Hyperspectral image classification uncertainty estimation method and system

Abstract

The invention relates to a hyperspectral image classification uncertainty estimation method and system, and the method comprises the steps: firstly constructing a hyperspectral image classification architecture based on a deep neural network, and enabling the architecture to be divided into a cascade structure of a feature extractor and a classifier; secondly, the space correlation Gaussian mixture model is embedded into an output layer, and the output layer based on the space correlation Gaussian mixture model is obtained and used for calculating the output probability; then, solving the space correlation Gaussian mixture model by using a variational Bayesian inference method, and estimating initial variation posteriori distribution; and finally, considering spatial relevance of the hyperspectral image, and improving a variational Bayesian inference method to obtain final variational posteriori distribution, namely uncertainty expression of classified output of the hyperspectral image. According to the method, the spatial correlation Gaussian mixture model is introduced to carry out probabilistic expression on the output feature space of the deep neural network, and the probability expression is used as confidence estimation output by the classification model, so that the credibility of the model is enhanced.

Description

Technical Field

[0001] The present invention belongs to the field of image processing, and particularly relates to a method and system for estimating the uncertainty of hyperspectral image classification. Background Art

[0002] In the current era of rapid technological development, hyperspectral image classification technology plays a crucial role in many fields. Hyperspectral images can obtain information of ground objects in continuous spectral bands, just like providing us with a fine "spectral fingerprint library", enabling us to more accurately identify and distinguish different ground object types. In the agricultural field, it can accurately monitor the growth status of crops, determine whether the crops are damaged by pests and diseases, and provide strong support for precision agriculture; in mineral resource exploration, it can help us quickly locate potential mineral resources; in environmental monitoring, it can also effectively monitor water pollution, vegetation cover changes, etc. Therefore, hyperspectral image classification technology is of irreplaceable significance for promoting the intelligent and refined development of various industries.

[0003] In recent years, deep neural networks have achieved significant performance improvements in the field of hyperspectral image classification. However, designing a sufficiently robust and reliable hyperspectral image classification method remains full of challenges. Deep neural networks often output definite prediction results, but do not provide the confidence intervals of the predicted values, so it is difficult to evaluate the uncertainty based on the output of the deep neural network itself to reflect the confidence level and quality level of the classification results. Therefore, uncertainty modeling of deep neural networks is of great significance. How to measure and quantify the uncertainty of deep neural networks and clarify the confidence level of hyperspectral image classification results has important value for improving the stability and reliability of hyperspectral image classification networks, and can also enable downstream algorithm models to make reasonable decisions when the uncertainty of input information is known. To solve this problem, uncertainty estimation is used to estimate the uncertainty of the output of deep neural networks, thereby further improving the accuracy and reliability of hyperspectral image classification results.

[0004] To achieve the purpose of estimating the output uncertainty of deep neural networks, recently, mainly through Monte Carlo sampling methods and distribution assumption-based methods. Among them, the distribution assumption-based methods mainly adopt some distribution forms, such as Gaussian distribution, Dirichlet distribution, Softmax distribution, etc., to model the output of deep neural networks and define uncertainty on these outputs. It should be noted that most of the above methods assume that the observed data are independent. However, for the hyperspectral image classification task, due to the spatial similarity of hyperspectral images, each pixel is highly correlated with its surrounding adjacent pixels. Therefore, most existing methods for estimating the output uncertainty of deep neural networks are difficult to well describe the information in hyperspectral images and there are certain biases.

[0005] To this end, the present invention establishes a method for estimating the output uncertainty of hyperspectral image classification based on a spatially associated Gaussian mixture model. The Gaussian mixture model is introduced to probabilistically represent the output feature space of the deep neural network, which is used as the confidence estimation of the classification model output. The variational Bayesian inference method is improved by considering the spatial association information in the hyperspectral image to solve the spatially associated Gaussian mixture model, thereby obtaining the probability distribution of the classification result uncertainty, which is used to characterize the confidence level and quality level of the classification result. Summary of the Invention

[0006] The present invention aims to design a method for estimating the output uncertainty of hyperspectral image classification based on a spatially associated Gaussian mixture model, which can measure the uncertainty of the output of a deep hyperspectral image classification network through a probability model and improve the credibility of the prediction result.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] The method for estimating the uncertainty of hyperspectral image classification includes the following steps:

[0009] Step S1, construct a hyperspectral image classification architecture based on a deep neural network;

[0010] Step S2, embed the Gaussian mixture model into the output layer to obtain an output layer based on the Gaussian mixture model for calculating the output probability;

[0011] Step S3, use the variational Bayesian inference method to solve the Gaussian mixture model and estimate the initial variational posterior distribution;

[0012] Step S4, use the spatial correlation of the hyperspectral image to improve the variational Bayesian inference method in Step S3, solve the spatially associated Gaussian mixture model, and obtain the final variational posterior distribution, that is, the uncertainty expression of the hyperspectral image classification output.

[0013] Further, the Step S1 includes:

[0014] Step S1-1, determine the model architecture, which can be divided into a cascaded structure of two parts, namely, the convolutional layer and the output layer, regarded as the feature extractor and classifier of the hyperspectral image respectively;

[0015] Step S1-2, design the feature extractor. According to the characteristics of the hyperspectral image, select a deep neural network to extract features and build multiple convolutional layers to extract the spatial features of the image;

[0016] Step S1-3, design the classifier. Design a fully connected layer to map the extracted features to the category space and design an output layer to generate the probability of each category.

[0017] Further, step S1 further includes:

[0018] Step S1-4: Use the sequence of hyperspectral images as the data input of the model;

[0019] Step S1-5: Perform data processing on the input data to calculate and output the prediction result.

[0020] Further, step S2 includes:

[0021] Step S2-1: Define the parameter structure of the Gaussian mixture model, including the mean vector, covariance matrix, and mixing weight of each Gaussian distribution;

[0022] Step S2-2: Design the structure of the output layer to output the parameters of the Gaussian mixture model;

[0023] Step S2-3: Calculate the probability output of the mixture Gaussian model;

[0024] Step S2-4: Determine the ground object category according to the probability in step S2-3. The method for determining the ground object category by probability also includes converting the probability into a representation of uncertainty.

[0025] Further, the probability calculation method in step S2 includes:

[0026] Determine the pixel, define as the pixel located in the m-th row and n-th column of the hyperspectral image, where L represents the spectral dimension, 1 ≤ m ≤ M, 1 ≤ n ≤ N;

[0027] After being processed by the feature extractor, each pixel is converted into a feature vector, is the feature after the feature extractor, where J represents the feature dimension;

[0028] When the hyperspectral image contains C categories, a Gaussian mixture model is established for each category, and the corresponding probability density function is:

[0029]

[0030] Among them, is the Gaussian distribution, K is the number of components of each Gaussian mixture model, μ = {μ ik}, ∑ = {∑ ik}, and η = {η ik} represent the parameter sets of the mean, covariance, and mixing weight respectively, ω = [ω 1 , ω 2 , …, ω C T represents the non-negative mixing weights of C Gaussian mixture models, and ω​i Its value can be directly determined by the proportion of each type of sample contained in the training set.

[0031] Furthermore, the step S3 includes:

[0032] Step S3-1: Define the model structure elements, where the parameter set is θ = {μ, ∑}; the latent variable tensor is The pixel feature z at the m-th row and n-th column mn is associated with the k-th mixture component, then otherwise D = {Z, E} is the data set, where represents the feature tensor of the hyperspectral image, represents the label map of the hyperspectral image;

[0033] Step S3-2: Determine the objective of variational Bayesian inference. The variational Bayesian inference estimation aims to solve the mixture model (1) to maximize the posterior distribution p(Z|θ) of the data D given the parameter θ:

[0034]

[0035] Furthermore, the step S3 also includes:

[0036] Step S3-3: Introduction of the variational distribution. To solve (2), the variational distribution q is used to approximate the original distribution p. In variational Bayesian inference, the variational distribution q comes from the variational family and approximates the posterior distribution by minimizing the KL divergence;

[0037] Step S3-4: Log-likelihood decomposition. According to the variational Bayesian inference theory, the log-likelihood of p(D|θ) can be decomposed into:

[0038]

[0039] where ELBO represents the expected log-likelihood, and KL(q(V)||p(V|D,θ)) is the KL divergence from q(V) to the prior distribution p(V|D,θ), and they can be expressed in the following forms:

[0040]

[0041] Step S3-5: Parameter optimization. According to the variational Bayesian inference theory, the two parameters V and θ can be optimized, and the representation form of the initial optimal variational posterior distribution q * (V) can be determined.

