A pixel-level classification method for hyperspectral images based on MADWA network

Abstract

A hyperspectral image pixel-level classification method based on a MADWA network belongs to the field of remote sensing image classification. The present invention solves the problem of poor classification accuracy of existing deep neural networks under the background of noisy labels. The present invention uses a multi-scale structure to fully learn the real label and noise label data information, so that the network's ability to learn information is improved; and a spectral spatial attention structure is designed to more effectively focus on the unique features to be learned in each dimension, so as to make more full use of the local features of hyperspectral data and improve classification accuracy. And by adopting an asymmetric convolution kernel, the model parameters can be greatly reduced, the running speed can be improved, and the computational complexity can be reduced. The method of the present invention can be applied to remote sensing image classification.

Description

Technical Field

[0001] The present invention belongs to the field of remote sensing image classification, and specifically relates to a hyperspectral image pixel-level classification method based on a MADWA network (multi-scale asymmetric attention network). Background Art

[0002] As deep learning methods are gradually applied to the field of scientific research, convolutional neural networks have become the most widely used structure in hyperspectral data classification algorithms. Convolutional neural networks can learn feature information from the original image data layer by layer and effectively extract useful feature information through local connections. Convolutional neural networks are supervision-oriented and do not require manual feature extraction. The shared weight mechanism of the convolution kernel parameters can effectively reduce the parameters of the network layer and reduce the size of the trained model. These characteristics lay the foundation for the application of convolutional neural networks in hyperspectral classification tasks. In recent years, three-dimensional convolutional neural networks have become mainstream due to their excellent feature extraction capabilities. However, the high dimensionality and limited training samples of hyperspectral images limit the improvement of their classification accuracy, and the traditional three-dimensional convolutional neural network parameters are large, which increases the computational complexity and running time. Secondly, although many deep learning methods have achieved remarkable results in hyperspectral data classification tasks, training data with precise labels is the key factor for deep neural networks to achieve excellent results. If there are label errors in the training data, the accuracy of the model in classification will be greatly reduced. Since most of the samples in the existing hyperspectral datasets are manually labeled, it is possible that there are deviations in the labeling process. In addition, when the instrument collects data, deviations caused by positioning or external environment (such as lighting, rain, etc.) also make accurate labeling more difficult. Therefore, in the hyperspectral data training samples, sample-by-sample labeling may cause a small number of samples to be incorrectly assigned to other categories. These unavoidable external problems may lead to the inclusion of some inaccurate labels in the training samples. Therefore, in the context of noisy labels, how to train a reliable and accurate model is particularly important. Summary of the invention

[0003] The purpose of the present invention is to solve the problem of poor classification accuracy of existing deep neural networks under the background of noisy labels, and to propose a hyperspectral image pixel-level classification method based on the MADWA network.

[0004] The technical solution adopted by the present invention to solve the above technical problems is:

[0005] A hyperspectral image pixel level classification method based on MADWA network, the method specifically comprises the following steps:

[0006] Step 1: Obtain a hyperspectral image dataset;

[0007] Step 2: preprocess the hyperspectral images in the acquired hyperspectral image dataset; the specific process of preprocessing is as follows:

[0008] Slice the hyperspectral image to obtain a sliced ​​hyperspectral image dataset, and then randomly divide the sliced ​​hyperspectral image dataset into two parts: a training set and a test set;

[0009] Step 3: construct a MADWA network, and use the training set to train the constructed MADWA network until the classification accuracy of the MADWA network on the test set no longer increases or the loss function value no longer decreases, and then stop training to obtain a trained MADWA network;

[0010] Step 4: Use the trained MADWA network to classify the hyperspectral image to be classified.

[0011] Furthermore, the working process of the MADWA network is as follows:

[0012] The input hyperspectral image first passes through the first three-dimensional convolution layer in the MADWA network with a convolution kernel size of 1×1×B, where B is the spectral dimension of the input hyperspectral image. The output of the first three-dimensional convolution layer is then input into three parallel branches respectively.

[0013] The output of the first three-dimensional convolutional layer passes through the second three-dimensional convolutional layer with a convolution kernel size of 1×1×1, and the output of the second three-dimensional convolutional layer is superimposed with the outputs of the three branches;

[0014] After the superposition result passes through the multi-scale asymmetric spatial-spectral attention unit, the output of the multi-scale asymmetric spatial-spectral attention unit passes through the average pooling layer and the Flatten layer in sequence, and the Flatten layer is used to output the classification result.

