A hyperspectral image unsupervised change detection method based on tensor convolution structure

Abstract

The application discloses a hyperspectral image unsupervised change detection method based on a tensor convolution structure and belongs to the technical field of remote sensing image intelligent analysis, and comprises the following steps: S1, constructing a training tensor for double-time-phase hyperspectral images; S2, performing Tucker decomposition on the training tensor to determine the spectral rank of a spectral mode factor matrix; S3, determining a layer-by-layer stacking structure according to the spectral rank of the spectral mode factor matrix, and outputting double-time-phase deep feature maps; S4, obtaining independent spectral difference maps and local spatial average maps according to the double-time-phase deep feature maps; and S5, obtaining a final change map according to the independent spectral difference maps and the local spatial average maps. The double-view change detector constructed by the application can effectively inhibit noise interference and has good anti-interference ability to illumination changes, seasonal fluctuations and system noise.

Description

Technical Field

[0001] This invention belongs to the field of intelligent analysis technology of remote sensing images, specifically relating to an unsupervised change detection method for hyperspectral images based on tensor convolution structure. Background Technology

[0002] Hyperspectral images, characterized by their high spectral resolution, can capture subtle spectral information within a certain wavelength range, making them invaluable for various applications. Hyperspectral change detection identifies changes in surface features by analyzing dual-temporal hyperspectral images of the same geographical area acquired at different times. Unsupervised learning paradigms, without relying on prior knowledge or manually labeled training data, can automatically discover and extract changes directly from image data. They demonstrate significant advantages in large-scale monitoring, sudden change events, and scenarios lacking ground-based ground data, and are widely used in environmental monitoring, agricultural management, and dynamic urban planning.

[0003] Existing unsupervised hyperspectral image change detection models mainly include traditional machine learning, deep learning methods, and machine learning-deep learning fusion methods. Deep learning methods typically employ convolutional neural networks (CNNs), Transformers, or other advanced architectures. To extract higher-level abstract features from bi-temporal hyperspectral images, these models often utilize pre-trained networks for knowledge transfer or transform the task into a supervised task by selecting high-confidence pseudo-labels from preprocessed samples. While deep learning models offer high accuracy, existing pre-trained models often do not match the structure of hyperspectral data, and the selection of pseudo-samples is difficult, resulting in a large number of model parameters and low efficiency. In contrast, traditional machine learning models based on difference image analysis, feature space transformation, and similarity metrics do not require pseudo-sample selection and are faster, but they rely on manually designed features and are susceptible to image noise, leading to lower accuracy. To overcome these problems, some studies combine deep learning with classic unsupervised machine learning algorithms, such as principal component analysis and slow feature analysis, as the core to construct an unsupervised change detection framework. This effectively integrates the advantages of both methods, reducing training samples, improving model efficiency, and simultaneously increasing change detection accuracy.

[0004] However, existing fusion models treat hyperspectral data, a three-dimensional structure, as a two-dimensional matrix. After projecting the three-dimensional data into two dimensions, features are extracted. This flattening process leads to the loss of key spectral-spatial features, resulting in the loss of spatial detail information in the transformation results. Summary of the Invention

[0005] To address the above problems, this invention proposes an unsupervised change detection method for hyperspectral images based on tensor convolution structures.

[0006] The technical solution of this invention is: an unsupervised change detection method for hyperspectral images based on tensor convolution structure, comprising the following steps:

[0007] S1, Construct a training tensor for the dual-temporal hyperspectral image;

[0008] S2. Perform Tucker decomposition on the training tensor to determine the spectral rank of the spectral mode factor matrix;

[0009] S3. Determine the layer-by-layer stacking structure based on the spectral rank of the spectral mode factor matrix and output the dual-temporal depth feature map;

[0010] S4. Based on the dual-temporal depth feature map, obtain the independent spectral difference map and the local spatial average map;

[0011] S5. Based on the obtained independent spectral difference map and local spatial average map, the final variation map is obtained.