[0042] Furthermore, the step S4 includes:

[0043] Step S4-1: Introduce spatial correlation to define the spatially correlated variational distribution. For a pixel in a hyperspectral image, there is a high correlation with its neighboring pixels. Therefore, consider the optimal variational distribution q * (v mn ) can be expressed as:

[0044] q * (v mn ) = q(v mn |v (m-1)n , v (m+1)n , v m(n-1) , v m(n+1) ) (6)

[0045] where v mn represents the latent variable vector at the m-th row and n-th column in the latent variable tensor V; further, the above equation is equivalent to:

[0046]

[0047] According to the conditional probability formula, Equation (7) can be further transformed into:

[0048]

[0049] where k', k'', k''' and k'''' respectively represent the mixture components with the maximum response of the data at the (m - 1)-th row and n-th column, the (m + 1)-th row and n-th column, the m-th row and (n - 1)-th column, and the m-th row and (n + 1)-th column;

[0050] Step S4-2: Define the joint probability. Considering the information of the mixture components of the data at the m-th row and n-th column and the data at the (m - 1)-th row and n-th column, the (m + 1)-th row and n-th column, the m-th row and (n - 1)-th column, and the m-th row and (n + 1)-th column, the following definition can be made:

[0051]

[0052] where

[0053]

[0054] represents the proportion of the k-th, k'-th, k''-th, k'''-th and k''''-th mixture components;

[0055] Step S4-3: Define the calculation of the marginal probability:

[0056]

[0057]

[0058] where

[0059]

[0060] Indicate the proportions of the k', k'', k''', and k''''th mixed components.

[0061] Furthermore, step S4 further includes:

[0062] Step S4-4: Calculate the final variational distribution q * (v mn ). Substituting equations (9) and (10) into equation (8), we can obtain the expression form of, that is, obtain the final variational distribution q * (v mn );

[0063] Step S4-5: Output the uncertainty estimate. This variational distribution is the uncertainty expression of the final hyperspectral image classification output.

[0064] Furthermore, an embodiment of the present invention also provides a hyperspectral image classification uncertainty estimation system, which includes:

[0065] A data preprocessing module for preprocessing the original hyperspectral image;

[0066] A feature extractor that automatically extracts the spatial and spectral features of the image using a deep neural network method. Input the preprocessed hyperspectral image and output the image feature vector;

[0067] A Gaussian mixture model embedding module that embeds the Gaussian mixture model into the output layer to calculate the probability distribution of each category. Input the feature vector output by the feature extractor and output the category probability prediction based on the Gaussian mixture model;

[0068] A variational Bayesian inference module that uses the variational Bayesian inference method to solve the Gaussian mixture model parameters and estimate the initial variational posterior distribution. Input the probability prediction output of the Gaussian mixture model and the training data, and output the initial variational posterior distribution;

[0069] A spatial correlation fusion module that considers the spatial correlation of the hyperspectral image, optimizes the variational Bayesian inference technology, and obtains the final variational posterior distribution. Input the initial variational posterior distribution and the spatial correlation information of the image, and output the final variational posterior distribution, expressing the uncertainty of classification;

[0070] A classification decision module that determines the category label of each pixel according to the output of the variational posterior distribution. Input the final variational posterior distribution and output the classification result and uncertainty estimate.

[0071] Compared with the prior art, the present invention adopts the above technical solutions and has the following beneficial effects:

[0072] (1) The hyperspectral image classification uncertainty estimation method and system provided by the present invention use uncertainty estimation to summarize the output of a deep neural network, thereby further improving the accuracy and reliability of hyperspectral image classification results.

[0073] (2) The hyperspectral image classification uncertainty estimation method and system provided by the present invention can measure and quantify the uncertainty of a deep neural network, clarify the confidence level of hyperspectral image classification results, have important value for improving the stability and reliability of hyperspectral image classification networks, and can also enable downstream algorithm models to make reasonable decisions when the uncertainty of input information is known.