[0015] Furthermore, the three branches of the MADWA network are:

[0016] The first branch:

[0017] The input of the first branch is used as the input of the third 3D convolution layer with a convolution kernel size of 1×1×7, the fourth 3D convolution layer with a convolution kernel size of 7×1×1, and the fifth 3D convolution layer with a convolution kernel size of 1×7×1;

[0018] After superimposing the outputs of the third 3D convolutional layer, the fourth 3D convolutional layer, and the fifth 3D convolutional layer, the superimposed results are used as inputs of the sixth 3D convolutional layer with a convolution kernel size of 1×1×7, the seventh 3D convolutional layer with a convolution kernel size of 7×1×1, and the eighth 3D convolutional layer with a convolution kernel size of 1×7×1, respectively;

[0019] Then, the outputs of the sixth three-dimensional convolutional layer, the seventh three-dimensional convolutional layer, and the eighth three-dimensional convolutional layer are superimposed to obtain a superposition result A, and the superposition result A is used as the output of the first branch;

[0020] The second branch:

[0021] The input of the second branch is used as the input of the ninth three-dimensional convolution layer with a convolution kernel size of 1×1×5, the tenth three-dimensional convolution layer with a convolution kernel size of 5×1×1, and the eleventh three-dimensional convolution layer with a convolution kernel size of 1×5×1 respectively;

[0022] After superimposing the outputs of the ninth 3D convolutional layer, the tenth 3D convolutional layer, and the eleventh 3D convolutional layer, the superimposed results are used as inputs of the twelfth 3D convolutional layer with a convolution kernel size of 1×1×5, the thirteenth 3D convolutional layer with a convolution kernel size of 5×1×1, and the fourteenth 3D convolutional layer with a convolution kernel size of 1×5×1, respectively;

[0023] Then, the outputs of the twelfth three-dimensional convolutional layer, the thirteenth three-dimensional convolutional layer, and the fourteenth three-dimensional convolutional layer are superimposed to obtain a superposition result B, and the superposition result B is used as the output of the second branch;

[0024] The third branch:

[0025] The input of the third branch is used as the input of the fifteenth three-dimensional convolution layer with a convolution kernel size of 1×1×3, the sixteenth three-dimensional convolution layer with a convolution kernel size of 3×1×1, and the seventeenth three-dimensional convolution layer with a convolution kernel size of 1×3×1 respectively;

[0026] After superimposing the outputs of the fifteenth three-dimensional convolutional layer, the sixteenth three-dimensional convolutional layer, and the seventeenth three-dimensional convolutional layer, the superimposed results are used as inputs of the eighteenth three-dimensional convolutional layer with a convolution kernel size of 1×1×3, the nineteenth three-dimensional convolutional layer with a convolution kernel size of 3×1×1, and the twentieth three-dimensional convolutional layer with a convolution kernel size of 1×3×1, respectively;

[0027] The outputs of the eighteenth three-dimensional convolutional layer, the nineteenth three-dimensional convolutional layer, and the twentieth three-dimensional convolutional layer are superimposed to obtain a superposition result C, and the superposition result C is used as the output of the third branch.

[0028] Furthermore, the working process of the multi-scale asymmetric spatial-spectral attention unit is:

[0029] The input of the multi-scale asymmetric spatial-spectral attention unit first passes through four parallel branches. The outputs of the four parallel branches are superimposed with the input of the multi-scale asymmetric spatial-spectral attention unit respectively. The superposition result then passes through the thirteenth point-by-point convolution layer with a convolution kernel size of 1×1. The output of the thirteenth point-by-point convolution layer is the output of the multi-scale asymmetric spatial-spectral attention unit.

