Hyperspectral image classification method based on Bayesian layer graph convolutional neural network

Abstract

The present invention relates to a hyperspectral image classification method based on a Bayesian layer graph convolutional neural network, comprising the following steps: Step 1: obtaining a hyperspectral image dataset, preprocessing the hyperspectral image dataset, and then dividing the dataset into a training set and a test set according to a preset ratio; Step 2: constructing a Bayesian layer graph convolutional neural network; Step 3: training and testing the Bayesian layer graph convolutional neural network using the training set and the test set to obtain a trained Bayesian layer graph convolutional neural network; Step 4: using the trained Bayesian layer graph convolutional neural network to identify and classify targets in the hyperspectral image to be tested, to obtain a recognition result. The method of the present invention achieves high-efficiency and high-accuracy classification of remote sensing images, solves the uncertainty assessment problem in the field of remote sensing image classification, realizes the quantification of the confidence of the classification results, and solves the problem of missed detection caused by overfitting when the neural network system faces small sample targets.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing image classification, and in particular relates to a hyperspectral image classification method based on a Bayesian layer graph convolutional neural network. Background Art

[0002] After the graph convolutional neural network was proposed, it has been widely used in the field of computer vision. At the same time, researchers have proposed a variety of neural network variants based on the graph convolution idea and classification strategies for remote sensing image classification. The main representatives are neural network architectures such as GCN (graph convolutional neural network), MDGCN (multi-scale dynamic graph convolutional neural network), and S2GCN (spatial-spectral graph convolutional neural network). The above technical solution extracts and aggregates spatial features from the remote sensing image feature matrix after superpixel segmentation to obtain the spatial adjacency matrix corresponding to the feature matrix. Combined with the graph convolution idea, it designs a neural network structure that matches different data sets for training and testing.

[0003] Machine learning training strategies based on Bayesian thinking have a long history of research. With the advent of deep neural networks, Bayesian thinking has been integrated with deep neural network weight matrices through a probabilistic distributed representation. By combining probabilistic modeling with neural networks, Bayesian neural networks enable confidence assessment of prediction results and ensure strong generalization of neural networks.

[0004] The Bayesian neural networks currently used in research are relatively rigid in form and have not yet been organically integrated with deep convolutional neural networks. Researchers often simply migrate and use the Bayesian forward propagation method and Bayesian loss function, while neglecting to adjust the network structure in light of practical problems. Because the weight matrix of a Bayesian neural network follows a certain distribution, whether applying variational inference or MCMC (Markov Chain Monte Carlo method), multiple sampling and approximate calculations are required during the forward propagation. When faced with the complex structure of a deep network, transplanting a Bayesian neural network will greatly increase the training time cost, the number of parameters, and the difficulty of fitting.

[0005] In the practical application of object classification in remote sensing images, the application of traditional neural networks, such as convolutional neural networks, often faces problems such as low classification accuracy, high model complexity, and missed detections in small sample classes due to overfitting. Furthermore, when Bayesian neural networks are used for classification, network training and actual classification efficiency are limited by the large number of parameters, resulting in relatively low real-time processing capabilities. Summary of the Invention

[0006] To address the above-mentioned problems in the prior art, the present invention provides a hyperspectral image classification method based on a Bayesian layer graph convolutional neural network. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0007] The present invention provides a hyperspectral image classification method based on a Bayesian layer graph convolutional neural network, comprising:

[0008] Step 1: Obtain a hyperspectral image dataset, preprocess the hyperspectral image dataset, and then divide it into a training set and a test set according to a preset ratio;

[0009] Step 2: Build a Bayesian layer graph convolutional neural network;

[0010] Step 3: Using the training set and the test set to train and test the Bayesian layer graph convolutional neural network to obtain a trained Bayesian layer graph convolutional neural network;

[0011] Step 4: Use the trained Bayesian layer graph convolutional neural network to identify and classify the targets in the hyperspectral image to obtain the recognition results;

[0012] The hyperspectral image dataset includes a plurality of original hyperspectral remote sensing images, each of which includes a plurality of identified targets and category labels corresponding to the targets;

[0013] The Bayesian layer graph convolutional neural network includes a cascaded fully connected module, a first Bayesian graph convolutional layer module, a Dropout layer, a second Bayesian graph convolutional layer module and a Softmax layer.

