Innovative Hyperspectral Image Classification Approach Using Optimized CNN and ELM

Abstract

In order to effectively extract features and improve classification accuracy for hyperspectral remote sensing images (HRSIs), the advantages of enhanced particle swarm optimization (PSO) algorithm, convolutional neural network (CNN), and extreme learning machine (ELM) are fully utilized to propose an innovative classification method of HRSIs (IPCEHRIC) in this paper. In the IPCEHRIC, an enhanced PSO algorithm (CWLPSO) is developed by improving learning factor and inertia weight to improve the global optimization performance, which is employed to optimize the parameters of the CNN in order to construct an optimized CNN model for effectively extracting the deep features of HRSIs. Then, a feature matrix is constructed and the ELM with strong generalization ability and fast learning ability is employed to realize the accurate classification of HRSIs. Pavia University data and actual HRSIs after Jiuzhaigou M7.0 earthquake are applied to test and prove the effectiveness of the IPCEHRIC. The experiment results show that the optimized CNN can effectively extract the deep features from HRSIs, and the IPCEHRIC can accurately classify the HRSIs after Jiuzhaigou M7.0 earthquake to obtain the villages, bareland, grassland, trees, water, and rocks. Therefore, the IPCEHRIC takes on stronger generalization, faster learning ability, and higher classification accuracy.

1. Introduction

Remote sensing image (RSI) classification is to divide the image into several regions by using specific rule or algorithm according to the spectral features, geometric texture features, or other features [1,2,3]. Each region is a set of ground and objects with the same characteristics, or a lot of RSIs are divided into several sets through some methods, and each set represents a kind of ground or object category. It is a very important basic problem and plays a very important position in the field of RSIs [4,5,6]. Therefore, the research on remote sensing image classification method has become an important direction, which has very important theoretical significance and practical application value.

In recent years, many classification methods of RSIs have been proposed, which can be divided into two categories of manual visual interpretation and computer classification [7]. The manual visual interpretation is the most traditional classification method, which has large workload, low efficiency, and requires rich professional knowledge and interpretation experiences [8,9,10]. With the rapid development of computer techniques, the automatic classification method of RSIs replaces the manual visual interpretation classification method. The more complex computer technology uses the spectral brightness value of pixels and the spatial relationship between pixels and their surrounding pixels to realize pixel classification. Tran et al. [11] presented a sub-pixel and per-pixel classification method to analyze the impact of land cover heterogeneity. Khodadadzadeh et al. [12] presented a new hyperspectral spectral-spatial classifier. Li et al. [13] presented a novel classification method of RSIs based on the probabilistic fusion of pixel-level and superpixel-level classifiers. Li et al. [14] presented a novel pixel-pair method. Mei et al. [15] presented a novel pixel-level perceptual subspace learning method. Pan et al. [16] presented a new central pixel selection strategy based on gradient information to realize texture image classification. Bey et al. [17] presented a new land cover assessment methodology. Yan et al. [18] presented a triple counter domain adaptation approach for learning domain invariant classifier. Li et al. [19] presented a novel multi-view active learning approach based on sub-pixel and super-pixel. Ma and Chang [20] presented a novel mixed pixel classification approach.