[0074] (3) The hyperspectral image classification uncertainty estimation method and system provided by the present invention consider the spatial correlation of pixels in hyperspectral images and improve the variational Bayesian inference method, making the classification results more accurate and reliable. Description of the Drawings

[0075] Figure 1 is the flowchart of the hyperspectral image classification uncertainty estimation method in an embodiment of the present invention;

[0076] Figure 2 is the flowchart of step S1 in the hyperspectral image classification uncertainty estimation method in an embodiment of the present invention;

[0077] Figure 3 is the flowchart of step S2 in the hyperspectral image classification uncertainty estimation method in an embodiment of the present invention;

[0078] Figure 4 is the flowchart of step S3 in the hyperspectral image classification uncertainty estimation method in an embodiment of the present invention;

[0079] Figure 5 is the flowchart of step S4 in the hyperspectral image classification uncertainty estimation method in an embodiment of the present invention Detailed Embodiments

[0080] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings of the specification. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0081] The hyperspectral image classification uncertainty estimation method, as Figure 1 shown, includes the following steps:

[0082] Step S1, construct a hyperspectral image classification architecture based on a deep neural network;

[0083] Step S2: Embed the Gaussian mixture model into the output layer to obtain an output layer based on the Gaussian mixture model for calculating the output probability;

[0084] Step S3: Use the variational Bayesian inference method to solve the Gaussian mixture model and estimate the initial variational posterior distribution;

[0085] Step S4: Utilize the spatial correlation of the hyperspectral image to improve the variational Bayesian inference method in Step S3, solve the spatially correlated Gaussian mixture model, and obtain the final variational posterior distribution, which is the uncertainty expression of the hyperspectral image classification output.

[0086] Further, as Figure 2 shown, the said Step S1 includes:

[0087] Step S1-1: Determine the model architecture, which can be divided into a cascaded structure of two parts, namely the convolutional layer and the output layer, regarded as the feature extractor and classifier of the hyperspectral image respectively;

[0088] Step S1-2: Feature extractor design. According to the characteristics of the hyperspectral image, select a deep neural network to extract features, and extract the spatial features of the image by building multiple convolutional layers;

[0089] Step S1-3: Classifier design. Design a fully connected layer to map the extracted features to the class space, and design an output layer to generate the probability of each class.

[0090] Further, as Figure 2 shown, the said Step S1 further includes:

[0091] Step S1-4: Use the sequence of the hyperspectral image as the data input of the model;

[0092] Step S1-5: Perform data processing on the input data to calculate the prediction result and output it.

[0093] Further, as Figure 3 shown, the said Step S2 includes:

[0094] Step S2-1: Define the parameter structure of the Gaussian mixture model, including the mean vector, covariance matrix, and mixing weight of each Gaussian distribution;

[0095] Step S2-2: Design the structure of the output layer to output the parameters of the Gaussian mixture model;

[0096] Step S2-3: Calculate the probability output of the mixture Gaussian model;

[0097] Step S2-4: Determine the ground object category according to the probability in the said step S2-3. The method for determining the ground object category by probability also includes converting the probability into an expression of uncertainty.

[0098] Furthermore, the method for calculating probability in the said step S2 includes:

[0099] Determine the pixel, and define as the pixel located at the m-th row and n-th column in the hyperspectral image, where L represents the spectral dimension, 1 ≤ m ≤ M, 1 ≤ n ≤ N;

[0100] After being processed by the feature extractor, each pixel is converted into a feature vector, is the feature after the feature extractor, where J represents the feature dimension;

[0101] When the hyperspectral image contains C categories, a Gaussian mixture model is established for each category, and then there is a corresponding probability density function as follows:

[0102]

[0103] Among them, is the Gaussian distribution, K is the number of components of each Gaussian mixture model, μ = {μ ik}, ∑ = {∑ ik} and η = {η ik} respectively represent the parameter sets of the mean, covariance and mixing weight, ω = [ω 1 , ω 2 , …, ω C T represents the non-negative mixing weights of C Gaussian mixture models, and The value of ω i can be directly determined by the proportion of each type of sample contained in the training set.

