CN111985543A - Construction method, classification method and system of hyperspectral image classification model - Google Patents

Abstract

The invention provides a method for constructing a hyperspectral image classification model, an image classification method and a system, wherein the method for constructing the hyperspectral image classification model comprises the following steps: acquiring a hyperspectral image, and standardizing the hyperspectral image to enable the hyperspectral image to accord with the Gaussian distribution of unit variance of 0 mean value; performing category calibration to obtain a calibration sample set and a label set; constructing a neighborhood data cube, and classifying according to a set proportion to obtain a training sample set, a verification sample set and a test sample set; the hyperspectral image classification method provided by the invention adopts a cascade network structure which extracts spectral attention features and then spatial attention features, so that the network focuses more on interested spatial regions and meaningful spectral bands, the spectral features and the spatial features are combined, abundant spectral information and spatial information are fully utilized, therefore, higher classification precision can be obtained, and the obtained classified images are more continuous visually.

Description

Construction method, classification method and system of a hyperspectral image classification model

technical field

The invention belongs to the technical field of hyperspectral image processing, and in particular relates to a construction method, a classification method and a system of a hyperspectral image classification model.

Background technique

A hyperspectral image is usually regarded as a three-dimensional data cube, in addition to the two dimensions of width and height in space, there are often hundreds of bands of spectral information. In addition to the visible spectrum, these bands also include information in ultraviolet and near-infrared bands. Using this rich information, people can find targets of interest that are not visible in visible light, and provide strong technical support for humans to learn the world and change the world. Therefore, there are There is an increasing number of studies on hyperspectral images.

Hyperspectral image classification is a research hotspot in hyperspectral image research, which refers to the pixel-by-pixel classification of images using the spatial and spectral information of hyperspectral images. Early methods include support vector machine-based classification methods, principal component analysis dimensionality reduction and reclassification methods, and sparse representation-based methods. Since these methods cannot make good use of the depth information of images, they often lead to low classification accuracy.

In recent years, deep learning-based methods have achieved great success in image classification and object detection, so many deep learning-based hyperspectral image classification methods have emerged. Commonly used deep learning networks include autoencoders, deep belief networks and convolutional neural networks, as well as networks that combine residual blocks and transfer learning. Although the existing deep learning-based methods can achieve better classification accuracy, most of the existing methods need more training samples, the network complexity is high, the network depth is deep, and there are many parameters that need to be trained. . In practical applications, the calibration samples of hyperspectral images are often limited, so the classification accuracy of the existing methods needs to be improved, and the robustness to different datasets needs to be enhanced.

SUMMARY OF THE INVENTION

Aiming at the above-mentioned deficiencies and defects of the prior art, the purpose of the present invention is to provide a method for constructing a hyperspectral image classification model, an image classification method and a system, so as to solve the problem that the method in the prior art is insufficient for hyperspectral image feature extraction, This leads to the problem of low accuracy of hyperspectral image classification.

In order to solve the above-mentioned technical problems, the present invention adopts the following technical solutions to achieve: a construction method of a hyperspectral image classification model, the construction method is:

Step 1, obtaining a hyperspectral image, and standardizing the hyperspectral image, so that the hyperspectral image conforms to a Gaussian distribution with 0 mean and unit variance;

Step 2, perform category calibration on the ground object category corresponding to the pixels of the standardized hyperspectral image, and obtain a calibration sample set and a label set;

Step 3: Construct a neighborhood data cube with the pixel of each sample in the calibration sample set as the center, and the dimension of the neighborhood data cube is w*w*B, where w*w represents the spatial neighborhood of the pixel, and B is the hyperspectral spectrum. The initial number of bands of the image, the constructed neighborhood data cube is divided according to the set ratio, and the training sample set, the verification sample set and the test sample set are obtained;

Step 4: Use the neighborhood data cube corresponding to each sample in the training sample set and the verification sample set as the input of the network, and use the label set as the expected output to train the network to obtain a trained network model;

The network includes a network input layer, a data dimensionality reduction module, a spectral attention module, a secondary data dimensionality reduction module, a spatial attention mechanism module, a pooling module and a fully connected network that are set in series in sequence;

The data dimensionality reduction module includes two convolution modules connected in sequence, and the convolution module includes a convolution layer, a batch normalization layer and an activation function layer connected in sequence;

The spectral attention module includes a plurality of spectral attention feature layers connected in sequence;

The secondary data dimensionality reduction module includes a convolution layer;

The spatial attention mechanism module includes a plurality of spatial attention feature layers connected in sequence.

