A hyperspectral image classification method based on SSBRID network

Abstract

A hyperspectral image classification method based on SSBRID network belongs to the technical field of remote sensing image classification. The present invention solves the problems of many parameters, large model size and low classification accuracy of existing convolutional neural networks. The present invention extends two-dimensional convolution to three-dimensional convolution, which can adapt to the three-dimensional characteristics of hyperspectral data; at the same time, the asymmetric convolution kernel is introduced into the three-dimensional convolution layer, which can reduce the number of model parameters and improve the running speed; finally, the output features of the three-dimensional convolution are reorganized and sent to the two-dimensional convolution, which can reduce the model size while reducing the parameters of the network, reduce spatial redundancy by sharing information, and improve classification accuracy. The method of the present invention can be applied to remote sensing image classification.

Description

Technical Field

[0001] The invention belongs to the technical field of remote sensing image classification, and in particular relates to a hyperspectral image classification method based on an SSBRID network. Background Art

[0002] Compared with conventional two-dimensional images, hyperspectral remote sensing images have an additional spectral dimension and can be considered as composed of two-dimensional images of multiple bands. By merging these two-dimensional images in different bands according to the order of the bands, we can obtain hyperspectral image data with a cubic structure. Hyperspectral image data has a unique high-dimensional structure compared to ordinary two-dimensional images. Spatial-spectral joint features are very important feature information in the classification task of hyperspectral image data. Spatial refers to the spatial information of hyperspectral images, and spectrum refers to the spectral information of hyperspectral images. Because hyperspectral image data has serious phenomena of different spectra for the same object and different objects for the same spectrum, the effect of classifying hyperspectral images based solely on spatial information or spectral information is not ideal. Fusion and classification of spatial information and spectral information can learn more feature information to improve classification performance.

[0003] Three-dimensional convolutional neural networks can perform spatial-spectral joint feature extraction on hyperspectral data, and their classification performance is good, but their parameters are large, which increases the computational complexity and running time. Compared with three-dimensional convolutional neural networks, two-dimensional convolutional neural networks have fewer parameters and a lower risk of overfitting, but they only have good performance for spatial feature extraction of hyperspectral data. The hyperspectral image classification method based on hybrid convolutional neural networks combines three-dimensional convolutional neural networks with two-dimensional convolutional neural networks to fully utilize the spatial-spectral joint feature information of hyperspectral data and obtain fewer parameters, thereby obtaining more accurate classification accuracy. However, with the development of convolutional neural networks, the number of network layers continues to deepen, the number of network parameters increases, and the model size increases, making it difficult to apply to mobile devices. Therefore, the design of lightweight networks is also particularly important. Moreover, the classification accuracy of existing convolutional neural networks is still low, and it is also necessary to improve the classification accuracy. Summary of the invention

[0004] The purpose of the present invention is to solve the problems of many parameters, large model size and low classification accuracy of existing convolutional neural networks, and to propose a hyperspectral image classification method based on SSBRID network.

[0005] The technical solution adopted by the present invention to solve the above technical problems is:

[0006] A hyperspectral image classification method based on SSBRID network, the method specifically comprises the following steps:

[0007] Step 1: Obtain a hyperspectral image dataset;

[0008] Step 2, performing dimensionality reduction processing on the hyperspectral image obtained in step 1 to obtain hyperspectral image data after dimensionality reduction processing;

[0009] Step 3, randomly divide the hyperspectral image data after dimensionality reduction into two parts: a training set and a test set;

[0010] Step 4: Build an SSBRID network, and use the training set to train the built SSBRID network until the classification accuracy of the SSBRID network on the test set is no longer improved, and then stop training to obtain a trained SSBRID network.

[0011] Step 5, after the dimensionality reduction processing is performed on the hyperspectral image to be classified, the hyperspectral image to be classified after the dimensionality reduction processing is input into the trained SSBRID network, and the classification result of the hyperspectral image to be classified is output through the SSBRID network;

[0012] The working process of the SSBRID network is as follows:

[0013] Starting from the input end, the input image passes through the first three-dimensional convolution layer with a convolution kernel size of 3×3×3, the first BN normalization layer, the first Mish activation function layer, and the second three-dimensional convolution layer with a convolution kernel size of 1×1×1.