[0104] Furthermore, as Figure 4 shown, the said step S3 includes:

[0105] Step S3-1: Define the model structure elements, where the parameter set is θ = {μ, ∑}; the latent variable tensor is The pixel feature z mn at the m-th row and n-th column is associated with the k-th mixing component, then Otherwise D = {Z, E} is the data set, where represents the feature tensor of the hyperspectral image, represents the label map of the hyperspectral image;

[0106] ​Step S3-2: Determine the objective of variational Bayesian inference. The variational Bayesian inference estimation aims to solve the mixture model (1) to maximize the posterior distribution p(Z|θ) of the data D given the parameter θ:

[0107]

[0108] Furthermore, as Figure 4 shown, the step S3 further includes:

[0109] Step S3-3: Introduction of variational distribution. To solve (2), a variational distribution q is used to approximate the original distribution p. In variational Bayesian inference, the variational distribution q comes from the variational family and approximates the posterior distribution by minimizing the KL divergence;

[0110] Step S3-4: Log-likelihood decomposition. According to the variational Bayesian inference theory, the log-likelihood of p(D|θ) can be decomposed as:

[0111]

[0112] where ELBO represents the expected log-likelihood, and KL(q(V)||p(V|D,θ)) is the KL divergence from q(V) to the prior distribution p(V|D,θ), and they can be expressed in the following forms:

[0113]

[0114] Step S3-5: Parameter optimization. According to the variational Bayesian inference theory, the two parameters V and θ can be optimized, and the representation form of the initial optimal variational posterior distribution q * (V) can be determined.

[0115] Furthermore, as Figure 5 shown, the step S4 includes:

[0116] Step S4-1: Introduce spatial correlation, and thus define a spatially correlated variational distribution. For a pixel in a hyperspectral image, it has a high correlation with its surrounding adjacent pixels. Therefore, consider the spatially correlated optimal variational distribution q * (v mn ) can be expressed as:

[0117] q * (v mn )=q(v mn |v (m-1)n ,v (m+1)n ,v m(n-1) ,v m(n+1) ) (6)

[0118] where v mnrepresents the latent variable vector at the m-th row and n-th column in the latent variable tensor V; further, the above equation is equivalent to:

[0119]

[0120] According to the conditional probability formula, Equation (7) can be further transformed into:

[0121]

[0122] where k', k'', k''' and k'''' respectively represent the mixed components with the maximum responsiveness of the data at the (m - 1)-th row and n-th column, the (m + 1)-th row and n-th column, the m-th row and (n - 1)-th column, and the m-th row and (n + 1)-th column;

[0123] Step S4-2: Define the joint probability. Considering the information of the mixed components of the data at the m-th row and n-th column and the data at the (m - 1)-th row and n-th column, the (m + 1)-th row and n-th column, the m-th row and (n - 1)-th column, and the m-th row and (n + 1)-th column, the following definitions can be made:

[0124]

[0125] where,

[0126]

[0127]

[0128] represents the proportions of the k-th, k'-th, k''-th, k'''-th, and k''''-th mixed components;

[0129] Step S4-3: Define and calculate the marginal probability:

[0130]

[0131] where,

[0132]

[0133] represents the proportions of the k'-th, k''-th, k'''-th, and k''''-th mixed components.

[0134] Further, as Figure 5 shown, the said Step S4 further includes:

[0135] Step S4-4: Calculate the final variational distribution q * (v mn ). Substituting Equation (9) and Equation (10) into Equation (8), the expression form of can be obtained, that is, the final variational distribution q * (v mn ) is obtained;

[0136] Step S4-5: Output the uncertainty estimate. This variational distribution is the expression of the uncertainty of the finally obtained hyperspectral image classification output.

