Remote sensing image urban landscape classification method and system

Abstract

The invention discloses a remote sensing image urban landscape classification method and system, and relates to the technical field of remote sensing image recognition and classification, and the method comprises the steps: obtaining a remote sensing image data set, and employing a pre-trained deep network model to extract a landscape convolution feature map; constructing a landscape attribute model based on each landscape convolution feature map by using a Gaussian mixture model and an expectation maximization estimation algorithm, and mixing all the landscape attribute models; adjusting parameters of the mixed landscape attribute model by using a maximum posteriori adaptive method to obtain a high-dimensional feature vector containing all landscape redundant attributes; analyzing the high-dimensional feature vector by adopting a factor analysis method to obtain a landscape vector; and obtaining a probability density function of the remote sensing image according to the landscape vector, and classifying the remote sensing image urban landscape according to the probability density function. The invention provides a compact and discriminative urban landscape representation method, and the effectiveness of urban landscape classification is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image recognition and classification, and more specifically, to a method and system for classifying urban landscapes in remote sensing images. Background Art

[0002] Remote sensing images provide detailed spatial and semantic information of ground objects and play an extremely important role in civilian and commercial aspects. For urban planning, in order to better understand the areas of interest on the urban ground, remote sensing image recognition and classification techniques are usually used to automatically label and classify urban landscapes. Urban landscape classification is to determine attribute labels for each urban area according to the semantic content of remote sensing images. Different from landscape classification in natural images, due to the large-scale macroscopic perception of remote sensing images and the complex distribution of ground coverings. An urban landscape often contains different ground entities. For example, an airport landscape includes airplanes, highways, runways, and buildings. In addition, due to changes in scale, pixel intensity, brightness, contrast, etc. and the non-uniformity of spatial, spectral, and radiation resolutions, it poses a great challenge to urban landscape classification.

[0003] Existing methods for classifying urban landscapes in remote sensing images have explored various shallow, middle, and high-level depth feature representations. Feature representation methods based on shallow features aim to understand specific features of a given set of remote sensing images, but are limited by the representation ability of handcrafted features. With the rapid development of deep learning, Convolutional Neural Network (CNN) has been proven to be able to effectively and comprehensively describe shallow and high-level landscape features. In particular, most CNN-based landscape classification methods have fully demonstrated the excellent performance of using CNN models pre-trained on ImageNet (such as AlexNet, VGGNet-16, and GoogLeNet and their fine-tuned versions) in remote sensing image landscape classification tasks.

[0004] Existing CNN architectures have limitations in dealing with intra-class differences, inter-class similarities, and the problem of class-imbalanced samples. To avoid the problems of intra-class diversity and inter-class similarities, few methods adopt metric learning to optimize intra-class and inter-class distances. Specifically, these methods focus on minimizing the confusion between different class landscapes. However, using distance functions to limit the similarity between samples makes it difficult for the network to summarize effective similarity and discriminative knowledge. In addition to intra-class and inter-class differences between landscapes, class-imbalanced samples also affect the performance of landscape classification. Recently, multi-granularity decoupled networks aim to solve the problem of class-imbalanced samples, thereby improving the performance of landscape classification. Existing methods attempt to solve the above two different problems separately, while ignoring the subtle differences between landscapes. In addition, the invariant depth compressible covariance method processes the changes in ground object types in landscape classification in a compact representation form. Therefore, how to learn the subtle differences between urban landscapes in remote sensing images and obtain a compact and discriminative feature representation for urban landscape classification in remote sensing images remains an important problem to be solved urgently. Summary of the Invention

[0005] An object of the present invention is to overcome the difficulty in learning the subtle differences between urban landscapes in remote sensing images and obtaining a compact and discriminative feature representation for urban landscape classification in remote sensing images, and to provide a method and system for urban landscape classification in remote sensing images.

[0006] To achieve the above object, the technical solution of the present invention is as follows:

[0007] In a first aspect, the present invention provides a method for urban landscape classification in remote sensing images, including the following steps:

[0008] Obtain a remote sensing image data set and extract landscape convolution features therefrom using a pre-trained deep network model;

[0009] Construct a landscape attribute model based on each of the landscape convolution features using a mixture of Gaussian models and the expectation-maximization estimation algorithm, and mix all the landscape attribute models;

[0010] Use the maximum a posteriori adaptation method to adjust the important attribute contributions in the mixed landscape attribute model to obtain a high-dimensional feature vector containing all landscape redundant attributes;

[0011] Perform factor analysis on the high-dimensional feature vector to obtain a landscape vector;

[0012] Use a support vector machine to obtain urban landscape category information from the landscape vector for classification.