If a boundary problem is encountered when constructing the neighborhood data cube in the third step, the boundary is expanded by using the method of filling the 0 boundary.

The concrete method of described step 4 is:

The neighborhood data cube corresponding to each sample in the training sample set and the validation sample set is used as the input of the network, and the label set is used as the expected output, and the cross entropy loss between the true label of the training sample and the predicted output value of the network is calculated, and the cross entropy loss will be minimized. The entropy loss is used as the optimization target, and the adaptive moment estimation algorithm is used to optimize the network, and the obtained network model is tested on the validation set. The model with the highest classification accuracy on the validation set obtains the trained network model, and uses the test sample set to evaluate the classification performance of the trained network model.

A hyperspectral image classification method, comprising the following steps:

Obtain the hyperspectral image set to be classified, input the network model, use the network model to classify each pixel of the hyperspectral image, and obtain the final classification result;

The network model is a hyperspectral image classification model constructed by the above-mentioned method for constructing a hyperspectral image classification model.

A hyperspectral image classification system, the classification system includes an image acquisition module, a data preprocessing module and a classification module;

The image acquisition module is used to acquire hyperspectral images;

The data preprocessing module is used to standardize the hyperspectral image, so that the input image conforms to the Gaussian distribution with 0 mean and unit variance;

The classification module adopts the aforementioned method to classify the hyperspectral image and outputs the classification result.

A method for visualizing spatial attention features, including the following steps:

Step A, according to the constructed hyperspectral image classification model and the constructed hyperspectral image classification system, extract the network layer label L where the spatial attention feature is located, and the label L-1 of the previous layer;

Step B, call the L-1 layer network parameter M1 and the L layer network parameter M2 in the hyperspectral image classification model, take the hyperspectral image as the input image, use the parameter M1 to obtain the image before the attention feature extraction, and use the parameter M2 to obtain The image after the attention feature is extracted to realize the visualization of the attention feature;

The hyperspectral image classification model is a hyperspectral image classification model constructed by the method for constructing the hyperspectral image classification model described above;

The hyperspectral image classification system is the above-mentioned image classification system.

Compared with the prior art, the present invention has the following technical effects:

(I) When classifying hyperspectral images, only using spectral features will be affected by "similar objects with different spectra" and "different objects with the same spectrum", and only using spatial features will lose rich spectral information, both of which will lead to Therefore, the hyperspectral image classification technology provided by the present invention adopts a cascade network structure that extracts spectral features first and then spatial features, combines spectral features and spatial features, and makes full use of the rich spectral information and space. Therefore, higher classification accuracy can be obtained, and the obtained classification images are more visually continuous;

(II) The attention mechanism adopted in the present invention enables the network to assign different weights to different spectral bands and spatial regions, which is more in line with the human visual mechanism, and more meaningful features can be obtained. More meaningful features, this feature makes the network more advantageous in the case of small sample training sets;

(III) The present invention adopts an end-to-end network structure. The entire network adopts 3D CNN in three modules: data dimensionality reduction, spectral attention feature extraction, and spatial attention feature extraction, making full use of the three-dimensional characteristics of hyperspectral images. , compared with the prior art using different technologies and networks in operations such as data dimensionality reduction, spectral feature extraction, spatial feature extraction, etc., the network structure of the present invention is simple and easy to implement;

(IV) The present invention visualizes the spatial attention feature in the deep network. The previous deep learning network is similar to a black box, so the role of each layer is not easy to understand. By visualizing the spatial attention feature, the present invention can be more Intuitively understand the function of the attention mechanism.

Description of drawings

1 is a frame diagram of a hyperspectral image classification model of the present invention;

Fig. 2 is the structure diagram of the spectral attention module of the present invention;

3 is a structural diagram of a spatial attention module of the present invention;

4 is a schematic diagram of the construction of a hyperspectral image classification model of the present invention;

Fig. 5 is the experimental result comparison diagram provided in the embodiment of the present invention;

FIG. 6 is an experimental result diagram of an embodiment of the visualization of spatial attention features proposed by the present invention.

The specific content of the present invention will be further explained in detail below in conjunction with the embodiments.

Detailed ways

Specific embodiments of the present invention are given below. It should be noted that the present invention is not limited to the following specific embodiments, and all equivalent transformations made on the basis of the technical solutions of the present application fall into the protection scope of the present invention.

