Remote sensing image classification method and system based on superpixel unsupervised linear discriminant analysis

Abstract

The invention belongs to the technical field of remote sensing data processing and charting, and discloses a remote sensing image classification method and system based on superpixel unsupervised linear discriminant analysis. Unsupervised global feature dimension reduction is realized through combination of superpixels and linear discriminant analysis, unsupervised local feature dimension reduction is realized through combination of superpixels and local Fisher discriminant analysis, finally extracted global features and local features are subjected to channel connection to obtain global-local features, and classification of remote sensing images is realized through combination of a support vector machine. The problems of difficult parameter adjustment, insufficient training samples and insufficient generalization ability of a denoising method in existing remote sensing image classification are solved. The problem that an existing dimension reduction method lacks data representation capability and damages the discrimination capability of features after dimension reduction is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing data processing and mapping, and in particular relates to a remote sensing image classification method and system based on superpixel unsupervised linear discriminant analysis. Background Art

[0002] Remote sensing images have the characteristics of multispectral, high resolution, temporal continuity and wide coverage. They can provide information content in multiple bands from visible light to microwaves, and regularly observe large areas to provide detailed ground details and time series change information. As a key link in remote sensing data processing and application, remote sensing image classification technology is widely used in land use and land cover change detection, crop monitoring, forest resource survey, urban planning and environmental protection. Common schemes for remote sensing image classification include machine learning methods, but these methods have high computational complexity when processing large-scale data. Deep learning methods have good automatic feature extraction effects, but require a large amount of labeled data and computing resources. Fusion methods combine multiple data and features to improve classification accuracy, but the processing process is complex. Graph convolutional networks are suitable for processing data with graph structures, but the composition process is complex. Transfer learning reduces the demand for training data, but the effect depends on the similarity between the pre-trained model and the target task. Data dimensionality reduction can effectively solve the problem of high computational complexity in machine learning classification, and denoising methods can be introduced to suppress noise interference. Therefore, when classifying remote sensing images, the two key problems that need to be solved are denoising and dimensionality reduction.

[0003] Common denoising methods are mainly divided into model-driven, data-driven and model-data-driven methods. Model-driven methods mainly use the intrinsic characteristics of remote sensing images to remove mixed noise. This type of method has good interpretability. Typical methods include non-local mean, total variation, low-rank matrix approximation, low-rank tensor decomposition and other methods. However, different remote sensing images have differences in data distribution, scenes and other aspects, which makes this type of method very sensitive to parameters and requires a lot of time to adjust parameters. Data-driven methods complete model parameter updates by designing loss functions for specific tasks. Therefore, this type of method can effectively solve the difficulties brought by parameter adjustment. Representative methods include spectral-spatial denoising residual network, autoencoder denoising network, 3D hole denoising convolutional neural network, deep spatial spectral global inference network and other methods. However, this type of method requires a large amount of clean remote sensing data as label samples. At the same time, due to the complexity of imaging mechanism and degradation process, the simulated noise distribution of clean remote sensing data is difficult to be consistent with the actual noise distribution. The existence of these problems greatly affects the versatility of data-driven methods. Model-data driven methods solve the problems of parameter adjustment and insufficient training samples to a certain extent by combining the advantages of model-driven and data-driven methods. Representative methods include subspace-based multidimensional sparse networks, joint low-rank spatial-spectral networks, deep spatial-spectral Bayesian posterior models, linear spectral mixture models, etc. However, these methods lack generalization capabilities for remote sensing images with different noises.

[0004] Commonly used dimensionality reduction methods are mainly divided into linear and nonlinear dimensionality reduction. Representative methods of nonlinear dimensionality reduction include manifold learning, stacked autoencoders, and convolutional autoencoders. These methods have powerful model learning capabilities, but are relatively sensitive to complex spectral changes, which impairs the discriminative ability of features after dimensionality reduction. Representative methods of linear dimensionality reduction include principal component analysis, linearized manifold learning, linear discriminant analysis, and local Fisher discriminant analysis. Although these methods can solve the problems existing in nonlinear dimensionality reduction methods, due to the limitations of model design, the model lacks data representation capabilities.

[0005] Through the above analysis, the problems and defects of the prior art are as follows:

[0006] (1) The existing denoising methods in remote sensing image classification have difficulties in parameter adjustment, insufficient training samples, and insufficient generalization ability.

[0007] (2) Existing dimensionality reduction methods lack data representation capabilities and impair the discriminative ability of features after dimensionality reduction. Summary of the invention

[0008] In view of the problems existing in the prior art, the present invention provides a remote sensing image classification method based on superpixel unsupervised linear discriminant analysis.

[0009] The present invention is implemented in this way: a remote sensing image classification method based on superpixel unsupervised linear discriminant analysis includes:

[0010] Step 1, superpixel generation;

[0011] Step 2, local reconstruction;

[0012] Step 3: Unsupervised global feature dimensionality reduction;

[0013] Step 4: unsupervised local feature dimensionality reduction;

[0014] Step 5: The support vector machine generates classification results.

[0015] Furthermore, the superpixel generation method:

[0016] The final deep features in the constructed superpixel generation network are expressed as Where H, W and C4 represent the height, width and number of channels of the image respectively, and are separated into and Two parts, For reconstructing a new image, It is used for superpixel generation;

[0017] The overall objective function consists of three parts: clustering term, spatial smoothing term and reconstruction term;

[0018]

[0019] Where l O Represents the objective function of the model as a whole; l C represents the clustering item; l S represents the spatial smoothness term; l R represents the reconstruction term; λ and are all constant coefficients, where λ = 2,

[0020] The defined cluster terms are calculated as follows:

[0021]

[0022] Where β = 2 is mainly used to control the number of superpixels generated; represents the average value of the category probability vector of all pixels; P (i，j)，n Represents the probability vector of pixel category n located in row i and column j, calculated as follows:

[0023]

[0024] Where n∈[0, C4-4] is rounded to the nearest integer. Indicates that the Softmax function is used to The channel dimension is converted into the category probability of the corresponding pixel;

