Hyperspectral remote sensing image change detection method and system based on multi-temporal unmixing

Abstract

The present invention relates to the technical field of remote sensing image change detection, and in particular to a hyperspectral remote sensing image change detection method and system based on multi-temporal unmixing, wherein the method comprises: obtaining a first remote sensing image and a second remote sensing image; the first remote sensing image and the second remote sensing image are hyperspectral remote sensing images collected at different time points in the same geographical area; extracting common end members of the first remote sensing image and the second remote sensing image; inputting the first remote sensing image, the second remote sensing image and the common end members into a trained multi-temporal unmixing network to obtain a first abundance map and a second abundance map; inputting the first abundance map and the second abundance map into a trained change detection network to obtain a change detection result between the two remote sensing images. The method can make full use of unmixing information for change detection, improve detection accuracy, and reduce computational complexity.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image change detection, and in particular to a hyperspectral remote sensing image change detection method and system based on multi-temporal unmixing. Background Art

[0002] Hyperspectral remote sensing image change detection is to detect changes in surface cover by analyzing hyperspectral images acquired at different times in the same geographic area. Since hyperspectral images have rich spectral information, they can distinguish different objects with fine spectral features and provide more accurate change detection capabilities than traditional RGB or multispectral data. Compared with traditional methods based on spectral or spatial features, unmixing-based change detection methods can provide more physically meaningful information on the composition of objects and improve the interpretability and robustness of detection.

[0003] Hyperspectral unmixing is an important task in the analysis of hyperspectral remote sensing data. Its goal is to decompose the observed mixed pixels into several end members (representing the spectrum of pure objects) and their corresponding abundance (indicating the proportion of each end member in the pixel). Due to the complexity of the surface landscape and the high-dimensional characteristics of hyperspectral data, pixel mixing is ubiquitous, especially in low spatial resolution or complex surface scenes. It is often difficult to accurately distinguish the changes of different objects by simply relying on spectral information for change detection. Therefore, by extracting abundance information through unmixing, the composition of objects and their temporal changes can be characterized at the physical level, thereby improving the accuracy of change detection.

[0004] However, traditional hyperspectral unmixing methods are usually performed on single-phase images, that is, the hyperspectral images of each phase are unmixed separately, and then the abundance map obtained by unmixing is used for change detection. This method does not consider the temporal information of the image in the unmixing stage, which may lead to inconsistent abundance extraction, thereby affecting the accuracy of change detection. In addition, existing change detection methods based on unmixing usually adopt the method of directly calculating the difference between the characteristics of the two phases, and lack of fine modeling of the change area. This method is difficult to fully utilize the information of the abundance map, resulting in blurred boundaries of the change area and affecting the accurate positioning of the change pixels. When faced with remote sensing scenes with complex land object categories and diverse change scales, existing methods often find it difficult to accurately extract change information, thereby affecting detection accuracy. Therefore, how to combine temporal information for unmixing and efficiently use abundance maps for change detection is a key challenge in current research. Summary of the invention

[0005] In order to solve the deficiencies of the prior art, the present invention provides a hyperspectral remote sensing image change detection method and system based on multi-temporal unmixing; the method first extracts end members through the VCA algorithm, and then uses the multi-temporal unmixing network to obtain the abundance map of different phases. Subsequently, the change detection network is used to extract features from the dual-phase abundance map, and the information of the change area is enhanced through the differential attention mechanism and the differential image branch, so as to obtain the final change detection result. This method can make full use of the unmixing information for change detection, improve the detection accuracy, and reduce the computational complexity.

[0006] On the one hand, a hyperspectral remote sensing image change detection method based on multi-temporal unmixing is provided;

[0007] Hyperspectral remote sensing image change detection method based on multi-temporal unmixing, including:

[0008] Acquire a first remote sensing image and a second remote sensing image; the first remote sensing image and the second remote sensing image are hyperspectral remote sensing images collected at different time points in the same geographical area;

[0009] For the first remote sensing image and the second remote sensing image, common end members of the two remote sensing images are extracted;

[0010] Inputting the first remote sensing image, the second remote sensing image and the common end member into the trained multi-temporal unmixing network to obtain a first abundance map and a second abundance map;

[0011] The first abundance map and the second abundance map are input into the trained change detection network to obtain the change detection result between the two remote sensing images.