[0014] In one embodiment of the present invention, preprocessing the hyperspectral image dataset includes:

[0015] For each of the original hyperspectral remote sensing images, a superpixel segmentation algorithm is used to segment the image to obtain a corresponding superpixel block;

[0016] For each superpixel block, the category with the most labels in the block is selected as the category label of the superpixel block, and the superpixel blocks with the category label of background are removed to obtain the superpixel blocks representing the ground object category;

[0017] Obtaining an adjacency matrix of the superpixel block representing the ground object category;

[0018] A feature vector corresponding to the super pixel block representing the ground object category is constructed, and a feature matrix corresponding to the super pixel block representing the ground object category is obtained based on the feature vector and the corresponding category label.

[0019] In one embodiment of the present invention, the feature vector includes spectral features and texture features.

[0020] In one embodiment of the present invention, the fully connected module includes a cascaded first fully connected layer, a ReLU activation layer, and a second fully connected layer.

[0021] In one embodiment of the present invention, the Bayesian graph convolution layer module includes a cascaded Bayesian graph convolution layer, a ReLU activation layer and a BN layer.

[0022] In one embodiment of the present invention, the Bayesian graph convolution layer is a fully connected layer in which weight parameters and biases are represented in a distributed form;

[0023] The Bayesian graph convolution layer is used to perform weighted calculation on the input matrix according to the weight parameter and then perform convolution calculation with the adjacency matrix to obtain a spatial fusion feature matrix.

[0024] In one embodiment of the present invention, the loss function of the Bayesian layer graph convolutional neural network is:

[0025] loss=KL[q(ω(||p(ω|Y,X)]-NLL_Loss;

[0026] Where, p(ω|Y, X) represents the posterior distribution, q(ω) is a simple distribution used to fit the posterior distribution, NLL_Loss represents the multi-classification task loss function, ω represents the weight parameter of the Bayesian graph convolutional layer, and X and Y represent the input and output of the Bayesian layer graph convolutional neural network, respectively.

[0027] In one embodiment of the present invention, the training process of the Bayesian layer graph convolutional neural network includes:

[0028] The training set is input into the constructed Bayesian layer graph convolutional neural network in batches. During each forward propagation process, the weight distribution of each Bayesian layer graph convolutional neural network is sampled once, and the output Y of the Bayesian layer graph convolutional neural network is obtained by forward propagation;

[0029] In each forward propagation process, the weight parameters of the Bayesian graph convolution layer are sampled multiple times to calculate the first term of the loss function;

[0030] The multi-classification task loss function is used to calculate the loss between the output Y of the Bayesian layer graph convolutional neural network and the label Y_label of the corresponding training set as the second term of the loss function;

[0031] According to the first and second terms of the loss function, a loss value of the loss function of the Bayesian layer graph convolutional neural network is obtained, and according to the loss value, the weight parameters of the Bayesian graph convolutional layer and other fully connected layer parameters are updated using a backpropagation algorithm;

[0032] The above training steps are repeated until the loss value calculated after a batch of training samples is input into the Bayesian layer graph convolutional neural network is less than the preset threshold, and the training is completed.

[0033] In one embodiment of the present invention, the recognition result includes the target recognition result of the hyperspectral image to be measured, the accuracy rate, and the confidence interval [a, b] of the accuracy rate. The confidence interval [a, b] of the accuracy rate is calculated according to the following formula:

[0034] a=μ Y -|z|×SE;

[0035] b=μ Y +|z|×SE;

[0036] Among them, μ Y represents the sample mean of the Bayesian layer graph convolutional neural network output, SE represents the standard error, and z represents the standard score.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] The hyperspectral image classification method based on the Bayesian layer graph convolutional neural network of the present invention combines the Bayesian deep learning idea, proposes the concept of Bayesian layer, realizes the quantification and output of uncertainty, and aims at the remote sensing image classification problem by combining the Bayesian layer with the graph convolutional neural network, proposes the concept of Bayesian graph convolutional layer, establishes the Bayesian layer graph convolutional neural network, and uses the trained Bayesian layer graph convolutional neural network to realize high-efficiency and high-accuracy classification of remote sensing images, solves the uncertainty assessment problem in the classification field of remote sensing images, realizes the quantification of the confidence of the classification results, and solves the problem of missed alarms that may be caused by overfitting when the neural network system faces small sample targets.