[0025] The spatial smoothness term is divided into two parts: the smoothness loss in the x-direction and the y-direction. The new image is re-arranged to obtain the image (The new image dimensions are H×W×C, image The dimension is C×H×W); the absolute value of the probability difference and the image The exponential function of the squared difference of the gradient is defined, and the average spatial smoothness loss of all pixels is calculated; the calculation method is as follows:

[0026]

[0027] Where n is the channel and the clustering term are consistent; ΔP x，ij Represents the pixel probability difference in the x direction of the probability map; Represents the pixel intensity difference in the x direction of the rearranged image; ΔP y，ij Represents the pixel probability difference in the y direction of the probability map; Represents the pixel intensity difference in the y direction of the rearranged image;

[0028] The reconstruction term is calculated as follows:

[0029]

[0030] In the formula is the rearranged image obtained when calculating the spatial smoothing term; represents the image generated from the first three channels separated from the deep features; ||·||2 indicates that the 2-norm is used for calculation.

[0031] Furthermore, the local reconstruction method:

[0032] The input remote sensing image is represented as I 1 ∈R W×H×B , where W, H and B represent width, height and number of spectra respectively, transforming the three-dimensional remote sensing data into a two-dimensional matrix X 1 ={x1, x2, ..., x HW}∈R B×WH , and the obtained superpixel is represented as in Indicates The i-th pixel in the superpixel, Indicates The total number of pixels in the superpixel; the reconstructed matrix is ​​expressed as The first All pixels of a local reconstructed superpixel are represented as So that it corresponds to the superpixel in the original data Each pixel is locally reconstructed using the surrounding k neighboring pixels belonging to the same superpixel block; For example, the k adjacent pixels belonging to the same superpixel block are The reconstructed data is represented as It is calculated as follows:

[0033]

[0034] In the formula w j express and The similarity weight between ,δ is calculated as follows:

[0035]

[0036] Furthermore, the unsupervised global feature dimensionality reduction method:

[0037] The pseudo labels generated by linear discriminant analysis and superpixel segmentation are used to perform unsupervised global feature dimensionality reduction on the original input remote sensing image and local reconstruction data; the principle of linear discriminant analysis is used to reduce the dimension of the original remote sensing data X 1 and local reconstruction data X 1* The global intra-class scatter matrix and inter-class scatter matrix are calculated in ; among them, the average vector u of the entire remote sensing data and the The average vector of superpixels The calculation method is as follows:

[0038]

[0039] Using the same method as above, the mean vector u of the entire reconstructed remote sensing image can be calculated * and The average vector of reconstructed superpixels The objective function for unsupervised global feature dimensionality reduction is defined as follows:

[0040]

[0041] Where A G is the projection matrix to be calculated, the global inter-class scatter matrix and the intra-class scatter matrix The calculation is as follows:

[0042]

[0043]

[0044] Using eigendecomposition to G Perform the calculation:

[0045]

[0046] Select the eigenvectors corresponding to the largest d eigenvalues ​​to construct the projection matrix A G ∈R B×d , global feature F G The calculation is as follows:

[0047] F G =A G T X 1* .

[0048] Furthermore, the unsupervised local feature dimensionality reduction method:

[0049] The local Fisher discriminant analysis and pseudo-labels are used to perform unsupervised local feature dimensionality reduction on the local data of each superpixel and its neighboring superpixels; for each reconstructed superpixel For , its neighboring superpixels are recorded as The corresponding number of pixels is Therefore, for the The total number of pixels in a superpixel and its neighboring superpixels is The objective function for unsupervised local feature dimensionality reduction is defined as follows:

[0050]

[0051] Where A L ∈R B×d is the projection matrix to be calculated, the local inter-class scatter matrix and the intra-class scatter matrix The calculation is as follows:

[0052]

[0053]

[0054] In the formula and The calculation is as follows:

[0055]

[0056]

[0057] In the formula express The superpixel label, A i，j express and The similarity between them is calculated as follows:

[0058]

[0059] In the formula, δ represents the parameter controlling affinity; finally, the corresponding A is solved by the characteristic decomposition method. L , for the The calculation method of local features in superpixels and the aggregation of local features of all superpixels is as follows:

[0060]

[0061]

[0062] Furthermore, the support vector machine generates classification results by:

[0063] The results of unsupervised global feature dimensionality reduction and unsupervised local feature dimensionality reduction are channel-connected to obtain the final global-local features for classification; finally, the best support vector machine model trained with this feature gives the classification result; the channel connection is defined as follows:

[0064] F=[F G ,F L ]∈R WH×2Dd .

[0065] Another object of the present invention is to provide a remote sensing image classification system based on superpixel unsupervised linear discriminant analysis, comprising:

[0066] The superpixel generation module is used to extract the red, green and blue bands from the input remote sensing image to form a new image and introduce pixel position information. The constructed superpixel generation network is then used to extract deep features and realize superpixel generation in combination with the overall objective function.

[0067] A local reconstruction module, used for locally reconstructing each pixel;

[0068] The dimensionality reduction module is used to perform unsupervised global feature dimensionality reduction on the original input remote sensing image and local reconstruction data using pseudo labels generated by linear discriminant analysis and superpixel segmentation; and to perform unsupervised local feature dimensionality reduction on the local data of each superpixel and its adjacent superpixels using local Fisher discriminant analysis and pseudo labels.

[0069] The classification result generation module is used to connect the results of unsupervised global feature dimensionality reduction and unsupervised local feature dimensionality reduction processing through channels to obtain the final global-local features used for classification; finally, the classification result is given by the optimal support vector machine model trained with this feature.

[0070] Another object of the present invention is to provide a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the remote sensing image classification method based on superpixel unsupervised linear discriminant analysis.

[0071] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to perform the steps of the remote sensing image classification method based on superpixel unsupervised linear discriminant analysis.

[0072] Another object of the present invention is to provide an information data processing terminal, which is used to implement the remote sensing image classification system based on superpixel unsupervised linear discriminant analysis.

