End-to-End Hyperspectral Image Change Detection Method and System Based on Unmixing

Abstract

The present invention relates to the field of remote sensing image processing, and discloses an end-to-end hyperspectral image change detection method and system based on unmixing. The method includes: obtaining dual-temporal hyperspectral images to be detected; subtracting the first-temporal hyperspectral image from the second-temporal hyperspectral image to obtain an image difference map to be detected; inputting the image difference map to be detected into a trained hyperspectral image change detection network to obtain a dual-temporal hyperspectral image change detection result. The trained hyperspectral image change detection network includes an unmixing sub-network and a change detection sub-network. The unmixing sub-network processes the image difference map to be detected to obtain an abundance difference map. The change detection sub-network processes the abundance difference map and the image difference map to be detected to obtain a dual-temporal hyperspectral image change detection result, significantly improving the change detection accuracy.

Description

Technical Field

[0001] The present invention relates to the technical fields of remote sensing image processing and computer vision, and particularly to an end-to-end hyperspectral image change detection method and system based on unmixing. Background Art

[0002] Hyperspectral remote sensing image change detection technology identifies the temporal evolution law of surface cover by analyzing multi-temporal imaging data of the same area. Different from conventional multi-spectral remote sensing image change detection, this technology makes full use of the rich spectral information of hyperspectral remote sensing images, can capture subtle differences in material composition, and has unique value in the identification of micro-changes such as vegetation succession and pollutant migration. In the existing technical system, the change detection method based on spectral unmixing effectively improves the ability to identify complex surface changes by analyzing the proportion of ground objects in mixed pixels. Its technical principle is as follows: each pixel of a hyperspectral image can be characterized as a linear or non-linear combination of several endmember spectra. By extracting the difference in endmember abundances in two multi-temporal images, the temporal change law of surface material composition can be revealed, overcoming the misjudgment problem of mixed pixels in traditional methods.

[0003] There are two key problems in the prior art: First, the change detection framework based on unmixing generally adopts a staged processing mode, that is, first extract the abundance map of the remote sensing image, and then perform change detection on the abundance map. This split processing method makes the unmixing process lack the constraint of the change detection result, resulting in small errors in the unmixing process being amplified in the change detection process, causing error accumulation and affecting the change detection result. Second, most change detection methods based on unmixing only use abundance features for discrimination, ignoring the complementary information contained in the original hyperspectral image, which limits the improvement of change detection accuracy. Summary of the Invention

[0004] To solve the deficiencies of the prior art, the present invention provides an end-to-end hyperspectral image change detection method and system based on unmixing; combine the unmixing network and the change detection network to form an end-to-end network, so that they can be co-trained. Among them, the unmixing network is responsible for extracting the abundance map, and the change detection network is responsible for fusing the information of the abundance map and the original image for change detection.

[0005] On the one hand, an end-to-end hyperspectral image change detection method based on unmixing is provided, including:

[0006] Obtain multi-temporal hyperspectral images to be detected; the multi-temporal hyperspectral images include: the first-temporal hyperspectral image and the second-temporal hyperspectral image;

[0007] Subtract the first-temporal hyperspectral image from the second-temporal hyperspectral image to obtain an image difference map to be detected;

[0008] Input the image difference map to be detected into the trained hyperspectral image change detection network to obtain the dual-temporal hyperspectral image change detection result. The trained hyperspectral image change detection network includes an unmixing sub-network and a change detection sub-network. The unmixing sub-network processes the image difference map to be detected to obtain an abundance difference map. The change detection sub-network processes the abundance difference map and the image difference map to be detected to obtain the dual-temporal hyperspectral image change detection result.

[0009] On the other hand, an end-to-end hyperspectral image change detection system based on unmixing is provided, including:

[0010] An acquisition module configured to acquire dual-temporal hyperspectral images to be detected. The dual-temporal hyperspectral images include a first-temporal hyperspectral image and a second-temporal hyperspectral image.

[0011] A subtraction module configured to subtract the first-temporal hyperspectral image from the second-temporal hyperspectral image to obtain an image difference map to be detected.