[0030] Furthermore, the four branches of the multi-scale asymmetric spatial-spectral attention unit are:

[0031] In the first branch, the input of the multi-scale asymmetric spatial-spectral attention unit is first divided into two parts, where the first part passes through the first point-by-point convolution layer with a convolution kernel size of 1×1, the first depth-wise separable convolution layer with a convolution kernel size of 3×1, and the second depth-wise separable convolution layer with a convolution kernel size of 1×3 in sequence; the second part passes through the second point-by-point convolution layer with a convolution kernel size of 1×1, and the outputs of the second point-by-point convolution layer and the second depth-wise separable convolution layer are multiplied, and the multiplication result passes through the third point-by-point convolution layer with a convolution kernel size of 1×1. The output of the third point-by-point convolution layer is the output of the first branch;

[0032] In the second branch, the input of the multi-scale asymmetric spatial-spectral attention unit is first divided into two parts, where the first part passes through the fourth point-by-point convolution layer with a convolution kernel size of 1×1, the third depth-wise separable convolution layer with a convolution kernel size of 5×1, and the fourth depth-wise separable convolution layer with a convolution kernel size of 1×5 in sequence; the second part passes through the fifth point-by-point convolution layer with a convolution kernel size of 1×1, and the outputs of the fifth point-by-point convolution layer and the fourth depth-wise separable convolution layer are multiplied. The multiplication result then passes through the sixth point-by-point convolution layer with a convolution kernel size of 1×1. The output of the sixth point-by-point convolution layer is the output of the second branch;

[0033] In the third branch, the input of the multi-scale asymmetric spatial-spectral attention unit is first divided into two parts, where the first part passes through the seventh point-by-point convolution layer with a convolution kernel size of 1×1, the fifth depth-wise separable convolution layer with a convolution kernel size of 7×1, and the sixth depth-wise separable convolution layer with a convolution kernel size of 1×7 in sequence; the second part passes through the eighth point-by-point convolution layer with a convolution kernel size of 1×1, and the outputs of the eighth point-by-point convolution layer and the sixth depth-wise separable convolution layer are multiplied, and the multiplication result passes through the ninth point-by-point convolution layer with a convolution kernel size of 1×1. The output of the ninth point-by-point convolution layer is the output of the third branch;

[0034] In the fourth branch, the input of the multi-scale asymmetric spatial-spectral attention unit is first divided into two parts. The first part passes through the tenth point-by-point convolution layer with a convolution kernel size of 1×1 and the seventh depth-wise separable convolution layer with a convolution kernel size of 1×1 in sequence; the second part passes through the eleventh point-by-point convolution layer with a convolution kernel size of 1×1, and the outputs of the eleventh point-by-point convolution layer and the seventh depth-wise separable convolution layer are multiplied. The multiplication result then passes through the twelfth point-by-point convolution layer with a convolution kernel size of 1×1. The output of the twelfth point-by-point convolution layer is the output of the fourth branch.

[0035] The beneficial effects of the present invention are:

[0036] The present invention uses a multi-scale structure to fully learn the real label and noise label data information, so that the network's ability to learn information is improved; and a spectral spatial attention structure is designed to more effectively focus on the unique features to be learned in each dimension, so as to make full use of the local features of hyperspectral data and improve classification accuracy. And by using an asymmetric convolution kernel, the model parameters can be greatly reduced, the running speed can be improved, and the computational complexity can be reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is a KSC data display chart;

[0038] In the figure, (a) is a false color image of hyperspectral data, and (b) is a true value image;

[0039] Figure 2 It is a UP data display chart;

[0040] In the figure, (a) is a false color image of hyperspectral data, and (b) is a true value image;

[0041] Figure 3 It is a BO data display chart;

[0042] In the figure, (a) is a false color image of hyperspectral data, and (b) is a true value image;

[0043] Figure 4 This is the MADWA network structure diagram;

[0044] Figure 5 This is the SSDWA attention structure diagram;

[0045] Figure 6 It is the MADWA network training flow chart;

[0046] Figure 7 This is the result of the MADWA network model classifying KSC data;

[0047] Figure 8 This is the result diagram of UP data classification by MADWA network model;

[0048] Fig. 9 This is the result of the MADWA network model classifying BO data. DETAILED DESCRIPTION

[0049] 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.

[0050] Specific implementation method 1. Combination Figure 6 The present embodiment describes a hyperspectral image pixel-level classification method based on a MADWA network, and the method specifically comprises the following steps:

[0051] Step 1: Obtain a hyperspectral image dataset (using a currently publicly available labeled dataset);

[0052] Step 2: preprocess the hyperspectral images in the acquired hyperspectral image dataset; the specific process of preprocessing is as follows:

[0053] Slice the hyperspectral image to obtain a sliced ​​hyperspectral image dataset, and then randomly divide the sliced ​​hyperspectral image dataset into two parts: a training set and a test set;

[0054] Step 3, construct a MADWA network, and use the training set to train the constructed MADWA network until the classification accuracy of the MADWA network on the test set is no longer improved or the loss function value (the loss function in the present invention only considers the joint influence of the pure label and the noise label) is no longer reduced, and the training is stopped to obtain a trained MADWA network;

[0055] Step 4: Use the trained MADWA network to classify the hyperspectral image to be classified.