First, the technical terms that appear in the present invention are explained to help better understand the technical content of the application:

Attention mechanism: The attention mechanism is inspired by the human visual mechanism. When humans visually observe a scene, they do not pay equal attention to all areas in the scene, but pay more attention to the target area of interest to obtain More detailed information about the target that needs to be focused on, while suppressing other useless information.

BN layer: Batch normalization layer, an effective regularization method. Like the activation function layer, convolution layer, fully connected layer, and pooling layer, it is also a layer of the network. The BN layer can solve the problem that the data distribution of the intermediate layer changes during the training process, can improve the gradient flowing through the network, greatly improve the training speed, and reduce the network's strong dependence on initialization.

In the present invention, the convolution layer is used many times during the construction of the deep network. In order to realize different functions at different stages of the network, different parameters need to be used when implementing the convolution operation. Now, the different convolution layers in the present invention are used. The main parameters are as follows:

Number of filters: that is, the number of feature maps obtained after the convolution operation, and the present invention adopts 24 uniformly;

Convolution kernel size: The present invention adopts 3D CNN, and the size of the convolution kernel is represented by a three-dimensional vector, and the three main functional modules of the network adopt different convolution kernel sizes, which are specifically described in the embodiments;

Step size: used to specify the convolution stride in 3 dimensions. In the present invention, only the first convolution layer of the data dimension reduction module adopts (1, 1, 2), see Figure 1, and the other convolution layers are uniformly adopted default(1,1,1);

Expansion rate: specifies the expansion rate for dilated convolution. In the present invention, only the convolutional layer of the spectral attention module adopts (1, 1, 2), see Figure 2, and the other convolutional layers uniformly use the default value (1, 1,1);

Boundary processing: whether to fill the boundary during the convolution operation. In the invention, both the spectral attention module and the spatial attention module adopt the boundary filling strategy, and the remaining convolution layers do not perform boundary filling by default.

In the specific implementation, the parameters that are not specially described are all set to default values.

Example 1:

A construction method of a hyperspectral image classification model, the construction method is:

Step 1, obtaining a hyperspectral image, the initial number of bands of the hyperspectral image is B, and standardizing the hyperspectral image so that the hyperspectral image conforms to a Gaussian distribution with 0 mean unit variance;

Step 2, perform category calibration on the ground object category corresponding to the pixels of the standardized hyperspectral image, and obtain a calibration sample set and a label set;

In this example, the data set used is a scene over the valley of Salinas, California, captured by an AVIRIS sensor with 224 bands, with high spatial resolution (3.7 m/pixel). The covered area consists of 512 rows of 217 samples each. The 20 water absorption bands [108-112], [154-167], 224 are deprecated and 204 bands are actually used, so the dimension of the actual dataset used is 521×217×204. It includes categories such as vegetables, bare land and vineyards. 16 categories are calibrated, and the total number of calibration samples is 54129. The specific category names and the number of calibration samples are shown in Table 1.

Table 1 Salinas dataset and classification of training set, validation set and test set

No. category Calibration sample Training samples Validation samples test sample 1 Brocoli_green_weeds_1 2009 twenty one twenty one 1967 2 Brocoli_green_weeds_2 3726 38 38 3650 3 Fallow 1976 20 20 1936 4 Fallow_rough_plow 1394 14 14 1366 5 Fallow_smooth 2678 27 27 2624 6 Stubble 3959 40 40 3879 7 Celery 3579 36 36 3507 8 Grapes_untrained 11271 113 113 11045 9 Soil_vinyard_develop 6203 63 63 6077 10 Corn_senesced_green_weeds 3278 33 33 3212 11 Lettuce_romaine_4wk 1068 11 11 1046 12 Lettuce_romaine_5wk 1927 20 20 1887 13 Lettuce_romaine_6wk 916 10 10 896 14 Lettuce_romaine_7wk 1070 11 11 1048 15 Vinyard_untrained 7268 73 73 7122 16 Vinyard_vertical_trellis 1807 19 19 1769 Total 54129 549 549 53031

Step 3: Construct a neighborhood data cube with the pixel of each sample in the calibration sample set as the center. The dimension of the neighborhood data cube is 7*7*204, where 7*7 represents the spatial neighborhood of the center pixel, and 204 is the initial The number of bands, divide the constructed neighborhood data cube according to the set ratio, and obtain the training sample set, the verification sample set and the test sample set;