[0012] On the other hand, a hyperspectral remote sensing image change detection system based on multi-temporal unmixing is provided;

[0013] Hyperspectral remote sensing image change detection system based on multi-temporal unmixing, including:

[0014] An acquisition module is configured to: acquire a first remote sensing image and a second remote sensing image; the first remote sensing image and the second remote sensing image are hyperspectral remote sensing images collected at different time points in the same geographical area;

[0015] An extraction module is configured to: extract common end members of the first remote sensing image and the second remote sensing image;

[0016] A processing module is configured to: input the first remote sensing image, the second remote sensing image and the common end member into a trained multi-temporal unmixing network to obtain a first abundance map and a second abundance map;

[0017] The output module is configured to: input the first abundance map and the second abundance map into the trained change detection network to obtain a change detection result between the two remote sensing images.

[0018] The above technical solution has the following advantages or beneficial effects:

[0019] A hyperspectral image change detection method based on multi-temporal unmixing is provided, which overcomes the problem of insufficient utilization of spatiotemporal information of hyperspectral images by existing unmixing algorithms, makes the unmixing network more suitable for change detection tasks, and improves the accuracy of change detection. By combining multi-temporal unmixing technology, the present invention makes full use of abundance information to characterize the category of objects and their change patterns, and effectively suppresses the interference of spectral mixing on the detection results. At the same time, the differential attention MDA mechanism is introduced to enhance the feature expression of the changed area, and by adding differential branches, the abundance map information is fully utilized to achieve accurate capture of image changes. Finally, the present invention not only improves the accuracy of change detection, but also effectively reduces the impact of spectral mixing on detection performance, providing a more reliable and effective technical solution for change detection of hyperspectral remote sensing images. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0021] Figure 1 This is a flow chart of the method of embodiment 1. DETAILED DESCRIPTION

[0022] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0023] Embodiment 1

[0024] This embodiment provides a hyperspectral remote sensing image change detection method based on multi-temporal unmixing;

[0025] like Figure 1 As shown, the hyperspectral remote sensing image change detection method based on multi-temporal unmixing includes:

[0026] S101: Acquire a first remote sensing image and a second remote sensing image; the first remote sensing image and the second remote sensing image are hyperspectral remote sensing images collected at different time points in the same geographical area;

[0027] S102: extracting common end members of the first remote sensing image and the second remote sensing image;

[0028] S103: inputting the first remote sensing image, the second remote sensing image and the common end member into the trained multi-temporal unmixing network to obtain a first abundance map and a second abundance map;

[0029] S104: Inputting the first abundance map and the second abundance map into the trained change detection network to obtain a change detection result between the two remote sensing images.

[0030] The present invention uses an unsupervised multi-temporal unmixing network to extract abundance information and combines it with a change detection network to extract change areas. This method can effectively improve the edge integrity and internal consistency of change detection, reduce noise interference, and improve the accuracy of change detection, which is of great significance to the fields of land use monitoring, ecological environment assessment, and disaster detection.

[0031] Furthermore, the spectral resolution is Spectral images within this order of magnitude range are called hyperspectral images.

[0032] Furthermore, the first remote sensing image and the second remote sensing image are acquired by using a hyperspectral imager.

[0033] Furthermore, the S102: extracting common endmembers of the first remote sensing image and the second remote sensing image; wherein, endmember refers to a pure spectrum containing only one type of land feature; land feature refers to a real-world object observed in a hyperspectral image; for example, in a hyperspectral image of a farmland, land feature information such as crops and soil is included. If the spectrum of a certain pixel comes only from crops or soil, its spectral feature can be regarded as the corresponding endmember.

[0034] Furthermore, the step S102: extracting common end members of the first remote sensing image and the second remote sensing image includes:

[0035] (1-1): The first remote sensing image and the second remote sensing image are spliced ​​along the width dimension to obtain a three-dimensional data cube;

[0036] (1-2): Input the three-dimensional data cube into the vertex component analysis algorithm to obtain the common end members between the first remote sensing image and the second remote sensing image.

[0037] The advantages of the above technical solution are: extracting a set of endmember spectra with cross-phase stability from the extended space-time dimension, and establishing a reliable spectral reference library for subsequent precise unmixing.