[0012] Furthermore, S1 includes the following sub-steps:

[0013] S11. Construct a three-dimensional local patch set for each pixel of the dual-temporal hyperspectral image;

[0014] S12. Randomly select several patches from the three-dimensional local patch set as the training patch set;

[0015] S13. Stack the training patch set along the sample dimension to construct the training tensor.

[0016] Furthermore, in S11, the three-dimensional local patch set The expression is:

[0017] ;

[0018] ;

[0019] in, Represents the real number field. This represents the side length of the local spatial patch, that is, the size of the local window that is truncated in the spatial dimension. Indicates spectral dimension, This represents the total number of 3D local patches extracted from the entire hyperspectral image. This indicates the spatial dimension (number of pixel rows) of the input hyperspectral image in the vertical direction. This indicates the spatial dimension (number of pixel columns) of the input hyperspectral image in the horizontal direction.

[0020] Furthermore, in S2, the spectral rank of the spectral mode factor matrix The expression is:

[0021] ;

[0022] in, This represents a predefined energy threshold used to control the retention ratio of the principal components of the spectrum. Indicates the preceding The cumulative energy ratio of each effective spectral component Indicates the number of effective spectral components retained. This means that the cumulative energy ratio is not lower than the threshold. The minimum spectral rank determined under the given conditions is used to construct the spectral mode factor matrix.

[0023] Furthermore, S3 includes the following sub-steps:

[0024] S31. Perform Tucker decomposition on the Patch set corresponding to the dual-temporal hyperspectral images, and obtain the first-level spectral projection matrix based on the spectral rank.

[0025] S32. Project the first layer spectral projection matrix to obtain the first layer output image;

[0026] S33. Based on the output diagram of the first layer, determine the output diagram of each layer;

[0027] S34. Based on the layer-by-layer stacking of the output maps of each layer, output a dual-temporal depth feature map.

[0028] Furthermore, in S32, the size of the first layer output map The expression is:

[0029] ;

[0030] in, This represents the hyperspectral image of phase t. This represents the first-level spectral projection matrix. Represents tensor multiplication. The time phase is indicated, with t=1 corresponding to the first time phase image and t=2 corresponding to the second time phase image.

[0031] In S33, the first Layer output image size The expression is:

[0032] ;

[0033] in, This represents the hyperspectral image of phase t. This represents the first-level spectral projection matrix. Represents tensor multiplication. Indicates phase, Corresponding to the first phase image, Corresponding to the second phase image;

[0034] In S34, the expression for the dual-temporal depth feature map is:

[0035] ;

[0036] ;

[0037] in, This represents the final depth feature map corresponding to the first phase hyperspectral image. This represents the final depth feature map corresponding to the second phase hyperspectral image. Indicates the first phase image The depth feature representation obtained after layer-by-layer feature extraction from the layer tensor convolutional structure. Indicates the second phase image The deep feature representation obtained after layer-by-layer feature extraction from the layer tensor convolution structure.

[0038] Furthermore, S4 includes the following sub-steps:

[0039] S41. Based on the dual-temporal depth feature map, calculate the spectral difference at each spatial location;

[0040] S42. Based on the spectral differences at each spatial location, calculate the spectral differences around each spatial location. Local average within the window;

[0041] S43. Based on the spectral differences at each spatial location and the surrounding environment The local average within the window is linearly normalized to obtain the independent spectral difference map and the local spatial average map.

[0042] Furthermore, in S41, the spectral differences in spatial location The expression is:

[0043] ;

[0044] in, The horizontal coordinate representing spatial location. The vertical coordinate represents the spatial location. Indicates the spatial location of the first phase hyperspectral image. The corresponding depth feature vector, which is expanded along the spectral dimension. Indicates the spatial location of the second phase hyperspectral image. The corresponding depth feature vector, which is expanded along the spectral dimension. Represents the L2 norm;

[0045] In S42, the spatial location around Local average within the window The expression is:

[0046] ;

[0047] in, This represents the size parameters of a local window, used to determine its spatial position. Centered Spatial neighborhood range This represents the relative offset index of the local window in the horizontal direction. This represents the relative offset index of the local window in the vertical direction. Indicates spatial location Spectral difference values ​​at the location.