[0039] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the following preferred embodiments are specifically cited and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 Schematic diagram of a hyperspectral image classification method based on a Bayesian layer graph convolutional neural network provided by an embodiment of the present invention;

[0041] Figure 2 Schematic diagram of the structure of the Bayesian layer graph convolutional neural network provided by an embodiment of the present invention;

[0042] Figure 3 2 is a schematic diagram of a Bayesian graph convolution layer provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0043] In order to further illustrate the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the following is a detailed description of a hyperspectral image classification method based on a Bayesian layer graph convolutional neural network proposed in accordance with the present invention, in combination with the accompanying drawings and specific embodiments.

[0044] The aforementioned and other technical contents, features, and effects of the present invention are clearly presented in the following detailed description of the specific embodiments in conjunction with the accompanying drawings. Through the description of the specific embodiments, a deeper and more specific understanding of the technical means and effects adopted by the present invention to achieve the intended purpose can be obtained. However, the accompanying drawings are provided for reference and illustration purposes only and are not intended to limit the technical solutions of the present invention.

[0045] Example 1

[0046] See Figure 1 , Figure 1 : This is a schematic diagram of a hyperspectral image classification method based on a Bayesian layer graph convolutional neural network provided by an embodiment of the present invention. As shown in the figure, the hyperspectral image classification method based on a Bayesian layer graph convolutional neural network of this embodiment includes:

[0047] Step 1: Obtain a hyperspectral image dataset, preprocess the hyperspectral image dataset, and then divide it into a training set and a test set according to a preset ratio;

[0048] In this embodiment, the hyperspectral image dataset includes a number of original hyperspectral remote sensing images, and each original hyperspectral remote sensing image includes a plurality of identified targets and category labels corresponding to the targets.

[0049] Optionally, the hyperspectral image dataset is one of four hyperspectral remote sensing datasets: Indian Pines, Salinas, Pavia University, and Houston University.

[0050] In this embodiment, the Indian Pines dataset is selected as an example to illustrate several steps of classifying a hyperspectral remote sensing image.

[0051] Specifically, in this implementation, the hyperspectral image dataset is preprocessed, including:

[0052] Step a: For each original hyperspectral remote sensing image, the superpixel segmentation (SLIC) algorithm is used to segment the image into corresponding superpixel blocks;

[0053] Step b: For each superpixel block, select the category with the most labels in the block as the category label of the superpixel block, remove the superpixel blocks with the category label of background, and obtain the superpixel blocks representing the ground object category;

[0054] Step c: Obtain the adjacency matrix of the superpixel block representing the ground object category;

[0055] Specifically, for the superpixel blocks representing the ground object categories, the adjacency relationship between the superpixel blocks representing the ground object categories is obtained based on the spatial adjacency relationship between the original pixel points, and is converted into an adjacency matrix.

[0056] Step d: Construct a feature vector corresponding to the superpixel block representing the ground object category, and obtain a feature matrix corresponding to the superpixel block representing the ground object category based on the feature vector and the corresponding category label.

[0057] In this embodiment, the feature vector includes spectral features and texture features. Specifically, the spectral features and texture features of the superpixel block are used to construct a superpixel block feature vector, and are saved in matrix form in combination with the superpixel block label to obtain a feature matrix corresponding to the superpixel block representing the ground feature category.

[0058] Step 2: Build a Bayesian layer graph convolutional neural network;

[0059] Please refer to Figure 2 The schematic diagram of the structure of the Bayesian layer graph convolutional neural network shown in the figure, in this embodiment, the Bayesian layer graph convolutional neural network includes a cascaded fully connected module, a first Bayesian graph convolutional layer module, a Dropout layer, a second Bayesian graph convolutional layer module and a Softmax layer.

[0060] Specifically, the fully connected module includes a cascaded first fully connected layer, a ReLU activation layer, and a second fully connected layer, wherein the size of the first fully connected layer is 401×512 and the size of the second fully connected layer is 512×128.

[0061] Furthermore, the Bayesian graph convolution layer module includes a cascaded Bayesian graph convolution layer, a ReLU activation layer, and a BN layer. In the first Bayesian graph convolution layer module, the Bayesian graph convolution layer has a size of 128×64, and the BN (Batch Normalization) layer has a size of 64 dimensions. In the second Bayesian graph convolution layer module, the Bayesian graph convolution layer has a size of 64×16, and the BN (Batch Normalization) layer has a size of 16 dimensions.