[0012] Input the image difference map to be detected into the trained hyperspectral image change detection network to obtain the dual-temporal hyperspectral image change detection result. The trained hyperspectral image change detection network includes an unmixing sub-network and a change detection sub-network. The unmixing sub-network processes the image difference map to be detected to obtain an abundance difference map. The change detection sub-network processes the abundance difference map and the image difference map to be detected to obtain the dual-temporal hyperspectral image change detection result.

[0013] The above technical solution has the following advantages or beneficial effects:

[0014] Through the integrated design of the unmixing network and the change detection network, dual-task collaborative optimization is achieved. The unmixing process directly serves the change detection task, solving the error accumulation problem in traditional step-by-step methods. Secondly, an information fusion module is designed to dynamically fuse the original hyperspectral image information and the abundance information of the abundance map using the cross-attention mechanism, alleviating the limitations of single-modal data. This solution enables the abundance estimation to automatically adapt to the change detection requirements through end-to-end training, significantly improving the change detection accuracy and providing a more advanced solution for change detection in complex mixed pixel scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The accompanying drawings forming a part of this invention are used to provide a further understanding of the invention. The schematic embodiments and descriptions thereof of the invention are used to explain the invention and do not constitute an improper limitation of the invention.

[0016] Figure 1 It is a flowchart of the method for the first embodiment. DETAILED DESCRIPTION

[0017] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the technical field to which the present invention belongs.

[0018] Embodiment 1

[0019] This embodiment provides an end-to-end hyperspectral image change detection method based on unmixing;

[0020] As Figure 1 shown, the end-to-end hyperspectral image change detection method based on unmixing includes:

[0021] S101: Obtain the dual-temporal hyperspectral images to be detected; the dual-temporal hyperspectral images include: the first-temporal hyperspectral image and the second-temporal hyperspectral image;

[0022] S102: Subtract the first-temporal hyperspectral image from the second-temporal hyperspectral image to obtain the image difference map to be detected;

[0023] S103: Input the image difference map to be detected into the trained hyperspectral image change detection network to obtain the dual-temporal hyperspectral image change detection result; wherein, the trained hyperspectral image change detection network includes: an unmixing sub-network and a change detection sub-network; the unmixing sub-network processes the image difference map to be detected to obtain the abundance difference map; the change detection sub-network processes the abundance difference map and the image difference map to be detected to obtain the dual-temporal hyperspectral image change detection result.

[0024] Further, in S101: Obtain the dual-temporal hyperspectral images to be detected, which is collected by an imaging spectrometer.

[0025] Further, the dual-temporal hyperspectral images include: the first-temporal hyperspectral image and the second-temporal hyperspectral image, wherein the acquisition time of the first-temporal hyperspectral image is earlier than that of the second-temporal hyperspectral image.

[0026] Further, S102: Subtract the first-temporal hyperspectral image from the second-temporal hyperspectral image to obtain the image difference map to be detected, including:

[0027]

[0028] wherein, represents taking the absolute value, represents the image difference map to be detected, represents the first-temporal hyperspectral image, represents the second-temporal hyperspectral image.

[0029] Further, in the actual application stage, the unmixing sub-network processes the image difference map to be detected to obtain the abundance difference map, specifically including:

[0030] (11) Input the image difference map to be detected into the unmixing sub-network for unmixing processing to obtain the abundance difference map ; The specific process is as follows: First, extract image features and reduce the image dimension through a two-dimensional convolution with a convolution kernel size of 3 ×3, then further constrain its data range through the ReLU activation function to obtain the intermediate features . Among them, the number of channels of the image difference map is C, and after convolution processing, the number of image channels is reduced to C / / 2, where / / represents integer division;

[0031] ;

[0032] (12) After obtaining the intermediate features , perform a two-dimensional convolution with a convolution kernel size of 3 ×3 again to further reduce its dimension so that its dimension is the same as the number of endmembers P; after convolution, the number of image channels is reduced from C / / 2 to P; then, constrain the output through the Softmax activation function to ensure that the value of each pixel in the abundance difference map is between 0 and 1, and obtain the abundance difference map ; ;

[0033] .