[0048] Furthermore, S5 includes the following sub-steps:

[0049] S51. The independent spectral difference map and the local spatial average map are fused to obtain the joint variance;

[0050] S52. Based on the joint variance, obtain the optimal threshold pair;

[0051] S53. Based on the relationship between the independent spectral difference map and the local spatial average map and the optimal threshold pair, the final change map is obtained.

[0052] Furthermore, in S51, joint variance The expression is:

[0053] ;

[0054] in, This represents the discrimination threshold parameter set for independent spectral difference views. This represents the discrimination threshold parameter set for a local spatial view. This represents the inter-class variance from independent spectral difference views. Represents the inter-class variance of a local spatial view;

[0055] In S52, the expression for the optimal threshold pair is:

[0056] ;

[0057] in, Indicates the joint variance The optimal spectral difference threshold for achieving the maximum value Indicates the joint variance The optimal local space threshold at which the maximum value is obtained;

[0058] In S53, the final change diagram is shown. The expression is:

[0059] ;

[0060] in, Indicates the difference in spectra After normalization or enhancement, in spatial location The spectral change response value obtained at the location, Represents the average variation of local space. After normalization or enhancement processing, in spatial location The spatial change response value obtained at the location.

[0061] The beneficial effects of this invention are:

[0062] (1) By fully preserving the three-dimensional tensor structure of hyperspectral data, this invention avoids the loss of spectral and spatial coupling information caused by two-dimensional flattening in traditional methods, thereby more fully exploring the intrinsic characteristics of the data.

[0063] (2) The present invention adopts a completely unsupervised design and combines an adaptive rank estimation mechanism, which significantly reduces human intervention and does not rely on pseudo-labels or complex iterative processes, thus improving the stability and reliability of the model.

[0064] (3) The dual-view change detector constructed in this invention can effectively suppress noise interference and has good resistance to changes in illumination, seasonal fluctuations and system noise. Attached Figure Description

[0065] Figure 1 This is a flowchart of an unsupervised change detection method for hyperspectral images based on tensor convolution structures;

[0066] Figure 2 This is a schematic diagram of the Tucker tensor convolution structure;

[0067] Figure 3 This is a schematic diagram of a multi-layer shared-weight Tucker convolutional network architecture;

[0068] Figure 4 This is a schematic diagram of a dual-view change detector. Detailed Implementation

[0069] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0070] like Figure 1 As shown, this invention provides an unsupervised change detection method for hyperspectral images based on tensor convolution structures, comprising the following steps:

[0071] S1, Construct a training tensor for the dual-temporal hyperspectral image;

[0072] S2. Perform Tucker decomposition on the training tensor to determine the spectral rank of the spectral mode factor matrix;

[0073] S3. Determine the layer-by-layer stacking structure based on the spectral rank of the spectral mode factor matrix and output the dual-temporal depth feature map;

[0074] S4. Based on the dual-temporal depth feature map, obtain the independent spectral difference map and the local spatial average map;

[0075] S5. Based on the obtained independent spectral difference map and local spatial average map, the final variation map is obtained.

[0076] This invention proposes a multilayer hyperspectral unsupervised change detection network based on a Tucker tensor convolution structure. Unlike previous frameworks, this invention uses tensor analysis as the core of network construction, extracting meaningful spectral-spatial coupling features through multilayer stacking, thereby improving change detection accuracy. Furthermore, unlike deep learning models that require extensive hyperparameter tuning, the framework of this invention has lower sensitivity to hyperparameters, reducing the need for manual tuning.