The ratio parameter train ratio=1% of the training samples is set, that is, fewer samples are used for training. This embodiment can measure the classification performance of the present invention under the small sample training set; the number of samples in the verification set is the same as the number of samples in the training set ;

The number of batch samples during training batchsize=50, and the number of iterations epoch=100;

The calibration samples (pixels) are classified according to the specified ratio, the training sample set accounts for 1%, the verification sample set accounts for 1%, and the remaining 98% of the calibration samples are divided into the test sample set;

Step 4: Use the neighborhood data cube corresponding to each sample in the training sample set and the verification sample set as the input of the network, and use the label set as the expected output to train the network to obtain a trained network model;

The neighborhood data cube corresponding to each sample in the training sample set and the validation sample set is used as the input of the network, and the label set is used as the expected output, and the cross entropy loss between the true label of the training sample and the predicted output value of the network is calculated, and the cross entropy loss will be minimized. The entropy loss is used as the optimization goal, and the adaptive moment estimation algorithm is used to optimize the network, and the obtained network model is tested on the validation set. If the obtained classification accuracy is better than the previous network model, the network model parameters are updated until the iteration is over, and saved in The model with the highest classification accuracy on the validation set is used to obtain the trained network model, and the test sample set is used to evaluate the classification performance of the trained network model.

If the boundary problem is encountered when building the neighborhood cube of the calibration sample, the boundary is expanded with 0 boundary filling to avoid the problem of boundary crossing.

The network includes a network input layer, a data dimensionality reduction module, a spectral attention module, a secondary data dimensionality reduction module, a spatial attention mechanism module, a pooling module and a fully connected network that are set in series in sequence;

The input of the network model is a 7*7*204 data cube, and its corresponding label is the class label corresponding to the center pixel of the cube, where 7*7 represents the spatial neighborhood of the center pixel, and 204 represents the initial number of bands of the hyperspectral image;

Perform the first convolution operation on the spectral dimension (3rd dimension) of the input 7*7*204 data cube, the size of the convolution kernel is 1*1*7, the step size of the convolution operation is straids=2, extract The first layer of shallow features of the data, and the spectral dimension of the data is reduced;

The second convolution operation is performed on the data after dimension reduction on the third dimension. The size of the convolution kernel is 1*1*7. No boundary padding is performed during convolution, and only valid data is convolved, so that the first convolution can be extracted. The second layer of shallow features, and the spectral dimension of the data is reduced again, the output of this layer is used as the input of the spectral attention module to extract the spectral dimension depth features;

The data dimensionality reduction module includes two convolution modules connected in sequence, and the convolution module includes a convolution layer, a BN layer and an activation function layer connected in sequence; the input 7*7*204 data cube is processed. Two 3-dimensional convolutions, the first convolution sets a step size greater than 1 in the spectral dimension (the third dimension), which greatly reduces the dimension; the convolution operations are all three-dimensional convolutions, and the size of the convolution kernel is 1*1* 7. The step size is straids=(1,1,2), extract the first layer of shallow features of the data, and reduce the spectral dimension of the data to (int(B-7+1)/2), int means downward Rounding;

Perform the second convolution operation on the spectral dimension of the above-mentioned dimensionally reduced data, and the size of the convolution kernel is 1*1*7, so that the second layer of shallow features can be extracted, and the spectral dimension of the data can be reduced to (int again) (B-7+1)/2-7+1), let TK=int(B-7+1)/2-7+1, the output is used as the input of the spectral attention module to extract deep spectral features; so The convolution operations described above are followed by a BN layer and a ReLu activation layer.

The spectral attention module includes a plurality of spectral attention feature layers connected in sequence; a 1*1*3 convolution kernel is used to extract spectral features for the input data, and a dilation rate of 2 is used in the third dimension during convolution. Dilated convolutions, so that fewer convolution kernels can be used to cover a larger convolutional field of view;

First perform hyperbolic tangent transformation on the extracted spectral features, and then convert the extracted features into a weight value from 0 to 1 through the softmax activation function, and the weight value represents the degree to which different spectral bands need to be paid attention to;

Perform element-level multiplication of the obtained weight value and the input value to obtain a layer of spectral attention features, after which a batch normalized BN layer and ReLU activation layer are added to obtain the output of the first layer of spectral attention network;

The output of the first layer of spectral attention network is used as the input of the next layer, and the above operations are repeated to obtain the output of the second layer of spectral attention network, and so on, the output of the multi-layer spectral attention network can be obtained, these values represent spectral attention features of different depths;