[0056] The present invention constructs a multi-scale asymmetric attention network based on noise labels, which can fully extract the features of hyperspectral data while reducing the influence of noise labels.

[0057] Specific implementation method 2: Combination Figure 4 This embodiment is different from the first embodiment in that the working process of the MADWA network is as follows:

[0058] The input hyperspectral image first passes through the first three-dimensional convolution layer in the MADWA network with a convolution kernel size of 1×1×B, where B is the spectral dimension of the input hyperspectral image. The output of the first three-dimensional convolution layer is then input into three parallel branches respectively.

[0059] The output of the first three-dimensional convolutional layer passes through the second three-dimensional convolutional layer with a convolution kernel size of 1×1×1, and the output of the second three-dimensional convolutional layer is superimposed with the outputs of the three branches;

[0060] After the superposition result passes through the multi-scale asymmetric spatial-spectral attention unit (SSDWA), the output of the multi-scale asymmetric spatial-spectral attention unit passes through the average pooling layer and the Flatten layer in sequence, and the Flatten layer is used to output the classification result.

[0061] The other steps and parameters are the same as those in the first embodiment.

[0062] Specific implementation method three: This implementation method is different from specific implementation methods one or two in that the three branches of the MADWA network are:

[0063] The first branch:

[0064] The input of the first branch is used as the input of the third 3D convolution layer with a convolution kernel size of 1×1×7, the fourth 3D convolution layer with a convolution kernel size of 7×1×1, and the fifth 3D convolution layer with a convolution kernel size of 1×7×1;

[0065] After superimposing the outputs of the third 3D convolutional layer, the fourth 3D convolutional layer, and the fifth 3D convolutional layer, the superimposed results are used as inputs of the sixth 3D convolutional layer with a convolution kernel size of 1×1×7, the seventh 3D convolutional layer with a convolution kernel size of 7×1×1, and the eighth 3D convolutional layer with a convolution kernel size of 1×7×1, respectively;

[0066] Then, the outputs of the sixth three-dimensional convolutional layer, the seventh three-dimensional convolutional layer, and the eighth three-dimensional convolutional layer are superimposed to obtain a superposition result A, and the superposition result A is used as the output of the first branch;

[0067] The second branch:

[0068] The input of the second branch is used as the input of the ninth three-dimensional convolution layer with a convolution kernel size of 1×1×5, the tenth three-dimensional convolution layer with a convolution kernel size of 5×1×1, and the eleventh three-dimensional convolution layer with a convolution kernel size of 1×5×1 respectively;

[0069] After superimposing the outputs of the ninth 3D convolutional layer, the tenth 3D convolutional layer, and the eleventh 3D convolutional layer, the superimposed results are used as inputs of the twelfth 3D convolutional layer with a convolution kernel size of 1×1×5, the thirteenth 3D convolutional layer with a convolution kernel size of 5×1×1, and the fourteenth 3D convolutional layer with a convolution kernel size of 1×5×1, respectively;

[0070] Then, the outputs of the twelfth three-dimensional convolutional layer, the thirteenth three-dimensional convolutional layer, and the fourteenth three-dimensional convolutional layer are superimposed to obtain a superposition result B, and the superposition result B is used as the output of the second branch;

[0071] The third branch:

[0072] The input of the third branch is used as the input of the fifteenth three-dimensional convolution layer with a convolution kernel size of 1×1×3, the sixteenth three-dimensional convolution layer with a convolution kernel size of 3×1×1, and the seventeenth three-dimensional convolution layer with a convolution kernel size of 1×3×1 respectively;

[0073] After superimposing the outputs of the fifteenth three-dimensional convolutional layer, the sixteenth three-dimensional convolutional layer, and the seventeenth three-dimensional convolutional layer, the superimposed results are used as inputs of the eighteenth three-dimensional convolutional layer with a convolution kernel size of 1×1×3, the nineteenth three-dimensional convolutional layer with a convolution kernel size of 3×1×1, and the twentieth three-dimensional convolutional layer with a convolution kernel size of 1×3×1, respectively;

[0074] The outputs of the eighteenth three-dimensional convolutional layer, the nineteenth three-dimensional convolutional layer, and the twentieth three-dimensional convolutional layer are superimposed to obtain a superposition result C, and the superposition result C is used as the output of the third branch.