[0073] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0074] First, the present invention realizes remote sensing image classification based on superpixel unsupervised linear discriminant analysis, with an overall classification accuracy of 97.51%, an average precision of 98.28%, and a Kappa coefficient of 97.14%. The method provided by the present invention uses the results of superpixel generation to perform local reconstruction to realize remote sensing image denoising, combines superpixels with linear discriminant analysis to realize unsupervised global feature dimensionality reduction, combines superpixels with local Fisher discriminant analysis to realize unsupervised local feature dimensionality reduction, and connects the global features and local features extracted at last to obtain global-local features through channel connection, and realizes remote sensing image classification in combination with support vector machine. It has the advantages of good noise resistance, strong feature representation ability, excellent feature discriminability, high classification accuracy, and strong generalization ability, and provides effective support for improving the efficiency and accuracy of remote sensing image classification processing.

[0075] The present invention realizes remote sensing image denoising by using the result of superpixel generation for local reconstruction, realizes unsupervised global feature dimensionality reduction by combining superpixel with linear discriminant analysis, realizes unsupervised local feature dimensionality reduction by combining superpixel with local Fisher discriminant analysis, connects the global features and local features finally extracted by channels to obtain global-local features, and realizes remote sensing image classification by combining support vector machine. The present invention solves the problems of difficulty in adjusting parameters, insufficient training samples and insufficient generalization ability of denoising methods in existing remote sensing image classification; and solves the problems of lack of data representation ability and damage to the discriminative ability of features after dimensionality reduction in existing dimensionality reduction methods.

[0076] Second, as auxiliary evidence of the inventiveness of the claims of the present invention, it is also reflected in the following important aspects:

[0077] (1) The expected benefits and commercial value of the technical solution of the present invention after transformation are:

[0078] The overall technical solution can be applied to remote sensing image classification and can realize the detection of land use and land cover changes in large areas.

[0079] (2) The technical solution of the present invention solves a technical problem that people have been eager to solve but have never been able to solve successfully:

[0080] The use of superpixels for local reconstruction effectively solves the problem of noise interference and can be used as a new denoising solution. It also effectively solves the problems of difficult parameter adjustment, insufficient training samples and insufficient generalization ability of denoising methods in existing remote sensing image classification. Both linear discriminant analysis and local Fisher discriminant analysis are supervised dimensionality reduction methods. In the present invention, pseudo labels generated by unsupervised superpixels are used to cleverly transform these two dimensionality reduction methods into unsupervised dimensionality reduction methods. Unsupervised global feature dimensionality reduction is combined with unsupervised local feature dimensionality reduction, which solves the problems of lack of data representation ability and damage to the discriminant ability of features after dimensionality reduction in existing dimensionality reduction methods.

[0081] (3) The technical solution of the present invention overcomes technical prejudice:

[0082] Researchers generally believe that when using support vector machines to classify remote sensing images, good classification results can only be achieved when the feature dimension is high enough, but a high feature dimension will bring higher computational complexity. The experimental results of this invention fully prove that the low feature dimension obtained by unsupervised global feature dimensionality reduction and unsupervised local feature dimensionality reduction can also be applied to support vector machines to achieve good classification results.

[0083] Third, the remote sensing image classification method based on superpixel unsupervised linear discriminant analysis provided by the present invention solves the problems of high computational complexity and low classification accuracy in the high-dimensional remote sensing image data processing process in the prior art. Traditional methods often rely on supervised learning, require a large amount of labeled data for training, and are easily affected by noise and redundant features when facing high-dimensional data, resulting in inaccurate classification results.

[0084] Through the technical solution of the present invention, the superpixel segmentation technology is used to effectively reduce the data dimension and processing complexity, while retaining the key edge information of the image. In addition, through unsupervised global and local feature dimensionality reduction, the most representative features can be automatically selected for classification without relying on pre-labeled category information, which greatly improves the adaptability and versatility of the method. The application of support vector machines further improves the accuracy and stability of classification.

[0085] The significant technical progress is that the present invention not only outperforms existing methods in classification accuracy, but also shows higher efficiency and robustness when processing complex high-dimensional remote sensing data. This unsupervised classification method is particularly suitable for remote sensing image analysis in large-scale and complex scenes, and provides a new and efficient solution for the automated classification of remote sensing images. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 It is a flow chart of a remote sensing image classification method based on superpixel unsupervised linear discriminant analysis provided by an embodiment of the present invention.

[0087] Figure 2 It is a structural block diagram of a remote sensing image classification system based on superpixel unsupervised linear discriminant analysis provided by an embodiment of the present invention.

[0088] Figure 3 It is a detailed flowchart of remote sensing image classification based on superpixel unsupervised linear discriminant analysis provided by an embodiment of the present invention.

[0089] Figure 4 It is a superpixel generation network diagram constructed according to an embodiment of the present invention.

[0090] Figure 5 This is a diagram of the structure change of remote sensing image data provided by an embodiment of the present invention.

[0091] Figure 6 It is a comparison diagram of the remote sensing image classification result and the real label provided by the embodiment of the present invention. DETAILED DESCRIPTION

[0092] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0093] like Figure 1 As shown, a remote sensing image classification method based on superpixel unsupervised linear discriminant analysis provided by an embodiment of the present invention includes the following steps:

[0094] S101, superpixel generation;

[0095] S102, local reconstruction;

[0096] S103, unsupervised global feature dimensionality reduction;

[0097] S104, unsupervised local feature dimensionality reduction;

[0098] S105, the support vector machine generates a classification result.

[0099] The remote sensing image classification method based on superpixel unsupervised linear discriminant analysis provided by the embodiment of the present invention first generates superpixels through step S101. The process of superpixel generation is to divide the remote sensing image into a group of interconnected small areas, which are called superpixels. The pixels inside each superpixel have similar spectral and spatial characteristics. Such preprocessing helps to reduce the dimension of the data, reduce the complexity of subsequent processing, and retain the edge information of the image.