[0034] Further, in the actual application stage, the change detection sub-network processes the abundance difference map and the image difference map to be detected to obtain the change detection result of the two-temporal hyperspectral image, including:

[0035] (11): The change detection sub-network respectively extracts features from the image difference map and the abundance difference map to obtain the image difference Figure 1 secondary processed features and the abundance difference Figure 1 secondary processed features;

[0036] (12): Respectively extract features from the image difference Figure 1 secondary processed features and the abundance difference Figure 1 secondary processed features to obtain the image difference map secondary processed features and the abundance difference map secondary processed features;

[0037] (13): Use the cross-attention module to process the image difference map secondary processed features and the abundance difference map secondary processed features to obtain the fused features;

[0038] (14) Classify the fused features to obtain the final classification result.

[0039] Further, the (11): The change detection sub-network respectively extracts features from the image difference map and the abundance difference map to obtain the image difference Figure 1 secondary processed feature and the abundance difference Figure 1 secondary processed feature, specifically including:

[0040] The input of the change detection sub-network is the image difference map and the abundance difference map , which respectively contain C and P image channels. Since and have different dimensions, which is not convenient for subsequent feature extraction and fusion. It is necessary to align the dimensions of and . The specific process is as follows: The image difference map passes through a 2D convolution with a convolution kernel size of 1 × 1, the image size remains unchanged, and the number of channels is reduced from C to 64, obtaining the image difference Figure 1 secondary processed feature ; The abundance difference map passes through a 2D convolution with a convolution kernel size of 1 × 1, the image size remains unchanged, and the number of channels is increased from P to 64, obtaining the abundance difference Figure 1 secondary processed feature :

[0041] ;

[0042] ;

[0043] Among them, represents the two-dimensional convolution operation using a 1 × 1 convolution kernel, which is used to adjust the number of channels of the input feature map. In this way, the present invention ensures and are consistent in the number of channels, providing a unified feature representation for subsequent feature extraction and fusion operations.

[0044] Further, the (12): Respectively extract features from the image difference Figure 1 secondary processed feature and the abundance difference Figure 1 secondary processed feature to obtain the image difference map secondary processed feature and the abundance difference map secondary processed feature, specifically including:

[0045] After dimension alignment, use a two-dimensional convolution with a convolution kernel size of 3 × 3 to and extract features to further obtain their spatial information. The specific process is as follows: After a two-dimensional convolution, the image size and the number of channels remain unchanged, and the secondary processed features of the image difference map are obtained. ; After a two-dimensional convolution, the image size and the number of channels remain unchanged, and the secondary processed features of the abundance difference map are obtained. :

[0046]

[0047]

[0048] Among them, represents a two-dimensional convolution operation using a 3 × 3 convolution kernel. In this way, the present invention not only ensures and are consistent in the number of channels, but also further enhances their spatial feature representation, providing richer information for subsequent feature fusion and change detection.

[0049] Furthermore, the (13): adopts a cross-attention module to process the secondary processed features of the image difference map and the secondary processed features of the abundance difference map to obtain fused features, specifically including:

[0050] The secondary processed features of the image difference map have rich spatial and spectral information, while the secondary processed features of the abundance difference map have clearer boundaries and less noise. The two both contain similar structures and have complementary information. Therefore, using the cross-attention module (Cross Attention), the complementary information in is fused into to obtain the fused feature . The specific process is as follows. First, is input into the fully connected layer Linear1 to calculate the query vector (Query); is input into the fully connected layer Linear2 and the fully connected layer Linear3 to calculate the key vector (Key) and the value vector (Value) respectively for subsequent calculation of the attention distribution. The above process can be formulated as:

[0051]

[0052] Then, the query vector Q is multiplied by the transpose of the value vector K and passed through the SoftMax activation function to obtain and the attention distribution of the complementary information between . The specific calculation formula is as follows:

[0053]

[0054] Among them, is the scaling factor of the feature dimension to maintain numerical stability.

[0055] Attention distribution for obtaining complementary information After that, by multiplying with the value vector and multiplying, the second - processed feature of the image difference map and the second - processed feature of the abundance difference map obtain the complementary information between them , and the specific process can be formulated as:

[0056]

[0057] Among them, represents matrix multiplication;

[0058] Finally, add the complementary information to the second - processed feature of the image difference map to obtain the final fused feature , and the above process can be formulated as:

[0059]

[0060] In this way, the present invention not only retains the spatial and spectral information of the image difference map, but also incorporates the complementary information between it and the abundance difference map, which helps to improve the accuracy and robustness of change detection.