[0013] Preferably, the pre-trained deep network model is a pre-trained CNN network model, including AlexNet and EfficientNet network models.

[0014] Preferably, the landscape convolution features include shallow, middle, and high-level landscape convolution features.

[0015] Preferably, the specific steps for extracting the landscape convolution features include:

[0016] Use AlexNet to perform feature mapping on the input remote sensing image to obtain a feature vector. Then, use EfficientNet to obtain deep landscape convolution features based on the feature vector.

[0017] Preferably, the defined landscape attribute model is as follows:

[0018]

[0019] where ω k , u k , σ k are respectively the weight, mean, and covariance of the k-th Gaussian model in the mixture Gaussian model, and its constraint condition is The feature vector s j represents a part of the remote sensing image s, and the entire remote sensing image is represented as s = {s1, s2, s3,..., s J}.

[0020] Preferably, the use of the maximum a posteriori adaptive method to adjust the important attribute contributions in the mixed landscape attribute model to obtain a high-dimensional feature vector containing all landscape redundant attributes specifically includes:

[0021] Given a set of landscape convolution feature vectors J describing the landscape image s, use the maximum a posteriori adaptive method to perform parameter adaptive adjustment on the mixed components of the landscape attribute model. All these feature vectors are arranged in each K-stage mixed component in the form of posterior probability, which is expressed as follows:

[0022]

[0023] where p(s j |k) represents the likelihood probability of the feature k extracted from the mixture k and ω , and ω k represents the landscape attribute model of the maximum a posteriori of the image-specific model, which is obtained by the convex combination of image-specific statistics;

[0024] The solution process for the adaptive weights and means of the landscape attribute model mixture k is as follows:

[0025]

[0026] In the formula, N k and Fk They are adaptive adjustment functions respectively, and the adapted means are superimposed as A high-dimensional landscape vector with a dimension of Kf×1 is obtained for each remote sensing image.

[0027] Preferably, factor analysis is used to analyze the high-dimensional feature vector to obtain the landscape vector. Specifically, it includes:

[0028] The low-dimensional representation of the high-dimensional landscape vector, that is, the landscape vector, is obtained by factor analysis decomposition:

[0029] g = m + Tw

[0030] Where m represents the mean of the landscape attribute model, T represents a low-rank total variation matrix of size Kf×q, w is a q-dimensional vector with a prior of standard Gaussian distribution N(0,1), and this q-dimensional vector is the landscape vector. After observing the remote sensing image s, the posterior distribution of w can be determined by the Baum-Welch statistic as P(w|s) ∝ P(s|w)N(0,1). Discarding the terms unrelated to w, we can get:

[0031]

[0032] In the formula, Σ is the covariance matrix, and the matrix A(s) is defined as A(s) = B -1 (s)T t Σ -1 F(s), F(s) is a Kf×1 vector obtained by connecting the centralized first-order Baum-Welch statistic F(s) = [F1(s), F2(s),..., F K (s)] t , and these first-order Baum-Welch statistics are obtained by Centralizing the mean of the landscape attribute model, N(s) is a Kf×Kf diagonal matrix with diagonal blocks N k (s), and t represents the transpose operation.

[0033] Preferably, the expectation-maximization (EM) algorithm is used to iteratively calculate the mean and covariance of the posterior distribution in the expectation step to obtain the updated T and Σ parameters in the maximization step. The expressions for the mean vector and covariance matrix are:

[0034] E[w(s)] = B -1 (s)T t Σ -1 F(s)

[0035] Cov(w(s), w(s)) = B -1 (s)

[0036] First, initialize \(m\) and \(\Sigma\) with the mean and covariance of the landscape attribute model. Use a randomly initialized matrix with an expected rank of \(q\) as the total variability matrix \(T\). Then, calculate \(E[w(s)]\) and \(Cov(w(s), w(s))\) using the expressions of the mean vector and covariance matrix respectively. The update expression for the total variability matrix \(T\) is as follows:

[0037]

[0038] where both \(F(s)\) and \(N(s)\) represent all available features in the image;

[0039] The estimation of the residual matrix \(\Sigma\) is defined as follows:

[0040]

[0041] where \(B\) k represents the \(K\)-th diagonal block in the \(K_d\times K_d\) matrix, and \(D\) k (s) represents the second-order Baum-Welch statistic of the image. The solution process is shown as follows:

[0042]

[0043] After \(M\) steps of iteration, use the estimated \(T\) and \(\sum\) matrices to calculate the posterior mean, specifically as follows

[0044] \(w(s)=(I + T\) t \(\Sigma\) -1 N(s)T)\) -1 T\) t \(\Sigma\) -1 F(s)\)

[0045] In the formula, \(w(s)\) represents the low-dimensional representation of the image, i.e., the landscape vector. The \(t\) matrix includes eigenvectors related to the dominant \(q -\)eigenvalues of the covariance matrix in the total variation space.

[0046] Preferably, the time complexity of completing the entire process of classifying and predicting urban landscapes in remote sensing images is \(O(Kfq+Kq\) 2 +q\) 3 ), where \(K\) represents the number of landscape attribute models in the mixed landscape attribute model, and \(f\) and \(q\) represent the dimensions of the feature vector and the landscape vector respectively.

[0047] In a second aspect, the present invention provides a remote sensing image urban landscape classification system for applying a remote sensing image urban landscape classification method described in the above technical solution. The system includes:

[0048] A feature extraction module for obtaining a remote sensing image data set and extracting landscape convolution features therefrom using a pre-trained deep network model;

[0049] A model construction and mixing module, which is used to construct a landscape attribute model based on each of the landscape convolution features by using a Gaussian mixture model and an expectation maximization estimation algorithm, and mix all the landscape attribute models;

[0050] A model adjustment module, which is used to adjust the important attribute contributions in the mixed landscape attribute model by using the maximum a posteriori adaptation method to obtain a high-dimensional feature vector containing all landscape redundant attributes;

[0051] A landscape vector generation module, which is used to analyze the high-dimensional feature vector by using factor analysis to obtain a landscape vector;

[0052] A remote sensing image urban landscape classification module, which is used to use a support vector machine to obtain urban landscape category information from the landscape vector for classification.

[0053] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0054] In the present invention, by taking the Gaussian distribution of deep learning features as a landscape attribute model, and using the expectation maximization algorithm to train the landscape attribute model to obtain implicit landscape attributes beneficial to remote sensing image recognition, and finally using the factor analysis method to learn the final compact representation, that is, the landscape vector, the present invention proposes a compact and discriminative urban landscape representation method for classification. By considering the Gaussian distribution of all landscapes of different categories, the problem of class imbalance samples is alleviated, and at the same time, the subtle differences between similar landscapes are learned to improve the accuracy of remote sensing image urban landscape classification. Description of the Drawings

[0055] Figure 1 It is a schematic diagram of the steps of a remote sensing image urban landscape classification method according to Embodiment 1 of the present application;

[0056] Figure 2 It is a schematic diagram of the framework of a remote sensing image urban landscape classification method according to Embodiment 1 of the present application;

[0057] Figure 3 It is a schematic diagram of the types of remote sensing image urban landscapes according to Embodiment 1 of the present application;

[0058] Figure 4 It is a t-SNE visualization schematic diagram of the building and dense residential landscape types respectively using (a) EfficientNet-B2 features and (b) the method of the present invention according to Embodiment 2 of the present application. Detailed Embodiments

[0059] In order to more clearly understand the above objects, features and advantages of the present invention, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the embodiments of the present application and the features in the embodiments may be combined with each other.

[0060] In the following description, many specific details are set forth in order to fully understand the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited by the specific embodiments disclosed below.

[0061] Embodiment 1

[0062] Please refer to Figure 1 and Figure 2 , Embodiment 1 of the present application provides a method for classifying urban landscapes in remote sensing images, including the following steps S1-S5:

[0063] S1: Obtain a remote sensing image data set and use a pre-trained deep network model to extract landscape convolution feature maps therefrom.

