CN111695467B - Spatial Spectral Fully Convolutional Hyperspectral Image Classification Method Based on Superpixel Sample Expansion - Google Patents

Abstract

The invention discloses a spatial spectrum full convolution hyperspectral image classification method based on super-pixel sample expansion, which inputs hyperspectral images; acquiring a training set and a testing set; carrying out principal component analysis and dimension reduction on the hyperspectral image; performing entropy rate segmentation on the dimension reduction result; generating a pseudo tag sample; updating the training set; carrying out data preprocessing on the hyperspectral image; inputting a convolutional neural network; training a convolutional neural network to classify hyperspectral images; repeating the above operation and voting; and outputting hyperspectral classification results. According to the method, the entropy rate super-pixel segmentation result is utilized to expand the pseudo-label sample, the priori features of the hyperspectral image are fully utilized, the number of samples is increased, the problems of network overfitting and slow network convergence speed are solved, and the accuracy of hyperspectral image classification under the condition of rare marked samples is improved.

Description

Spatial Spectral Fully Convolutional Hyperspectral Image Classification Method Based on Superpixel Sample Expansion

technical field

The invention belongs to the technical field of image processing, and in particular relates to a method for classifying hyperspectral images with spatial spectrum full convolution based on superpixel sample expansion.

Background technique

With the advancement of science and technology, hyperspectral remote sensing technology has been greatly developed. Hyperspectral data can be represented as a hyperspectral data cube, which is a three-dimensional data structure. Hyperspectral data can be regarded as a three-dimensional image, and one-dimensional spectral information is added to the ordinary two-dimensional image. Its spatial image describes the two-dimensional spatial characteristics of the earth's surface, and its spectral dimension reveals the spectral curve characteristics of each pixel in the image, thus realizing the organic fusion of remote sensing data spatial dimension and spectral dimension information. Hyperspectral remote sensing images contain rich spectral information, can provide spatial domain information and spectral domain information, have the characteristics of "integration of map and spectrum", can realize accurate identification and detail extraction of ground objects, and provide favorable conditions for understanding the objective world. Due to the unique characteristics of hyperspectral images, hyperspectral remote sensing technology has been widely used in different fields. In the civilian field, hyperspectral remote sensing images have been used in urban environmental monitoring, surface soil monitoring, geological exploration, disaster assessment, agricultural output estimation, crop analysis, etc. Hyperspectral remote sensing technology has been widely used in people's daily life. Therefore, designing a practical and efficient hyperspectral image classification method has become an indispensable scientific and technological requirement in modern society.

At present, researchers have proposed many classic classification methods for hyperspectral image classification, representative classification methods are Support Vector Machine (SVM) and Neural Network (CNN). SVM achieves better classification results in few-shot classification by maximizing the class boundaries. F.Melgani and L.Bruzzone introduced SVM into hyperspectral image classification in "Classification of hyperspectral remote sensing images with support vector machines", and achieved the best classification results at that time, but SVM does not require empirical judgment. Choosing an inappropriate kernel function can lead to poor classification performance. At the same time, with the rise of deep learning, convolutional neural networks have also been applied to the field of hyperspectral image classification. However, since the training of convolutional neural networks requires a large number of labeled samples as training samples, and the cost of hyperspectral image labeling is very expensive, how to solve the problem of small samples is currently a hot direction.

Contents of the invention

The technical problem to be solved by the present invention is to provide a space-spectrum full-convolution hyperspectral image classification method based on superpixel sample expansion, which uses the segmentation results as prior information to improve the hyperspectral image. classification effect.