[0100] Next, in step S102, the generated superpixels are locally reconstructed. Local reconstruction constructs its representation in the original image by analyzing the local features within each superpixel. The main purpose of this step is to extract features from each superpixel to capture its structure and texture information in the local area. These features will provide basic data for subsequent feature dimensionality reduction and classification.

[0101] In step S103, unsupervised global feature dimensionality reduction is performed. This process analyzes the feature distribution of the entire image, selects the features that can most effectively distinguish different categories for retention, and discards redundant features. The unsupervised method does not rely on pre-labeled category information, but performs dimensionality reduction based on the distribution characteristics of the data itself, thereby improving the adaptability and robustness of the algorithm.

[0102] Then, step S104 is entered to perform unsupervised local feature dimensionality reduction. Different from global dimensionality reduction, local feature dimensionality reduction focuses on dimensionality reduction in the local feature space of each superpixel. By retaining the local features that contribute most to classification, this process can further reduce the dimension of the data while maximally retaining the unique information of each superpixel to ensure the accuracy of classification.

[0103] Finally, in step S105, support vector machine (SVM) is used for classification. SVM is a commonly used supervised learning algorithm that performs well in processing high-dimensional data. In the present invention, SVM receives feature data that has undergone dimensionality reduction processing, and generates a final classification result based on its built-in classification model. SVM uses a hyperplane to separate superpixels of different categories, thereby achieving accurate classification of the entire remote sensing image.

[0104] In summary, the present invention can effectively reduce the dimensionality of high-dimensional remote sensing image data through multi-step processing of superpixel generation, local reconstruction, unsupervised global and local feature dimensionality reduction, and support vector machine classification, and finally achieve high-precision image classification. This method is particularly suitable for processing complex remote sensing image data and has high accuracy and efficiency in practical applications.

[0105] The superpixel generation method provided by the embodiment of the present invention:

[0106] In order to balance the effectiveness and complexity of the model, the three bands of red, green and blue are extracted from the input remote sensing image to form a new image for superpixel generation and the pixel position information is introduced. The constructed four-layer sequential network is used to extract deep features. Each layer of the network consists of reflection filling, convolution, normalization and ReLU activation. The use of reflection filling can better retain the image boundary information and avoid the problem of undesirable boundary effects in the segmentation results. At the same time, in order to enable the network to better capture important features and improve the representation ability of the network, a channel attention mechanism is introduced after the four-layer sequential network to enhance the model's attention to different channels. The edge complexity is calculated according to the new image to determine the number of feature channels extracted by the four-layer sequential network for each pixel. The final deep features in the constructed superpixel generation network are expressed as Where H, W and C4 represent the height, width and number of channels of the image respectively, and are separated into and Two parts, For reconstructing a new image, It is used for superpixel generation;

[0107] In order to encourage deterministic superpixel allocation and make the generated superpixel size as uniform as possible, a mutual information term similar to the maximization of regular information is introduced to construct the clustering term; at the same time, a spatial smoothing term is introduced as the main prior for image processing tasks and used to quantify the differences between adjacent pixels; and a reconstruction term is constructed to ensure that the final extracted deep features come from the image itself; therefore, the overall objective function consists of three parts: clustering term, spatial smoothing term and reconstruction term;

[0108]

[0109] Where l O Represents the objective function of the model as a whole; l C represents the clustering item; l S represents the spatial smoothness term; l R represents the reconstruction term; λ and are all constant coefficients, where λ = 2,

[0110] The defined cluster terms are calculated as follows:

[0111]

[0112] Where β = 2 is mainly used to control the number of superpixels generated; represents the average value of the category probability vector of all pixels; P (i,j),n Represents the probability vector of pixel category n located in row i and column j, calculated as follows:

[0113]

[0114] Where n∈[0,C4-4] is rounded to an integer. Indicates that the Softmax function is used to The channel dimension is converted into the category probability of the corresponding pixel;

[0115] The spatial smoothness term is divided into two parts: the smoothness loss in the x-direction and the y-direction. The new image is re-arranged to obtain the image (The new image dimensions are H×W×C, image The dimension is C×H×W); the absolute value of the probability difference and the image The exponential function of the squared difference of the gradient is defined, and the average spatial smoothness loss of all pixels is calculated; the calculation method is as follows:

[0116]

[0117] Where n is the channel and the clustering term are consistent; ΔP x,ij Represents the pixel probability difference in the x direction of the probability map; Represents the pixel intensity difference in the x direction of the rearranged image; ΔP y，ij Represents the pixel probability difference in the y direction of the probability map; Represents the pixel intensity difference in the y direction of the rearranged image;

[0118] The reconstruction term is calculated as follows:

[0119]

[0120] In the formula is the rearranged image obtained when calculating the spatial smoothing term; represents the image generated from the first three channels separated from the deep features; ||·||2 indicates that the 2-norm is used for calculation.

[0121] The local reconstruction method provided by the embodiment of the present invention:

[0122] To achieve remote sensing image denoising, the input remote sensing image is represented as I 1 ∈R W×H×B , where W, H and B represent width, height and number of spectra respectively, transforming the three-dimensional remote sensing data into a two-dimensional matrix X 1 ={x1, x2, ..., x HW}∈R B×WH , and the obtained superpixel is represented as in Indicates The i-th pixel in the superpixel, Indicates The total number of pixels in the superpixel; the reconstructed matrix is ​​expressed as The first All pixels of a local reconstructed superpixel are represented as So that it corresponds to the superpixel in the original data Each pixel is locally reconstructed using the surrounding k neighboring pixels belonging to the same superpixel block; For example, the k adjacent pixels belonging to the same superpixel block are The reconstructed data is represented as It is calculated as follows:

[0123]

[0124] In the formula w j express and The similarity weight between ,δ is calculated as follows:

[0125]

[0126] The unsupervised global feature dimensionality reduction method provided by the embodiment of the present invention is as follows:

[0127] The pseudo labels generated by linear discriminant analysis and superpixel segmentation are used to perform unsupervised global feature dimensionality reduction on the original input remote sensing image and local reconstruction data; the principle of linear discriminant analysis is used to reduce the dimension of the original remote sensing data X 1 and local reconstruction data X 1* The global intra-class scatter matrix and inter-class scatter matrix are calculated in ; among them, the average vector u of the entire remote sensing data and the The average vector of superpixels The calculation method is as follows:

[0128]

[0129] Using the same method as above, the mean vector u* and the first The average vector of reconstructed superpixels The objective function for unsupervised global feature dimensionality reduction is defined as follows:

[0130]

[0131] Where A G is the projection matrix to be calculated, the global inter-class scatter matrix and the intra-class scatter matrix The calculation is as follows:

[0132]

[0133]

[0134] Using eigendecomposition to G Perform the calculation:

[0135]

[0136] Select the eigenvectors corresponding to the largest d eigenvalues ​​to construct the projection matrix A G ∈R B×d , global feature F G The calculation is as follows:

[0137] F G =A G T X 1* .

[0138] The unsupervised local feature dimensionality reduction method provided by the embodiment of the present invention:

[0139] Similar to unsupervised global feature dimensionality reduction, unsupervised local feature dimensionality reduction is performed on the local data of each superpixel and its neighboring superpixels using local Fisher discriminant analysis and pseudo labels. For , its neighboring superpixels are recorded as The corresponding number of pixels is Therefore, for the The total number of pixels in a superpixel and its neighboring superpixels is The objective function for unsupervised local feature dimensionality reduction is defined as follows:

[0140]

[0141] Where A L ∈R B×d is the projection matrix to be calculated, the local inter-class scatter matrix and the intra-class scatter matrix The calculation is as follows:

[0142]

[0143]

[0144] In the formula and The calculation is as follows:

[0145]

[0146]

[0147] In the formula express The superpixel label, A i，j express and The similarity between them is calculated as follows:

[0148]

[0149] In the formula, δ represents the parameter controlling affinity; finally, the corresponding A is solved by the characteristic decomposition method. L , for the The calculation method of local features in superpixels and the aggregation of local features of all superpixels is as follows:

[0150]

[0151]

[0152] The support vector machine method for generating classification results provided by the embodiment of the present invention is as follows:

[0153] In order to make the features used by the support vector machine have better representation and discrimination capabilities, the results of the unsupervised global feature dimensionality reduction and the unsupervised local feature dimensionality reduction are channel-connected to obtain the final global-local features used for classification; finally, the best support vector machine model trained by this feature gives the classification result; the channel connection is defined as follows:

[0154] F=[F G ,F L ]∈R WH×2Dd .

[0155] like Figure 2 As shown, a remote sensing image classification system based on superpixel unsupervised linear discriminant analysis provided by an embodiment of the present invention includes:

[0156] The superpixel generation module is used to extract the red, green and blue bands from the input remote sensing image to form a new image and introduce pixel position information. The constructed superpixel generation network is then used to extract deep features and realize superpixel generation in combination with the overall objective function.

[0157] A local reconstruction module, used for locally reconstructing each pixel;

[0158] The dimensionality reduction module is used to perform unsupervised global feature dimensionality reduction on the original input remote sensing image and local reconstruction data using pseudo labels generated by linear discriminant analysis and superpixel segmentation; and to perform unsupervised local feature dimensionality reduction on the local data of each superpixel and its adjacent superpixels using local Fisher discriminant analysis and pseudo labels.

[0159] The classification result generation module is used to connect the results of unsupervised global feature dimensionality reduction and unsupervised local feature dimensionality reduction processing through channels to obtain the final global-local features used for classification; finally, the classification result is given by the optimal support vector machine model trained with this feature.

[0160] Another object of the present invention is to provide a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the remote sensing image classification method based on superpixel unsupervised linear discriminant analysis.

[0161] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to perform the steps of the remote sensing image classification method based on superpixel unsupervised linear discriminant analysis.

[0162] Another object of the present invention is to provide an information data processing terminal, which is used to implement the remote sensing image classification system based on superpixel unsupervised linear discriminant analysis.

[0163] Principles involved in steps S101 to S105:

[0164] Superpixel generation principle:

[0165] The goal of superpixel generation is to divide an image into several sets of pixels with similar features. Compared with a single pixel, a superpixel can more effectively represent the local structural information in an image. The edge complexity is calculated from a given image using the Sobel operator, which is used to guide the sequential network to finally extract the number of feature channels for each pixel, transforming low-level features into deep features, and then using the channel attention mechanism to dynamically adjust the weight of each channel of the deep feature, and minimize the objective function during inference time, thereby achieving superpixel generation.

[0166] Principle of local reconstruction:

[0167] The pixels within the superpixel area are reconstructed using a local reconstruction method. By analyzing the difference between the pixel values ​​of the input remote sensing image and the predicted values ​​of the local model, the pixels that are most affected by noise can be identified, thereby determining the distribution and intensity of the noise. Combining the superpixel segmentation results with the original data for local reconstruction can effectively remove noise and improve the visual quality of remote sensing images while retaining the details of the remote sensing image. This method utilizes the local similarity characteristics of superpixels, making the denoising process more accurate and efficient. It can effectively solve the problems of difficult parameter adjustment, insufficient training samples, and insufficient generalization ability in traditional denoising methods.

[0168] Unsupervised global feature dimensionality reduction principle:

[0169] Each superpixel region generated has a specific pseudo-label. The original input remote sensing image and the result of local reconstruction are used as input data for unsupervised global feature dimensionality reduction, and the pseudo-label is used as the class label. Linear discriminant analysis is applied for dimensionality reduction. Among them, linear discriminant analysis finds the optimal linear transformation by maximizing the difference between classes and minimizing the difference within classes, and projects high-dimensional features into low-dimensional space. This method retains the local structural information of the original input remote sensing image, improves the discriminative ability of global features and the efficiency of dimensionality reduction, and enhances the accuracy and efficiency of subsequent classification while retaining details.