[0075] The other steps and parameters are the same as those in the first or second embodiment.

[0076] In this implementation, a three-branch parallel convolutional neural network is used, and each branch uses a three-dimensional convolution kernel of different sizes to learn the features of the input data with different receptive fields to avoid blind spots and unrecognizable areas in the receptive field. At the same time, an asymmetric convolution kernel is used to greatly reduce the number of parameters while maintaining accuracy, allowing the network to capture longer temporal dependencies.

[0077] Specific implementation method four: Combination Figure 5 This embodiment is different from the first to third embodiments in that the working process of the multi-scale asymmetric spatial-spectral attention unit (SSDWA) is as follows:

[0078] The input of the multi-scale asymmetric spatial-spectral attention unit first passes through four parallel branches. The outputs of the four parallel branches are superimposed with the input of the multi-scale asymmetric spatial-spectral attention unit respectively. The superposition result then passes through the thirteenth point-by-point convolution layer with a convolution kernel size of 1×1. The output of the thirteenth point-by-point convolution layer is the output of the multi-scale asymmetric spatial-spectral attention unit (SSDWA).

[0079] The other steps and parameters are the same as those in Specific Embodiments 1 to 3.

[0080] Specific implementation mode 5: This implementation mode is different from any one of the specific implementation modes 1 to 4 in that the four branches of the multi-scale asymmetric spatial-spectral attention unit are:

[0081] In the first branch, the input of the multi-scale asymmetric spatial-spectral attention unit is first divided into two parts, where the first part passes through the first point-by-point convolution layer with a convolution kernel size of 1×1, the first depth-wise separable convolution layer with a convolution kernel size of 3×1, and the second depth-wise separable convolution layer with a convolution kernel size of 1×3 in sequence; the second part passes through the second point-by-point convolution layer with a convolution kernel size of 1×1, and the outputs of the second point-by-point convolution layer and the second depth-wise separable convolution layer are multiplied, and the multiplication result passes through the third point-by-point convolution layer with a convolution kernel size of 1×1. The output of the third point-by-point convolution layer is the output of the first branch;

[0082] In the second branch, the input of the multi-scale asymmetric spatial-spectral attention unit is first divided into two parts, where the first part passes through the fourth point-by-point convolution layer with a convolution kernel size of 1×1, the third depth-wise separable convolution layer with a convolution kernel size of 5×1, and the fourth depth-wise separable convolution layer with a convolution kernel size of 1×5 in sequence; the second part passes through the fifth point-by-point convolution layer with a convolution kernel size of 1×1, and the outputs of the fifth point-by-point convolution layer and the fourth depth-wise separable convolution layer are multiplied. The multiplication result then passes through the sixth point-by-point convolution layer with a convolution kernel size of 1×1. The output of the sixth point-by-point convolution layer is the output of the second branch;

[0083] In the third branch, the input of the multi-scale asymmetric spatial-spectral attention unit is first divided into two parts, where the first part passes through the seventh point-by-point convolution layer with a convolution kernel size of 1×1, the fifth depth-wise separable convolution layer with a convolution kernel size of 7×1, and the sixth depth-wise separable convolution layer with a convolution kernel size of 1×7 in sequence; the second part passes through the eighth point-by-point convolution layer with a convolution kernel size of 1×1, and the outputs of the eighth point-by-point convolution layer and the sixth depth-wise separable convolution layer are multiplied, and the multiplication result passes through the ninth point-by-point convolution layer with a convolution kernel size of 1×1. The output of the ninth point-by-point convolution layer is the output of the third branch;

[0084] In the fourth branch, the input of the multi-scale asymmetric spatial-spectral attention unit is first divided into two parts. The first part passes through the tenth point-by-point convolution layer with a convolution kernel size of 1×1 and the seventh depth-wise separable convolution layer with a convolution kernel size of 1×1 in sequence; the second part passes through the eleventh point-by-point convolution layer with a convolution kernel size of 1×1, and the outputs of the eleventh point-by-point convolution layer and the seventh depth-wise separable convolution layer are multiplied. The multiplication result then passes through the twelfth point-by-point convolution layer with a convolution kernel size of 1×1. The output of the twelfth point-by-point convolution layer is the output of the fourth branch.