[0038] It should be understood that for the first remote sensing image and the second remote sensing image, the common endmembers of the two remote sensing images are extracted, including: firstly, the dual-phase hyperspectral images (size is C×H×W, where C is the number of spectral channels; H and W are the image height and width respectively) are spliced ​​along the spatial width (W) dimension to construct an enhanced three-dimensional data cube (size is C×H×2W) with joint characterization capability. This splicing strategy breaks through the dimensional limitation of traditional single-phase endmember extraction, so that when the Vertex Component Analysis (VCA) algorithm is used later, the common endmembers in the two-phase data can be mined simultaneously. Through the orthogonal subspace projection and maximum projection vector iteration mechanism, VCA extracts the endmember spectrum set with cross-phase stability from the extended space-time dimension, and establishes a reliable spectral reference library for subsequent precise unmixing.

[0039] Further, S103: inputting the first remote sensing image, the second remote sensing image and the common end member into the trained multi-temporal unmixing network to obtain a first abundance map and a second abundance map, wherein the trained multi-temporal unmixing network is used to:

[0040] (2-1): First, the first remote sensing image and the second remote sensing image of different time phases are spliced ​​according to the time dimension to form a spatiotemporal data cube;

[0041] (2-2): For the spatiotemporal data cube, two three-dimensional convolution operations are used in sequence to extract the spatiotemporal information, and the spatiotemporal information is separated to obtain the first spatiotemporal feature and the second spatiotemporal feature;

[0042] (2-3): The first spatiotemporal feature is subjected to dimensionality reduction processing by three two-dimensional convolutional layers in sequence to obtain a first dimensionality reduction processing result, and then the value of each pixel of the first dimensionality reduction processing result is constrained to obtain a first abundance map;

[0043] (2-4): The second spatiotemporal feature is subjected to dimensionality reduction processing of three two-dimensional convolutional layers in sequence to obtain a second dimensionality reduction processing result, and then the value of each pixel of the second dimensionality reduction processing result is constrained to obtain a second abundance map.

[0044] (2-5): multiplying the first abundance map by the common end member to obtain a first reconstructed image; constructing a first loss function according to the difference between the first reconstructed image and the first remote sensing image;

[0045] Multiplying the second abundance map by the common end member to obtain a second reconstructed image; constructing a second loss function according to the difference between the second reconstructed image and the second remote sensing image;

[0046] The sum of the first loss function and the second loss function is the total loss function value of the multi-temporal unmixing network.

[0047] It should be understood that S103: inputting the first remote sensing image, the second remote sensing image and the common end member into the trained multi-temporal unmixing network to obtain the first abundance map and the second abundance map includes:

[0048] First, images of different phases are spliced ​​according to the time dimension to form a spatiotemporal data cube (size is C×T×H×W, where T is the time dimension); then the spatiotemporal information is extracted through 3D convolution operation to capture the changing characteristics of the objects in time and space. Finally, the extracted features are separated according to the time dimension and restored to the original size; the above process can be formulated as:

[0049]

[0050]

[0051] in, Indicates that images are spliced ​​along the time dimension, It means to separate the spliced ​​data and restore the original size;

[0052] After obtaining the spatiotemporal features, they are gradually reduced in dimension through multiple 2D convolution layers, where the number of input image channels is C, the number of output channels of the first 2D convolution is C / / 2, and the number of output channels of the second 2D convolution is C / / 4, where / / represents integer division, and finally the output with the number of channels P (number of end members) is obtained through the third 2D convolution; then, the output is processed by the Softmax activation function to ensure that the value of each pixel in the abundance map is between 0 and 1, satisfying the constraints of the abundance map. This process effectively combines spatiotemporal information and realizes high-precision abundance map generation. The above process can be formulated as:

[0053]

[0054] in, represents the abundance map corresponding to the input image at time T1, represents the abundance map corresponding to the input image at time T2, Represents the convolution kernel as 2D convolution of size, is the abundance map constraint;

[0055] Finally, the abundance map is multiplied by the common end member to reconstruct the input image , ,get , , and thus calculate the reconstruction loss , the calculation formula is as follows:

[0056]

[0057]

[0058]

[0059]

[0060]

[0061] in, , Represent the first abundance map and the second abundance map matrix respectively, represents the endmember matrix, represents matrix multiplication, , Represents the first OK The pixel value of the column, , Respectively represent the first OK The pixel value of the column.