Summing the obtained multi-layer spectral attention features, the result fuses spectral attention features of different depths, which helps to improve the classification accuracy of images;

The second data dimension reduction module includes a convolution layer; the second convolution does not perform boundary filling in the spectral dimension, and reduces the dimension slightly. The output dimension of the multi-layer spectral attention network is 7*7*TK, and the convolution operation using the convolution kernel size (1,1,TK) can convert the third dimension of the data to 1, that is, all spectral features are concentrated In one dimension, the dimension of the resulting data is 7*7*1, which is ready for subsequent spatial feature extraction;

Perform a convolution operation, which converts the third dimension of the data to 1, that is, concentrating all spectral features on one band, and the resulting data has a dimension of 7*7*1, which is ready for subsequent spatial feature extraction; The convolution operations are all three-dimensional convolutions, the size of the convolution kernel is 1*1*TK, and the dilation rate of dilation rate>1 is used in the third dimension during convolution. The output dimension of the previous step is 7*7*TK, and the convolution operation of the convolution kernel size (1,1,TK) can convert the third dimension of the data to 1, that is, focus all spectral features on one dimension , the dimension of the resulting data is 7*7*1, which is ready for the subsequent spatial feature extraction;

The spatial attention mechanism module includes a plurality of spatial attention feature layers connected in sequence;

Perform a 3-dimensional convolution operation on the input data, and use a 3*3*1 convolution kernel to extract spatial features;

First perform hyperbolic tangent transformation on the extracted spatial features, and then convert the extracted spatial features into weight values from 0 to 1 through the softmax activation function. These values represent the different regions (or pixels) in space under the attention mechanism. The degree of attention, and these values are called the weights of spatial attention features;

Multiply the obtained weight value and the input value of the extracted spatial feature pixel by pixel to obtain a layer of spatial attention features, after which a BN layer and a ReLu activation layer are added to obtain the first layer of spatial attention network. output;

The output of the first layer of spatial attention network is used as the input of the next layer, and the above operation is repeated to obtain the output of the second layer of spatial attention network, and so on, the output of the multi-layer spatial attention network can be obtained. These values represent spatial attention features of different depths;

The obtained multi-layer spatial attention features are summed, and the spatial attention features of different depths are fused; this helps to improve the classification accuracy of the image; the convolution operations are all three-dimensional convolutions, and the size of the convolution kernel is 3*3*1.

Average pooling is performed on the above results, and the pooling results are input into the fully connected network to obtain the final classification result.

Further, the convolution operations are all three-dimensional convolutions, and the convolution operations are followed by a batch normalized BN layer and a ReLu activation layer.

This embodiment also discloses a hyperspectral image classification method, comprising the following steps:

Obtain the hyperspectral image set to be classified, input the network model, use the network model to classify each pixel of the hyperspectral image, and obtain the final classification result;

The network model is a hyperspectral image classification model constructed by the method for constructing a hyperspectral image classification model described above.

The classification quality evaluation indicators mainly use OA (Overall Accuracy), AA (Average Accuracy) and Kappa coefficient to evaluate the percentage of overall correctly classified pixels, the average and The ratio of error reductions produced by classification versus completely random classification, the higher the better for all three metrics. Table 2 shows the results provided by the present invention and the results of other comparison methods. Each result is the average value and variance of 10 random experiments. The higher the average value, the better the classification performance, and the smaller the variance, the better the classification method. more stable. It can be seen from the comparison that the present invention has obtained good results in various indicators.

Table 2 Comparison of classification effects obtained by different methods

SSFCN HbridSN SSRN CDSCN Spe_AN Spa_AN this invention AA 95.13±0.93 97.84±0.46 97.94±0.44 97.54±1.08 97.41±2.68 98.69±0.27 98.91±0.34 Kappa 94.75±1.23 96.38±0.54 95.11±1.59 94.91±1.59 96.57±1.55 97.77±0.44 98.17±0.44 OA 95.29±1.1 96.75±0.49 95.61±1.41 95.43±1.43 96.92±1.39 98±0.39 98.35±0.39

The classification results are shown in Figure 5. In Fig. 5, Fig. 5(a) is the false color image of the hyperspectral image, Fig. 5(b) is the calibration image, Fig. 5(c) is the classification image of the SSFCA (spectral-spatial fully convolutional neural network) method, and Fig. 5(d) is the image classified by the HbridSN (Hybrid Spectral Convolutional Neural Network) method, Figure 5(e) is the image classified by the SSRN (Spectral-Spatial Residual Network) method, and Figure 5(f) is the CDSCN (Cascaded Dual Scale Crossover). Figure 5(g) is the image classified by the Spe_AN (only spectral attention feature network) method, Figure 5(h) is the image classified by the Spa_AN (only spatial attention feature network) method, Figure 5(i) ) is the classification method provided by the present invention to classify images, and it can be clearly seen that the classification results provided by the present invention are continuous, have fewer noise points, and achieve better classification results.