The present invention adopts following technical scheme:

A method for classifying hyperspectral images based on spatial-spectrum full-convolution hyperspectral image expansion based on superpixel samples, including the following steps:

S1, input hyperspectral image PaviaU, obtain training sample X _t and test sample X _e from hyperspectral image PaviaU;

S2. Normalize the hyperspectral data set, and for each training sample, take n labels that are in its neighborhood and are the same as its segmentation label from the segmentation label matrix and add them to the training sample as pseudo-label samples;

S3. Construct the spectral feature extraction module and the spatial spectral feature extraction module respectively, construct the spectral feature map and the spatial spectral feature map weighted fusion module, use the space-spectrum combination feature as input and set it through two convolutional layers, and construct a high-level A fully convolutional neural network with space-spectrum combination for spectral classification;

S4, building the loss function of the fully convolutional neural network in step S3, and training the neural network;

S5. Obtain a final classification result map after multiple training and voting, and realize image classification.

Specifically, step S1 is specifically:

U，V，C分别为高光谱图像的空间长度，空间宽度和光谱通道数，高光谱图像包含N个像素点，每个像素点有C个光谱波段，N＝U×V；S101. Record the three-dimensional hyperspectral image PaviaU as

U, V, and C are the spatial length, spatial width, and number of spectral channels of the hyperspectral image, respectively. The hyperspectral image contains N pixels, and each pixel has C spectral bands, N=U×V;

S102. Randomly select 30 samples of each class label 1 to 9 in X to form an initial training sample set X _t , and use the rest as test samples X _e .

Specifically, step S2 is specifically:

S201. Perform PCA dimensionality reduction processing on the 3D hyperspectral image, and the number of channels of the image after dimensionality reduction is 1;

S202. Carry out entropy rate superpixel segmentation on the image after PCA dimension reduction, the segmentation result is 50 blocks, and the segmentation label matrix will be obtained as

分割标签矩阵为/>

(x₀,y₀)处的训练样本的真实标签为/>

分割图中分割标签为/>

选择(x,y)为中心的7×7的空间中满足

的任意n个样本，生成伪标签为/>

符合上述标准的伪标签样本进行扩充，此时训练样本数量变成原来的n+1倍，测试样本维持不变。S203, setting the real label matrix

Split label matrix as />

The true label of the training sample at (x ₀ ,y ₀ ) is />

The segmentation label in the segmentation map is />

Choose (x,y) as the center of the 7×7 space to satisfy

For any n samples of , the generated pseudo-label is />

The pseudo-label samples that meet the above criteria are expanded. At this time, the number of training samples becomes n+1 times the original, and the test samples remain unchanged.

Specifically, step S3 is specifically:

S301, build a spectral feature extraction module, the spectral feature extraction module includes three convolutional layers and a merging layer, each convolutional layer is followed by a relu activation function and batch normalization processing;

S302. Construct a spatial spectral feature extraction module. The spatial spectral feature extraction module includes 1×1 convolution, relu activation, batch normalization, multi-scale spatial feature fusion layer, 3×3 dilated convolution, relu activation, batch normalization , 2×2 average pooling and a pooling layer;

S303. Construct a weighted fusion module of the spectral feature map and the spatial spectral feature map;

S304. The feature of space-spectrum combination is used as input through two convolutional layers;

S305. The feature map after convolution is subjected to PCA dimensionality reduction to 5 dimensions for use in subsequent CRF processing;

S306. Perform a Softmax operation on the convoluted feature map to output a classification probability matrix, and output the dimension with the largest value in the classification probability matrix as a predicted category label to obtain a classification result.

Further, in step S301, the batch normalization processing parameters are: momentum=0.8, the convolution kernel size is 1, the step size is 1, and the number of channels after all convolutions is 64. After three consecutive convolutions, the convolution The results are summed to obtain a spectral profile.

Further, in step S302, the first convolutional layer uses 1×1 convolution with a step size of 1; the dilation rate of 3×3 atrous convolution is 2 and the step size is 1; the number of channels of all convolution results is 64 ; All batch normalization processing parameters are: momentum=0.8, the merge layer is the addition of the feature maps of the three convolution layers, and the number of channels remains 64.