[0062] Furthermore, the S104: inputs the first abundance map and the second abundance map into the trained change detection network to obtain a change detection result between the two remote sensing images.

[0063] The input to the change detection network is a 5×5 image patch.

[0064] Therefore, the present invention cuts the two abundance matrices of the dual-phase image into image patches of size 5 × 5 as input to extract difference features. Small image patches can ensure full utilization of local spatial information. Then, 1% of the data is randomly selected for training, 1% of the data is used for verification, and the rest of the data is used for testing.

[0065] Further, the S104: inputting the first abundance map and the second abundance map into the trained change detection network to obtain a change detection result between the two remote sensing images, wherein the trained change detection network includes:

[0066] Subtract the first abundance map from the second abundance map to obtain the abundance map difference map ;

[0067] Abundance difference map Perform convolution calculation Get the convolution result ;

[0068] ;

[0069] Adopt the first multi-temporal information fusion module The first abundance map and the second abundance map are fused, and the information of images at different times is fused through the attention mechanism, so as to significantly improve the expression of the changed area and obtain the first phase enhanced features. and the second phase once enhanced features ;

[0070] ;

[0071] ;

[0072] Enhance the features of the first phase and the second phase once enhanced features Perform subtraction processing to obtain the first subtraction result, and then add the first subtraction result to the convolution result Perform summation to obtain the first enhancement result ;

[0073] ;

[0074] The first enhancement result Perform convolution calculation Get the convolution result ;

[0075]

[0076] Adopt the second multi-temporal information fusion module Enhance the features of the first phase and the second phase once enhanced features Perform information fusion to obtain the secondary enhancement features of the first phase and the second phase secondary enhancement features ;

[0077] ;

[0078] ;

[0079] Secondary enhancement of the first phase and the second phase secondary enhancement features Perform subtraction processing to obtain the second subtraction result, and then add the second subtraction result to the convolution result. Perform summation to obtain the second enhancement result ;

[0080] ;

[0081] The second enhancement result Perform convolution calculation Get the convolution result ;

[0082]

[0083] Adopt the third multi-temporal information fusion module Secondary enhancement of the first phase and the second phase secondary enhancement features Perform information fusion to obtain the three enhanced features of the first phase and the third enhancement feature of the second phase ;

[0084] ;

[0085] ;

[0086] Three enhancement features for the first phase and the third enhancement feature of the second phase Perform subtraction processing to obtain the third subtraction result, and then add the third subtraction result to the convolution result. Perform summation to obtain the third enhancement result ;

[0087] ;

[0088] Finally, the features are fused through cascade splicing. In order to further extract high-level features and remove redundant information, the final change representation is extracted through convolution operation. , providing high-quality input for subsequent change detection:

[0089]

[0090] in Represents 1×1 convolution, which is used to further refine features and improve the discriminative ability of the model. Represents concatenating images by channel dimension.

[0091] It should be understood that change detection combines 3×3 convolution and multi-temporal information fusion module (MDA) to fully extract the change characteristics of multi-temporal abundance maps.

[0092] Furthermore, the trained change detection network further includes:

[0093] Representation of the final changes First perform a global average pooling operation, then perform a full connection operation to obtain the final feature ;

[0094] The final features , the activation function is used to process and obtain the change probability of each pixel;

[0095] The change probability of each pixel is compared with a set threshold to obtain a detection result, which includes: pixels that have changed and pixels that have not changed.

[0096] It should be understood that in order to obtain high quality variation characteristics After that, the classification layer is used to classify the changes at the pixel level. Since the final result of change detection needs to classify the pixels into two categories (i.e., changed pixels and unchanged pixels), it is necessary to reduce the dimension of the features through the pooling layer and the fully connected layer to adapt to the subsequent classification processing. The above process can be formulated as:

[0097]

[0098] in, The function is to perform global average pooling on the feature map to aggregate global information, while reducing the amount of calculation and improving the generalization ability of the model; Represents a fully connected layer, which is used to adjust the feature dimension.

[0099] Finally, the output is normalized using the activation function to obtain the change probability of each pixel.