[0077] In this embodiment of the invention, S1 includes the following sub-steps:

[0078] S11. Construct a three-dimensional local patch set for each pixel of the dual-temporal hyperspectral image;

[0079] S12. Randomly select several patches from the three-dimensional local patch set as the training patch set;

[0080] S13. Stack the training patch set along the sample dimension to construct the training tensor.

[0081] In this embodiment of the invention, in S11, the three-dimensional local patch set The expression is:

[0082] ;

[0083] ;

[0084] in, Represents the real number field. This represents the side length of the local spatial patch, that is, the size of the local window that is truncated in the spatial dimension. Indicates spectral dimension, This represents the total number of 3D local patches extracted from the entire hyperspectral image. This indicates the spatial dimension (number of pixel rows) of the input hyperspectral image in the vertical direction. This indicates the spatial dimension (number of pixel columns) of the input hyperspectral image in the horizontal direction.

[0085] from Randomly selected from each patch One as the training patch set This is used for subsequent model training; among which, This indicates the first training patch randomly selected from the entire patch set. This indicates that the second training patch is randomly selected from the entire patch set. This indicates that the first patch is randomly selected from the entire patch set. One training patch.

[0086] Stack the training patches along the sample dimension to construct a third-order tensor. ,in, For the sample size, For local spatial dimensions, For spectral dimensions.

[0087] In this embodiment of the invention, in S2, the spectral rank of the spectral mode factor matrix... The expression is:

[0088] ;

[0089] in, This represents a predefined energy threshold used to control the retention ratio of the principal components of the spectrum. Indicates the preceding The cumulative energy ratio of each effective spectral component Indicates the number of effective spectral components retained. This means that the cumulative energy ratio is not lower than the threshold. The minimum spectral rank determined under the given conditions is used to construct the spectral mode factor matrix.

[0090] For training tensors The goal of performing Tucker decomposition is to find the core tensor. and three factor matrices , so that:

[0091] ;

[0092] in, For the core tensor, For the sample pattern factor matrix, This is the spatial pattern factor matrix. This is the spectral mode factor matrix;

[0093] Since the training patches already contain local spatial information, and the stacking of patches enhances nonlocal associations, this invention fixes the spatial mode rank as follows:

[0094] ;

[0095] ;

[0096] in, Let represent the rank of the first mode in the tensor decomposition, and represent the spatial dimension of the stacked training patches. This represents the rank of the second mode in tensor decomposition, which is the spatial dimension of a patch. That is, no dimensionality reduction is performed on the spatial dimension.

[0097] Spectral Rank For adaptive estimation, consider the spectral mode expansion matrix as follows:

[0098] ;

[0099] Perform SVD decomposition on it:

[0100] ;

[0101] in, Representation matrix The left singular vector matrix obtained by performing singular value decomposition. Represents a matrix The right singular vector matrix obtained by performing singular value decomposition. Represents a matrix The singular value diagonal matrix obtained by performing singular value decomposition has diagonal elements that are the corresponding singular values.

[0102] Singular value sequences can be obtained. :

[0103] ;

[0104] in, This represents the first singular value in the singular value sequence, corresponding to the principal component with the highest energy. This represents the second singular value in the singular value sequence, corresponding to the second largest principal component. Represents the singular value sequence of the th There are singular values, among which This represents the total number of non-zero singular values;

[0105] Based on singular values, the cumulative energy ratio is defined. :

[0106] ;

[0107] in, Represents the first singular value in the sequence. Each singular value corresponds to a magnitude of energy in that direction of the spectral mode expansion matrix. This represents the total number of non-zero singular values ​​in the singular value sequence.

[0108] Therefore, the spectral rank is determined. The top [rank] in SVD is then [determined]. The eigenvectors are used as the projection matrix. :

[0109] ;

[0110] in, Indicates from matrix Select column 1 to column 2 from all rows A submatrix composed of columns, which is formed by the previous columns. It consists of eigenvectors corresponding to each singular value.

[0111] This matrix is ​​used to perform spectral dimensionality reduction and construct Tucker convolution kernels.