The single pixel spectral classification method can obtain the hyperspectral spectral-spatial classification results, but they still exist at low classification accuracy and high time complexity. The signal processing method on computer has the characteristics of large amount of calculation and can obtain high classification accuracy. However, the high-resolution RSIs have high spatial resolution and complexity. It is very difficult to classify high-resolution RSIs by using traditional classification methods. Therefore, it is urgent to deeply study a fast classification approach that can be effectively applied to high-resolution RSIs [21,22]. As a field of artificial intelligence, deep learning has attracted extensive attention, and has gradually become one of the important technologies to promote the development of artificial intelligence. Therefore, many scholars have applied deep learning to remote sensing image classification and proposed many features extraction and classification methods. Romero et al. [23] presented a sparse feature unsupervised learning approach based on greedy hierarchical unsupervised pretraining method. Sharma et al. [24] presented a new deep patch-based CNN. Maggiori et al. [25] presented a dense pixel-level classification model. Wang et al. [26] presented a HRSI classification method using principal component analysis (PCA) and guided filtering, deep learning architecture. Ji et al. [27] presented a novel three-dimensional CNN to automatically classify crops. Ben et al. [28] presented 3-D deep learning approach. Xu et al. [29] presented a novel RSI classification model using generative adversarial network. Tao et al. [30] presented a novel reinforced deep neural network (DNN) with depth and width. Liang et al. [31] presented a new RSI classification approach using stacked denoising autoencoder. Li et al. [32] presented a novel region-wise depth feature extraction model. Li et al. [33] presented an adaptive multiscale deep fusion residual network. Yuan et al. [34] presented a classification approach based on rearranged local features. Zhang et al. [35] presented a new dense network with multi-scales. Zhang et al. [36] presented a new feature aggregation model based on 3-D CNN. Chen et al. [37] presented a novel deep Boltzmann machine based on the conjugate gradient update algorithm. Xiong et al. [38] presented a novel deep multi-feature fusion network based on two different deep architecture branches. Tong et al. [39] presented a channel-attention-based DenseNet network. Zhu et al. [40] presented a new deep network with dual-branch attention fusion. Raza et al. [41] presented a four-layer classification network based on visual attention mechanisms. Li et al. [42] presented a classification approach by combining generative adversarial network (GAN), CNN with long short-term memory. Gu et al. [43] presented a pseudo labeled sample generation method. Guo et al. [44] presented a novel self-supervised gated self-attention GAN. Li et al. [45] presented a novel locally preserving deep cross embedded classification network. Lei et al. [46] presented a novel deep convolutional capsule network using spectral-spatial features. Cui et al. [47] presented a dual-channel deep learning recognition model. Peng et al. [48] presented an efficient search framework to discover optimal network architectures. Guo et al. [49] presented a novel semi-supervised scene classification method using GAN. Dong et al. [50] presented a pixel cluster CNN. Li et al. [51] presented a new RSI classification approach using error-tolerant deep learning. Li et al. [52] presented a gated recursive neural network. Dong et al. [53] explored the potential of the reference-based super-resolution method. Wu et al. [54] presented a self-paced dynamic infinite mixture model. Karadal et al. [55] presented automated classification of remote sensing images based on multileveled MobileNetV2 and DWT. Ma et al. [56] presented a novel adaptive hybrid fusion network for multiresolution remote sensing images classification. Cai et al. [57] presented a novel cross-attention mechanism and graph convolution integration algorithm. Zhang et al. [58] presented a convolutional neural architecture for remote sensing image scene classification. Hilal et al. [59] presented a new deep transfer learning-based fusion model for remote-sensing image classification. Li et al. [60] presented a multi-scale fully convolutional network to exploit discriminative representations. In addition, some new optimization algorithms are proposed [61,62,63,64,65,66,67,68,69,70,71,72], which can optimize the parameters of classification models.

Because the CNN has good feature extraction ability, these classification methods based on CNN have obtained better classification effects. It has attracted extensive attention and has been widely applied in RSIs. However, the structure and parameter selection of the CNN seriously affect its learning accuracy. Therefore, the enhanced PSO algorithm with global optimization ability is employed to optimize and determine the parameters of the CNN to obtain the optimized parameter values for constructing an optimized CNN, which is applied to effectively extract the multi-layer features of HRSIs to form a multi-feature fusion matrix. Then, the ELM is employed to realize the classification of HRSIs. The effectiveness is verified by typical data set and actual HRSIs after Jiuzhaigou M7.0 earthquake.

The main contributions of this paper are described as follows.

(1)

For the slow convergence and low accuracy of the PSO, an enhanced PSO based on fusing multi-strategy (CWLPSO) is proposed by adding new acceleration factor strategy and inertia weight linear decreasing strategy.

(2)

For the difficultly determining the parameters of the CNN, an optimized CNN model using CWLPSO is developed to effectively extract the deep features of HRSIs.

(3)

The ELM with strong generalization ability, fast learning ability, and the constructed feature vector are combined to realize the accurate classification of HRSIs.

(4)

An innovative classification method of HRSIs based on CWLPSO, CNN, and ELM, namely, IPCEHRIC is proposed.

2. Basic Methods

2.1. CNN

The CNN is a feedforward neural network, which includes convolution calculation and representative algorithm. It has the representation learning ability and can classify the input information according to its hierarchical structure. The CNN includes input layer, hidden layer, and output layer, which is shown in Figure 1.

The structure of the CNN is described in detail as follows.

Input layer. It can deal with multidimensional data, and the input features need to be standardized.

Hidden layer. It includes convolution operation, pooling operation, and full connection layer. The convolution layer is used to extract features from input data through the convolution operation of multiple convolution cores to obtain and construct the feature map. The pooling layer is to select features and filter information from the feature map to retain important features, and preset the pooling function. The full connection layer is equivalent to the hidden layer in the network. The output is obtained.