[0137] Furthermore, the embodiment of the present invention also provides a hyperspectral image classification uncertainty estimation system. The hyperspectral image classification uncertainty estimation system includes:

[0138] A data preprocessing module for preprocessing the original hyperspectral image;

[0139] A feature extractor that automatically extracts the spatial and spectral features of the image using a deep neural network method. It inputs the preprocessed hyperspectral image and outputs an image feature vector;

[0140] A Gaussian mixture model embedding module that embeds the Gaussian mixture model into the output layer to calculate the probability distribution of each class. It inputs the feature vector output by the feature extractor and outputs the class probability prediction based on the Gaussian mixture model;

[0141] A variational Bayesian inference module that uses the variational Bayesian inference method to solve the Gaussian mixture model parameters and estimate the initial variational posterior distribution. It inputs the probability prediction output of the Gaussian mixture model and the training data, and outputs the initial variational posterior distribution;

[0142] A spatial correlation fusion module that considers the spatial correlation of the hyperspectral image, optimizes the variational Bayesian inference technology, and obtains the final variational posterior distribution. It inputs the initial variational posterior distribution and the spatial correlation information of the image, and outputs the final variational posterior distribution, expressing the uncertainty of classification;

[0143] A classification decision module that determines the class label of each pixel according to the output of the variational posterior distribution. It inputs the final variational posterior distribution and outputs the classification result and the uncertainty estimate.

[0144] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. Hyperspectral image classification uncertainty estimation method, characterized by: The steps include: Step S1, constructing a hyperspectral image classification architecture based on deep neural network; Step S2: embedding the Gaussian mixture model into the output layer to obtain an output layer based on the Gaussian mixture model, which is used to calculate the output probability; Step S3, using the variational Bayesian inference method to solve the Gaussian mixture model and estimate the initial variational posterior distribution; Step S4: Using the spatial correlation of the hyperspectral image, the variational Bayesian inference method in step S3 is improved to solve the spatial correlation Gaussian mixture model and obtain the final variational posterior distribution, that is, the uncertainty expression of the hyperspectral image classification output.

2. The hyperspectral image classification uncertainty estimation method according to claim 1, characterized in that: The step S1 comprises: Step S1-1, determine the model architecture, which can be divided into a cascade structure of two parts, namely, a convolutional layer and an output layer, which are regarded as feature extractors and classifiers of hyperspectral images respectively; Step S1-2, feature extractor design, according to the characteristics of hyperspectral images, select deep neural network to extract features, and extract the spatial features of the image by building multiple convolutional layers; Step S1-3, classifier design, design the fully connected layer to map the extracted features to the category space, and design the output layer to generate the probability of each category.

3. The hyperspectral image classification uncertainty estimation method according to claim 1, characterized in that: The step S1 further includes: Step S1-4, using the sequence of hyperspectral images as data input for the model; Step S1-5: Process the input data, calculate the prediction result and output it.

4. The hyperspectral image classification uncertainty estimation method according to claim 1, characterized in that: The step S2 comprises: Step S2-1, defining the parameter structure of the Gaussian mixture model, including the mean vector, covariance matrix and mixing weight of each Gaussian distribution; Step S2-2, designing the structure of the output layer and outputting the parameters of the Gaussian mixture model; Step S2-3, calculating the probability output of the mixed Gaussian model; Step S2-4: determining the type of the ground object according to the probability in step S2-3. The method for determining the type of the ground object by probability also includes converting the probability into an expression of uncertainty.

5. The hyperspectral image classification uncertainty estimation method according to claim 1, characterized in that: The probability calculation method in step S2 includes: Determine the pixel, define is the pixel located in the mth row and nth column in the hyperspectral image, where L represents the spectral dimension, 1≤m≤M, 1≤n≤N; After being processed by the feature extractor, each pixel is converted into a feature vector. is the feature after the feature extractor, where J represents the feature dimension; When the hyperspectral image contains C categories, a Gaussian mixture model is established for each category, and the corresponding probability density function is: in, is a Gaussian distribution, K is the number of components in each Gaussian mixture model, μ = {μ ik },∑={∑ ik } and η={η ik } respectively represent the parameter set of mean, covariance and mixing weight, ω=[ω1,ω2,…,ω C ] T represents the non-negative mixing weights of C Gaussian mixture models, and ω i The value of can be directly determined by the proportion of samples of each class contained in the training set.