[0064] Convolutional features depend on the neural network structure and can directly obtain representations from images. Therefore, the convolutional features of the deep CNN architecture describe more semantic information, leaving the complexity of feature design to the network architecture. In addition, the pre-trained CNN model is effective in comprehensively describing shallow, middle, and high-level landscape features. Embodiment 1 of the present invention utilizes the convolutional feature maps generated by CNN architectures of different dimensions. First, for remote sensing images with a size of 227×227×3, AlexNet is used to perform feature mapping on the input images to obtain a feature vector with a dimension of 13×13×256. Then, EfficientNet is used to obtain deep landscape convolutional features, and these features describe regions of different scales in the landscape through its compound scaling mechanism. Thus, shallow, middle, and high-level landscape convolutional features are extracted from the original remote sensing images using AlexNet and EfficientNet.

[0065] S2: Use a mixture of Gaussian models and the expectation maximization estimation algorithm to construct a landscape attribute model based on each of the landscape convolution feature maps, and mix all the landscape attribute models.

[0066] The present invention constructs a Landscape Attribute Model (LAM) to learn implicit landscape description attributes. Let each representative landscape image be the value of a sample function in a random process for realizing landscape generation. The proximity of the sample functions of two landscapes is determined by the parameters of their probability density functions, thereby quantifying the similarity. To estimate the underlying probability density function using a Gaussian Mixture Model (GMM), a large number of mixture models are required to understand the landscape variability of remote sensing images. Therefore, a Gaussian Mixture Model (GMM) is trained by considering the images of all landscapes to construct a mixed Landscape Attribute Model (LAM), which contains a sufficient number of landscape attribute models to model different landscape attributes. Since these attributes are shared among different landscapes, a reasonable number of mixtures is sufficient to obtain an effective representation.

[0067] The Landscape Attribute Model (LAM) is defined as follows:

[0068]

[0069] where ω k 、u k 、σ k are respectively the weight, mean, and covariance of the k-th Gaussian model in the mixed landscape attribute model, and its constraint condition is The feature vector s j represents a part of the image s, and the entire image is represented as s = {s1, s2, s3,..., s J}. In this work, the feature maps generated from various CNN architectures are used to classify the urban landscapes of remote sensing images. Using the Expectation-Maximization (EM) estimation algorithm, a Landscape Attribute Model (LAM) is constructed for the convolutional features of each CNN architecture respectively. In the case of successful training, each component in the Landscape Attribute Model (LAM) captures the attributes unique to one landscape or a part of several landscapes. To obtain the landscape description features and generate the probability density function (pdf) of the image, the present invention adopts the Maximum A Posteriori (MAP) adaptive method to adjust the landscape attribute model parameters.

[0070] S3: Use the Maximum A Posteriori adaptive method to adjust the important attribute contributions in the mixed landscape attribute model, and obtain a high-dimensional feature vector containing all landscape redundant attributes.

[0071] In Embodiment 1 of the present application, a set of feature vectors J describing the landscape image S is given, and the Maximum A Posteriori (MAP) adaptive method is used to perform parameter adaptation on the mixed components of the mixed landscape attribute model. All these feature vectors are arranged in the posterior probability manner into each K-stage mixed component, as shown below:

[0072]

[0073] In the formula, p(s j |k) represents the mixture k and ω k The features extracted from Likelihood probability, ω k The landscape attribute model of the MAP representing the image-specific model is obtained by a convex combination of image-specific statistics. The adaptive weights and means of the landscape attribute model mixture k are solved as follows:

[0074]

[0075] Where N k and F k They are adaptive adjustment functions respectively.

[0076] The lack of data in the image limits the adaptation of the covariance matrix of the landscape attribute model. However, by superimposing the adapted means as A high-dimensional landscape vector (HDSV) with (Kf×1) dimensions is obtained for each image. Although the HDSV representation is a high-dimensional representation composed of many attributes, most of the attributes do not contribute to the image. At the same time, the mean of the mixture (attributes) of those LAMs that contribute to the image is modified. Therefore, the present invention adopts a suitable decomposition mechanism to convert the high-dimensional landscape vector (HDSV) into a low-dimensional landscape vector.

[0077] S4: Analyze the high-dimensional feature vector using factor analysis to obtain a landscape vector.