[0085] Its steps and parameters are the same as those of the specific implementation modes 1 to 4.

[0086] This implementation method performs parallel feature extraction on the input data in four branches. Each branch first divides the input data into two parts, and performs dense linear transformation through point-by-point convolution and depth-wise separable convolution respectively. After the two parts after linear transformation are multiplied element by element, the product is superimposed on the input data. Three of the branches are spatial attention, and asymmetric convolution is added to weight the transformation results. The other branch is spectral attention, and point-by-point grouped convolution is added to focus on the spectral information of hyperspectral data. Finally, the outputs of the four branches are superimposed to further improve the feature representation.

[0087] Example

[0088] The present invention proposes a hyperspectral image pixel-level classification method based on the MADWA network. The specific algorithm process is shown in Table 1:

[0089] Table 1 MADWA network structure algorithm flow

[0090]

[0091] The specific implementation steps are as follows:

[0092] Step 1: Obtain hyperspectral image data (the data obtained is the currently publicly available labeled data).

[0093] Step 2: Preprocess the hyperspectral image data. Randomly divide the training set and the test set according to the number of labeled pixel samples in the hyperspectral image and store them separately. Then, slice the two data sets to obtain images of the required size.

[0094] Step 3: Training of MADWA network

[0095] Step 3.1: Construct a MADWA network. The training set and test set images used in the present invention both adopt a 7×7 resolution. Figure 4 As shown, H×W represents the spatial dimension, B represents the third dimension, i.e., the band, and a series of 7×7×B images are obtained after preprocessing.

[0096] Step 3.2, the MADWA network structure is mainly divided into the three-dimensional multi-scale asymmetric convolution in the upper part, and the three-dimensional multi-scale asymmetric spatial-spectral attention module before the output. The three-dimensional multi-scale asymmetric convolution mainly includes 1×1×3, 1×3×1, 3×1×1, 1×1×5, 1×5×1, 5×1×1, 1×1×7, 1×7×1 and 7×1×1, which can fully extract feature information while reducing the number of parameters. The output features of each branch are superimposed and sent to the spatial-spectral spatial-spectral attention module.

[0097] Step 3.3, the space-spectral attention module is as follows Figure 5 As shown in the figure, it consists of three spatial attention branches and one spectral attention branch. Asymmetric convolution weights the transformation results, controls the number of channels through point-by-point convolution, and extracts spatial-spectral joint information through asymmetric depth-separable convolution. Finally, the outputs of the four branches are merged and the final classification result is obtained after the robustness loss function.

[0098] Step 4: Input the image to be classified and output the classification result.

[0099] Experimental Section

[0100] The computer configuration used in the present invention to train the network model and verify the classification results is as follows: CPU is Intel Core i9-12900K, memory is 32GB DDR5 5200, graphics card is NVIDIA GeForce RTX 3090Ti, hard disk is 1TB SSD plus 4T HDD, and the system is Windows 11 Professional Edition. The three hyperspectral datasets used are Kennedy Space Center (KSC), University of Pavia (UP) and Botswana (BO) datasets. The datasets are as follows Figure 1 , Figure 2 and Figure 3 As shown, different colors represent different types of land objects.

[0101] 1. The KSC dataset was imaged at the Kennedy Space Center in Florida, USA in March 1996, and a 512×614-sized hyperspectral dataset was captured for annotation. It was taken by the Airborne Visible Infrared Imaging Spectrometer (AVIRIS), which has a wavelength range of 400-2500nm and a spatial resolution of 18 meters. After removing the water absorption band and the low signal-to-noise ratio band, the actual bands used for training are 176, and there are 13 types of land features in the study area.

[0102] 2. The UP dataset was acquired by the ROSIS sensor during a flight over Pavia in northern Italy. The spectral range is 430-860nm, with a total of 115 bands and a spatial resolution of 1.3 meters. The UP dataset actually used for training is 103 bands with a size of 610×340 and a total of 9 types of ground objects.

[0103] 3. The BO dataset was acquired by NASA's EO-1 satellite in the Okavango Delta, Botswana in May 2001, with an image size of 1476 × 56. The wavelength range of the sensor on EO-1 is 400-2500nm, and the spatial resolution is about 20 meters. Among the 242 bands, the noise bands (1-9, 56-81, 98-101, 120-133, 165-186) were removed, and 145 bands were actually used for training. There are 14 types of land features in the study area.