[0112] Apply the spectral projection matrix to the spectral dimension of each patch:

[0113] ;

[0114] in, The original hyperspectral image, It is a Tucker tensor convolution structure. This is the result after convolution.

[0115] In this embodiment of the invention, S3 includes the following sub-steps:

[0116] S31. Perform Tucker decomposition on the Patch set corresponding to the dual-temporal hyperspectral images, and obtain the first-level spectral projection matrix based on the spectral rank.

[0117] S32. Project the first layer spectral projection matrix to obtain the first layer output image;

[0118] S33. Based on the output diagram of the first layer, determine the output diagram of each layer;

[0119] S34. Based on the layer-by-layer stacking of the output maps of each layer, output a dual-temporal depth feature map.

[0120] By stacking Tucker convolutions layer by layer, the output of the previous layer becomes the input of the next layer (concatenated structure). The spectral dimension is compressed or remapped layer by layer, and finally a deeper and more stable feature representation is extracted. This structure constitutes a Siamese Tucker convolutional network with multiple weights shared.

[0121] Network stacking After layers, the final output is a dual-temporal depth feature representation:

[0122] In this embodiment of the invention, in S32, the size of the first layer output image The expression is:

[0123] ;

[0124] in, This represents the hyperspectral image of phase t. This represents the first-level spectral projection matrix. Represents tensor multiplication. Indicates phase, Corresponding to the first phase image, Corresponding to the second phase image;

[0125] In S33, the first Layer output image size The expression is:

[0126] ;

[0127] in, Indicates the first The size of the layer output image, Indicates the first Layer spectral projection matrix;

[0128] In S34, the expression for the dual-temporal depth feature map is:

[0129] ;

[0130] ;

[0131] in, This represents the final depth feature map corresponding to the first phase hyperspectral image. This represents the final depth feature map corresponding to the second phase hyperspectral image. Indicates the first phase image The depth feature representation obtained after layer-by-layer feature extraction from the layer tensor convolutional structure. Indicates the second phase image The deep feature representation obtained after layer-by-layer feature extraction from the layer tensor convolution structure.

[0132] In this embodiment of the invention, S4 includes the following sub-steps:

[0133] S41. Based on the dual-temporal depth feature map, calculate the spectral difference at each spatial location;

[0134] S42. Based on the spectral differences at each spatial location, calculate the spectral differences around each spatial location. Local average within the window;

[0135] S43. Based on the spectral differences at each spatial location and the surrounding environment The local average within the window is linearly normalized to obtain the independent spectral difference map and the local spatial average map.

[0136] In this embodiment of the invention, in S41, the spectral differences of spatial location The expression is:

[0137] ;

[0138] in, The horizontal coordinate representing spatial location. The vertical coordinate represents the spatial location. Indicates the spatial location of the first phase hyperspectral image. The corresponding depth feature vector, which is expanded along the spectral dimension. Indicates the spatial location of the second phase hyperspectral image. The corresponding depth feature vector, which is expanded along the spectral dimension. Represents the L2 norm;

[0139] This is a pixel-level intensity map, reflecting the spatiotemporal feature differences between two phases in the deep feature space.

[0140] In S42, to enhance spatial consistency and suppress isolated noise, at position Surrounding Calculate the local average within the window. (Spatial location surroundings) Local average within the window The expression is:

[0141] ;

[0142] in, The size parameter represents the local window, used to determine the center position (x, y). Spatial neighborhood range This represents the relative offset index of the local window in the horizontal direction. This represents the relative offset index of the local window in the vertical direction. Indicates spatial location Spectral difference values ​​at the location.

[0143] Linearly normalize the two views to obtain the independent spectral difference plot: Local spatial average plot: This ensures that their numerical range remains consistent, facilitating threshold selection. Among other things, This represents an independent spectral difference map, used to describe the intensity of spectral changes in two-temporal images at various spatial locations. This represents the spectral change response map obtained after normalizing the independent spectral difference map 𝑅. This represents a spatial variation statistical graph calculated from the average of the local neighborhood. Represents the local spatial average plot The spatial variation response map obtained after normalization.