[0100]

[0101] Furthermore, in the training process, in order to optimize the prediction ability of the model, the trained change detection network adopts a cross-entropy loss function to measure the difference between the prediction result and the actual change label. The expression of the cross-entropy loss function is as follows:

[0102]

[0103] in, represents the number of labeled samples, and Respectively represent The true labels and predicted labels of samples.

[0104] Furthermore, the first multi-temporal information fusion module MDA includes:

[0105] a first input terminal and a second input terminal, wherein the first input terminal is used to input a first input value; the second input terminal is used to input a second input value; the first input value of the first multi-temporal information fusion module MDA is a first abundance map; the second input value of the first multi-temporal information fusion module MDA is a second abundance map;

[0106] The first input value and the second input value are respectively subjected to depthwise separable convolution to obtain the query vector (Query), key vector (Key) and value vector (Value) for subsequent attention calculation. Specifically, the input features can be expressed as follows after depthwise separable convolution:

[0107]

[0108] in, and denote the query, key, and value of the first and second abundance graphs, respectively, Represents depth-wise separable convolution, which can reduce computational complexity and prevent overfitting.

[0109] Then, the difference between the query vector of the first abundance map and the second abundance map is calculated to extract the spatiotemporal variation information:

[0110]

[0111] in, It is input into the subsequent attention calculation module as a change-sensitive feature to enhance the representation ability of the changed area.

[0112] Then, the attention of the first abundance map and the second abundance map to the changed area is calculated respectively through the differential attention mechanism. The specific calculation method is as follows:

[0113]

[0114]

[0115] in, is the scaling factor of the feature dimension to maintain numerical stability.

[0116] By calculating the differential attention weight, we finally divide the attention weight into the value vector Multiply by , and get the enhanced change feature output:

[0117]

[0118] It should be understood that multi-temporal information fusion is implemented through the MDA module, which aims to extract the change information of multi-temporal remote sensing images and enhance the feature expression of the changed area through the differential attention mechanism. By constructing a spatiotemporal differential attention mechanism, the difference information between the features of the two temporal phases can be effectively modeled, thereby improving the accuracy of change detection.

[0119] Embodiment 2, this embodiment provides a hyperspectral remote sensing image change detection system based on multi-temporal unmixing, including: an acquisition module, which is configured to: acquire a first remote sensing image and a second remote sensing image; the first remote sensing image and the second remote sensing image are hyperspectral remote sensing images collected at different time points in the same geographical area;

[0120] An extraction module is configured to: extract common end members of the first remote sensing image and the second remote sensing image;

[0121] A processing module is configured to: input the first remote sensing image, the second remote sensing image and the common end member into a trained multi-temporal unmixing network to obtain a first abundance map and a second abundance map;

[0122] The output module is configured to: input the first abundance map and the second abundance map into the trained change detection network to obtain a change detection result between the two remote sensing images.

[0123] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A hyperspectral remote sensing image change detection method based on multi-temporal unmixing, characterized by: include: Acquire a first remote sensing image and a second remote sensing image; The first remote sensing image and the second remote sensing image are hyperspectral remote sensing images collected at different time points in the same geographical area; For the first remote sensing image and the second remote sensing image, common end members of the two remote sensing images are extracted; The first remote sensing image, the second remote sensing image and the common end member are input into the trained multi-temporal unmixing network to obtain a first abundance map and a second abundance map, wherein the trained multi-temporal unmixing network is used to: (2-1): First, the first remote sensing image and the second remote sensing image of different time phases are spliced ​​according to the time dimension to form a spatiotemporal data cube; (2-2): For the spatiotemporal data cube, two three-dimensional convolution operations are used in sequence to extract the spatiotemporal information, and the spatiotemporal information is separated to obtain the first spatiotemporal feature and the second spatiotemporal feature; (2-3): The first spatiotemporal feature is subjected to dimensionality reduction processing by three two-dimensional convolutional layers in sequence to obtain a first dimensionality reduction processing result, and then the value of each pixel of the first dimensionality reduction processing result is constrained to obtain a first abundance map; (2-4): The second spatiotemporal feature is subjected to dimensionality reduction processing of three two-dimensional convolutional layers in sequence to obtain a second dimensionality reduction processing result, and then the value of each pixel of the second dimensionality reduction processing result is constrained to obtain a second abundance map; The first abundance map and the second abundance map are input into the trained change detection network to obtain the change detection result between the two remote sensing images.