6. The hyperspectral image classification uncertainty estimation method according to claim 1, characterized in that: The step S3 comprises: Step S3-1, define the model structure elements, where the parameter set is θ = {μ, ∑}; the latent variable tensor is The pixel feature z of the mth row and nth column mn is associated with the kth mixing component, then otherwise D={Z,E} is the data set, where Represents the feature tensor of the hyperspectral image, A label map representing a hyperspectral image; Step S3-2, determine the goal of variational Bayesian inference. Variational Bayesian inference estimation aims to solve the mixture model (1) to maximize the posterior distribution p(Z|θ) of the data D on the given parameter θ:

7. The hyperspectral image classification uncertainty estimation method according to claim 1, characterized in that: The step S3 further includes: Step S3-3, introduction of variational distribution. To solve (2), variational distribution q is used to approximate the original distribution p. In variational Bayesian inference, variational distribution q comes from the variational family Approximate the posterior distribution by minimizing the KL divergence; Step S3-4, log-likelihood decomposition. According to variational Bayesian inference theory, the log-likelihood of p(D|θ) can be decomposed into: Among them, ELBO represents the expected log likelihood, KL(q(V)||p(V|D,θ)) is the KL divergence from q(V) to the prior distribution p(V|D,θ), which can be expressed as follows: Step S3-5, parameter optimization. According to the variational Bayesian inference theory, the two parameters V and θ can be optimized and the initial optimal variational posterior distribution q can be determined. * (V) representation.

8. The hyperspectral image classification uncertainty estimation method according to claim 1, characterized in that: The step S4 comprises: Step S4-1: Introduce spatial correlation to define the variational distribution of spatial correlation. For a hyperspectral image pixel, it has a high correlation with the surrounding adjacent pixels. Therefore, the optimal variational distribution q considering spatial correlation is * (v mn ) can be expressed as: q * (in mn )=q(v mn |v (m-1)n ,v (m+1)n ,v m(n-1) ,v m(n+1) ) (6) Among them, v mn represents the latent variable vector in the mth row and nth column of the latent variable tensor V; further, the above formula is equivalent to: According to the conditional probability formula, formula (7) can be further transformed into: Wherein, k', k", k'' and k''' respectively represent the mixed components of the maximum responsivity of the data of the m-1th row and nth column, the m+1th row and nth column, the mth row and n-1th column and the mth row and n+1th column; Step S4-2, define the joint probability, considering the information of the mixed components of the data in the mth row and nth column and the data in the m-1th row and nth column, the m+1th row and nth column, the mth row and n-1th column, and the mth row and n+1th column, which can be defined as follows: in, represents the proportion of the kth, k', k", k"' and k"" mixed components; Step S4-3, define and calculate the edge probability: in, Indicates the proportion of the k', k", k"' and k"" mixed components.

9. The hyperspectral image classification uncertainty estimation method according to claim 1, characterized in that: The step S4 further includes: Step S4-4, calculate the final variational distribution q * (v mn ), substituting equations (9) and (10) into equation (8), we can obtain The expression form of the final variational distribution q * (v mn ); Step S4-5: Output uncertainty estimation. The variational distribution is the uncertainty expression of the hyperspectral image classification output finally obtained.

10. Hyperspectral image classification uncertainty estimation system, characterized by: The hyperspectral image classification uncertainty estimation system is applied to the hyperspectral image classification uncertainty estimation method according to any one of claims 1 to 9, and the hyperspectral image classification uncertainty estimation system comprises: Data preprocessing module, used to preprocess the original hyperspectral image; Feature extractor, which uses deep neural network method to automatically extract spatial and spectral features of images, inputs preprocessed hyperspectral images, and outputs image feature vectors; Gaussian mixture model embedding module, which embeds the Gaussian mixture model into the output layer to calculate the probability distribution of each category, inputs the feature vector output by the feature extractor, and outputs the category probability prediction based on the Gaussian mixture model; The variational Bayesian inference module uses the variational Bayesian inference method to solve the parameters of the Gaussian mixture model, estimate the initial variational posterior distribution, input the probability prediction output and training data of the Gaussian mixture model, and output the initial variational posterior distribution; The spatial correlation fusion module considers the spatial correlation of hyperspectral images, optimizes the variational Bayesian inference technology, obtains the final variational posterior distribution, inputs the initial variational posterior distribution and the spatial correlation information of the image, and outputs the final variational posterior distribution to express the uncertainty of classification; The classification decision module determines the category label of each pixel based on the output of the variational posterior distribution, inputs the final variational posterior distribution, and outputs the classification result and uncertainty estimate.

Images