[0014] The output of the second 3D convolutional layer is used as the input of the third 3D convolutional layer with a convolution kernel size of 1×1×3 and the fourth 3D convolutional layer with a convolution kernel size of 3×1×1. The output of the fourth 3D convolutional layer is used as the input of the fifth 3D convolutional layer with a convolution kernel size of 1×3×1.

[0015] The output of the third 3D convolutional layer is concatenated with the output of the fifth 3D convolutional layer. The concatenated result is then passed through the second BN normalization layer and the second Mish activation function layer. The output of the second Mish activation function layer is resized and used as the input of the first 2D convolutional layer with a convolution kernel size of 3×3.

[0016] The output of the first two-dimensional convolutional layer is used as the input of the AOS module. The output of the AOS module then passes through the Softmax layer, and the classification result is output through the Softmax layer.

[0017] Furthermore, the dimensionality reduction process for the hyperspectral image acquired in step 1 is performed by using a principal component analysis method.

[0018] Furthermore, before randomly dividing the hyperspectral image data after the dimensionality reduction processing into a training set and a test set, it is necessary to perform slicing processing on the hyperspectral image after the dimensionality reduction processing.

[0019] Furthermore, the working process of the AOS module is as follows:

[0020] The input of the AOS module is divided into two parts. One part passes through the second two-dimensional convolution layer with a convolution kernel size of 1×1, and the output of the second two-dimensional convolution layer is used as the input of the first ASF submodule; at the same time, the other part passes through the first maximum pooling layer with a size of 2×2 and the third two-dimensional convolution layer with a convolution kernel size of 1×1, and the output of the third two-dimensional convolution layer is used as the input of the second ASF submodule;

[0021] The output of the first ASF submodule is divided into two parts. One part of the output of the first ASF submodule is used as the input of the third ASF submodule, and the other part of the output of the first ASF submodule is used as the input of the second maximum pooling layer of size 2×2. The output of the second maximum pooling layer is then passed through the fourth ASF submodule.

[0022] The output of the second ASF submodule is divided into two parts, one part of the output of the second ASF submodule is used as the input of the fifth ASF submodule, the other part of the output of the second ASF submodule is used as the input of the sixth ASF submodule, and the output of the fifth ASF submodule is then passed through the first upsampling layer;

[0023] After the output of the third ASF submodule is superimposed with the output of the first upsampling layer, the superimposed result is passed through the third maximum pooling layer of size 2×2;

[0024] After the output of the fourth ASF submodule is superimposed with the output of the sixth ASF submodule, the superimposed result is passed through the fourth maximum pooling layer of size 2×2;

[0025] The output of the third maximum pooling layer is divided into two parts. One part of the output of the third maximum pooling layer passes through the seventh ASF submodule. Meanwhile, the other part of the output of the third maximum pooling layer passes through the fourth convolution layer with a convolution kernel size of 1×1, the fifth maximum pooling layer with a size of 2×2, and the eighth ASF submodule in sequence.

[0026] The output of the fourth maximum pooling layer is divided into two parts. One part of the output of the fourth maximum pooling layer passes through the ninth ASF submodule. Meanwhile, the other part of the output of the fourth maximum pooling layer passes through the fifth convolution layer with a convolution kernel size of 1×1 and the tenth ASF submodule in sequence.

[0027] After the output of the eighth ASF submodule and the output of the tenth ASF submodule are superimposed, the superimposed result passes through the eleventh ASF submodule;

[0028] After the output of the ninth ASF submodule passes through the second upsampling layer, the output of the second upsampling layer is superimposed with the output of the seventh ASF submodule, and the superposition result passes through the twelfth ASF submodule;

[0029] The output of the eleventh ASF submodule passes through the third upsampling layer, and then the output of the third upsampling layer is superimposed with the output of the twelfth ASF submodule as the output of the AOS module.

[0030] Furthermore, the working process of the first ASF submodule is as follows:

[0031] The input of the first ASF submodule first passes through two parallel branches, where the first branch includes a two-dimensional convolution layer with a convolution kernel size of 1×1, and the second branch includes a two-dimensional convolution layer with a convolution kernel size of 1×3 and a two-dimensional convolution layer with a convolution kernel size of 3×1. The output of the first branch and the output of the second branch are then spliced, and the splicing result is the output of the first ASF submodule.