[0061] Furthermore, the (14) classifies the fused feature to obtain the final classification result, including:

[0062] After obtaining the fused feature , use the classification layer to classify each pixel in the image. Since the final result of change detection needs to divide the pixels into two categories (i.e., changed pixels and unchanged pixels), it is necessary to reduce the dimension of the feature through the pooling layer and the fully - connected layer to adapt to the subsequent classification process. The above process can be formulated as:

[0063]

[0064] Among them, functions to perform global average pooling on the feature map to aggregate global information; represents the fully - connected layer, which is used to adjust the feature dimension.

[0065] Further, in step S103: input the image difference map to be detected into the trained hyperspectral image change detection network to obtain the change detection result of the two-temporal hyperspectral image; wherein, the training process of the trained hyperspectral image change detection network includes:

[0066] Construct a dataset, where the dataset is a two-phase hyperspectral image with known change detection results; subtract the two-phase hyperspectral images and take the absolute value of the subtraction result to obtain a hyperspectral image difference map; perform endmember extraction on the hyperspectral image difference map to obtain endmembers; divide the dataset into a training set and a validation set;

[0067] Input the training set into the hyperspectral image change detection network to train the network. When the value of the total loss function of the network no longer decreases, stop training to obtain the trained hyperspectral image change detection network; use the validation set to verify the trained hyperspectral image change detection network. If the verification passes, obtain the finally trained hyperspectral image change detection network; if the verification fails, replace the dataset and retrain;

[0068] The total loss function is the sum of the cosine loss function and the cross-entropy loss function;

[0069] Assume that the training process consists of a total of 2N rounds of training. During the first N rounds of training, the numerical decrease rate of the cross-entropy loss function is slower than that of the cosine loss function; during the last N rounds of training, the numerical decrease rate of the cosine loss function is slower than that of the cross-entropy loss function.

[0070] Further, the step of subtracting the two-phase hyperspectral images and taking the absolute value of the subtraction result to obtain a hyperspectral image difference map includes:

[0071]

[0072] wherein, denotes taking the absolute value, denotes the hyperspectral image difference map, denotes the third-temporal hyperspectral image, denotes the fourth-temporal hyperspectral image.

[0073] Further, the step of performing endmember extraction on the hyperspectral image difference map to obtain endmembers is implemented by using the vertex component analysis algorithm.

[0074] It should be understood that first subtract the two-temporal hyperspectral images X1 and X2 and take the absolute value to obtain the hyperspectral image difference map. The specific calculation formula is 。Then, the endmembers E of the difference map are extracted by the Vertex Component Analysis (VCA) algorithm and saved for input to the subsequent network. VCA is an unsupervised endmember extraction algorithm based on convex geometry theory. It identifies spectral features representing pure land cover classes from hyperspectral images by iteratively finding the maximum projection vector.

[0075] It should be understood that the present invention crops the two-temporal hyperspectral images into image patches of size 7 × 7 as the input to the network. Small patches can avoid interference from weakly correlated pixels while ensuring rich spatial information is included. Then, 9% of the data is randomly selected for training, 1% for validation, and the remaining data for testing.

[0076] Furthermore, during the training process, the unmixing sub-network processes the image difference map to obtain the abundance difference map, specifically including:

[0077] (21) Input the image difference map into the unmixing sub-network for unmixing processing to obtain the abundance difference map ; The specific process is as follows: First, two-dimensional convolution with a convolution kernel size of 3 × 3 is used to extract image features and reduce the image dimension. Then, the ReLU activation function is used to further constrain its data range to obtain intermediate features . Among them, the number of channels of the image difference map is C, and after convolution processing, the number of image channels is reduced to C / / 2, where / / represents integer division;

[0078] ;

[0079] (22) After obtaining the intermediate features , perform two-dimensional convolution with a convolution kernel size of 3 × 3 again to further reduce its dimension to be the same as the number of endmembers P; after convolution, the number of image channels is reduced from C / / 2 to P; then, the Softmax activation function is used to constrain the output to ensure that the value of each pixel in the abundance difference map is between 0 and 1, obtaining the abundance difference map ;

[0080] .