A visualization method for spatial attention features, which is performed as follows:

Step A: According to the constructed hyperspectral image classification model and the constructed hyperspectral image classification system, the best saved network model is denoted by M, the label of the network layer where the spatial attention feature is extracted is denoted by L, and its previous layer network is denoted by M. Use L-1 for the label;

In step B, the L-1 layer network parameter M1 and the L layer network parameter M2 are called in the hyperspectral image classification model, the hyperspectral image is used as the input image, and the parameter M1 is used to obtain the image before attention feature extraction, see Figure 6. (a) Use the parameter M2 to obtain the image after the attention feature extraction to realize the visualization of the attention feature; see Figure 6.(b), it can be seen from the figure that the attention feature is more concerned with the boundary lines or the boundaries of different regions in the image. Rich feature regions such as inflection points.

Claims (5)

1. A construction method of a hyperspectral image classification model is characterized by comprising the following steps:

acquiring a hyperspectral image, and standardizing the hyperspectral image to enable the hyperspectral image to accord with the Gaussian distribution of unit variance of 0 mean value;

secondly, performing category calibration on the ground object categories corresponding to the pixels of the standardized hyperspectral image to obtain a calibration sample set and a label set;

constructing a neighborhood data cube by taking the pixel of each sample in the calibration sample set as a center, wherein the dimension of the neighborhood data cube is w W B, w represents the spatial neighborhood of the pixel, B is the initial wave band number of the hyperspectral image, and dividing the constructed neighborhood data cube according to a set proportion to obtain a training sample set, a verification sample set and a test sample set;

taking a neighborhood data cube corresponding to each sample in the training sample set and the verification sample set as the input of the network, taking the label set as the expected output, and training the network to obtain a trained network model;

the network comprises a network input layer, a data dimension reduction module, a spectrum attention module, a secondary data dimension reduction module, a space attention mechanism module, a pooling module and a full-connection network which are sequentially arranged in a cascade manner;

the data dimension reduction module comprises two convolution modules which are connected in sequence, and each convolution module comprises a convolution layer, a batch normalization layer and an activation function layer which are connected in sequence;

the spectral attention module comprises a plurality of spectral attention feature layers which are connected in sequence;

the secondary data dimension reduction module comprises a convolution layer;

the space attention mechanism module comprises a plurality of space attention characteristic layers which are connected in sequence.

2. The construction method according to claim 1, wherein, when a boundary problem is encountered during the construction of the neighborhood data cube in the third step, the boundary is extended by using a 0 boundary filling method.

3. The construction method according to claim 1, wherein the concrete method of the fourth step is as follows:

taking a neighborhood data cube corresponding to each sample in a training sample set and a verification sample set as the input of a network, taking a label set as expected output, calculating the cross entropy loss between a real label of the training sample and a network prediction output value, taking the minimized cross entropy loss as an optimization target, optimizing the network by adopting an adaptive moment estimation algorithm, testing an obtained network model on the verification set, updating network model parameters until iteration is finished if the obtained classification accuracy is superior to that of the previous network model, storing the model with the highest classification accuracy on the verification set, obtaining the trained network model, and evaluating the classification performance of the trained network model by adopting the test sample set.

4. A hyperspectral image classification method is characterized by comprising the following steps:

acquiring a hyperspectral image set to be classified, inputting the hyperspectral image set into a network model, and classifying each pixel of the hyperspectral image by using the network model to obtain a final classification result;

the network model is a hyperspectral image classification model constructed by the hyperspectral image classification model construction method according to any one of claims 1 to 3.

5. A hyperspectral image classification system is characterized by comprising an image acquisition module, a data preprocessing module and a classification module;

the image acquisition module is used for acquiring a hyperspectral image;

the data preprocessing module is used for carrying out standardization processing on the hyperspectral image so that the input image conforms to Gaussian distribution of unit variance of 0 mean value;

the classification module classifies the hyperspectral images by adopting the method of claim 4 and outputs a classification result.

Images