[0078] The low-dimensional representation of the high-dimensional feature vector HDSV is obtained by factor analysis:

[0079] g=m+Tw

[0080] Where m represents the mean of the landscape attribute model (LAM), T represents the low-rank total variation matrix of size Kf×q, and w is a q-dimensional vector with a standard Gaussian distribution N(0,1) prior, which is called the landscape vector. After observing image s, the posterior distribution of w can be determined using the Baum-Welch statistic as P(w|s)∝P(s|w)N(0,1). Discarding items that are not related to w, we can obtain:

[0081]

[0082] Where Σ is the covariance matrix, and the matrix A(s) is defined as A(s) = B -1 (s)T t Σ -1 F(s), where F(s) is the Kf×1 vector by connecting the centralized first-order Baum-Welch statistics F(s) = [F1(s), F2(s), ..., F K (s)] t。These first-order Baum-Welch statistics are centered by taking the mean of the Landscape Attribute Model (LAM). N(s) is a Kf×Kf diagonal matrix with diagonal blocks N k (s), and t represents the transpose operation.

[0083] S5: Use a support vector machine to obtain urban landscape category information from the landscape vectors for classification.

[0084] The support vector machine (SVM) is a supervised learning algorithm mainly used for classification and regression tasks. Its core idea is to find an optimal hyperplane to separate data of different classes and maximize the margin between data of different classes to achieve classification.

[0085] Please refer to Figure 3 , Figure 3 which is a schematic diagram of the urban landscape type of the remote sensing image for Embodiment 1 of this application.

[0086] Embodiment 2

[0087] Based on Embodiment 1, Embodiment 2 of this application further elaborates on the solution of the landscape vectors in step S4 of Embodiment 1, as follows:

[0088] The expressions for the mean vector and covariance matrix are:

[0089] E[w(s)] = B -1 (s)T t Σ -1 F(s)

[0090] Cov(w(s), w(s)) = B -1 (s)

[0091] The present invention uses the expectation-maximization estimation algorithm to obtain the updated T and Σ parameters in the maximization step by iteratively calculating the mean and covariance of the posterior distribution in the expectation step.

[0092] First, initialize m and Σ with the mean and covariance of the landscape attribute model. Use a randomly initialized matrix with an expected rank of q as the total variability matrix T, and then calculate E[w(s)] and Cov(w(s), w(s)) respectively using the expressions for the mean vector and covariance matrix. The update expression for the total variability matrix T is as follows:

[0093]

[0094] where F(s) and N(s) both represent all available features in the image.

[0095] The estimation of the residual matrix Σ can be defined as follows:

[0096]

[0097] wherein, B k represents the Kth diagonal block in the Kd×Kd matrix, and D k (s) represents the second-order Baum-Welch statistics of the image, and the solution process is as follows:

[0098]

[0099] After M steps of iteration, the present invention uses the estimated T and Σ matrices to calculate the posterior mean, specifically as follows

[0100] w(s) = (I + T t Σ -1 N(s)T) -1 T t Σ -1 F(s)

[0101] wherein, w(s) represents the landscape vector, i.e., the low-dimensional representation of the image. The t matrix includes the eigenvectors related to the dominant q-eigenvalues of the covariance matrix in the total variation space. Assuming that the dominant eigenvalues come from the mixture of the landscape attribute model (LAM), the model of the present invention helps to model the landscape attributes in the remote sensing image. By using T to project the original HDSV onto the q-dimensional landscape vector, a learned discriminative representation is obtained.

[0102] The time complexity of the classification method of the present invention is O(Kfq + Kq 2 + q 3 ), where K represents the number of mixtures in the landscape attribute model (LAM), and f and q respectively represent the dimensions of the feature vector and the landscape vector.

[0103] Please refer to Figure 4 , Figure 4 which is the t-SNE visualization schematic diagram of the building and dense residential landscape using (a) EfficientNet-B2 features and (b) the method of the present invention in Embodiment 2 of the present application.

[0104] Other technical details and steps of Embodiment 2 are the same as those of Embodiment 1, and will not be elaborated here.