[0062] In this embodiment, a Dropout layer is provided between the first Bayesian graph convolution layer module and the second Bayesian graph convolution layer module to reduce overfitting. A Softmax layer is used to normalize the output of the second Bayesian graph convolution layer module.

[0063] Furthermore, in this embodiment, the Bayesian graph convolution layer is a fully connected layer in which both weight parameters and bias are represented in a distributed form. The Bayesian graph convolution layer is used to perform a weighted calculation on the input matrix according to the weight parameters and then convolve it with the adjacency matrix to obtain a spatial fusion feature matrix.

[0064] like Figure 3 Figure 1 shows a schematic diagram of the Bayesian graph convolution layer. For the Indian Pines hyperspectral image dataset, the corresponding feature matrix is obtained through preprocessing. After the Bayesian layer operation, the feature matrix outputs feature matrices with dimensions of 3880×64 and 3880×16 respectively. The adjacency matrix with dimension of 3880×3880 obtained during the preprocessing process is used to perform convolution operation with the two feature matrices to combine and extract the spatial features of the remote sensing image to obtain the spatial fusion feature matrix, thereby realizing a complete Bayesian graph convolution layer in structure based on the Bayesian layer.

[0065] In this embodiment, the loss function of the Bayesian layer graph convolutional neural network is:

[0066] loss=KL[q(ω)||p(ω|Y,X)]-NLL_Loss (1);

[0067] Where, p(ω|Y, X) represents the posterior distribution, q(ω) is a simple distribution used to fit the posterior distribution, NLL_Loss represents the multi-classification task loss function, ω represents the weight parameter of the Bayesian graph convolutional layer, and X and Y represent the input and output of the Bayesian layer graph convolutional neural network, respectively.

[0068] Furthermore, the design of the loss function of the above-mentioned Bayesian layer graph convolutional neural network is specifically described as follows.

[0069] The weight parameter ω of the Bayesian neural network is expressed in the form of distribution, Y and X represent the neural network output and the corresponding neural network input as training data, respectively, x * Represents the feature vector as the test data, then the Bayesian neural network outputs the corresponding predicted value y * The probability P(y * |x * , Y, X) can be calculated by formula (2),

[0070] P(y * |x * , Y, X) = ∫[(y* |x * ,ω)p(ω|Y,X) (2),

[0071] It is necessary to make an approximate inference of the posterior probability distribution p(ω|Y, X) of the weight parameter expressed in formula (3).

[0072]

[0073] Using the variational inference method, we use the simple distribution q(ω) to approximate the posterior distribution p(ω|Y, X). In order to make the fitting effect as good as possible, we need to minimize KL[q(ω)||p(ω|Y, X)],

[0074]

[0075] In order to maximize the uncertainty, it is assumed that both the approximate distribution and the posterior distribution obey Gaussian distribution, that is,

[0076] q(ω)=q(ω|θ), θ=(μ,σ 2 ) (5);

[0077]

[0078] Among them, μ and σ 2 is the mean and variance of the parameters θ in the approximate distribution q(ω), μ i and σ i 2 is the weight w i The mean and variance of the distribution obeyed, the original problem, that is, minimizing KL[q(ω)||p(ω|Y,X)], can be equivalent to optimizing the selection of the optimal parameter θ * :

[0079]

[0080] The loss function of the Bayesian neural network is defined as:

[0081]

[0082] Apply the reparameterization technique, for the weights There is ω i =μ i +σ i ⊙∈ i ,in ⊙ represents the Hadamard product operation.

[0083] For the back propagation of the loss function,

[0084]

[0085] In order to ensure that the range of the θ parameter value includes the entire real axis, it is necessary to resample σ, so that σ=log(1+e ρ ), then θ=(μ, ρ).

[0086] In the loss function of the Bayesian layer graph convolutional neural network of this embodiment, the multi-classification task loss is used to replace the likelihood function term (the second term) in the loss function of the above-mentioned Bayesian neural network (Formula (8)). Under the premise of a given output y, the likelihood function L(ω|y) about the parameter ω is numerically equal to the probability p(Y=y|ω) that the output is y after the given parameter ω. Therefore, the multi-classification task loss function NLL_Loss of the traditional neural network is used to replace the likelihood function term to obtain the loss function of the Bayesian layer graph convolutional neural network of this embodiment.