[0081] Finally, the abundance difference map is multiplied by the extracted endmember E to obtain the reconstructed image difference map , thereby calculating the cosine similarity loss function between the image difference map and the reconstructed image difference map ​ , the above process can be formulated as:

[0082]

[0083] Cosine similarity loss function , and the specific expression is as follows:

[0084]

[0085] Where represents the abundance difference map matrix, represents the endmember matrix, represents matrix multiplication, represents the th element value in the reconstructed image spectral vector, represents the th element value in the original image spectral vector, and C represents the vector dimension.

[0086] Furthermore, during the training process, the change detection sub-network processes the abundance difference map and the image difference map to obtain the change detection results of the two-temporal hyperspectral images, including:

[0087] The input of the change detection sub-network is the image difference map and the abundance difference map , which contain C and P image channels respectively. Since and have different dimensions, which is not convenient for subsequent feature extraction and fusion. It is necessary to align the dimensions of and . The specific process is as follows: The image difference map passes through a 2D convolution with a kernel size of 1 ×1, and the image size remains unchanged while the number of channels is reduced from C to 64, obtaining the image difference Figure 1 after the first processing feature ; The abundance difference map passes through a 2D convolution with a kernel size of 1 ×1, and the image size remains unchanged while the number of channels is increased from P to 64, obtaining the abundance difference Figure 1 after the first processing feature :

[0088] ;

[0089] ;

[0090] Where represents the two-dimensional convolution operation using a 1 ×1 convolution kernel, which is used to adjust the number of channels of the input feature map. In this way, the present invention ensures that and Consistency in the number of channels provides a unified feature representation for subsequent feature extraction and fusion operations;

[0091] After dimension alignment, a two-dimensional convolution with a convolution kernel size of 3 3 is used to and perform feature extraction to further obtain their spatial information. The specific process is as follows: After passing through a two-dimensional convolution, the image size and the number of channels remain unchanged, and the second-processed feature of the image difference map is obtained ; After passing through a two-dimensional convolution, the image size and the number of channels remain unchanged, and the second-processed feature of the abundance difference map is obtained :

[0092]

[0093]

[0094] Among them, represents the two-dimensional convolution operation using a 3 3 convolution kernel. In this way, the present invention not only ensures and the consistency in the number of channels, but also further enhances their spatial feature representation, providing richer information for subsequent feature fusion and change detection.

[0095] After the above processing, the second-processed feature of the image difference map has rich spatial and spectral information, while the second-processed feature of the abundance difference map has clearer boundaries and less noise. The two both contain similar structures and have complementary information. Therefore, using the Cross Attention module, the complementary information in is fused into to obtain the fused feature . The specific process is as follows. First, is input into the fully connected layer Linear1 to calculate the query vector (Query); is input into the fully connected layer Linear2 and the fully connected layer Linear3 to calculate the key vector (Key) and the value vector (Value) respectively for subsequent calculation of the attention distribution. The above process can be formulated as:

[0096]

[0097] Then, the query vector Q is multiplied by the transpose of the value vector K and passed through the SoftMax activation function to obtain Attention distribution of complementary information between and . The specific calculation formula is as follows:

[0098]

[0099] where is the scaling factor of the feature dimension to maintain numerical stability.

[0100] After obtaining the attention distribution of complementary information , by multiplying with the value vector , the secondary processed feature of the image difference map and the secondary processed feature of the abundance difference map are multiplied to obtain the complementary information between them. The specific process can be formulated as:

[0101]

[0102] where represents matrix multiplication;

[0103] Finally, the complementary information is added to the secondary processed feature of the image difference map to obtain the final fused feature . The above process can be formulated as:

[0104]

[0105] In this way, the present invention not only retains the spatial and spectral information of the image difference map, but also incorporates the complementary information between it and the abundance difference map, which helps to improve the accuracy and robustness of change detection.

[0106] After obtaining the fused feature , a classification layer is used to classify each pixel in the image. Since the final result of change detection needs to divide the pixels into two categories (i.e., changed pixels and unchanged pixels), it is necessary to reduce the dimension of the features through a pooling layer and a fully connected layer to adapt to the subsequent classification process. The above process can be formulated as:

[0107]

[0108] where is used to perform global average pooling on the feature map to aggregate global information; represents the fully connected layer, which is used to adjust the feature dimension.