[0105] Embodiment 3

[0106] Based on Embodiments 1 and 2, Embodiment 3 of the present application provides a remote sensing image urban landscape classification system for applying the remote sensing image urban landscape classification method described in Embodiments 1 and 2. The system includes:

[0107] A feature extraction module, configured to obtain a remote sensing image dataset and extract landscape convolution features therefrom using a pre-trained deep network model;

[0108] A model construction and mixing module, configured to construct a landscape attribute model based on each of the landscape convolution features using a Gaussian mixture model and an expectation maximization estimation algorithm, and mix all the landscape attribute models;

[0109] A model adjustment module, configured to adjust the important attribute contributions in the mixed landscape attribute model using the maximum a posteriori adaptation method to obtain a high-dimensional feature vector containing all landscape redundant attributes;

[0110] A landscape vector generation module, configured to analyze the high-dimensional feature vector using factor analysis to obtain a landscape vector;

[0111] A remote sensing image urban landscape classification module, configured to obtain urban landscape category information from the landscape vector using a support vector machine for classification.

[0112] Other technical details and steps of this Embodiment 3 are the same as those of Embodiment 1 or Embodiment 2, and will not be elaborated herein.

[0113] In summary of the above embodiments, the present invention takes the Gaussian distribution of deep learning features as the landscape attribute model, trains the landscape attribute model using the expectation maximization algorithm to obtain implicit landscape attributes beneficial to remote sensing image recognition, and finally uses the factor analysis method to learn the final compact representation, i.e., the landscape vector. The landscape attribute model (LAM) is proposed to represent the implicit landscape attributes of remote sensing images. Each component of the LAM captures the attributes of the landscape, which may be specific to the landscape or shared among different classes of landscapes. By utilizing this inter-class similarity and applying MAP to a specific landscape simultaneously, the class imbalance problem is overcome. To effectively classify the urban landscape of remote sensing images, a compact and discriminative landscape vector representation method is proposed. By considering the Gaussian distributions of all landscapes of different classes, the class imbalance sample problem is alleviated, and at the same time, the subtle differences between similar landscapes are learned to improve the accuracy of urban landscape classification of remote sensing images.

Claims

1. A method for classifying urban landscapes in remote sensing images, characterized in that, It includes the following steps: Obtain a remote sensing image dataset and use a pre-trained deep network model to extract landscape convolutional feature maps therefrom; Use a Gaussian mixture model and an expectation maximization estimation algorithm to construct a landscape attribute model based on each of the landscape convolutional feature maps, and mix all the landscape attribute models; Use the maximum a posteriori adaptive method to adjust the parameters of the mixed landscape attribute model to obtain a high-dimensional feature vector containing all landscape redundant attributes; Adopt factor analysis to analyze the high-dimensional feature vector to obtain a landscape vector; Use a support vector machine to obtain urban landscape category information from the landscape vector for classification.

2. The method for classifying urban landscapes in remote sensing images according to claim 1, wherein The pre-trained deep network model is a pre-trained CNN network model, including AlexNet and EfficientNet network models.

3. A method for classifying urban landscapes in remote sensing images according to claim 2, characterized in that The landscape convolutional features include shallow, middle, and high-level landscape convolutional features.

4. A method for classifying urban landscapes in remote sensing images according to claim 3, characterized in that, The specific steps for extracting landscape convolutional features include: Use AlexNet to perform feature mapping on the input remote sensing image to obtain a feature vector, and then use EfficientNet to obtain deep landscape convolutional features according to the feature vector.

5. A method for classifying urban landscapes in remote sensing images according to claim 1, characterized in that, The defined landscape attribute model is as follows: Among them, ω k , u k , σ k are respectively the weight, mean and covariance of the k-th Gaussian model in the mixed landscape attribute model, and its constraint condition is The feature vector s j represents a part of the remote sensing image s, and the entire remote sensing image is represented as s = {s1, s2, s3,..., s J}.

6. A method for classifying urban landscapes in remote sensing images according to claim 1, characterized in that, The use of the maximum a posteriori adaptive method to adjust the important attribute contributions in the mixed landscape attribute model to obtain a high-dimensional feature vector containing all landscape redundant attributes specifically includes: Give a set of landscape convolutional feature vectors J describing the landscape image s, and use the maximum a posteriori adaptive method to perform parameter adaptive adjustment on the mixed components of the landscape attribute model. All these feature vectors are arranged in each K-stage mixed component in a posterior probability manner, expressed as follows: Among them, p(s j |k) represents the mixture k and ω k Features extracted from The likelihood probability, ω k A maximum a posteriori landscape attribute model representing an image-specific model is obtained by a convex combination of image-specific statistics; The solution process for the adaptive weights and means of the k-th mixture of the landscape attribute model is as follows: where N k and F k are adaptive adjustment functions respectively, and the adapted means are superimposed as to obtain a high-dimensional landscape vector of Kf×1 dimension for each remote sensing image.