Further, in step S303, the weighted superimposition of the spectral feature map and the spatial spectral feature map is as follows:

C _unite = λ _spectral C _spectral + λ _spatial C _spatial

Among them, C _unite is the weighted feature map, λ _spectral and λ _spatial are the weight coefficients of the trainable spectral features and spatial features in the network, respectively, and C _spectral and C _spatial are the spectral feature map and spatial spectral feature map, respectively.

Specifically, in step S4, the loss function is:

L＝L ₁ +L ₂

和/>

表示第i个训练样本的标签和预测标签，j取1或2代表样本为原始样本或伪标签样本。Among them, L is the final loss function, L ₁ and L ₂ are the labeled samples and pseudo-labeled samples in the training set, respectively,

and />

Indicates the label and predicted label of the i-th training sample, and j takes 1 or 2 to indicate that the sample is an original sample or a pseudo-label sample.

Specifically, step S5 is specifically:

S501, adding the classification result obtained by inputting the convolutional feature map obtained by PCA dimension reduction to 5 dimensions in step S3 and the hyperspectral data normalized in step S4 into the network, and adding the conditional random field;

S502. Obtain a classification result of one training through a conditional random field;

S503. Repeat the above m times of pseudo sample expansion and network training operations on the same training sample to obtain m times of classification results, and output the predicted class label with the most occurrences of each pixel as the final predicted class label.

Further, in step S501, the energy function of the conditional random field is as follows:

Among them, ψ _u (y _i ) and ψ _p (y _i , y _j ) are the unary function part and the binary function part respectively.

Compared with the prior art, the present invention has at least the following beneficial effects:

The present invention is a space-spectrum full-convolution hyperspectral image classification method based on superpixel sample expansion, which utilizes the results generated by hyperspectral image segmentation to guide the generation of pseudo-label samples, and effectively utilizes the prior information of hyperspectral images to expand Training samples, so that it can still maintain a good classification accuracy in the case of small samples; feature extraction using the combination of space and spectrum can more fully extract the spectral and space spectral features of hyperspectral images, thereby improving the performance of hyperspectral images. Classification accuracy rate; In the spatial feature extraction module, dilated convolutions with different dilation rates are used to achieve multi-scale feature fusion, and the spatial features of hyperspectral images are extracted on multiple scales; before the final classification result, add The voting device enhances the robustness of the entire structure, making the classification results more stable and reliable; using a fully convolutional neural network without introducing a fully connected layer, it can achieve end-to-end image classification for hyperspectral images of any size; fully convolutional The input of the neural network is the preprocessed hyperspectral data of the same size as the original hyperspectral image, which avoids the highly redundant training data caused by using the Patch of each pixel as the input data commonly used in existing methods.

Furthermore, setting the entropy rate superpixel to obtain the segmentation label effectively utilizes the prior information of the hyperspectral image, and adds samples similar to the training sample to the training set without knowing the classification label, effectively solving the hyperspectral image problem. Images available for training have difficulty with the scarcity of labeled samples.

Furthermore, the fully convolutional neural network used to construct a spatial-spectral combined fully convolutional neural network for hyperspectral classification does not have a fully connected layer, so it is very convenient to accept hyperspectral images of any size as input. The space-spectrum combination mode can combine the spectral information and spatial information in the hyperspectral image into a new feature that is better than a single spatial feature or spectral feature.

Further, the feature extraction of the spectral module uses continuous 1×1 convolutional layers to complete the extraction of spectral features while affecting the spatial information as little as possible, and the existence of the residual module allows the gradient information to be preserved so that the model can be better. of convergence.

Furthermore, the feature extraction of the spatial spectrum module uses dilated convolutions with different dilation rates, which can expand the receptive field and extract multi-scale spatial information at the same time.

Further, for the proposed pseudo-label sample expansion method, the loss function of constructing the fully convolutional neural network is improved accordingly. Adding the cross-entropy of pseudo-label samples to the loss function can make the network converge better.

Furthermore, in view of the instability of the proposed pseudo-label sample expansion method due to its own limitations, the final classification map obtained by multiple training votes effectively increases the robustness of the model.