[0109] Finally, through the Softmax activation function, the prediction label of whether each pixel has changed is obtained.

[0110]

[0111] The cross-entropy loss function is adopted, which is used to measure the difference between the predicted label and the true label. The expression of the cross-entropy loss function is as follows:

[0112]

[0113] Where represents the number of labeled samples, and respectively represent the true label and the predicted label of the -th sample.

[0114] Cosine loss function is added to the cross-entropy loss function to obtain the loss function of the final network , and the expression is as follows:

[0115]

[0116] In the initial stage of network training, has a large value and plays a dominant role in network training, and the convergence speed of the unmixing sub-network is relatively fast. In the middle and late stages of network training, the unmixing sub-network tends to be stable, has a small decline in value, plays a dominant role, the convergence speed of the change detection sub-network is accelerated, and the unmixing sub-network will be fine-tuned under the common constraints of the loss functions and to further improve the unmixing effect, make the abundance maps of the invariant regions more similar, and the abundance maps of the changed regions more different. Thus, the collaborative training of the unmixing sub-network and the change detection sub-network is realized.

[0117] Taking the hyperspectral image difference map as the input, first extract the endmembers through the Vertex Component Analysis (VCA) algorithm, and then obtain the abundance difference map by using the unmixing sub-network of the network. Finally, send the image difference map and the abundance difference map into the change detection sub-network of the network to obtain the final change detection result. In the change detection sub-network, the convolution operation is responsible for aligning the feature dimensions and extracting the spatial information, and the cross-attention module is responsible for fusing the information of the image difference map and the abundance difference map. This method can effectively overcome the limitations of using a single feature for change detection, and realizes the collaborative training of unmixing and change detection through an end-to-end network, effectively improving the accuracy and flexibility of hyperspectral change detection.

[0118] Embodiment 2

[0119] This embodiment provides an end-to-end hyperspectral image change detection system based on unmixing, including:

[0120] An acquisition module, which is configured to: acquire a dual-temporal hyperspectral image to be detected; the dual-temporal hyperspectral image includes: a first-temporal hyperspectral image and a second-temporal hyperspectral image;

[0121] A subtraction module, which is configured to: subtract the first-temporal hyperspectral image from the second-temporal hyperspectral image to obtain a difference image to be detected;

[0122] Input the difference image to be detected into a trained hyperspectral image change detection network to obtain a dual-temporal hyperspectral image change detection result; wherein, the trained hyperspectral image change detection network includes: an unmixing sub-network and a change detection sub-network; the unmixing sub-network processes the difference image to be detected to obtain an abundance difference image; the change detection sub-network processes the abundance difference image and the difference image to be detected to obtain a dual-temporal hyperspectral image change detection result.

[0123] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and changes. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. An end-to-end hyperspectral image change detection method based on unmixing, characterized in that Including: Obtain a dual-temporal hyperspectral image to be detected; The dual-temporal hyperspectral image includes a first-temporal hyperspectral image and a second-temporal hyperspectral image; Subtract the first-temporal hyperspectral image from the second-temporal hyperspectral image to obtain an image difference map to be detected; Input the image difference map to be detected into a trained hyperspectral image change detection network to obtain a dual-temporal hyperspectral image change detection result; wherein, the trained hyperspectral image change detection network includes an unmixing sub-network and a change detection sub-network; the unmixing sub-network processes the image difference map to be detected to obtain an abundance difference map, specifically including: (11) Input the image difference map to be detected into the unmixing sub-network for unmixing processing to obtain the abundance difference map ; The specific process is as follows: First, extract image features and reduce the image dimension through a two-dimensional convolution with a convolution kernel size of 3 × 3, and then further constrain its data range through the ReLU activation function to obtain intermediate features , where the number of channels of the image difference map is C, and after convolution processing, the number of image channels is reduced to C / / 2, where / / represents integer division; ； (12) Obtain intermediate features After that, perform two-dimensional convolution again with a convolution kernel size of 3×3 to further reduce its dimension so that its dimension is the same as the number of endmembers P; after convolution, the number of image channels is reduced from C / / 2 to P; then, the output is constrained by the Softmax activation function to ensure that the value of each pixel in the abundance difference map is between 0 and 1, obtaining the abundance difference map ; ； The change detection sub-network processes the abundance difference map and the image difference map to be detected to obtain a dual-temporal hyperspectral image change detection result, including: (11): The change detection sub-network respectively extracts features from the image difference map and the abundance difference map to obtain a first-processed feature of the image difference map and a first-processed feature of the abundance difference map; (12): Respectively extract features from the first-processed feature of the image difference map and the first-processed feature of the abundance difference map to obtain a second-processed feature of the image difference map and a second-processed feature of the abundance difference map; (13): Use a cross-attention module to process the second-processed feature of the image difference map and the second-processed feature of the abundance difference map to obtain a fused feature; (14) Classify the fused feature to obtain a final classification result.