[0032] The beneficial effects of the present invention are:

[0033] The present invention extends two-dimensional convolution to three-dimensional convolution, which can adapt to the three-dimensional characteristics of hyperspectral data; at the same time, the asymmetric convolution kernel is introduced into the three-dimensional convolution layer, which can reduce the number of model parameters and improve the running speed; finally, the output features of the three-dimensional convolution are reorganized and sent to the two-dimensional convolution, which can reduce the model size while reducing the network parameters, reduce spatial redundancy by sharing information, and improve classification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 is an image in the LK dataset;

[0035] From top to bottom, the categories represented by each color are corn, cotton, sesame, round-leaf soybean, long-leaf soybean, rice, water, houses and roads, and mixed weeds. The number of pixels in each category is 34511, 8374, 3031, 63212, 4151, 11854, 67056, 7124, and 5229, respectively;

[0036] Figure 2 is an image in the HC dataset;

[0037] From top to bottom, the categories represented by each color are strawberry, cowpea, soybean, sorghum, spinach, watermelon, vegetable garden, tree, grassland, red roof, gray roof, mulch, bare soil, road, highlighted object, and water body. The number of pixels in each category is 44735, 22753, 10287, 5353, 1200, 4533, 5903, 17978, 9469, 10516, 16911, 3679, 9116, 18560, 1136, and 75401 respectively;

[0038] Figure 3 is an image in the HH dataset;

[0039] From top to bottom, the categories represented by each color are red roof, road, bare soil, cotton, cottonwood, rape, cabbage, Chinese cabbage, cabbage, mustard, cauliflower, green vegetables, green vegetables, lettuce, green lettuce, lettuce covered with film, romaine lettuce, carrot, cabbage, garlic sprouts, beans, and persimmon trees. The number of pixels in each category is 14041, 3512, 21821, 163285, 6218, 44557, 24103, 4054, 10819, 12394, 11015, 8954, 22507, 7356, 1002, 7262, 3010, 3217, 8712, 3486, 1328, and 4040 respectively;

[0040] Figure 4 This is the SSBRID network structure diagram;

[0041] Figure 5 This is the training flow chart of the SSBRID network;

[0042] Figure 6 This is the AOS module structure diagram;

[0043] Figure 7 This is the ASF submodule structure diagram;

[0044] Figure 8(a) shows the classification result of the SSBRID network model on the LK dataset;

[0045] Figure 8(b) shows the classification result of the SSBRID network model on the HC dataset;

[0046] Figure 8(c) shows the classification result of the SSBRID network model on the HH dataset. DETAILED DESCRIPTION

[0047] The present application is further described in detail below through specific implementations in combination with the accompanying drawings. Obviously, the described implementations are only part of the implementations of the present invention, not all of the implementations. Based on the implementations in the present invention, other implementations obtained by ordinary technicians in the field without making creative work are all within the scope of protection of the present invention.

[0048] Specific implementation method 1. Combination Figure 5 The present embodiment is described as follows. The present embodiment describes a hyperspectral image classification method based on an SSBRID network, and the method specifically comprises the following steps:

[0049] Step 1: Obtain a hyperspectral image dataset (the data obtained is currently publicly available labeled data);

[0050] Step 2, performing dimensionality reduction processing on the hyperspectral image obtained in step 1 to obtain hyperspectral image data after dimensionality reduction processing;

[0051] Step 3, randomly divide the hyperspectral image data after dimensionality reduction into two parts: a training set and a test set;

[0052] Step 4: Build an SSBRID network, and use the training set to train the built SSBRID network until the classification accuracy of the SSBRID network on the test set is no longer improved, and then stop training to obtain a trained SSBRID network.

[0053] Step 5: After the dimensionality reduction processing is performed on the hyperspectral image to be classified, the hyperspectral image to be classified after the dimensionality reduction processing is input into the trained SSBRID network, and the classification result of the hyperspectral image to be classified is output through the SSBRID network.

[0054] The method of the present invention can realize pixel-level classification and has a lower parameter amount while fully learning data features.

[0055] Specific implementation method 2: This implementation method is different from specific implementation method 1 in that the dimension reduction processing of the hyperspectral image acquired in step 1 adopts the principal component analysis method.

[0056] The other steps and parameters are the same as those in the first embodiment.