2. The hyperspectral remote sensing image change detection method based on multi-temporal unmixing as claimed in claim 1, characterized in that: For the first remote sensing image and the second remote sensing image, common end members of the two remote sensing images are extracted, including: (1-1): The first remote sensing image and the second remote sensing image are spliced ​​along the width dimension to obtain a three-dimensional data cube; (1-2): Input the three-dimensional data cube into the vertex component analysis algorithm to obtain the common end members between the first remote sensing image and the second remote sensing image.

3. The hyperspectral remote sensing image change detection method based on multi-temporal unmixing as claimed in claim 1, characterized in that: The method further includes: (2-5): multiplying the first abundance map by the common end member to obtain a first reconstructed image; constructing a first loss function according to the difference between the first reconstructed image and the first remote sensing image; Multiplying the second abundance map by the common end member to obtain a second reconstructed image; constructing a second loss function according to the difference between the second reconstructed image and the second remote sensing image; The sum of the first loss function and the second loss function is the total loss function value of the multi-temporal unmixing network.

4. The hyperspectral remote sensing image change detection method based on multi-temporal unmixing as claimed in claim 1, characterized in that: Inputting the first remote sensing image, the second remote sensing image and the common end member into the trained multi-temporal unmixing network to obtain a first abundance map and a second abundance map, including: First, images of different phases are spliced ​​according to the time dimension to form a spatiotemporal data cube; then, the spatiotemporal information is extracted through 3D convolution operation to capture the changing characteristics of the objects in time and space; finally, the extracted features are separated according to the time dimension and restored to the original size; the above process can be formulated as: ； ； in, Indicates that images are spliced ​​along the time dimension, It means to separate the spliced ​​data and restore the original size; After obtaining the spatiotemporal features, they are gradually reduced in dimension through multiple 2D convolution layers, where the number of input image channels is C, the number of output channels of the first 2D convolution is C / / 2, and the number of output channels of the second 2D convolution is C / / 4, where / / represents integer division, and finally the output with the number of channels P is obtained through the third 2D convolution; then, the output is processed by the Softmax activation function to ensure that the value of each pixel in the abundance map is between 0 and 1, satisfying the constraints of the abundance map; the above process can be formulated as: ； ； in, represents the abundance map corresponding to the input image at time T1, represents the abundance map corresponding to the input image at time T2, The convolution kernel is 2D convolution of size, is the abundance map constraint.

5. The hyperspectral remote sensing image change detection method based on multi-temporal unmixing as claimed in claim 4, characterized in that: The method further comprises: multiplying the abundance map with the common end member to reconstruct the input image , ,get , , and thus calculate the reconstruction loss , the calculation formula is as follows: ； ； ； ； ； in, , Represent the first abundance map and the second abundance map matrix respectively, represents the endmember matrix, represents matrix multiplication, , Represents the first OK The pixel value of the column, , Respectively represent the first OK The pixel value of the column.

6. The hyperspectral remote sensing image change detection method based on multi-temporal unmixing as claimed in claim 1, characterized in that: The first abundance map and the second abundance map are input into the trained change detection network to obtain the change detection result between the two remote sensing images, wherein the trained change detection network includes: Subtract the first abundance map from the second abundance map to obtain the abundance map difference map ; Abundance difference map Perform convolution calculation Get the convolution result ; ； Adopt the first multi-temporal information fusion module The first abundance map and the second abundance map are fused, and the information of images at different times is fused through the attention mechanism, so as to significantly improve the expression of the changed area and obtain the first phase enhanced features. and the second phase once enhanced features ; ； ； Enhance the features of the first phase and the second phase once enhanced features Perform subtraction processing to obtain the first subtraction result, and then add the first subtraction result to the convolution result Perform summation to obtain the first enhancement result ; ； The first enhancement result Perform convolution calculation Get the convolution result ; ； Adopt the second multi-temporal information fusion module Enhance the features of the first phase and the second phase once enhanced features Perform information fusion to obtain the secondary enhancement features of the first phase and the second phase secondary enhancement features ; ； ； Secondary enhancement of the first phase and the second phase secondary enhancement features Perform subtraction processing to obtain the second subtraction result, and then add the second subtraction result to the convolution result. Perform summation to obtain the second enhancement result ; ； The second enhancement result Perform convolution calculation Get the convolution result ; ； Adopt the third multi-temporal information fusion module Secondary enhancement of the first phase and the second phase secondary enhancement features Perform information fusion to obtain the three enhanced features of the first phase and the third enhancement feature of the second phase ; ； ； Three enhancement features for the first phase and the third enhancement feature of the second phase Perform subtraction processing to obtain the third subtraction result, and then add the third subtraction result to the convolution result. Perform summation to obtain the third enhancement result ; ； Finally, the features are fused through cascade splicing. In order to further extract high-level features and remove redundant information, the final change representation is extracted through convolution operation. , providing high-quality input for subsequent change detection: ； in Represents 1×1 convolution, which is used to further refine features and improve the discriminative ability of the model. Represents concatenating images by channel dimension.