[0170] Principle of unsupervised local feature dimensionality reduction:

[0171] In order to solve the problem of dimensionality limitation and mixed objects in superpixels in the dimensionality reduction process, it is necessary to further consider the local information between superpixels. Local Fisher discriminant analysis combines local and global information to reduce dimensionality, extracts feature vectors from the locally reconstructed data, builds a neighborhood graph to maintain the local structure, and calculates intra-class and inter-class divergence matrices to maximize inter-class differences and minimize intra-class differences. By optimizing the objective function, the optimal linear transformation is found, and the high-dimensional features are projected into a low-dimensional space, so that the data after dimensionality reduction retains the local structure while enhancing the separability between different categories.

[0172] The principle of generating classification results by support vector machine:

[0173] In order to make the features used by the support vector machine have better representation and discrimination capabilities, the results of the unsupervised global feature dimensionality reduction and the unsupervised local feature dimensionality reduction are channel-connected to obtain the global-local features used for classification. This can solve the problem that the existing dimensionality reduction methods lack data representation capabilities and damage the discrimination capabilities of the features after dimensionality reduction. Among them, the support vector machine maps the data to a high-dimensional space by using a radial basis function kernel, thereby finding the optimal hyperplane in the high-dimensional space to achieve nonlinear classification. The parameters of the radial basis function kernel and the penalty parameters of the support vector machine are systematically optimized using grid search. By training and evaluating multiple models within a predefined parameter range, the best parameter combination is selected to improve the classification accuracy of the final remote sensing image and the generalization ability of the model.

[0174] In summary, the present invention realizes remote sensing image denoising by using the results of superpixel generation for local reconstruction, realizes unsupervised global feature dimensionality reduction by combining superpixel with linear discriminant analysis, realizes unsupervised local feature dimensionality reduction by combining superpixel with local Fisher discriminant analysis, connects the global features and local features extracted at last to obtain global-local features through channel connection, and realizes remote sensing image classification by combining support vector machine. The problems of difficulty in adjusting parameters, insufficient training samples and insufficient generalization ability of denoising methods in existing remote sensing image classification are solved; the problems of lack of data representation ability and damage to the discriminative ability of features after dimensionality reduction in existing dimensionality reduction methods are solved.

[0175] The present invention is specifically implemented:

[0176] A remote sensing image classification method based on superpixel unsupervised linear discriminant analysis (such as Figure 3 As shown), in this embodiment, the superpixel generation part uses the deep learning framework PyTorch and the programming language is Python; in this example, local reconstruction, unsupervised global feature dimensionality reduction, unsupervised local feature dimensionality reduction, and support vector machine generation of classification results are implemented using MATLAB software.

[0177] Step 1: Input remote sensing image Indian Pines dataset is collected by airborne visible light and infrared imaging spectrometer (AVIRIS), with spatial resolution of 20m, wavelength range of 0.4-2.5μm, effective spectrum number after removing water absorption band is 200, image size is 145×145, select red, green and blue bands from effective spectrum to form new image. And use super pixel generation network (such as Figure 4 As shown), the composed new image is applied to the written __preprocess function, the position information is introduced and normalized to obtain a tensor of shape. The edge complexity is calculated as 167 using the calculate_edge_ratio function and the calculate_kernel_size function, and the fast one-dimensional convolution kernel size in the channel attention mechanism is calculated as 3. Then, the deep features of shape are obtained through the four-layer sequential network and the channel attention mechanism, and separated into a tensor of shape for calculating the reconstruction term, and another tensor of shape for calculating the clustering term and the spatial smoothing term. Iterate through the defined optimize function, and use the Adam optimizer to update the model parameters to minimize the objective function. Among them, the number of iterations is set to 500 times, the learning rate is set to 0.01, and finally 54 superpixels are generated.

[0178] Step 2: Standardize the original input remote sensing image, use the reshape function in MATLAB to convert the three-dimensional input into a two-dimensional array, and convert the pseudo labels generated by the superpixel into a one-dimensional row vector. Use the findConstruct function to locally reconstruct each pixel for the 35 neighboring pixels in the same superpixel block.

[0179] Step 3. Use the getsw2 function to calculate the global inter-class scatter matrix for the original remote sensing data and the local reconstruction data, and use the getsb2 function to calculate the global intra-class scatter matrix. Use the Find_K_Max_Eigen function to solve the 10 largest eigenvalues, and use the eigenvectors corresponding to the 10 eigenvalues ​​to construct the projection matrix. The projection matrix and local reconstruction can be used to calculate the result of unsupervised global feature dimensionality reduction, and then transform the result into a two-dimensional array, and finally use the data_normalization function for standardization.

[0180] Step 4: The local reconstructed data, pseudo labels generated by superpixels, and the hyperparameter 10 used to determine the unsupervised local feature dimensionality reduction are used as inputs to the local_LFDA function to find the adjacent superpixels of each reconstructed superpixel and the number of pixels corresponding to each. Then, the LFDA function is used to calculate the local inter-class scatter matrix and intra-class scatter matrix for each reconstructed superpixel and its adjacent reconstructed superpixels, thereby achieving the overall unsupervised local feature dimensionality reduction. As in step 3, the data_normalization function is finally used to normalize the results of the unsupervised local feature dimensionality reduction.

[0181] Step 5: Connect the global features and local features extracted in steps 3 and 4 to obtain global-local features. Figure 5 As shown in the figure, the global-local features are transformed from a two-dimensional array of size to a three-dimensional array of size using the reshape function. Define an array trainsets containing the size of the training set, and use the samplesdivide function to divide the input remote sensing image Indian Pines dataset into a training set and a test set. When training the support vector machine, the radial basis function kernel is selected, and the parameters of the radial basis function kernel and the penalty parameters of the support vector machine are systematically optimized by combining grid search. First, the svmtrain function is used to complete the model training, and then the svmpredict function is used to predict the test set. The prediction results are input into the confusion_matrix function to calculate the confusion matrix, thereby calculating the overall classification accuracy, average precision, and Kappa coefficient used to evaluate the performance.