Convolution kernel. When the convolution kernel works, it will regularly scan input features, multiply and sum the input features, and superimpose the deviation. The output of the $l+1$ layer is described as follow.

$$\begin{gathered} Z^{l+1}\left(i,j\right)=[Z^l\otimes w^{l+1}]\left(i,j\right) + b={\displaystyle \sum }_{k=1}^{K_l}{\displaystyle \sum }_{x=1}^f{\displaystyle \sum }_{y=1}^f[Z_k^l\left(s_{0}i+x, s_{0}j+y\right)w_k^{l+1}\left(x,y\right)] + b \\ \left(i,j\right) \in \left\{0,1,\dots ,L_{l+1}\right\} L_{l+1}=\frac{L_l+2p-f}{s_0}+1 \end{gathered}$$

(1)

where, $b$ is the offset, $Z^l$ and $Z^{l+1}$ represents the convolution input and output of the $l+1$ layer, $L_{l+1}$ is the size of $Z_{l+1}$. In here, it is assumed that the length and width of the characteristic graph are the same. $Z\left(i,j\right)$ corresponds the pixels of the feature map, $K$ is the number of channels, $f$, $s_0$ and $p$ are the convolution layer parameters, which correspond to the kernel size, convolution step size and number of filling layers. Especially, when the kernel is f = 1, the step size is $s_0$ = 1, and when a filled unit convolution kernel is not included, the cross-correlation calculation is equivalent to matrix multiplication, and a fully connected network is established between the convolution layers.

$$Z^{l+1}={\displaystyle \sum }_{k=1}^{K_l}{\displaystyle \sum }_{x=1}^L{\displaystyle \sum }_{y=1}^L(Z_{i,j}^l{w}_k^{l+1})+b=w_{l+1}^T{Z}_{l+1}+b, L^{l+1}=L$$

(2)

Output layer. The output layer is the same, and the output result is obtained.

2.2. PSO

The PSO is an intelligent algorithm, which was proposed by Eberhart and Kennedy in 1995 [73]. At first, it was to study the predation behavior of birds. Inspired by this, it carried out modeling research on bird activities. In PSO, the update formula of the particle velocity and position are described as follows.

$$v_{m+1}=\omega v_m+c_1{r}_1\left(pbes{t}_m-x_m\right)+c_2{r}_2\left(gbes{t}_m-x_m\right)$$

(3)

$$x_{m+1}=x_m+v_{m+1}$$

(4)

where, $v_{m+1}$ represents the velocity of particles, $\omega$ is the inertia weight factor, $c_1$ and $c_2$ are learning factors, $\omega , c_1,$ and $c_2$ are usually preseted in advance. $r_1$ and $r_2$ represent a random number, $pbes{t}_m$ is the optimal value of individual, $gbes{t}_m$ is the optimal value of swarm. The function used to evaluate the fitness value of particles is called fitness function, i.e., objective function. In most cases, the fitness value is smaller, the particle is better. The optimal value of the individual and the optimal value of swarm are generally updated by the following formula.

$$pbes{t}_{m+1}=\left\{\begin{array}{l}{x}_{m+1},f\left(x_{m+1}\right)<f\left(pbes{t}_m\right)\\ pbes{t}_m,otherwise\end{array}\right.$$

(5)

$$gbes{t}_{m+1}=\left\{\begin{array}{l}pbes{t}_{m+1},f\left(pbes{t}_{m+1}\right)<f\left(gbes{t}_{m+1}\right)\\ gbes{t}_{m+1},otherwise\end{array}\right.$$

(6)

If the value of $x_{m+1}$ is smaller than the value of the individual extreme value, then $pbes{t}_{m+1}$ is equal to $x_{m+1}$. On the contrary, the individual extreme value is not updated. If the value of $gbes{t}_{m+1}$ is greater than the value of the individual extreme value, then $gbes{t}_{m+1}$ is equal to $gbes{t}_{m+1}$.

2.3. ELM

The ELM is one of the commonly used neural network models in machine learning. Its essence is a machine learning method based on single-hidden layer feed forward network (SLFN). Compared with back propagation (BP) neural network model that uses gradient descent algorithm to update the weight in the field of machine learning, the ELM can randomly generate the threshold value. It has low computational complexity and less time-consuming. In the classification and regression problems, the structure of the ELM model is generally divided into the input layer, hidden and output layers. The specific structure is shown in Figure 2.