In summary, the space-spectrum full-convolution hyperspectral image classification method based on superpixel sample expansion proposed by the present invention effectively utilizes the prior information of hyperspectral images to realize pseudo-sample expansion and solve the scarcity of marked samples in hyperspectral images At the same time, the spatial-spectrum full-volume classification network also makes full use of multi-scale spatial and spectral features to achieve higher classification accuracy.

The technical solutions of the present invention will be described in further detail below with reference to the accompanying drawings and embodiments.

Description of drawings

Fig. 1 is the realization flowchart of the present invention;

Fig. 2 is the flow chart diagram that the sample expansion of pseudo-label among the present invention realizes;

Fig. 3 is a multi-scale spatial feature fusion module in the present invention.

Detailed ways

The invention provides a space-spectrum full-convolution hyperspectral image classification method based on superpixel sample expansion, input hyperspectral images; obtain training sets and test sets; perform principal component analysis and dimensionality reduction on hyperspectral images; dimensionality reduction As a result, perform entropy superpixel segmentation; generate pseudo-label samples; update the training set; perform data preprocessing on hyperspectral images; input fully convolutional neural networks; train fully convolutional neural networks to classify hyperspectral images; repeat the above operations And vote; output hyperspectral classification results. The present invention uses entropy rate superpixel segmentation results to expand pseudo-label samples, fully utilizes the spatial prior information of hyperspectral images, increases the number of samples, alleviates the problem of network over-fitting, and effectively improves high-resolution images under small-sample conditions. Spectral Image Classification Accuracy, Classification Efficiency and Classification Performance.

Please refer to Fig. 1, a kind of space spectrum full convolution hyperspectral image classification method based on superpixel sample expansion of the present invention, comprises the following steps:

S1, input hyperspectral image PaviaU, this hyperspectral three-dimensional image, is the data used in the experiment of the present invention, obtains training sample X _t and test sample X _e from PaviaU hyperspectral image;

其中U，V，C分别是高光谱图像的空间长度，空间宽度和光谱通道数，高光谱图像包含N个像素点，每个像素点有C个光谱波段，其中N＝U×V。PaviaU数据集中N＝207400个样本，U＝610,V＝340,C＝103。类别标签为1到9的样本共有42776个。对X进行归一化，将其中数据数值保持在[0，1]之间，具体如下：S101. Record the three-dimensional hyperspectral image PaviaU as

Where U, V, and C are the spatial length, spatial width, and number of spectral channels of the hyperspectral image, respectively. The hyperspectral image contains N pixels, and each pixel has C spectral bands, where N=U×V. N=207400 samples in PaviaU dataset, U=610, V=340, C=103. There are a total of 42776 samples with class labels from 1 to 9. Normalize X and keep the data value in it between [0, 1], as follows:

；

;

S102. Randomly select 30 samples from X with category labels 1 to 9 for each category to form an initial training sample set X _t , and use the rest as test samples X _e .

S2. After normalizing the hyperspectral data set, use the entropy rate superpixel to generate segmentation labels to generate pseudo-label samples;

S201. Perform PCA dimensionality reduction processing on the 3D hyperspectral image, and the number of channels of the image after dimensionality reduction is 1;

S202. Carry out entropy rate superpixel segmentation on the image after PCA dimension reduction, the segmentation result is 50 blocks, and the segmentation label matrix will be obtained as

分割标签矩阵为/>

(x₀,y₀)处的训练样本的真实标签为/>

分割图中分割标签为/>

选择(x,y)为中心的7×7的空间中满足

的任意n个样本，为其生成伪标签为/>

符合上述标准的伪标签样本进行扩充，此时训练样本数量变成了原来的n+1倍，测试样本维持不变，如图2所示。S203, setting the real label matrix

Split label matrix as />

The true label of the training sample at (x ₀ ,y ₀ ) is />

The segmentation label in the segmentation map is />

Choose (x,y) as the center of the 7×7 space to satisfy

For any n samples of , generate a pseudo-label for it

The pseudo-label samples that meet the above criteria are expanded, and the number of training samples becomes n+1 times the original, while the test samples remain unchanged, as shown in Figure 2.