[0087] It should be noted that, in this embodiment, the Bayesian graph convolution layer is defined as a fully connected layer in which all weight parameters and biases are expressed in the form of distributions. At the same time, it is stipulated that the loss function must have a KL divergence form optimized by the variational inference method. When performing gradient calculation and backpropagation to update parameters of the constructed Bayesian layer graph convolutional neural network, since the prior and posterior only involve the weights within the Bayesian graph convolutional layer, the multi-classification loss function that replaces the likelihood function includes all weight parameters of the entire Bayesian layer graph convolutional neural network. Therefore, the parameter updates for the Bayesian graph convolutional layer and the non-Bayesian graph convolutional layer can be performed sequentially under the same process without interfering with each other.

[0088] Step 3: Use the training set and test set to train and test the Bayesian layer graph convolutional neural network to obtain the trained Bayesian layer graph convolutional neural network;

[0089] Specifically, in this embodiment, the training process of the Bayesian layer graph convolutional neural network includes:

[0090] Step ①: Input the training set into the constructed Bayesian layer graph convolutional neural network in batches. During each forward propagation process, the weight distribution of each Bayesian layer graph convolutional neural network is sampled once, and the output Y of the Bayesian layer graph convolutional neural network is obtained by forward propagation.

[0091] Step 2: During each forward propagation process, the weight parameters of the Bayesian graph convolution layer are sampled multiple times to calculate the first term of the loss function.

[0092] In this embodiment, during each forward propagation process, the weight parameters of the Bayesian graph convolution layer are sampled 30 times to calculate the first term of the loss function, the KL divergence.

[0093] Step 3: Use the multi-classification task loss function to calculate the loss between the output Y of the Bayesian layer graph convolutional neural network and the label Y_label of the corresponding training set as the second term of the loss function;

[0094] Step 4: According to the first and second terms of the loss function, the loss value of the Bayesian layer graph convolutional neural network loss function is obtained. According to the loss value, the weight parameters of the Bayesian graph convolutional layer and other fully connected layer parameters are updated using the back propagation algorithm;

[0095] In this embodiment, the first and second terms of the calculated loss function are subtracted as the loss value of the overall loss function, and the weight parameters (μ and σ) of the Bayesian graph convolution layer are updated using the back propagation algorithm. 2 ) and other fully connected layer parameters.

[0096] Step ⑤, repeat steps ①-④ until the loss value calculated after a batch of training samples is input into the Bayesian layer graph convolutional neural network is less than the preset threshold, and the training is completed.

[0097] For the trained Bayesian layer graph convolutional neural network, the test set is used to test it to verify the network performance.

[0098] Step 4: Use the trained Bayesian layer graph convolutional neural network to identify and classify the targets in the hyperspectral image to obtain the recognition results;

[0099] In this embodiment, the recognition result includes the target recognition result of the hyperspectral image to be tested, the accuracy rate, and the confidence interval [a, b] of the accuracy rate. The confidence interval [a, b] of the accuracy rate is calculated according to the following formula:

[0100] a=μ Y -|z|×SE;

[0101] b=μ Y +|z|×SE;

[0102] Among them, μ Y represents the sample mean of the Bayesian layer graph convolutional neural network output, SE represents the standard error, and z represents the standard score.

[0103] In this embodiment, the multi-classification task loss calculated based on the Bayesian layer graph convolutional neural network output and the corresponding labels is defined as the accuracy of the neural network, and statistical knowledge is applied to calculate the confidence interval of the accuracy. For the case of a large sample size of the Indian Pines dataset, a Z test is used to determine whether there is a significant difference between the sample mean and the overall mean of the neural network output during the same forward propagation process. First, calculate the sample mean μ Yand standard error SE, then determine the confidence level and obtain the standard score z. When the confidence level is set to 95%, the standard score z is numerically equal to 1.96.

[0104] It should be noted that the above-mentioned accuracy confidence interval calculation formula can be used to calculate the confidence level of the final output accuracy, and the classification results after processing by the Bayesian layer graph convolutional neural network can also be calculated under the premise of a given confidence level.