3. Improved Learning Factor and Inertia Weight

Although many researchers have proposed some effective researches and improvements on the shortcomings of PSO, the PSO still has the problems of slow convergence, high time complexity, and low accuracy. Therefore, the acceleration factor strategy and the inertia weight linear decreasing strategy are introduced to propose an enhanced PSO(CWLPSO) in this paper. That is, aiming at the slow convergence speed, a fast convergence strategy with small deviation angle of particle speed and position is adopted to accelerate the convergence of particles. Aiming at the poor search ability, a new improvement strategy of learning factor is proposed in here. That is, different $c_1$ and $c_2$ values are selected in order to improve the local search ability of particles in the early stage, enhance the optimization ability of particle swarm and strengthen the overall search ability of particles in the later stage. Aiming at the premature in the later stage, a new linear decreasing strategy of inertia weight is adopted to linearly reduce the inertia weight from the maximum value to the minimum value, so as to avoid the premature and the oscillation in the later stage of the algorithm.

3.1. Improve Learning Factors

The learning factors $c_1$ and $c_2$ in the PSO represent the function of the particle itself and the remaining particles removed from the particle itself on the motion route of moving particles. At the same time, they also represent the information exchange between particles, which result in different motion trajectories of particles. Therefore, an improvement strategy of learning factor is designed to improve the local search ability of particles, enhance the optimization ability of particle swarm and strengthen the overall search ability of particles in here. That is, in the early stage of the algorithm, the $c_1$ value is larger and the $c_2$ value is smaller, so that the particles can enhance the ability of self-cognition and weaken the swarm cognition of the particles. However, in the later stage of the algorithm, the $c_1$ value decreases and the $c_2$ value increases, it can improve the search ability by increasing the $c_1$ value in the early stage, the proportion of particle swarm will be strengthened in the later stage, so that more particles can learn from the swarm optimum. At the same time, the fewer particles can learn from individual optimum, which is conducive to enhancing the optimization ability, and strengthening the overall search ability of particles. The improved strategy of learning factor is described as follows.

$$c_1=c_{1max}+\left(c_{1max}-c_{1min}\right)\ast \left(\frac{i}{k}\right)$$

(7)

$$c_2=c_{2min}+\left(c_{2max}-c_{2min}\right)\ast \left(\frac{i}{k}\right)$$

(8)

where, $c_{1max}$ and $c_{1min}$ represent the maximum and minimum values of learning factor $c_1.c_{2min}$ and $c_{2max}$ represent the maximum and minimum values of learning factor $c_2$, i represents the current iterations, and k represents the maximum iterations.

3.2. Linear Decreasing of Inertia Weight

Inertia weight plays an important role in PSO. Generally, the inertia weight is generally set to a fixed value between 0.6 and 0.9. The improper selection of inertia weight will cause errors. If the inertia weight is larger, on the one hand, it will help to jump out from the local minimum point and facilitate the global search, on the other hand, it will weaken the local search ability. Therefore, for the premature in the later stage of the algorithm, a new linear decreasing strategy of inertia weight is developed. That is, the inertia weight is linearly reduced from ${\omega }_{max}$ to ${\omega }_{min}$, which is described as follows.

$$\omega ={\omega }_{max}-\left(\frac{i\ast \left({\omega }_{max}-{\omega }_{min}\right)}{k}\right)$$

(9)

where, $\omega$ is inertia weight, ${\omega }_{max}$ is maximum value of inertia weight, ${\omega }_{min}$ is minimum value of inertia weight, $i$ is current iteration, and $k$ is maximum iterations.

4. Optimize CNN Using CWLPSO

4.1. Optimized Idea for CNN

The CNN with combining weight sharing and local area connection reduces the complexity of the model and the values of parameters. However, the selection of the number of filters, activation function, and learning rate of the CNN seriously affects the learning accuracy. The parameters of the CNN are trained by the steepest gradient descent method, which has a great impact on the learning performance. The proposed CWLPSO has the characteristics of global search ability, population diversity, and fast convergence. Therefore, the CWLPSO is employed to optimize the parameters of the CNN, and an optimized CNN model based on the CWLPSO algorithm is developed in this paper. That is, each particle is a network structure of the CNN. After the CNN calculates the error between the expected value and the actual value, each particle considered the number of filters, activation function learning rate, initial weight, and initial offset of the CNN as particle dimensions. The obtained test error is taken as the fitness function value, the optimal CNN model is selected through the iteration of the CWLPSO.

4.2. Model of Optimized CNN

The optimization process of the CNN using CWLPSO is shown in Figure 3.

The specific optimization process of the CNN using CWLPSO are described as follows.

Step 1. Initialize the parameters of the CNN, which include the number of nodes in hidden layer, the learning rate, and so on.