S3. Construct a fully convolutional neural network for space-spectrum combination for hyperspectral classification;

S301. The spectral feature extraction module is composed of three convolutional layers, each convolutional layer is followed by a relu activation function and batch normalization processing.

The batch normalization processing parameters are:

momentum=0.8, the size of the convolution kernel is 1, the step size is 1, and the number of channels after all convolutions is 64. After three consecutive convolutions, the convolution results are added to obtain the spectral feature map.

S302. The spatial spectral feature extraction module is composed of three convolutional layers. The convolution kernel size of the first convolutional layer is 1, and the step size is 1; the second convolutional layer is used to realize multi-scale feature fusion, and the structure is as follows Figure 3 is obtained by adding three convolution kernels with a size of 3, a dilation rate of 2, 3, 4, and a step size of 1; a 2×2 average pooling layer is added after the third convolution layer.

After each convolution is completed, the relu activation function and batch normalization are processed; the convolution kernel size of the third convolution layer is 3, the dilation rate is 2, and the step size is 1.

All batch normalization processing parameters are:

momentum=0.8, the number of channels of all convolution results is 64, and the convolution results are added after three consecutive convolutions to obtain a spatial spectral feature map.

S303. Weighted and superimposed the spectral characteristic map and the spatial spectral characteristic map as follows:

C _unite = λ _spectral C _spectral + λ _spatial C _spatial

Among them, C _unite is the weighted feature map, and its channel is still 64, λ _spectral and λ _spatial are the weight coefficients of the spectral features and spatial features that can be trained in the network, respectively, C _spectral and C _spatial are the spectral feature map and empty spectrogram;

S304. The spatial-spectral combination feature is used as input through two 1×1 convolutional layers, and relu activation is performed after convolution. The convolution kernel size is 1, the step size is 1, and the number of channels of the first convolution result is 64. , the second is 128;

S305. The feature map after convolution is subjected to PCA dimensionality reduction to 5 dimensions for use in subsequent CRF processing;

S306. Perform Softmax operation on the convoluted feature map to output a classification probability matrix of 610×340×9, and output the dimension with the largest value among the 9 dimensions as the predicted category label to obtain a classification result with a size of 610×340.

S4. Construct the loss function of the fully convolutional neural network and train the neural network;

S401. The loss function uses cross entropy to calculate the cross entropy of the expanded training sample prediction label and the training sample label, as shown in the following formula:

L＝L ₁ +L ₂

和/>

表示第i个训练样本的标签和预测标签，j取1或2代表该样本为原始样本或伪标签样本；Among them, L is the final loss function, L ₁ and L ₂ are the labeled samples and pseudo-labeled samples in the training set, respectively,

and />

Indicates the label and predicted label of the i-th training sample, and j takes 1 or 2 to indicate that the sample is an original sample or a pseudo-label sample;

S402. Input the normalized hyperspectral data into the network, and iterate 1000 times to generate a predicted label map.

S5. Multiple training votes to obtain a final classification result map.

S501, adding the output of step S305 and step S402 into the conditional random field, the energy function of the conditional random field is as follows,

ψ _u (y _i ) and ψ _p (y _i ,y _j ) are the unary function part and the binary function part respectively.

In the present invention, the unary part is calculated as ψ _u (y _i )=-log P(y _i ), where P(y _i ) is the label assignment probability of pixel i given by the proposed fully convolutional network .

The binary function part is defined as:

Among them, if y _i =y _j , μ(y _i ,y _j )=1, otherwise zero; k _m is a Gaussian kernel; f _i and f _j are the feature vectors of pixels i and j in any feature space; ω _m is the corresponding weight; in order to fully utilize the deep spectral spatial features, the first five principal components of C _unite in S305 are used as the features of each pixel. The full form of the Gaussian kernel is written as:

；

;

S502. Obtain a classification result of one training through a conditional random field;

S503. Repeat the above operations m times on the same training sample to obtain m times of classification results, and output the predicted class label with the most occurrences of each pixel as the final predicted class label.