[0105] The hyperspectral image classification method based on the Bayesian layer graph convolutional neural network of this embodiment combines the Bayesian deep learning idea, proposes the concept of Bayesian layer, realizes the quantification and output of uncertainty, and for the remote sensing image classification problem, combines the Bayesian layer with the graph convolutional neural network, proposes the concept of Bayesian graph convolutional layer, establishes a Bayesian layer graph convolutional neural network, and uses the trained Bayesian layer graph convolutional neural network to achieve high-efficiency and high-accuracy classification of remote sensing images, solves the uncertainty assessment problem in the classification field of remote sensing images, realizes the quantification of the confidence of the classification results, and solves the problem of missed alarms that may be caused by overfitting when the neural network system faces small sample targets.

[0106] Example 2

[0107] This embodiment illustrates the effect of the hyperspectral image classification method based on the Bayesian layer graph convolutional neural network of embodiment one through simulation experiments. In this embodiment, the Bayesian layer graph convolutional neural network of the present invention is designed to be compared with the remote sensing image classification processing model commonly used in the current research field, and SVM (support vector machine), 3D-CNN (three-dimensional convolutional neural network), GCN, S2GCN, MDGCN and other algorithms are selected as comparison algorithms. The codes of the comparison algorithms are all open source on the relevant websites. Four commonly used hyperspectral remote sensing image data sets, such as Indian Pines, Salinas, Pavia University and Houston University, were selected for testing. The selected training set samples account for 10% of all samples, and the neural network learning rate is set to 1×10 -3 , and set the Dropout and weight loss ratios to 0.2 and 5×10 respectively -4 .

[0108] OA (overall classification accuracy), AA (average classification accuracy), and the Kappa coefficient were selected as evaluation indicators. Overall classification accuracy refers to the percentage of samples correctly classified by the classifier over the total number of samples in a given test dataset. Class classification accuracy refers to the percentage of correctly classified samples in each class over the total number of samples in that class. The Kappa coefficient is calculated by multiplying the total number of correctly classified pixels in all objects by the sum of the diagonal of the confusion matrix, minus the product of the total number of ground object pixels in a certain class and the total number of classified pixels in that class. This sum is then divided by the square of the total number of pixels minus the product of the total number of ground object pixels in a certain class and the total number of classified pixels in that class.

[0109] The training and classification results are listed in Tables 1 to 4. It can be seen that the proposed classification method achieves better classification results on all four datasets than the control algorithm. This means that the proposed method can effectively improve the accuracy of ground object classification for remote sensing image classification problems.

[0110] At the same time, during the forward propagation of the neural network, the forward propagation calculation was repeated 30 times, and the classification results obtained were averaged as the classification result obtained in that iteration. Based on the above, it can be estimated that the final classification result obtained by this method has a 95% confidence level. This effectively avoids the problem of missed detections due to the small amount of data for specific sample types.

[0111] Table 1 Classification results of Indian Pines dataset training

[0112]

[0113] Table 2 Classification results of Pavia University dataset training

[0114]

[0115] Table 3 Salinas dataset training classification results

[0116]

[0117] Table 4 Houston University dataset training classification results

[0118]

[0119]

[0120] It should be noted that, in this document, relational terms such as first and second are used solely to distinguish one entity or operation from another, and do not necessarily require or imply any actual relationship or order between these entities or operations. Furthermore, the terms "comprise," "include," or any other variations thereof are intended to encompass non-exclusive inclusion, such that an article or device comprising a list of elements includes not only those elements but also other elements not explicitly listed. Without further limitation, an element defined by the phrase "comprising a..." does not preclude the presence of additional identical elements in the article or device comprising the element. Terms such as "connected" or "connected" are not limited to physical or mechanical connections but may include electrical connections, whether direct or indirect. References to orientations or positional relationships, such as "upper," "lower," "left," and "right," are based on the orientations or positional relationships shown in the accompanying drawings and are intended solely to facilitate description and simplify the description of the present invention. They do not indicate or imply that the device or element referred to must have, be constructed, or operate in a specific orientation, and are therefore not to be construed as limiting the present invention.

[0121] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should be considered to fall within the scope of protection of the present invention.