Step 2. Initialize the parameters of the CWLPSO, which include the number of the population, the maximum number of iterations, and the initial learning factor and inertia weight, and so on.

Step 3. Construct the optimization objective function.

Step 4. Calculate the individual fitness values in the population in order to obtain the initial fitness values of the population.

Step 5. Determine whether the end condition is met. If the end condition is met, then the optimal individual is regarded as the optimal parameter value of the CNN and loop Step 7. Otherwise execute Step 6.

Step 6. The velocity and position are updated, then the learning factor and the weight factor is updated. Then return to Step 4.

Step 7. Obtain the optimal parameter values of the CNN and an optimized CNN model is output.

5. An Innovative Classification Method of HRSIs Using Optimized CNN and ELM

Classification accuracy is important indicators to evaluate the classification model for HRSIs. Therefore, the effective feature extraction of HRSIs is the key factor for affecting classification accuracy. As a deep learning method, the CNN can effectively mine the multi-layer representation feature information. Different levels of representation correspond to different feature attributes of the recognition object. For example, the shallow network mainly represents the texture, edge and other local information of the recognition object, while the deep network represents the more abstract semantics, structure, and other global information. This feature matrix composes of the multi-layer feature attributes of the HRSIs. As a fast machine learning algorithm, the weight parameters of the ELM and the offset parameters on the hidden layer do not need to be adjusted repeatedly through iteration, which can reduce the amount of calculation and shorten the training time. Therefore, in order to make full use of the feature extraction ability of the optimized CNN, the comprehensiveness of multi-layer features and the fast-training speed of the ELM, an innovative classification model of HRSIs based on combining the optimized CNN and ELM, namely IPCEHRIC is developed to improve the robustness and classification effect of the model. The classification process of HRSIs is shown in Figure 4.

The classification process of the IPCEHRIC is described as follows.

(1)

Preprocess HRSIs

Some preprocessed methods, such as whitening processing, normalization processing, gray transformation, image smoothing, interpolation method, and so on are used to eliminate irrelevant information in hyperspectral remote sensing images, restore useful real information, enhance the detectability of relevant information, and simplify the data to the greatest extent, including image denoising, enhancement, smoothing, and sharpening, so as to improve the reliability of feature extraction, image matching, and so on.

(2)

Optimize parameters of CNN

The CWLPSO with global optimization capability is employed to optimize and determine the parameters of the CNN, including the number of filters, activation function, learning rate, initial weight, and initial bias as particle dimension. The optimized parameter values are obtained, and an optimized CNN model is constructed.

(3)

Extract features

The optimized CNN is essentially a multi-layer perceptron, which is mainly characterized by its local connection and weight sharing mode. When the input data are images, the alternated convolution layer and maximum pool layer by layer are used to automatically complete the feature extraction.

(4)

Construct feature matrix

The extracted local features are input into the full connection layer of the first layer in order to form the global features. These images are taken from different feature ranges. Then these extracted features are selected to construct a feature matrix in order to provide feature matrix for the classifier.

(5)

Establish ELM classifier

The feature matrix is taken as the input of the ELM, elmtrain( ) function and training sets are created to train the ELM. Then, the trained parameters and elmpredict( ) function are used to test the test set, and finally the classification results are obtained.

6. Experiment Verification and Result Analysis

6.1. Experimental Environment and Parameter Setting

The experimental environment is Intel i7-11700 HQ_CPU_@ _ 2.5GHz, 16G RAM with Windows 10, and the programming language is Matlab 2018b. The IPCEHRIC network structure consists of two convolution layers, two pooling layers and an ELM classifier. The nonlinear activation function of CNN is RELU function, and the ELM classifier uses Sigmoid function. The initial parameters of CWLPSO are $c_{1max}=2.0$, $c_{1min}=0.5$, ${\omega }_{max}=0.9$, $\otimes$, maximum number of iterations K = 200. The initial parameters of the CNN are the number of convolution kernels (6), and the size of convolution kernels (1 * 3). The initial parameters of the ELM are σ = 0.1, regularity coefficient C = 0.5.

6.2. Pavia University Data

6.2.1. Data Description

Pavia University data set is a hyperspectral remote sensing image data set collected from the University of Pavia in northern Italy by using the airborne reflection optical spectrum imager of Germany. The size of image is 610 × 340, including 42,776 pixels and 9 types of features through excluding a large number of backgrounds. Basic information of Pavia University data is shown in

Table 1. Basic information of Pavia University data.

