objects
PaviaU, Botswana, and Cuprite database image recognition
in remote-sensing hyperspectral images are classi-
are 0.883, 0.944, and 0.972, respectively, and the Kappa
fied, which lays a foundation for increasing the
coefficients all exceed 0.75; hence, the classification model of
classification accuracies of various algorithms.
hyperspectral image features that is based on DBN has high
classification accuracy. ,is is consistent with the research (iii) ,e DBN model that is constructed in this study can
results of Li et al. [33]. In addition, the Kappa coefficients of effectively extract features from hyperspectral im-
the SVM and ELM models exceed 0.75; thus, these two ages and classify various types of ground objects.
methods can also effectively classify ground objects, but their (iv) ,e DBN model that is constructed in this study
classification accuracies are lower than that of DBN. outperforms the SVM and ELM models in terms of
,erefore, DBN has higher robustness for spectral feature classification performance in the classification of
8483, 2020, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2020/8886932, Wiley Online Library on [21/07/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
12 Computational Intelligence and Neuroscience

ground objects in remote-sensing hyperspectral [8] A. Rajkomar, J. Dean, and I. Kohane, “Machine learning in
images. medicine,” New England Journal of Medicine, vol. 380, no. 14,
pp. 1347–1358, 2019.
However, strong spatial dimensional texture informa- [9] A. Garcı́a-Floriano, Á. Ferreira-Santiago, O. Camacho-Nieto,
tion and more noise are present in hyperspectral images, and and C. Yáñez-Márquez, “A machine learning approach to
the impacts of these factors on the classification performance medical image classification: detecting age-related macular
have not been considered. ,erefore, it is necessary to degeneration in fundus images,” Computers & Electrical
combine filtering and texture enhancement to increase the Engineering, vol. 75, pp. 218–229, 2019.
classification accuracy of the model. ,e results of this study [10] P. J. Sudharshan, C. Petitjean, F. Spanhol, L. E. Oliveira,
can provide a theoretical basis for increasing the efficiency of L. Heutte, and P. Honeine, “Multiple instance learning for
terrain classification in remote-sensing hyperspectral histopathological breast cancer image classification,” Expert
Systems with Applications, vol. 117, pp. 103–111, 2019.
images.
[11] S. Li, W. Song, L. Fang, Y. Chen, P. Ghamisi, and
J. A. Benediktsson, “Deep learning for hyperspectral image
Data Availability classification: an overview,” IEEE Transactions on Geoscience
and Remote Sensing, vol. 57, no. 9, pp. 6690–6709, 2019.
,e data used to support the findings of this study are [12] F. Lv and M. Han, “Hyperspectral image classification based
available from the corresponding author upon request. on multiple reduced kernel extreme learning machine,” In-
ternational Journal of Machine Learning and Cybernetics,
Conflicts of Interest vol. 10, no. 12, pp. 3397–3405, 2019.
[13] J. M. Murphy and M. Maggioni, “Spectral-spatial diffusion
,e authors declare that they have no conflicts of interest. geometry for hyperspectral image clustering,” IEEE Geo-
science and Remote Sensing Letters, vol. 17, no. 7, pp. 1243–
1247, 2019.
Acknowledgments [14] L. Zhang, L. Zhang, B. Du, J. You, and D. Tao, “Hyperspectral
image unsupervised classification by robust manifold matrix
,is work was supported by the Major Science and Tech- factorization,” Information Sciences, vol. 485, pp. 154–169,
nology Innovation Projects of Shandong Province (no. 2019.
2019JZZY020103), National Innovation Training Program [15] M. Shadman Roodposhti, A. Lucieer, A. Anees, and B. Bryan,
for College Students (no. 201910424028), and Innovation “A robust rule-based ensemble framework using mean-shift
Training Program for College Students in Shandong Prov- segmentation for hyperspectral image classification,” Remote
ince (nos. S201910424041 and S201910424051). Sensing, vol. 11, no. 17, p. 2057, 2019.
[16] P. Duan, X. Kang, S. Li, and P. Ghamisi, “Noise-robust
hyperspectral image classification via multi-scale total vari-
References ation,” IEEE Journal of Selected Topics in Applied Earth Ob-
[1] Z. Chen, X. Li, C. Yang et al., “A data-driven ground fault servations and Remote Sensing, vol. 12, no. 6, pp. 1948–1962,
detection and isolation method for main circuit in railway 2019.
electrical traction system,” ISA Transactions, vol. 87, [17] P. Mishra, A. Karami, A. Nordon, D. N. Rutledge, and
pp. 264–271, 2019. J.-M. Roger, “Automatic de-noising of close-range hyper-
[2] K. Karila, X. Yu, M. Vastaranta, M. Karjalainen, E. Puttonen, spectral images with a wavelength-specific shearlet-based
and J. Hyyppä, “TanDEM-X digital surface models in boreal image noise reduction method,” Sensors and Actuators B:
forest above-ground biomass change detection,” ISPRS Chemical, vol. 281, pp. 1034–1044, 2019.
Journal of Photogrammetry and Remote Sensing, vol. 148, [18] P. Zhou, J. Han, G. Cheng, and B. Zhang, “Learning compact
pp. 174–183, 2019. and discriminative stacked autoencoder for hyperspectral
[3] S. Liu, B. Yang, Z. Zhang et al., “Influence of polarized re- image classification,” IEEE Transactions on Geoscience and
flection on airborne remote sensing of canopy foliar nitrogen Remote Sensing, vol. 57, no. 7, pp. 4823–4833, 2019.
content,” International Journal of Remote Sensing, vol. 41, [19] R. Hang, Q. Liu, D. Hong, and P. Ghamisi, “Cascaded re-
no. 13, pp. 1–22, 2020. current neural networks for hyperspectral image classifica-
[4] A. Baez, H. Fabelo, S. Ortega et al., “High-level synthesis of tion,” IEEE Transactions on Geoscience and Remote Sensing,
multiclass SVM using code refactoring to classify brain cancer vol. 57, no. 8, pp. 5384–5394, 2019.
from hyperspectral images,” Electronics, vol. 8, no. 12, p. 1494, [20] R. Dian, S. Li, L. Fang, and Q. Wei, “Multispectral and
2019. hyperspectral image fusion with spatial-spectral sparse rep-
[5] A.-K. Mahlein, M. T. Kuska, S. ,omas et al., “Quantitative resentation,” Information Fusion, vol. 49, pp. 262–270, 2019.
and qualitative phenotyping of disease resistance of crops by [21] W. Deng, R. Yao, H. Zhao, X. Yang, and G. Li, “A novel
hyperspectral sensors: seamless interlocking of phytopathol- intelligent diagnosis method using optimal LS-SVM with
ogy, sensors, and machine learning is needed!,” Current improved PSO algorithm,” Soft Computing, vol. 23, no. 7,
Opinion in Plant Biology, vol. 50, pp. 156–162, 2019. pp. 2445–2462, 2019.
[6] J. Marrs and W. Ni-Meister, “Machine learning techniques for [22] Z. M. Yaseen, S. O. Sulaiman, R. C. Deo, and K.-W. Chau, “An
tree species classification using co-registered LiDAR and enhanced extreme learning machine model for river flow
hyperspectral data,” Remote Sensing, vol. 11, no. 7, p. 819, forecasting: state-of-the-art, practical applications in water
2019. resource engineering area and future research direction,”
[7] S. Zhang, L. Yao, A. Sun, and Y. Tay, “Deep learning based Journal of Hydrology, vol. 569, pp. 387–408, 2019.
recommender system,” ACM Computing Surveys, vol. 52, [23] K. Tan, F. Wu, Q. Du, P. Du, and Y. Chen, “A parallel
no. 1, pp. 1–38, 2019. Gaussian-Bernoulli restricted Boltzmann machine for mining
8483, 2020, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2020/8886932, Wiley Online Library on [21/07/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Computational Intelligence and Neuroscience 13