[0104] The objective classification results of the MADWA network model used in the present invention on three data sets are shown in Table 2, Table 3 and Table 4:

[0105] Table 2 Classification results of KSC data by MADWA network model

[0106]

[0107] In the table, OA is the overall classification accuracy, AA is the average classification accuracy, and K is the Kappa coefficient;

[0108] Table 3 Classification results of UP data by MADWA network model

[0109]

[0110]

[0111] Table 4 Classification results of MADWA network model for BO data

[0112]

[0113] From left to right, the first column is the number of true labels; the second column is the number of noise labels; the third column is the OA data of the classification results; the fourth column is the AA data of the classification results; the fifth column is the K data of the classification results; the sixth column is the training time of the classification; and the seventh column is the test time of the classification. Figure 7 , Figure 8 and Fig. 9 As shown, (a) represents the classification effect under pure labels, (b), (c), and (d) represent the classification effects with different numbers of noisy labels. The higher the accuracy, the less salt and pepper noise in the image.

[0114] 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 pixel-level classification method based on MADWA network, characterized in that: The method specifically comprises the following steps: Step 1: Obtain a hyperspectral image dataset; Step 2: preprocess the hyperspectral images in the acquired hyperspectral image dataset; the specific process of preprocessing is as follows: Slice the hyperspectral image to obtain a sliced ​​hyperspectral image dataset, and then randomly divide the sliced ​​hyperspectral image dataset into a training set and a test set. Step 3: construct a MADWA network, and use the training set to train the constructed MADWA network until the classification accuracy of the MADWA network on the test set no longer increases or the loss function value no longer decreases, and then stop training to obtain a trained MADWA network; The working process of the MADWA network is as follows: The input hyperspectral image first passes through the first three-dimensional convolution layer in the MADWA network with a convolution kernel size of 1×1×B, where B is the spectral dimension of the input hyperspectral image. The output of the first three-dimensional convolution layer is then input into three parallel branches respectively. The output of the first three-dimensional convolutional layer passes through the second three-dimensional convolutional layer with a convolution kernel size of 1×1×1, and the output of the second three-dimensional convolutional layer is superimposed with the outputs of the three branches; After the superposition result passes through the multi-scale asymmetric spatial-spectral attention unit, the output of the multi-scale asymmetric spatial-spectral attention unit passes through the average pooling layer and the Flatten flattening layer in sequence, and the classification result is output using the Flatten flattening layer; The three branches of the MADWA network are: The first branch: The input of the first branch is used as the input of the third 3D convolution layer with a convolution kernel size of 1×1×7, the fourth 3D convolution layer with a convolution kernel size of 7×1×1, and the fifth 3D convolution layer with a convolution kernel size of 1×7×1; After superimposing the outputs of the third 3D convolutional layer, the fourth 3D convolutional layer, and the fifth 3D convolutional layer, the superimposed results are used as inputs of the sixth 3D convolutional layer with a convolution kernel size of 1×1×7, the seventh 3D convolutional layer with a convolution kernel size of 7×1×1, and the eighth 3D convolutional layer with a convolution kernel size of 1×7×1, respectively; Then, the outputs of the sixth three-dimensional convolutional layer, the seventh three-dimensional convolutional layer, and the eighth three-dimensional convolutional layer are superimposed to obtain a superposition result A, and the superposition result A is used as the output of the first branch; The second branch: The input of the second branch is used as the input of the ninth three-dimensional convolution layer with a convolution kernel size of 1×1×5, the tenth three-dimensional convolution layer with a convolution kernel size of 5×1×1, and the eleventh three-dimensional convolution layer with a convolution kernel size of 1×5×1 respectively; After superimposing the outputs of the ninth 3D convolutional layer, the tenth 3D convolutional layer, and the eleventh 3D convolutional layer, the superimposed results are used as inputs of the twelfth 3D convolutional layer with a convolution kernel size of 1×1×5, the thirteenth 3D convolutional layer with a convolution kernel size of 5×1×1, and the fourteenth 3D convolutional layer with a convolution kernel size of 1×5×1, respectively; Then, the outputs of the twelfth three-dimensional convolutional layer, the thirteenth three-dimensional convolutional layer, and the fourteenth three-dimensional convolutional layer are superimposed to obtain a superposition result B, and the superposition result B is used as the output of the second branch; The third branch: The input of the third branch is used as the input of the fifteenth three-dimensional convolution