[0144] For any view (e.g.) ), set threshold The pixels were then divided into "changed" and "unchanged" classes, and the inter-class variance was... for:

[0145] ;

[0146] It characterizes the threshold. The separability between the two categories of "changed / unchanged". Among them, The pixel ratio to the left of the threshold The proportion of pixels to the right of the threshold; , This is the mean of the two classes.

[0147] In this embodiment of the invention, S5 includes the following sub-steps:

[0148] S51. The independent spectral difference map and the local spatial average map are fused to obtain the joint variance;

[0149] S52. Based on the joint variance, obtain the optimal threshold pair;

[0150] S53. Based on the relationship between the independent spectral difference map and the local spatial average map and the optimal threshold pair, the final change map is obtained.

[0151] In this embodiment of the invention, in S51, the joint variance The expression is:

[0152] ;

[0153] in, This represents the discrimination threshold parameter set for independent spectral difference views. This represents the discrimination threshold parameter set for a local spatial view. This represents the inter-class variance from independent spectral difference views. Represents the inter-class variance of a local spatial view;

[0154] In S52, the expression for the optimal threshold pair is:

[0155] ;

[0156] in, Indicates the joint variance The optimal spectral difference threshold for achieving the maximum value Indicates the joint variance The optimal local space threshold at which the maximum value is obtained;

[0157] In S53, the final change diagram is shown. The expression is:

[0158] ;

[0159] in, Indicates the difference in spectra After normalization or enhancement, in spatial location The spectral change response value obtained at the location, Represents the average variation of local space. After normalization or enhancement processing, in spatial location The spatial change response value obtained at the location.

[0160] like Figure 2 The diagram illustrates the Tucker tensor convolution feature extraction process in this invention. First, multiple features of varying sizes are extracted from the input hyperspectral image using a sliding window approach in the spatial dimension. Three-dimensional local patches are generated to form a set of three-dimensional local patches. Then, these three-dimensional patches are stacked along a certain dimension of the space to construct a training tensor for subsequent tensor decomposition modeling. Tucker decomposition is performed on the stacked training tensor. During the decomposition process, an adaptive estimation mechanism for the spectral rank is introduced, determining the spectral rank based on the singular value energy accumulation ratio of the spectral mode expansion matrix. This process yields the core tensor and its corresponding factor matrix. The spatial mode factor remains unchanged in dimensionality, while only the spectral mode is compressed and modeled. The resulting spectral factor matrix is ​​used as the spectral projection matrix. Modal multiplication is then performed along the spectral dimension on the original hyperspectral image or the previously obtained feature tensor to achieve Tucker tensor convolution feature mapping. This maps the original high-dimensional spectral features to a new low-dimensional spectral subspace, enabling efficient representation of the spatial-spectral joint features of the hyperspectral image and providing stable input for subsequent change detection or feature analysis.

[0161] like Figure 3 The diagram illustrates the overall architecture of the dual-temporal hyperspectral change detection based on Tucker tensor convolution of this invention. The overall network adopts a dual-branch shared structure, extracting features from two temporal hyperspectral images, including T1 and T2 phases, and performing change discrimination in a high-level feature space. At the input end, the dual-temporal hyperspectral images are fed into the network as two separate inputs. Each branch consists of cascaded layers of Tucker tensor convolution modules. Each layer performs tensor convolution operations on the input features along the spectral dimension, achieving layer-by-layer mapping and compression of spatial-spectral joint features. Through hierarchical stacking, the network can progressively extract multi-scale feature representations from local spectral information to high-level semantic structures from shallow to deep layers. The two branches use the same Tucker tensor convolution structure and parameter form, ensuring that the dual-temporal images are represented in the same feature space and avoiding spurious changes introduced due to inconsistent feature distributions. At the end of the network, the depth features output from the two branches are input into the dual-view change detector. Through dual-view feature comparison and discrimination, the final dual-temporal change detection result image is generated.