Data	Pavia University
Collection location	Northern Italy
Acquisition equipment	ROSIS
Spectral coverage (μm)	0.43–0.86
Data size (pixel)	610 × 340
Spatial resolution (m)	1.3
Number of bands	115
Number of bands after denoising	103
Sample size	42,776
Number of categories	9

. A total of 20% of the samples are randomly selected as the training set and 80% of the samples are used as the test set. The number of samples for training and test is shown in

Table 2. The number of samples in Pavia University.

Types	Class	Training Samples	Test Samples	Samples
1	Asphalt	1326	5305	6631
2	Meadows	3722	14,927	18,649
3	Gravel	418	1681	2099
4	Trees	612	2452	3064
5	Painted metal sheets	268	1077	1345
6	Bare Soil	1004	4025	5029
7	Bitumen	266	1064	1330
8	Self-Blocking Bricks	736	2946	3682
9	Shadows	188	759	947
Total		8540	34,236	42,776

, and the describing of the HRSIs is shown in Figure 5.

6.2.2. Experimental Results and Analysis

To verify the effectiveness of the IPCEHRIC, the CNN, local binary pattern (LBP) and CNN (LBP-CNN), CNN and ELM (CNN-ELM), LBP, CNN and ELM (LBP-CNN-ELM), LBP, PCA, CNN and ELM (LBP-PCA-CNN-ELM) are selected in here. The experiment results of the Pavia university data are shown in

Table 3. The experiment results of the Pavia University data (%).

Types	Class	CNN	LBP-CNN	CNN-ELM	LBP-CNN-ELM	LBP-PCA-CNN-ELM	IPCEHRIC
1	Asphalt	90.00	89.64	94.72	95.72	99.92	99.96
2	Meadows	89.99	89.79	93.00	95.00	99.12	99.67
3	Gravel	89.70	91.63	99.94	99.94	100.00	100.00
4	Trees	88.90	87.10	89.15	94.17	96.88	99.84
5	Painted metal sheets	86.00	89.91	92.26	96.68	99.72	100.00
6	Bare Soil	88.15	89.90	95.00	96.27	100.00	100.00
7	Bitumen	90.45	92.00	94.15	96.15	99.15	99.82
8	Self-Blocking Bricks	89.83	91.86	93.25	95.01	99.66	100.00
9	Shadows	87.50	93.87	90.90	97.74	97.94	99.15
OA (%)		85.67	88.75	92.63	95.64	98.95	99.21
AA (%)		88.95	90.63	93.60	96.30	99.15	99.83
STD		1.467	1.939	3.022	1.722	1.075	0.279

. The overall accuracy (OA), average accuracy (AA), and standard deviation (STD) of classification results are calculated for each algorithm.

It can be seen from Table 3 that the IPCEHRIC method obtains the classification accuracies of OA and AA are 99.21 and 99.83%, which are best classification results among the CNN, LBP-CNN, CNN-ELM, LBP-CNN-ELM, LBP-PCA-CNN-ELM, and IPCEHRIC methods. The STD of the IPCEHRIC is 0.279, which is also the least STD among these methods. Among other comparison methods, the LBP-PCA-CNN-ELM method obtains the classification accuracies of OA and AA as 98.95 and 99.15%. While the CNN-ELM method obtains the classification accuracies for OA and AA of 92.63 and 93.60%. Compared with the CNN-ELM, the classification accuracies of OA and AA of the IPCEHRIC are improved by 6.58 and 6.23% than those of the CNN-ELM. This shows that the feature extraction ability of the optimized CNN is better than that of the CNN, which explains the global optimization ability of the CWLPSO algorithm. Therefore, the classification performance of the IPCEHRIC method is significantly better than those of the CNN, LBP-CNN, CNN-ELM, LBP-CNN-ELM, and LBP-PCA-CNN-ELM. The experiment results show that the IPCEHRIC method has higher classification accuracy than other comparison methods. The IPCEHRIC is an effective classification method for HRSIs.

6.3. Actual HRSI after Jiuzhaigou M7.0 Earthquake

6.3.1. Description of HRSI after Jiuzhaigou 7.0 Earthquake

Jiuzhaigou is located in Zhangzha Town, Jiuzhaigou County, Sichuan Province. It is located in the transition zone. It is more than 400 km away from Chengdu. It is a mountain valley with a depth of more than 50 km, with a total area of 64,297 hm² and a forest coverage rate of more than 80%. The hyperspectral remote sensing image after Jiuzhaigou M7.0 earthquake on 8 August 2017 is shown in Figure 6.