[0057] Dimensionality reduction processing can reduce the number of bands in the third dimension of the three-dimensional original data, reduce data redundancy, and speed up the running time. The training set and test set images used in the present invention are both 15×15 resolution. Figure 4 As shown in the figure, H×W represents the spatial dimension, B represents the third-dimensional band, and after PCA dimensionality reduction and data block preprocessing, a series of 15×15×L images are obtained, and L represents the third-dimensional band after dimensionality reduction.

[0058] Specific implementation method three: This implementation method is different from specific implementation method one or two in that before randomly dividing the hyperspectral image data after dimensionality reduction processing into a training set and a test set, the hyperspectral image data after dimensionality reduction processing needs to be sliced.

[0059] The other steps and parameters are the same as those in the first or second embodiment.

[0060] Specific implementation method four: Combination Figure 4 This embodiment is different from the first to third embodiments in that the working process of the SSBRID network is as follows:

[0061] Starting from the input end, the input image passes through the first three-dimensional convolution layer with a convolution kernel size of 3×3×3, the first BN normalization layer, the first Mish activation function layer, and the second three-dimensional convolution layer with a convolution kernel size of 1×1×1.

[0062] The output of the second 3D convolutional layer is used as the input of the third 3D convolutional layer with a convolution kernel size of 1×1×3 and the fourth 3D convolutional layer with a convolution kernel size of 3×1×1. The output of the fourth 3D convolutional layer is used as the input of the fifth 3D convolutional layer with a convolution kernel size of 1×3×1.

[0063] The output of the third 3D convolutional layer is concatenated with the output of the fifth 3D convolutional layer. The concatenated result is then passed through the second BN normalization layer and the second Mish activation function layer. The output of the second Mish activation function layer is resized and used as the input of the first 2D convolutional layer with a convolution kernel size of 3×3.

[0064] The output of the first two-dimensional convolutional layer is used as the input of the AOS module. The output of the AOS module then passes through the Softmax layer, and the classification result is output through the Softmax layer.

[0065] The other steps and parameters are the same as those in Specific Embodiments 1 to 3.

[0066] The present invention adopts a three-dimensional convolution layer with an asymmetric convolution kernel, which mainly includes a 1×1×3 spectral branch and a 1×3×1, 3×1×1 spatial branch, which can reduce model parameters and improve the running speed while extracting spatial-spectral joint feature information; after splicing the output features of the two branches, the reorganized features are input into the AOS module, which can reduce network parameters and reduce spatial redundancy by sharing information.

[0067] Specific implementation method five: Combination Figure 6This embodiment is different from the first to fourth embodiments in that the working process of the AOS module is as follows:

[0068] The input of the AOS module is divided into two parts. One part passes through the second two-dimensional convolution layer with a convolution kernel size of 1×1, and the output of the second two-dimensional convolution layer is used as the input of the first ASF submodule; at the same time, the other part passes through the first maximum pooling layer with a size of 2×2 and the third two-dimensional convolution layer with a convolution kernel size of 1×1, and the output of the third two-dimensional convolution layer is used as the input of the second ASF submodule;

[0069] The output of the first ASF submodule is divided into two parts. One part of the output of the first ASF submodule is used as the input of the third ASF submodule, and the other part of the output of the first ASF submodule is used as the input of the second maximum pooling layer of size 2×2. The output of the second maximum pooling layer is then passed through the fourth ASF submodule.

[0070] The output of the second ASF submodule is divided into two parts, one part of the output of the second ASF submodule is used as the input of the fifth ASF submodule, the other part of the output of the second ASF submodule is used as the input of the sixth ASF submodule, and the output of the fifth ASF submodule is then passed through the first upsampling layer;

[0071] After the output of the third ASF submodule is superimposed with the output of the first upsampling layer, the superimposed result is passed through the third maximum pooling layer of size 2×2;

[0072] After the output of the fourth ASF submodule is superimposed with the output of the sixth ASF submodule, the superimposed result is passed through the fourth maximum pooling layer of size 2×2;

[0073] The output of the third maximum pooling layer is divided into two parts. One part of the output of the third maximum pooling layer passes through the seventh ASF submodule. Meanwhile, the other part of the output of the third maximum pooling layer passes through the fourth convolution layer with a convolution kernel size of 1×1, the fifth maximum pooling layer with a size of 2×2, and the eighth ASF submodule in sequence.