layer with a convolution kernel size of 1×1×3, the sixteenth three-dimensional convolution layer with a convolution kernel size of 3×1×1, and the seventeenth three-dimensional convolution layer with a convolution kernel size of 1×3×1 respectively; After superimposing the outputs of the fifteenth three-dimensional convolutional layer, the sixteenth three-dimensional convolutional layer, and the seventeenth three-dimensional convolutional layer, the superimposed results are used as inputs of the eighteenth three-dimensional convolutional layer with a convolution kernel size of 1×1×3, the nineteenth three-dimensional convolutional layer with a convolution kernel size of 3×1×1, and the twentieth three-dimensional convolutional layer with a convolution kernel size of 1×3×1, respectively; Then, the outputs of the eighteenth three-dimensional convolutional layer, the nineteenth three-dimensional convolutional layer, and the twentieth three-dimensional convolutional layer are superimposed to obtain a superposition result C, and the superposition result C is used as the output of the third branch; The working process of the multi-scale asymmetric spatial-spectral attention unit is: The input of the multi-scale asymmetric space-spectral attention unit first passes through four parallel branches, and the outputs of the four parallel branches are superimposed with the input of the multi-scale asymmetric space-spectral attention unit respectively. The superposition result then passes through the thirteenth point-by-point convolution layer with a convolution kernel size of 1×1. The output of the thirteenth point-by-point convolution layer is the output of the multi-scale asymmetric space-spectral attention unit; The four branches of the multi-scale asymmetric spatial-spectral attention unit are: In the first branch, the input of the multi-scale asymmetric spatial-spectral attention unit is first divided into two parts, where the first part passes through the first point-by-point convolution layer with a convolution kernel size of 1×1, the first depth-wise separable convolution layer with a convolution kernel size of 3×1, and the second depth-wise separable convolution layer with a convolution kernel size of 1×3 in sequence; the second part passes through the second point-by-point convolution layer with a convolution kernel size of 1×1, and the outputs of the second point-by-point convolution layer and the second depth-wise separable convolution layer are multiplied, and the multiplication result passes through the third point-by-point convolution layer with a convolution kernel size of 1×1. The output of the third point-by-point convolution layer is the output of the first branch; In the second branch, the input of the multi-scale asymmetric spatial-spectral attention unit is first divided into two parts, where the first part passes through the fourth point-by-point convolution layer with a convolution kernel size of 1×1, the third depth-wise separable convolution layer with a convolution kernel size of 5×1, and the fourth depth-wise separable convolution layer with a convolution kernel size of 1×5 in sequence; the second part passes through the fifth point-by-point convolution layer with a convolution kernel size of 1×1, and the outputs of the fifth point-by-point convolution layer and the fourth depth-wise separable convolution layer are multiplied. The multiplication result then passes through the sixth point-by-point convolution layer with a convolution kernel size of 1×1. The output of the sixth point-by-point convolution layer is the output of the second branch; In the third branch, the input of the multi-scale asymmetric spatial-spectral attention unit is first divided into two parts, where the first part passes through the seventh point-by-point convolution layer with a convolution kernel size of 1×1, the fifth depth-wise separable convolution layer with a convolution kernel size of 7×1, and the sixth depth-wise separable convolution layer with a convolution kernel size of 1×7 in sequence; the second part passes through the eighth point-by-point convolution layer with a convolution kernel size of 1×1, and the outputs of the eighth point-by-point convolution layer and the sixth depth-wise separable convolution layer are multiplied, and the multiplication result passes through the ninth point-by-point convolution layer with a convolution kernel size of 1×1. The output of the ninth point-by-point convolution layer is the output of the third branch; In the fourth branch, the input of the multi-scale asymmetric spatial-spectral attention unit is first divided into two parts, where the first part passes through the tenth point-by-point convolution layer with a convolution kernel size of 1×1 and the seventh depth-wise separable convolution layer with a convolution kernel size of 1×1 in sequence; the second part passes through the eleventh point-by-point convolution layer with a convolution kernel size of 1×1, and the outputs of the eleventh point-by-point convolution layer and the seventh depth-wise separable convolution layer are multiplied, and the multiplication result passes through the twelfth point-by-point convolution layer with a convolution kernel size of 1×1. The output of the twelfth point-by-point convolution layer is the output of the fourth branch; Step 4: Use the trained MADWA network to classify the hyperspectral image to be classified.

Images