The HRSI after Jiuzhaigou M7.0 earthquake is saved as *. mat file, which determined the coordinates of different areas by manual frame drawing. Then a matrix consistent with the size of the picture is constructed. The corresponding positions of the matrix with different numbers is marked according to the coordinates of different areas, so as to mark different labels on different areas of the picture, save and generate *.mat file with labels. A data set containing four types of samples is made, which include villages, water, grassland, and trees in the HRSIs after Jiuzhaigou M7.0 earthquake. The number of samples and four types are shown in

Table 4. The number of samples and four types.

Types	Class	Samples
1	Villages	12,575
2	Water	14,953
3	Grassland	38,790
4	Trees	39,159
Total		105,477

.

According to the gray value of pixels, the color function is used to set the threshold. The different areas of HRSIs after Jiuzhaigou M7.0 earthquake are marked by different colors. A matrix consistent with the image size is constructed, and the different areas are marked with color. A data set with six types of samples is made, which include the villages, bareland, grassland, trees, water, and rocks in the HRSIs after Jiuzhaigou M7.0 earthquake. The number of samples and six types are shown in

Table 5. The number of samples and six types.

Types	Class	Samples
1	Villages	1608
2	Bareland	25
3	Grassland	376,651
4	Trees	110,409
5	Water	5558
6	Rocks	2469
Total		495,087

.

6.3.2. Experimental Results and Analysis

To prove the ability of the IPCEHRIC to solve practical engineering problems, the hyperspectral remote sensing images after Jiuzhaigou M7.0 earthquake is used for the experimental comparison and analysis. Similarly, the CNN, LBP-CNN, CNN-ELM, LBP-CNN-ELM, and LBP-PCA-CNN-ELM are selected to compare in here. Each algorithm is executed ten times independently. The classification results of HRSI after Jiuzhaigou 7.0 earthquake for four types are shown in

Table 6. The classification results of HRSIs for 10 times for four types (%).

Times	CNN	LBP-CNN	CNN-ELM	LBP-CNN-ELM	LBP-PCA-CNN-ELM	IPCEHRIC
1	41.47	36.68	69.80	64.38	65.67	89.76
2	41.80	36.68	75.84	64.25	65.16	88.96
3	41.75	36.68	75.98	64.16	65.33	89.99
4	41.73	36.68	75.38	64.40	65.14	89.26
5	41.85	37.02	61.45	64.47	65.47	89.76
6	41.70	37.02	75.80	64.12	65.56	90.58
7	41.86	37.02	74.04	63.83	65.40	91.64
8	41.77	37.02	60.02	64.44	65.19	92.12
9	41.78	36.68	74.80	64.38	65.49	90.99
10	41.76	37.02	75.46	64.10	65.81	89.94
AA (%)	41.75	36.85	71.86	64.25	65.42	90.30
STD	0.109	0.179	6.145	0.201	0.223	1.019

and

Table 7. The classification results of HRSIs for four types (%).

Types	Class	CNN	LBP-CNN	CNN-ELM	LBP-CNN-ELM	LBP-PCA-CNN-ELM	IPCEHRIC
1	Villages	50.47	46.76	79.16	74.64	75.70	92.46
2	Water	41.80	35.43	78.37	70.47	73.28	90.73
3	Grassland	39.26	33.58	73.78	63.19	72.45	89.15
4	Trees	40.73	36.29	76.12	69.24	75.42	91.48
OA (%)		41.75	36.85	71.86	64.25	65.42	90.30
AA (%)		43.07	38.02	76.86	69.39	74.21	90.96
STD		5.046	5.939	2.422	4.733	1.597	1.396

. The classification results of HRSI after Jiuzhaigou M7.0 earthquake for six types are shown in

Table 8. The classification results of HRSIs for 10 times for six types (%).

Times	CNN	LBP-CNN	CNN-ELM	LBP-CNN-ELM	LBP-PCA-CNN-ELM	IPCEHRIC
1	79.77	79.83	99.21	85.12	85.12	99.99
2	79.78	79.85	99.78	84.80	84.14	100.0
3	79.84	79.84	99.99	84.14	84.01	99.98
4	79.78	79.86	99.26	84.80	84.22	99.78
5	79.86	79.84	99.99	85.12	85.46	100.0
6	79.87	79.81	99.21	84.76	86.13	99.77
7	79.88	79.59	99.98	85.46	84.57	100.0
8	79.87	79.59	99.27	86.08	85.12	99.98
9	79.86	79.80	99.98	85.46	86.02	100.0
10	79.86	79.84	99.77	86.43	84.80	99.99
AA (%)	79.84	79.79	99.64	85.22	84.96	99.95
STD	0.043	0.104	0.360	0.672	0.753	0.092

and

Table 9. The classification results of HRSIs for six types (%).