[0074] The output of the fourth maximum pooling layer is divided into two parts. One part of the output of the fourth maximum pooling layer passes through the ninth ASF submodule. Meanwhile, the other part of the output of the fourth maximum pooling layer passes through the fifth convolution layer with a convolution kernel size of 1×1 and the tenth ASF submodule in sequence.

[0075] After the output of the eighth ASF submodule and the output of the tenth ASF submodule are superimposed, the superimposed result passes through the eleventh ASF submodule;

[0076] After the output of the ninth ASF submodule passes through the second upsampling layer, the output of the second upsampling layer is superimposed with the output of the seventh ASF submodule, and the superposition result passes through the twelfth ASF submodule;

[0077] The output of the eleventh ASF submodule passes through the third upsampling layer, and then the output of the third upsampling layer is superimposed with the output of the twelfth ASF submodule as the output of the AOS module.

[0078] Its steps and parameters are the same as those of the specific implementation modes 1 to 4.

[0079] This implementation can reduce network parameters while fully learning data features.

[0080] Specific implementation method six: Combination Figure 7 The present embodiment is different from the first to fifth embodiments in that the working process of the first ASF submodule is as follows:

[0081] The input of the first ASF submodule first passes through two parallel branches, where the first branch includes a two-dimensional convolution layer with a convolution kernel size of 1×1, and the second branch includes a two-dimensional convolution layer with a convolution kernel size of 1×3 and a two-dimensional convolution layer with a convolution kernel size of 3×1. The output of the first branch and the output of the second branch are then spliced, and the splicing result is the output of the first ASF submodule.

[0082] The other steps and parameters are the same as those in Specific Implementation Methods 1 to 5.

[0083] The present invention adopts asymmetric convolution 1×3 and 3×1 to form an ASF module, which is used to reduce module parameters. The working process of other ASF submodules is the same as the working process of the first ASF submodule.

[0084] Experimental Section

[0085] The configuration of the computer used in the present invention for training the SSBRID network model and verifying the classification results is as follows: the CPU is Intel Core i9-12900K, the memory is 32GB DDR5 5200, the graphics card is NVIDIA GeForce RTX 3090Ti, the hard disk is 1TB SSD plus 4T HDD, and the system is Windows 11 Professional Edition. The three hyperspectral image datasets used are WHU-Hi-LongKou (LK), WHU-Hi-HanChuan (HC) and WHU-Hi-HongHu (HH) datasets. The datasets are respectively as follows Figure 1 , Figure 2 and Figure 3 As shown, different colors represent different types of land objects.

[0086] 1. The LK dataset was collected in Longkou Town, Hubei Province, China in 2018 using an 8mm focal length Headwall Nano-Hyperspec imaging sensor mounted on a DJI matrix 600Pro (DJI M600 Pro) drone platform. The study area is a simple agricultural scene containing six crop varieties: corn, cotton, sesame, broad-leaved soybean, narrow-leaved soybean, and rice. The drone flew at an altitude of 500m, with an image size of 550×400 pixels, 270 bands from 400 to 1000nm, and the spatial resolution of the drone-borne hyperspectral image is about 0.463m.

[0087] 2. The HC dataset was collected in Hanchuan, Hubei Province, China in 2016, with a 17mm focal length Headwall Nano-Hyperspec imaging sensor mounted on a Leica Aibot X6 drone V1 platform. The study area is an urban-rural junction with buildings, water, and cultivated land, including seven crops: strawberry, cowpea, soybean, sorghum, water spinach, watermelon, and green vegetables. The drone flew at an altitude of 250m, with an image size of 1217×303 pixels, 274 bands from 400 to 1000nm, and a spatial resolution of about 0.109m for the drone-borne hyperspectral image. There are many shadow-covered areas in the image.

[0088] 3. The HH dataset was collected in Honghu City, Hubei Province, China in 2017 using a HeadwallNano-Hyperspec imaging sensor with a focal length of 17 mm, mounted on a DJI matrix 600Pro drone platform. The experimental area is a complex agricultural scene with many types of crops, and different varieties of the same crop are also grown in the area, including Chinese cabbage and Chinese cabbage, as well as Brassica rapa and Brassica rapa. Different varieties of the same crop type are grown in the area, for example, Chinese cabbage / Chinese cabbage and Brassica rapa / Brassica rapa. The drone flew at an altitude of 100m, with an image size of 940×475 pixels, 270 bands from 400 to 1000nm, and the spatial resolution of the drone-borne hyperspectral image is about 0.043m.