In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments It is a part of embodiments of the present invention, but not all embodiments. The components of the embodiments of the invention generally described and illustrated in the drawings herein may be arranged and designed in a variety of different configurations. Accordingly, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the invention. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

The three evaluation indicators of the simulation experiment are as follows:

The total accuracy OA indicates the proportion of correctly classified samples to all samples, and the larger the value, the better the classification effect. The average accuracy AA represents the average value of the classification accuracy of each category, and the larger the value, the better the classification effect. The chi-square coefficient Kappa represents different weights in the confusion matrix, and the larger the value, the better the classification effect.

The prior art comparative classification method used in the present invention is as follows:

The hyperspectral image classification method proposed by W.Song et al. in "Hyperspectral Image Classification With Deep FeatureFusion Network, IEEE Trans.Geosci.Remote Sens., vol.56, no.6, pp.3173-3184, June2018", referred to as DFFN method for deep feature fusion.

Table 1 is the classification result quantitative analysis table of the present invention (choose PaviaU data set, every kind of 30 marked samples are used as training set:

In summary, the present invention is a space-spectrum full-convolution hyperspectral image classification method based on superpixel sample expansion, which effectively utilizes the prior information of hyperspectral images to realize pseudo-sample expansion and solves the problem of marked samples in hyperspectral images. At the same time, the spatial-spectrum full-volume classification network also makes full use of multi-scale spatial and spectral features to achieve higher classification accuracy.

The above content is only to illustrate the technical ideas of the present invention, and cannot limit the protection scope of the present invention. Any changes made on the basis of the technical solutions according to the technical ideas proposed in the present invention shall fall within the scope of the claims of the present invention. within the scope of protection.

Claims (9)

1. The space-spectrum full-convolution hyperspectral image classification method based on superpixel sample expansion, is characterized in that, comprises the following steps: S1, input hyperspectral image PaviaU, obtain training sample X _t and test sample X _e from hyperspectral image PaviaU; S2. Normalize the hyperspectral data set, and for each training sample, take n labels that are in its neighborhood and are the same as its segmentation label from the segmentation label matrix and add them to the training sample as pseudo-label samples. Specifically: S201. Perform PCA dimensionality reduction processing on the 3D hyperspectral image, and the number of channels of the image after dimensionality reduction is 1; S202. Carry out entropy rate superpixel segmentation on the image after PCA dimension reduction, the segmentation result is 50 blocks, and the segmentation label matrix will be obtained as

S203, setting the real label matrix

Split label matrix as />

The true label of the training sample at (x ₀ ,y ₀ ) is />

The segmentation label in the segmentation map is />

Choose (x,y) as the center of the 7×7 space to satisfy />

For any n samples of , the generated pseudo-label is />

The pseudo-label samples that meet the above criteria are expanded. At this time, the number of training samples becomes n+1 times the original, and the test samples remain unchanged; S3. Construct the spectral feature extraction module and the spatial spectral feature extraction module respectively, construct the spectral feature map and the spatial spectral feature map weighted fusion module, use the space-spectrum combination feature as input and set it through two convolutional layers, and construct a high-level A fully convolutional neural network with space-spectrum combination for spectral classification; S4, building the loss function of the fully convolutional neural network in step S3, and training the neural network; S5. Obtain a final classification result map after multiple training and voting, and realize image classification. 2. The space-spectrum full-convolution hyperspectral image classification method based on superpixel sample expansion according to claim 1, wherein step S1 is specifically: S101. Record the three-dimensional hyperspectral image PaviaU as