[0162] like Figure 4 The diagram illustrates the processing flow of the dual-view change detection method in this invention. This method analyzes change information from two complementary perspectives: a spectral difference perspective and a spatial statistical perspective, and obtains the final change result through a fusion strategy. First, in the independent spectral difference view, features at the same spatial location in both temporal images are compared pixel-by-pixel to obtain a difference distribution map reflecting the intensity of change, namely feature map F1 and feature map F2. This view primarily characterizes the degree of change in spectral features of each pixel, highlighting potential change areas. Second, in the local spatial average view, spatial statistics are performed on the spectral difference results within the neighborhood of each pixel to obtain the local average change response. This includes the view... Figure 1 Independent spectral difference plots and visual Figure 2The local spatial average map is used. This view incorporates spatial neighborhood information, making the changing areas more spatially coherent while effectively suppressing interference from isolated noise points. Based on this, a dual-view fusion strategy is employed, comprehensively utilizing the discriminative information from the spectral difference view and the local spatial average view to jointly determine changing and non-changing areas. Only locations showing significant changes in both views are ultimately identified as changing areas, thus avoiding false positives or false negatives caused by a single view.

[0163] The following description is based on specific embodiments.

[0164] Assuming a given two-phase hyperspectral image The input size is, for example, :

[0165] S1: Get Patch Size For each pixel, obtain The local tensor is normalized by removing the mean, and tensor X is formed by randomly sampling Ns=10000 patches.

[0166] S2: Constructing a third-order Patch tensor Perform higher-order singular value decomposition to automatically determine the spectral rank R3:

[0167] ;

[0168] in The entire image is obtained by performing Tucker projection. The dimensionality-reduced spectral-spatial coupling feature map was obtained.

[0169] S3: Configure a three-layer Siamese network with 3×3 kernels, shared weights between layers, and process the two temporal projection maps separately to output depth feature maps. .

[0170] S4: Calculate the characteristic Euclidean distance to obtain the spectral characteristic difference R; within a 3×3 neighborhood, calculate the local spatial analysis μ to measure the consistency of homogeneous regions.

[0171] S5: Employ a two-class variance maximization strategy to obtain the threshold. The final binary image is generated by fusing two thresholds.

[0172] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for unsupervised change detection in hyperspectral images based on tensor convolution structure, characterized in that, Includes the following steps: S1, Construct a training tensor for the dual-temporal hyperspectral image; S2. Perform Tucker decomposition on the training tensor to determine the spectral rank of the spectral mode factor matrix; S3. Determine the layer-by-layer stacking structure based on the spectral rank of the spectral mode factor matrix and output the dual-temporal depth feature map; S4. Based on the dual-temporal depth feature map, obtain the independent spectral difference map and the local spatial average map; S5. Based on the obtained independent spectral difference map and local spatial average map, the final variation map is obtained; S5 includes the following sub-steps: S51. The independent spectral difference map and the local spatial average map are fused to obtain the joint variance; S52. Based on the joint variance, obtain the optimal threshold pair; S53. Based on the relationship between the independent spectral difference map and the local spatial average map and the optimal threshold pair, the final change map is obtained; In S51, the joint variance The expression is: ； in, This represents the discrimination threshold parameter set for independent spectral difference views. This represents the discrimination threshold parameter set for a local spatial view. This represents the inter-class variance from independent spectral difference views. Represents the inter-class variance of a local spatial view; In S52, the expression for the optimal threshold pair is: ； in, Indicates the joint variance The optimal spectral difference threshold for achieving the maximum value. Indicates the joint variance The optimal local space threshold at which the maximum value is obtained; In S53, the final change diagram is represented as follows: The expression is: ； in, Indicates the difference in spectra After normalization or enhancement processing, in spatial location The spectral change response value obtained at the location, Represents the average variation of local space. After normalization or enhancement processing, in spatial location The spatial change response value obtained at the location.