7. The hyperspectral remote sensing image change detection method based on multi-temporal unmixing as claimed in claim 6, characterized in that: The trained change detection network further includes: Representation of the final changes First perform a global average pooling operation, then perform a full connection operation to obtain the final feature ; The final features , the activation function is used to process and obtain the change probability of each pixel; The change probability of each pixel is compared with a set threshold to obtain a detection result, which includes: pixels that have changed and pixels that have not changed.

8. The hyperspectral remote sensing image change detection method based on multi-temporal unmixing as claimed in claim 6, characterized in that: The first multi-temporal information fusion module includes: A first input terminal and a second input terminal, wherein the first input terminal is used to input a first input value; the second input terminal is used to input a second input value; the first input value of the first multi-temporal information fusion module is a first abundance map; the second input value of the first multi-temporal information fusion module is a second abundance map; The first input value and the second input value are respectively subjected to depthwise separable convolution to obtain the query vector, key vector and value vector for subsequent attention calculation; specifically, the input features can be expressed as follows after depthwise separable convolution: ； ； in, and denote the query, key, and value of the first and second abundance graphs, respectively, represents depthwise separable convolution; Then, the difference between the query vector of the first abundance map and the second abundance map is calculated to extract the spatiotemporal variation information: ； in, It is used as a change-sensitive feature and then fed into the subsequent attention calculation module to enhance the representation ability of the change area. Then, the attention of the first abundance map and the second abundance map to the changed area is calculated respectively through the differential attention mechanism. The specific calculation method is as follows: ； ； in, is the scaling factor of the feature dimension; By calculating the differential attention weight, we finally divide the attention weight into the value vector Multiply by , and get the enhanced change feature output: ; .

9. A hyperspectral remote sensing image change detection system based on multi-temporal unmixing, characterized by: include: An acquisition module is configured to: acquire a first remote sensing image and a second remote sensing image; The first remote sensing image and the second remote sensing image are hyperspectral remote sensing images collected at different time points in the same geographical area; An extraction module is configured to: extract common end members of the first remote sensing image and the second remote sensing image; A processing module is configured to: input the first remote sensing image, the second remote sensing image and the common end member into a trained multi-temporal unmixing network to obtain a first abundance map and a second abundance map, wherein the trained multi-temporal unmixing network is used to: (2-1): First, the first remote sensing image and the second remote sensing image of different time phases are spliced ​​according to the time dimension to form a spatiotemporal data cube; (2-2): For the spatiotemporal data cube, two three-dimensional convolution operations are used in sequence to extract the spatiotemporal information, and the spatiotemporal information is separated to obtain the first spatiotemporal feature and the second spatiotemporal feature; (2-3): The first spatiotemporal feature is subjected to dimensionality reduction processing by three two-dimensional convolutional layers in sequence to obtain a first dimensionality reduction processing result, and then the value of each pixel of the first dimensionality reduction processing result is constrained to obtain a first abundance map; (2-4): The second spatiotemporal feature is subjected to dimensionality reduction processing of three two-dimensional convolutional layers in sequence to obtain a second dimensionality reduction processing result, and then the value of each pixel of the second dimensionality reduction processing result is constrained to obtain a second abundance map; The output module is configured to: input the first abundance map and the second abundance map into the trained change detection network to obtain a change detection result between the two remote sensing images.

Images