2. The end-to-end hyperspectral image change detection method based on unmixing according to claim 1, wherein Subtracting the first-temporal hyperspectral image from the second-temporal hyperspectral image to obtain an image difference map to be detected includes: ； Among them, represents taking the absolute value, represents the image difference map to be detected, represents the hyperspectral image of the first phase, represents the hyperspectral image of the second phase.

3. The end-to-end hyperspectral image change detection method based on unmixing according to claim 1, wherein, The change detection sub-network respectively extracts features from the image difference map and the abundance difference map to obtain a first-processed feature of the image difference map and a first-processed feature of the abundance difference map, specifically including: The input of the change detection sub-network is the image difference map and the abundance difference map , which contain C and P image channels respectively. Dimension alignment is performed on and as follows: The image difference map is passed through a 2D convolution with a kernel size of 1 ×1. After that, the image size remains unchanged, and the number of channels is reduced from C to 64, obtaining the first processed feature of the image difference map ; The abundance difference map is passed through a 2D convolution with a kernel size of 1 ×1. After that, the image size remains unchanged, and the number of channels is increased from P to 64, obtaining the first processed feature of the abundance difference map : ； ； Among them, represents a two-dimensional convolution operation using 1 1 convolution kernel, which is used to adjust the number of channels of the input feature map.

4. The end-to-end hyperspectral image change detection method based on unmixing according to claim 3, characterized in that Respectively extract features from the first-processed feature of the image difference map and the first-processed feature of the abundance difference map to obtain a second-processed feature of the image difference map and a second-processed feature of the abundance difference map, specifically including: After dimension alignment, two-dimensional convolution with a convolution kernel size of 3 × 3 is used for and feature extraction. The specific process is as follows: After one two-dimensional convolution, the image size and the number of channels remain unchanged, and the secondary processed feature of the image difference map is obtained ; After one two-dimensional convolution, the image size and the number of channels remain unchanged, and the secondary processed feature of the abundance difference map is obtained : ； ； Among them, represents a two-dimensional convolution operation using a 3 × 3 convolution kernel.

5. The end-to-end hyperspectral image change detection method based on unmixing according to claim 4, characterized in that, Use a cross-attention module to process the second-processed feature of the image difference map and the second-processed feature of the abundance difference map to obtain a fused feature, specifically including: Using a cross-attention module, fuse the complementary information in into to obtain the fused feature ; The specific process is as follows. First, input into the fully connected layer Linear1 to calculate the query vector; input into the fully connected layer Linear2 and the fully connected layer Linear3 to calculate the key vector and the value vector respectively for the subsequent calculation of the attention distribution: ； Then, the query vector Q is multiplied by the transpose of the value vector K and passed through the SoftMax activation function to obtain the attention distribution of the complementary information between , and the specific calculation formula is as follows: ； Among them, is the scaling factor of the feature dimension; Attention distribution for obtaining complementary information After that, by multiplying with the value vector the second - processed feature of the image difference map is obtained by multiplication and the second - processed feature of the abundance difference map the complementary information between them : ； Among them, represents matrix multiplication; Finally, the complementary information is added to the secondary processing feature of the image difference map to obtain the final fusion feature : 。 6. The end-to-end hyperspectral image change detection method based on unmixing according to claim 5, characterized in that, Classifying the fused feature to obtain a final classification result, including: After obtaining the fused features the features are dimensionally reduced through a pooling layer and a fully connected layer: ； Among them, Its function is to perform global average pooling on the feature map to aggregate global information; Represents a fully connected layer used to adjust the feature dimension.