2. The unsupervised change detection method for hyperspectral images based on tensor convolution structure according to claim 1, characterized in that, S1 includes the following sub-steps: S11. Construct a three-dimensional local patch set for each pixel of the dual-temporal hyperspectral image; S12. Randomly select several patches from the three-dimensional local patch set as the training patch set; S13. Stack the training patch set along the sample dimension to construct the training tensor.

3. The unsupervised change detection method for hyperspectral images based on tensor convolution structure according to claim 2, characterized in that, In S11, the three-dimensional local patch set The expression is: ； ； in, Represents the real number field. This represents the side length of the local space patch. Indicates spectral dimension, This represents the total number of 3D local patches extracted from the entire hyperspectral image. This indicates the spatial dimensions of the input hyperspectral image in the vertical direction. This indicates the spatial dimensions of the input hyperspectral image in the horizontal direction.

4. The unsupervised change detection method for hyperspectral images based on tensor convolution structure according to claim 1, characterized in that, In S2, the spectral rank of the spectral mode factor matrix The expression is: ； in, This represents a predefined energy threshold. Indicates the preceding The cumulative energy ratio of each effective spectral component This represents the minimum spectral rank determined under the condition that the cumulative energy ratio is not lower than the threshold τ.

5. The unsupervised change detection method for hyperspectral images based on tensor convolution structure according to claim 1, characterized in that, S3 includes the following sub-steps: S31. Perform Tucker decomposition on the Patch set corresponding to the dual-temporal hyperspectral images, and obtain the first-level spectral projection matrix based on the spectral rank. S32. Project the first layer spectral projection matrix to obtain the first layer output image; S33. Based on the output diagram of the first layer, determine the output diagram of each layer; S34. Based on the layer-by-layer stacking of the output maps of each layer, output a dual-temporal depth feature map.

6. The unsupervised change detection method for hyperspectral images based on tensor convolution structure according to claim 5, characterized in that, In S32, the size of the first layer output image The expression is: ； in, Indicates phase Hyperspectral imagery This represents the first-level spectral projection matrix. Represents tensor multiplication; In S33, the first Layer output image size The expression is: ； in, Indicates the first The size of the layer output image. Indicates the first Layer spectral projection matrix; In S34, the expression for the dual-temporal depth feature map is: ； ； in, This represents the final depth feature map corresponding to the first phase hyperspectral image. This represents the final depth feature map corresponding to the second phase hyperspectral image. Indicates the first phase image The depth feature representation obtained after layer-by-layer feature extraction from the layer tensor convolutional structure. Indicates the second phase image The deep feature representation obtained after layer-by-layer feature extraction from the layer tensor convolution structure.

7. The unsupervised change detection method for hyperspectral images based on tensor convolution structure according to claim 1, characterized in that, S4 includes the following sub-steps: S41. Based on the dual-temporal depth feature map, calculate the spectral difference at each spatial location; S42. Based on the spectral differences at each spatial location, calculate the spectral differences around each spatial location. Local average within the window; S43. Based on the spectral differences at each spatial location and the surrounding environment The local average within the window is linearly normalized to obtain the independent spectral difference map and the local spatial average map.

8. The unsupervised change detection method for hyperspectral images based on tensor convolution structure according to claim 7, characterized in that, In S41, the spectral differences at spatial locations The expression is: ； in, The horizontal coordinate representing spatial location. The vertical coordinate represents the spatial location. This represents the depth feature vector corresponding to the first phase hyperspectral image at spatial location (x, y). This represents the depth feature vector corresponding to the spatial location (x, y) in the second time-phase hyperspectral image. Represents the L2 norm; In S42, the spatial location around Local average within the window The expression is: ； in, This represents the size parameters of a local window. This represents the relative offset index of the local window in the horizontal direction. This represents the relative offset index of the local window in the vertical direction. Indicates spatial location Spectral difference values ​​at the location.

Images