Types	Class	CNN	LBP-CNN	CNN-ELM	LBP-CNN-ELM	LBP-PCA-CNN-ELM	IPCEHRIC
1	Villages	82.34	85.46	99.46	87.45	90.35	99.98
2	Bareland	86.05	86.04	99.64	89.62	93.46	100.0
3	Grassland	79.98	85.32	99.06	87.17	90.67	100.0
4	Trees	78.46	84.14	99.31	86.43	89.86	99.81
5	Water	83.49	87.25	99.78	87.69	90.34	100.0
6	Rocks	82.16	85.68	99.34	88.03	92.05	99.85
OA (%)		79.84	79.79	99.64	85.22	84.96	99.95
AA (%)		82.08	85.65	99.43	87.73	91.12	99.94
STD		2.658	1.013	0.256	1.072	1.367	0.086

.

As can be seen from Table 6, Table 7, Table 8 and Table 9 that the IPCEHRIC obtains the classification accuracies of AA are 90.30% for four types and 99.95% for six types, respectively, which are best classification results among the CNN, LBP-CNN, CNN-ELM, LBP-CNN-ELM, LBP-PCA-CNN-ELM, and IPCEHRIC methods. The STD of the IPCEHRIC is 1.396 for four types and 0.086 for six types, which are also the least STD among these methods. Among other comparison methods, for four types of the samples, the overall classification effect of these methods is not ideal. Especially, the classification accuracies of the CNN and LBP-CNN are very unsatisfactory. For six types of the samples, the overall classification effect of these methods is better. Especially, the classification accuracies of the CNN-ELM are ideal among CNN, LBP-CNN, CNN-ELM, LBP-CNN-ELM, and LBP-PCA-CNN-ELM. Compared with the CNN-ELM, the classification accuracy of AA of the IPCEHRIC method are improved by 18.44 and 0.31%, which indicate that the optimized CNN has better feature extraction ability and classification performance, and the CWLPSO has better global optimization ability. Therefore, the experiment results show that the classification accuracy of the IPCEHRIC is better than that of other comparison methods. The CWLPSO can optimize and determine the parameters of the CNN in order to construct an optimized CNN model, which can effectively extract the deep features of HRSIs after Jiuzhaigou 7.0 earthquake, so as to obtain a better classification result. It can effectively classify the HRSIs after Jiuzhaigou 7.0 earthquake to obtain the villages, bareland, grassland, trees, water, and rocks in HRSIs after Jiuzhaigou 7.0 earthquake.

The HRSIs after Jiuzhaigou 7.0 earthquake are divided into four types and six types. The classification effects of HRSIs are shown in Figure 7.

As can be seen from Figure 7, the classification effects of six types by using the IPCEHRIC for the HRSIs after Jiuzhaigou M7.0 earthquake is ideal. For actual HRSIs, the IPCEHRIC method has higher classification accuracy, and it is an effective classification method for actual HRSIs.

7. Conclusions

In this paper, an innovative hyperspectral remote sensing image classification method based on combining CWLPSO, CNN, and ELM, namely IPCEHRIC is proposed to obtain the accurate classification results. The CWLPSO with fusing multi-strategy is proposed to optimize the parameters of the CNN. Then the deep features are extracted from HRSIs, which are input into the ELM to realize the accurate classification of HRSIs. Pavia University data and actual HRSIs after Jiuzhaigou 7.0 earthquake are selected to verify the effectiveness of the IPCEHRIC. The experiment results show that the IPCEHRIC obtains the classification accuracies of 99.21% for Pavia University data, 90.30 and 99.95% for actual HRSIs after Jiuzhaigou 7.0 earthquake. The classification results of the IPCEHRIC are better than those of the CNN, LBP-CNN, CNN-ELM, LBP-CNN-ELM, and LBP-PCA-CNN-ELM methods. Compared with the CNN-ELM, the classification accuracies of the IPCEHRIC are improved by 6.58, 21.44, and 0.31%, respectively. This shows that the CWLPSO algorithm can effectively optimize the parameters and obtain reasonable parameter values for CNN to improve the feature extraction ability. Therefore, the IPCEHRIC has certain advantages on classification effect of the HRSIs. Especially, the IPCEHRIC can obtain accurate classification accuracy for actual HRSIs after Jiuzhaigou M7.0 earthquake. It can effectively classify the villages, bareland, grassland, trees, water, and rocks in the HRSIs after Jiuzhaigou M7.0 earthquake and achieve good classification result.