7. The end-to-end hyperspectral image change detection method based on unmixing according to claim 1, wherein Input the image difference map to be detected into a trained hyperspectral image change detection network to obtain a dual-temporal hyperspectral image change detection result; wherein, the training process of the trained hyperspectral image change detection network includes: Construct a data set, the data set is a dual-phase hyperspectral image with known change detection results; subtract the dual-phase hyperspectral image and take the absolute value of the subtraction result to obtain a hyperspectral image difference map; perform endmember extraction on the hyperspectral image difference map to obtain endmembers; divide the data set into a training set and a validation set; Input the training set into the hyperspectral image change detection network to train the network. When the total loss function value of the network no longer decreases, stop training to obtain a trained hyperspectral image change detection network; use the validation set to validate the trained hyperspectral image change detection network. If the validation passes, obtain the finally trained hyperspectral image change detection network; if the validation fails, replace the data set and retrain; The total loss function is the sum of the cosine loss function and the cross-entropy loss function; Assume that the training process consists of 2N rounds of training. During the first N rounds of training, the numerical decrease rate of the cross-entropy loss function is slower than that of the cosine loss function. During the last N rounds of training, the numerical decrease rate of the cosine loss function is slower than that of the cross-entropy loss function; Cosine similarity loss function , and the specific expression is as follows: ； Among them, represents the abundance difference map matrix, represents the endmember matrix, represents matrix multiplication, represents the value of the -th element in the reconstructed image spectral vector, represents the value of the -th element in the original image spectral vector, and C represents the vector dimension; The expression of the cross-entropy loss function is as follows: ； Among them, represents the number of labeled samples, and respectively represent the true label and the predicted label of the th sample.

8. An end-to-end hyperspectral image change detection system based on unmixing, characterized in that, including: An acquisition module, which is configured to acquire a dual-temporal hyperspectral image to be detected; The dual-temporal hyperspectral image includes a first-temporal hyperspectral image and a second-temporal hyperspectral image; A subtraction module, which is configured to subtract the first-temporal hyperspectral image from the second-temporal hyperspectral image to obtain an image difference map to be detected; Input the image difference map to be detected into the trained hyperspectral image change detection network to obtain the dual-temporal hyperspectral image change detection result. Among them, the trained hyperspectral image change detection network includes an unmixing sub-network and a change detection sub-network. The unmixing sub-network processes the image difference map to be detected to obtain an abundance difference map, specifically including: (11) Input the image difference map to be detected into the unmixing sub-network for unmixing processing to obtain the abundance difference map ; The specific process is as follows: First, extract image features through a two-dimensional convolution with a convolution kernel size of 3 × 3 and reduce the image dimension, and then further constrain its data range through the ReLU activation function to obtain intermediate features , where the number of channels of the image difference map is C, and after convolution processing, the number of image channels is reduced to C / / 2, where / / represents integer division; ； (12) Obtain intermediate features After that, perform two-dimensional convolution again with a convolution kernel size of 3×3 to further reduce its dimension so that its dimension is the same as the number of endmembers P; after convolution, the number of image channels is reduced from C / / 2 to P; then use the Softmax activation function to constrain the output to ensure that the value of each pixel in the abundance difference map is between 0 and 1, and obtain the abundance difference map ; ； The change detection sub-network processes the abundance difference map and the image difference map to be detected to obtain the dual-temporal hyperspectral image change detection result, including: (11): The change detection sub-network respectively extracts features from the image difference map and the abundance difference map to obtain the first-processed features of the image difference map and the first-processed features of the abundance difference map; (12): Respectively extract features from the first-processed features of the image difference map and the first-processed features of the abundance difference map to obtain the second-processed features of the image difference map and the second-processed features of the abundance difference map; (13): Use a cross-attention module to process the second-processed features of the image difference map and the second-processed features of the abundance difference map to obtain fused features; (14) Classify the fused features to obtain the final classification result.

Images