area classification with hyperspectral imagery,” IEEE Journal


of Selected Topics in Applied Earth Observations and Remote
Sensing, vol. 12, no. 2, pp. 627–636, 2019.
[24] J. Zheng, X. Fu, and G. Zhang, “Research on exchange rate
forecasting based on deep belief network,” Neural Computing
and Applications, vol. 31, no. 1, pp. 573–582, 2019.
[25] M. Wendorff and A. Świa˛der, “Lithostratigraphic classifica-
tion of the tsodilo hills group: a palaeo-to mesoproterozoic
metasedimentary succession in NW Botswana,” Geology,
Geophysics and Environment, vol. 45, no. 4, p. 305, 2019.
[26] M. Bevilacqua and Y. Berthoumieu, “Multiple-feature kernel-
based probabilistic clustering for unsupervised band selec-
tion,” IEEE Transactions on Geoscience and Remote Sensing,
vol. 57, no. 9, pp. 6675–6689, 2019.
[27] F. Samadi, G. Akbarizadeh, and H. Kaabi, “Change detection
in SAR images using deep belief network: a new training
approach based on morphological images,” IET Image Pro-
cessing, vol. 13, no. 12, pp. 2255–2264, 2019.
[28] M. Ahmad, D. Ai, G. Xie et al., “Deep belief network modeling
for automatic liver segmentation,” IEEE Access, vol. 7,
pp. 20585–20595, 2019.
[29] J. Gong, J. Yan, and D. Li, “Comparison of micro-Doppler
signatures registered using RBM of helicopters and WSM of
vehicles,” IET Radar, Sonar & Navigation, vol. 13, no. 11,
pp. 1951–1955, 2019.
[30] H. I. Fawaz, G. Forestier, J. Weber, L. Idoumghar, and
P. A. Muller, “Deep learning for time series classification: a
review,” Data Mining and Knowledge Discovery, vol. 33, no. 4,
pp. 917–963, 2019.
[31] K. Kourou, G. Rigas, C. Papaloukas, M. Mitsis, and
D. I. Fotiadis, “Cancer classification from time series
microarray data through regulatory dynamic bayesian net-
works,” Computers in Biology and Medicine, vol. 116, Article
ID 103577, 2020.
[32] T.-S. Tsou, “A robust likelihood approach to inference about
the kappa coefficient for correlated binary data,” Statistical
Methods in Medical Research, vol. 28, no. 4, pp. 1188–1202,
2019.
[33] C. Li, Y. Wang, X. Zhang, H. Gao, Y. Yang, and J. Wang,
“Deep belief network for spectral-spatial classification of
hyperspectral remote sensor data,” Sensors, vol. 19, no. 1,
p. 204, 2019.
[34] J. Maggu, H. K. Aggarwal, and A. Majumdar, “Label-con-
sistent transform learning for hyperspectral image classifi-
cation,” IEEE Geoscience and Remote Sensing Letters, vol. 16,
no. 9, pp. 1502–1506, 2019.
[35] Z. Lyu, M. Han, and D. Li, “Denoising of uncertain type noise
images by spatial feature classification in nonsubsampled
shearlet transform,” IEEE Access, vol. 8, pp. 5009–5021, 2019.

You might also like