[0182] This example processes the remote sensing image Indian Pines dataset. Figure 6 This is a comparison chart of the remote sensing image classification results and the true labels of the embodiment.

[0183] It can be seen from the above embodiments that the present invention realizes remote sensing image classification based on superpixel unsupervised linear discriminant analysis, with an overall classification accuracy of 97.51%, an average precision of 98.28%, and a Kappa coefficient of 97.14%. The method provided by the present invention utilizes the results of superpixel generation to perform local reconstruction to realize remote sensing image denoising, realizes unsupervised global feature dimensionality reduction by combining superpixels with linear discriminant analysis, realizes unsupervised local feature dimensionality reduction by combining superpixels with local Fisher discriminant analysis, connects the global features and local features extracted at last through channels to obtain global-local features, and realizes remote sensing image classification by combining support vector machine. It has the advantages of good noise resistance, strong feature representation ability, excellent feature discriminability, high classification accuracy, and strong generalization ability, and provides effective support for improving the efficiency and accuracy of remote sensing image classification processing.

[0184] 1. Specific application fields or related products of the present invention.

[0185] It is used in land use and land cover change detection, crop monitoring, forest resource survey, urban planning and environmental protection related fields.

[0186] 2. Relevant evidence of the technical effects obtained by the embodiments of the present invention.

[0187] The Indian Pines dataset is collected by the Airborne Visible and Infrared Imaging Spectrometer (AVIRIS) with a spatial resolution of 20 m and a wavelength range of 0.4 to 2.5 μm. The number of effective spectra after removing the water absorption band is 200 and the image size is 145 × 145. Table 1 shows the 16 categories and the corresponding number of pixels excluding the ground background.

[0188] Table 1: Indian Pines dataset

[0189]

[0190] The Adam optimizer is used to update the model parameters in the superpixel generation network to minimize the objective function. The number of iterations is set to 500, the learning rate is set to 0.01, 54 superpixels are generated after processing the Indian Pines data, the number of spatial neighbors in the superpixel reconstruction is set to 35, and the superpixels are combined with linear discriminant analysis to obtain 10-dimensional global features. The superpixels are combined with local Fisher discriminant analysis to obtain 10-dimensional local features. The extracted global features and local features are channel-connected to obtain 20-dimensional global-local features for support vector machines.

[0191] For the training of the support vector machine model, the radial basis function kernel is selected, and the parameters of the radial basis function kernel are systematically optimized in combination with grid search, and the penalty factor is set to c = 100000, and the best searched parameters are used to determine the final support vector machine model. For the 16 categories in the Indian Pines dataset, the number of samples provided to the support vector machine for training is shown in the second column of Table 2, and the accuracy of the final support vector machine model in predicting each category is shown in the last column of Table 2. The last three rows of Table 2 show the overall classification accuracy (OA), average accuracy (AA) and Kappa coefficient.

[0192] Table 2: Classification results of Indian Pines dataset

[0193]

[0194] Only the classification accuracy of the Grass-pasture-mowed feature category is slightly lower. The reason for this is that the number of samples in this category is too small, which limits the learning ability of the support vector machine model to a certain extent. Figure 6 The comparison between the remote sensing image classification results of the embodiment and the real labels is shown in the figure. It can be seen from the comparison chart that the misclassified places in the results mostly appear at the junction of different categories. The reason for this is that the real labels cannot fully represent the spatial distribution of the actual categories on the ground, or because there are certain boundary errors in the process of superpixel generation, and local reconstruction using the generated superpixels will produce error transmission and affect the final classification accuracy. The experimental results of the present invention prove that the interference of noise can be effectively solved by local reconstruction, and fully prove that the low feature dimension obtained by unsupervised global feature dimensionality reduction and unsupervised local feature dimensionality reduction can be applied to support vector machines to achieve good classification results.

[0195] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. It can be understood by a person of ordinary skill in the art that the above-mentioned devices and methods can be implemented using computer executable instructions and / or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0196] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. A remote sensing image classification method based on superpixel unsupervised linear discriminant analysis, characterized in that: The following steps are involved: a) Superpixel generation: The remote sensing image is segmented into multiple superpixel regions, and the pixels within each superpixel have similar spectral and spatial characteristics; b) Local reconstruction: extract features from the generated superpixels and construct a representation of the superpixels in the original image to capture their local structure and texture information; c) Unsupervised global feature dimensionality reduction: By analyzing the feature distribution of the entire image, the features that can effectively distinguish different categories are selected and retained, and global feature dimensionality reduction is performed; d) Unsupervised local feature dimensionality reduction: further reduce the dimensionality in the local feature space of superpixels and retain the local features that contribute most to classification; e) Support vector machine generates classification results: Use the support vector machine algorithm to classify the feature data after dimensionality reduction processing to generate the final remote sensing image classification results.

2. The remote sensing image classification method based on superpixel unsupervised linear discriminant analysis according to claim 1, characterized in that: The unsupervised global feature dimensionality reduction and the unsupervised local feature dimensionality reduction do not rely on pre-labeled category information, but perform dimensionality reduction based on the distribution characteristics of the data itself, thereby improving the adaptability and robustness of the classification algorithm.