Claims

1. A hyperspectral image classification method based on Bayesian layer graph convolutional neural network, characterized in that: include: Step 1: Obtain a hyperspectral image dataset, preprocess the hyperspectral image dataset, and then divide it into a training set and a test set according to a preset ratio; Preprocessing the hyperspectral image dataset includes: For each of the original hyperspectral remote sensing images, a superpixel segmentation algorithm is used to segment the image to obtain a corresponding superpixel block; For each superpixel block, the category with the most labels in the block is selected as the category label of the superpixel block, and the superpixel blocks with the category label of background are removed to obtain the superpixel blocks representing the ground object category; Obtaining an adjacency matrix of the superpixel block representing the ground object category; Constructing a feature vector corresponding to the superpixel block representing the ground object category, and obtaining a feature matrix corresponding to the superpixel block representing the ground object category based on the feature vector and the corresponding category label; Step 2: Build a Bayesian layer graph convolutional neural network; Step 3: Using the training set and the test set to train and test the Bayesian layer graph convolutional neural network to obtain a trained Bayesian layer graph convolutional neural network; Step 4: Use the trained Bayesian layer graph convolutional neural network to identify and classify the targets in the hyperspectral image to obtain the recognition results; The hyperspectral image dataset includes a plurality of original hyperspectral remote sensing images, each of which includes a plurality of identified targets and category labels corresponding to the targets; The Bayesian layer graph convolutional neural network includes a cascaded fully connected module, a first Bayesian graph convolutional layer module, a Dropout layer, a second Bayesian graph convolutional layer module and a Softmax layer; the Bayesian graph convolutional layer module includes a cascaded Bayesian graph convolutional layer, a ReLU activation layer and a BN layer; the Bayesian graph convolutional layer is a fully connected layer in which weight parameters and biases are expressed in a distributed form; the Bayesian graph convolutional layer is used to perform a weighted calculation on the input matrix according to the weight parameters and then perform a convolution calculation with the adjacency matrix to obtain a spatial fusion feature matrix.

2. The hyperspectral image classification method based on Bayesian layer graph convolutional neural network according to claim 1 is characterized in that The feature vector includes spectral features and texture features.

3. The hyperspectral image classification method based on Bayesian layer graph convolutional neural network according to claim 1 is characterized in that The fully connected module includes a cascaded first fully connected layer, a ReLU activation layer, and a second fully connected layer.

4. The hyperspectral image classification method based on Bayesian layer graph convolutional neural network according to claim 1 is characterized in that The loss function of the Bayesian layer graph convolutional neural network is: ； Where, , represents the posterior distribution, is a simple distribution used to fit the posterior distribution, represents the multi-classification task loss function, represents the weight parameters of the Bayesian graph convolution layer, and Represent the input and output of the Bayesian layer graph convolutional neural network respectively.

5. The hyperspectral image classification method based on Bayesian layer graph convolutional neural network according to claim 4 is characterized in that The training process of the Bayesian layer graph convolutional neural network includes: The training set is input into the constructed Bayesian layer graph convolutional neural network in batches. During each forward propagation process, the weight distribution of each Bayesian layer graph convolutional neural network is sampled once, and the output Y of the Bayesian layer graph convolutional neural network is obtained by forward propagation; In each forward propagation process, the weight parameters of the Bayesian graph convolution layer are sampled multiple times to calculate the first term of the loss function; The multi-classification task loss function is used to calculate the loss between the output Y of the Bayesian layer graph convolutional neural network and the label Y_label of the corresponding training set as the second term of the loss function; According to the first and second terms of the loss function, a loss value of the loss function of the Bayesian layer graph convolutional neural network is obtained, and according to the loss value, the weight parameters of the Bayesian graph convolutional layer and other fully connected layer parameters are updated using a backpropagation algorithm; The above training steps are repeated until the loss value calculated after a batch of training samples is input into the Bayesian layer graph convolutional neural network is less than the preset threshold, and the training is completed.

6. The hyperspectral image classification method based on Bayesian layer graph convolutional neural network according to claim 1 is characterized in that The recognition result includes the target recognition result, accuracy rate and confidence interval of the hyperspectral image to be tested. , the confidence interval of the accuracy Calculated according to the following formula: ； ； in, represents the sample mean of the Bayesian layer graph convolutional neural network output, SE represents the standard error, and z represents the standard score.

Images