[0089] For the evaluation indicators of the model, we select the three most widely used objective evaluation indicators in the industry: Overall accuracy (OA), average accuracy (AA) and Kappa coefficient (K). These three evaluation indicators are calculated based on the confusion matrix. The following is an introduction to these evaluation indicators:

[0090] (1) Confusion Matrix: The confusion matrix is ​​an error matrix that is expressed in a standard form of computational accuracy and consists of a matrix of n rows and n columns. We can get the details of the composition of this matrix from Table 1 below. It is a specific matrix used to visualize the performance of the algorithm. Each column of the confusion matrix represents the predicted value and each row represents the true category. The confusion matrix can clearly indicate whether there is confusion between multiple categories, that is, whether one category is predicted as another category.

[0091] Table 1 Composition of confusion matrix

[0092]

[0093] (2) Overall classification accuracy (OA): OA is a basic evaluation indicator, which is expressed as the ratio of the number of correctly classified samples to the total number of samples, that is, the overall classification accuracy.

[0094] The overall classification accuracy represents the classification accuracy of the evaluation algorithm as a whole. It refers to the ratio of the number of correctly classified category pixels to the total number of categories:

[0095]

[0096] (3) Average classification accuracy (AA): AA is an evaluation index of classification accuracy in the classification algorithm. First, the classification accuracy of each type of object is summed up, and then the average is calculated as the classification accuracy. The average classification accuracy is different from the overall classification accuracy. It focuses on evaluating the classification accuracy of each type of object in the algorithm:

[0097]

[0098] (4) Kappa coefficient (K): The Kappa coefficient indicates the reduction ratio of completely random classification error. It is also an evaluation index of classification accuracy. As a supplement to the error matrix, which cannot clearly show the classification accuracy, it can clearly reflect the effect of classification accuracy:

[0099]

[0100] Among them, n represents the number of rows and columns of the classification matrix, m ij Represents the value of the i-th row and j-th column in the confusion matrix, m i+ is the row sum of the confusion matrix, and m +i It is the column integration of the confusion matrix, and N represents all the elements contained in the confusion matrix.

[0101] The objective classification results of the SSBRID network model used in the present invention on three data sets are shown in Table 2:

[0102] Table 2 Classification results of SSBRID network model for different data

[0103]

[0104] From left to right, the first column is the data set; the second column is the OA data of the classification results; the third column is the AA data of the classification results; the fourth column is the K data of the classification results; the fifth column is the training time of the classification; the sixth column is the test time of the classification; and the seventh column is the model parameter quantity.

[0105] The subjective classification results are shown in Figure 8(a), Figure 8(b), and Figure 8(c). Figure 8(a) represents the classification effect of the LK dataset, Figure 8(b) represents the classification effect of the HC dataset, and Figure 8(c) represents the classification effect of the HH dataset. The higher the accuracy, the less salt and pepper noise in the image.

[0106] The above calculation examples of the present invention are only used to explain the calculation model and calculation process of the present invention in detail, and are not intended to limit the implementation methods of the present invention. For ordinary technicians in the relevant field, other different forms of changes or modifications can be made based on the above description. It is impossible to list all the implementation methods here. All obvious changes or modifications derived from the technical solution of the present invention are still within the scope of protection of the present invention.