3. The remote sensing image classification method based on superpixel unsupervised linear discriminant analysis as claimed in claim 1, characterized in that: The super pixel generation method: The final deep features in the constructed superpixel generation network are expressed as Where H, W and C4 represent the height, width and number of channels of the image respectively, and are separated into and Two parts, For reconstructing a new image, It is used for superpixel generation; The overall objective function consists of three parts: clustering term, spatial smoothing term and reconstruction term; Where l O Represents the objective function of the model as a whole; l C represents the clustering item; l S represents the spatial smoothness term; l R represents the reconstruction term; λ and are all constant coefficients, where λ = 2, The defined cluster terms are calculated as follows: Where β = 2 is mainly used to control the number of superpixels generated; represents the average value of the category probability vector of all pixels; P (i，j)，n Represents the probability vector of pixel category n located in row i and column j, calculated as follows: Where n∈[0, C4-4] is rounded to the nearest integer. Indicates that the Softmax function is used to The channel dimension is converted into the category probability of the corresponding pixel; The spatial smoothness term is divided into two parts: the smoothness loss in the x-direction and the y-direction. The new image is re-arranged to obtain the image (The new image dimensions are H×W×C, image The dimension is C×H×W); the absolute value of the probability difference and the image The exponential function of the squared difference of the gradient is defined, and the average spatial smoothness loss of all pixels is calculated; the calculation method is as follows: Where n is the channel and the clustering term are consistent; ΔP x，ij Represents the pixel probability difference in the x direction of the probability map; Represents the pixel intensity difference in the x direction of the rearranged image; ΔP y，ij Represents the pixel probability difference in the y direction of the probability map; Represents the pixel intensity difference in the y direction of the rearranged image; The reconstruction term is calculated as follows: In the formula is the rearranged image obtained when calculating the spatial smoothing term; represents the image generated from the first three channels separated from the deep features; ||·||2 indicates that the 2-norm is used for calculation.

4. The remote sensing image classification method based on superpixel unsupervised linear discriminant analysis as claimed in claim 1, characterized in that: The local reconstruction method: The input remote sensing image is represented as I 1 ∈R W×H×B , where W, H and B represent width, height and number of spectra respectively, transforming the three-dimensional remote sensing data into a two-dimensional matrix X 1 ={x1, x2, ..., x HW }∈R B×WH , and the obtained superpixel is represented as in Indicates The i-th pixel in the superpixel, Indicates The total number of pixels in the superpixel; the reconstructed matrix is ​​expressed as The first All pixels of a local reconstructed superpixel are represented as So that it corresponds to the superpixel in the original data Each pixel is locally reconstructed using the surrounding k neighboring pixels belonging to the same superpixel block; For example, the k adjacent pixels belonging to the same superpixel block are The reconstructed data is represented as It is calculated as follows: In the formula wj means and The similarity weight between ,δ is calculated as follows:

5. The remote sensing image classification method based on superpixel unsupervised linear discriminant analysis as claimed in claim 1, characterized in that: The unsupervised global feature dimensionality reduction method: Linear discriminant analysis and pseudo labels generated by superpixel segmentation are used to perform unsupervised global feature dimensionality reduction on the original input remote sensing images and local reconstruction data; Using the principle of linear discriminant analysis, from the original remote sensing data X 1 and local reconstruction data X 1* The global intra-class scatter matrix and inter-class scatter matrix are calculated in ; among them, the average vector u of the entire remote sensing data and the The average vector of superpixels The calculation method is as follows: Using the same method as above, the mean vector u of the entire reconstructed remote sensing image can be calculated * and The average vector of reconstructed superpixels The objective function for unsupervised global feature dimensionality reduction is defined as follows: Where AG is the projection matrix to be calculated, and the global inter-class scatter matrix and the intra-class scatter matrix The calculation is as follows: Using eigendecomposition to G Perform the calculation: Select the eigenvectors corresponding to the largest d eigenvalues ​​to construct the projection matrix A G ∈R B×d , global feature F G The calculation is as follows: F G =A G T X 1* 。 6. The remote sensing image classification method based on superpixel unsupervised linear discriminant analysis as claimed in claim 1, characterized in that: The unsupervised local feature dimensionality reduction method: The local Fisher discriminant analysis and pseudo-labels are used to perform unsupervised local feature dimensionality reduction on the local data of each superpixel and its neighboring superpixels; for each reconstructed superpixel For , its neighboring superpixels are recorded as The corresponding number of pixels is Therefore, for the The total number of pixels in a superpixel and its neighboring superpixels is n t*k ; The objective function for unsupervised local feature dimensionality reduction is defined as follows: Where AL∈RB×d is the projection matrix to be calculated, and the local inter-class scatter matrix and the intra-class scatter matrix The calculation is as follows: In the formula and The calculation is as follows: In the formula express The superpixel label of A i，j express and The similarity between them is calculated as follows: In the formula, δ represents the parameter controlling affinity; finally, the corresponding A is solved by the characteristic decomposition method. L , for the The calculation method of local features in superpixels and the aggregation of local features of all superpixels is as follows:

7. The remote sensing image classification method based on superpixel unsupervised linear discriminant analysis as claimed in claim 1, characterized in that: The support vector machine generates classification results: The results of unsupervised global feature dimensionality reduction and unsupervised local feature dimensionality reduction are channel-connected to obtain the final global-local features for classification; finally, the best support vector machine model trained with this feature gives the classification result; the channel connection is defined as follows: F=[F G ,F L ]∈R WH×2Dd 。 8. A remote sensing image classification system based on superpixel unsupervised linear discriminant analysis that implements the remote sensing image classification method based on superpixel unsupervised linear discriminant analysis as claimed in any one of claims 1 to 7, characterized in that: The remote sensing image classification system based on superpixel unsupervised linear discriminant analysis includes: The superpixel generation module is used to extract the red, green and blue bands from the input remote sensing image to form a new image and introduce pixel position information. The constructed superpixel generation network is then used to extract deep features and realize superpixel generation in combination with the overall objective function. A local reconstruction module, used for locally reconstructing each pixel; The dimension reduction module is used to perform unsupervised global feature dimension reduction on the original input remote sensing image and local reconstruction data using linear discriminant analysis and pseudo labels generated by superpixel segmentation; and to perform unsupervised local feature dimension reduction on the local data of each superpixel and its neighboring superpixels using local Fisher discriminant analysis and pseudo labels; The classification result generation module is used to connect the results of unsupervised global feature dimensionality reduction and unsupervised local feature dimensionality reduction processing through channels to obtain the final global-local features used for classification; finally, the classification result is given by the optimal support vector machine model trained with this feature.

9. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the remote sensing image classification method based on superpixel unsupervised linear discriminant analysis as described in any one of claims 1 to 7.

10. An information data processing terminal, characterized in that: The information data processing terminal is used to implement the remote sensing image classification system based on superpixel unsupervised linear discriminant analysis as described in claim 8.

Images