U, V, and C are the spatial length, spatial width, and number of spectral channels of the hyperspectral image, respectively. The hyperspectral image contains N pixels, and each pixel has C spectral bands, N=U×V; S102. Randomly select 30 samples of each class label 1 to 9 in X to form an initial training sample set X _t , and use the rest as test samples X _e . 3. The space-spectrum full-convolution hyperspectral image classification method based on superpixel sample expansion according to claim 1, wherein step S3 is specifically: S301, build a spectral feature extraction module, the spectral feature extraction module includes three convolutional layers and a merging layer, each convolutional layer is followed by a relu activation function and batch normalization processing; S302. Construct a spatial spectral feature extraction module. The spatial spectral feature extraction module includes 1×1 convolution, relu activation, batch normalization, multi-scale spatial feature fusion layer, 3×3 dilated convolution, relu activation, batch normalization , 2×2 average pooling and a pooling layer; S303. Construct a weighted fusion module of the spectral feature map and the spatial spectral feature map; S304. The feature of space-spectrum combination is used as input through two convolutional layers; S305. The feature map after convolution is subjected to PCA dimensionality reduction to 5 dimensions for use in subsequent CRF processing; S306. Perform a Softmax operation on the convoluted feature map to output a classification probability matrix, and output the dimension with the largest value in the classification probability matrix as a predicted category label to obtain a classification result. 4. The space-spectrum full-convolution hyperspectral image classification method based on superpixel sample expansion according to claim 3, characterized in that, in step S301, the batch normalization processing parameters are: momentum=0.8, convolution kernel size All are 1, the step size is 1, and the number of channels after all convolutions is 64. After three consecutive convolutions, the convolution results are added to obtain the spectral feature map. 5. The space-spectrum full-convolution hyperspectral image classification method based on superpixel sample expansion according to claim 3, wherein in step S302, the first convolutional layer uses 1×1 convolution with a step size of 1; the dilationrate of 3×3 hole convolution is 2, and the step size is 1; the number of channels of all convolution results is 64; all batch normalization processing parameters are: momentum=0.8, and the merge layer is three convolution layers The feature maps are added, and the number of channels remains 64. 6. The space-spectrum full-convolution hyperspectral image classification method based on superpixel sample expansion according to claim 3, characterized in that, in step S303, the spectral feature map and the space-spectrum feature map are weighted and superimposed as follows: C _unite = λ _spectral C _spectral + λ _spatial C _spatial Among them, C _unite is the weighted feature map, λ _spectral and λ _spatial are the weight coefficients of the trainable spectral features and spatial features in the network, respectively, and C _spectral and C _spatial are the spectral feature map and spatial spectral feature map, respectively. 7. The space-spectrum full-convolution hyperspectral image classification method based on superpixel sample expansion according to claim 1, wherein, in step S4, the loss function is: L＝L ₁ +L ₂

Among them, L is the final loss function, L ₁ and L ₂ are the labeled samples and pseudo-labeled samples in the training set, respectively, y _j ⁽ⁱ⁾ and

Indicates the label and predicted label of the i-th training sample, and j takes 1 or 2 to indicate that the sample is an original sample or a pseudo-label sample. 8. The space-spectrum full-convolution hyperspectral image classification method based on superpixel sample expansion according to claim 1, wherein step S5 is specifically: S501, adding the classification result obtained by inputting the convolutional feature map obtained by PCA dimension reduction to 5 dimensions in step S3 and the hyperspectral data normalized in step S4 into the network, and adding the conditional random field; S502. Obtain a classification result of one training through a conditional random field; S503. Repeat the above m times of pseudo sample expansion and network training operations on the same training sample to obtain m times of classification results, and output the predicted class label with the most occurrences of each pixel as the final predicted class label. 9. The space-spectrum full-convolution hyperspectral image classification method based on superpixel sample expansion according to claim 8, characterized in that, in step S501, the energy function of the conditional random field is as follows:

Among them, ψ _u (y _i ) and ψ _p (y _i , y _j ) are the unary function part and the binary function part respectively.

Images