Claims

1. A hyperspectral image classification method based on SSBRID network, characterized in that: The method specifically comprises the following steps: Step 1: Obtain a hyperspectral image dataset; Step 2: Perform dimensionality reduction processing on the hyperspectral image obtained in step 1 to obtain hyperspectral image data after dimensionality reduction processing; Step 3, randomly divide the hyperspectral image data after dimensionality reduction into two parts: a training set and a test set; Step 4: Build an SSBRID network, and use the training set to train the built SSBRID network until the classification accuracy of the SSBRID network on the test set is no longer improved, and then stop training to obtain a trained SSBRID network. Step 5, after the dimensionality reduction processing is performed on the hyperspectral image to be classified, the hyperspectral image to be classified after the dimensionality reduction processing is input into the trained SSBRID network, and the classification result of the hyperspectral image to be classified is output through the SSBRID network; The working process of the SSBRID network is as follows: Starting from the input end, the input image passes through the first three-dimensional convolution layer with a convolution kernel size of 3×3×3, the first BN normalization layer, the first Mish activation function layer, and the second three-dimensional convolution layer with a convolution kernel size of 1×1×1. The output of the second 3D convolutional layer is used as the input of the third 3D convolutional layer with a convolution kernel size of 1×1×3 and the fourth 3D convolutional layer with a convolution kernel size of 3×1×1. The output of the fourth 3D convolutional layer is used as the input of the fifth 3D convolutional layer with a convolution kernel size of 1×3×1. The output of the third 3D convolutional layer is concatenated with the output of the fifth 3D convolutional layer. The concatenated result is then passed through the second BN normalization layer and the second Mish activation function layer. The output of the second Mish activation function layer is resized and used as the input of the first 2D convolutional layer with a convolution kernel size of 3×3. The output of the first two-dimensional convolutional layer is used as the input of the AOS module. The output of the AOS module then passes through the Softmax layer, and the classification result is output through the Softmax layer. The working process of the AOS module is as follows: The input of the AOS module is divided into two parts. One part passes through the second two-dimensional convolution layer with a convolution kernel size of 1×1, and the output of the second two-dimensional convolution layer is used as the input of the first ASF submodule; at the same time, the other part passes through the first maximum pooling layer with a size of 2×2 and the third two-dimensional convolution layer with a convolution kernel size of 1×1, and the output of the third two-dimensional convolution layer is used as the input of the second ASF submodule; The output of the first ASF submodule is divided into two parts. One part of the output of the first ASF submodule is used as the input of the third ASF submodule, and the other part of the output of the first ASF submodule is used as the input of the second maximum pooling layer of size 2×2. The output of the second maximum pooling layer is then passed through the fourth ASF submodule. The output of the second ASF submodule is divided into two parts, one part of the output of the second ASF submodule is used as the input of the fifth ASF submodule, the other part of the output of the second ASF submodule is used as the input of the sixth ASF submodule, and the output of the fifth ASF submodule is then passed through the first upsampling layer; After the output of the third ASF submodule is superimposed with the output of the first upsampling layer, the superimposed result is passed through the third maximum pooling layer of size 2×2; After the output of the fourth ASF submodule is superimposed with the output of the sixth ASF submodule, the superimposed result is passed through the fourth maximum pooling layer of size 2×2; The output of the third maximum pooling layer is divided into two parts. One part of the output of the third maximum pooling layer passes through the seventh ASF submodule. Meanwhile, the other part of the output of the third maximum pooling layer passes through the fourth convolution layer with a convolution kernel size of 1×1, the fifth maximum pooling layer with a size of 2×2, and the eighth ASF submodule in sequence. The output of the fourth maximum pooling layer is divided into two parts. One part of the output of the fourth maximum pooling layer passes through the ninth ASF submodule. Meanwhile, the other part of the output of the fourth maximum pooling layer passes through the fifth convolution layer with a convolution kernel size of 1×1 and the tenth ASF submodule in sequence. After the output of the eighth ASF submodule and the output of the tenth ASF submodule are superimposed, the superimposed result passes through the eleventh ASF submodule; After the output of the ninth ASF submodule passes through the second upsampling layer, the output of the second upsampling layer is superimposed with the output of the seventh ASF submodule, and the superposition result passes through the twelfth ASF submodule; The output of the eleventh ASF submodule passes through the third upsampling layer, and then the output of the third upsampling layer is superimposed with the output of the twelfth ASF submodule as the output of the AOS module; The working process of the first ASF submodule is as follows: The input of the first ASF submodule first passes through two parallel branches, where the first branch includes a two-dimensional convolution layer with a convolution kernel size of 1×1, and the second branch includes a two-dimensional convolution layer with a convolution kernel size of 1×3 and a two-dimensional convolution layer with a convolution kernel size of 3×1. The output of the first branch and the output of the second branch are then spliced, and the splicing result is the output of the first ASF submodule; The working process of other ASF submodules is the same as the working process of the first ASF submodule.

2. The hyperspectral image classification method based on SSBRID network according to claim 1 is characterized in that: The dimensionality reduction process for the hyperspectral image acquired in step 1 adopts the principal component analysis method.

3. The hyperspectral image classification method based on SSBRID network according to claim 2 is characterized in that: Before randomly dividing the hyperspectral image data after the dimensionality reduction processing into a training set and a test set, the hyperspectral image after the dimensionality reduction processing needs to be sliced.

Images