7. A method for classifying urban landscapes in remote sensing images according to claim 1, characterized in that, The adoption of factor analysis to analyze the high-dimensional feature vector to obtain a landscape vector specifically includes: Obtain the low-dimensional representation of the high-dimensional landscape vector through factor analysis decomposition, that is, the landscape vector: g = m + Tw where m represents the mean of the landscape attribute model, T represents a low-rank total variability matrix of size Kf×q, w is a q-dimensional vector with a prior of standard Gaussian distribution N(0,1), this q-dimensional vector is the landscape vector. After observing the remote sensing image s, the posterior distribution of w can be determined by the Baum-Welch statistic as P(w|s) ∝ P(s|w)N(0,1). Discarding the terms irrelevant to w, we can obtain: where Σ is the covariance matrix, and the matrix A(s) is defined as A(s) = B -1 (s)T t Σ -1 F(s), where F(s) is a Kf×1 vector formed by concatenating the centered first-order Baum-Welch statistics F(s) = [F1(s), F2(s),..., F K (s)] t , and these first-order Baum-Welch statistics are obtained by centering the mean of the landscape attribute model. N(s) is a Kf×Kf diagonal matrix with diagonal blocks N k (s), and t represents the transpose operation.

8. A method for classifying urban landscapes in remote sensing images according to claim 1, characterized in that, Use the expectation maximization algorithm to iteratively calculate the mean and covariance of the posterior distribution in the expectation step to obtain the updated T and Σ parameters in the maximization step. The expressions for the mean vector and covariance matrix are: E[w(s)] = B -1 (s)T t Σ -1 F(s) Cov(w(s), w(s)) = B -1 (s) First, initialize m and Σ with the mean and covariance of the landscape attribute model, use a randomly initialized matrix with an expected rank of q as the total variability matrix T, and then calculate E[w(s)] and Cov(w(s), w(s)) respectively using the expressions for the mean vector and covariance matrix. Then the update expression for the total variability matrix T is as follows: where F(s) and N(s) both represent all the available features in the image; The estimation of the residual matrix Σ is defined as follows: Among them, B k represents the K-th diagonal block in the Kd×Kd matrix, D k (s) represents the second-order Baum-Welch statistics of the image, and the solution process is shown as follows: After M steps of iteration, the posterior mean is calculated using the estimated T and Σ matrices as follows w(s) = (I + T t Σ -1 N(s)T) -1 T t Σ -1 F(s) where w(s) represents the low-dimensional representation of the image, i.e., the landscape vector, and the t matrix includes the eigenvectors associated with the dominant q-eigenvalues of the covariance matrix in the total variation space.

9. A method for classifying urban landscapes in remote sensing images according to any one of claims 1-8, characterized in that, The time complexity for completing the entire process of classifying and predicting urban landscapes in remote sensing images is O(Kfq + Kq 2 + q 3 ), where K represents the number of landscape attribute models in the mixed landscape attribute model, and f and q represent the dimensions of the feature vector and the landscape vector respectively.

10. A remote sensing image urban landscape classification system for applying the remote sensing image urban landscape classification method according to any one of claims 1-9, characterized in that, The system includes: A feature extraction module for obtaining a remote sensing image dataset and extracting landscape convolution features therefrom using a pre-trained deep network model; A model construction and mixing module for constructing a landscape attribute model based on each of the landscape convolution features using a Gaussian mixture model and an expectation maximization estimation algorithm, and mixing all of the landscape attribute models; A model adjustment module for adjusting the important attribute contributions in the mixed landscape attribute model using a maximum a posteriori adaptive method to obtain a high-dimensional feature vector containing all landscape redundant attributes; A landscape vector generation module for analyzing the high-dimensional feature vector using factor analysis to obtain a landscape vector; A remote sensing image urban landscape classification module for obtaining urban landscape category information from the landscape vector using a support vector machine for classification.

Images