Hyperspectral image domain adaptive change detection method, device and software product

The depth characteristics and difference characteristics of the hyperspectral image are extracted through the twin network and channel attention mechanism, and the characteristics are fusionized by the Kullback-Leibler divergence loss function, which solves the problem of feature distribution differences and overprediction in hyperspectral image change detection. Improved the accuracy and reliability of change detection.

CN118918458AActive Publication Date: 2024-11-08WUHAN UNIV
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202410903674.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-08
Publication Date
2024-11-08
Estimated Expiration
2044-07-08

AI Technical Summary

Technical Problem

The existing hyperspectral image change detection methods have different feature distribution differences and over-prediction capabilities when processing different phase images, and ignore the inherent differences between the original images, making it difficult to deeply explore subtle changes information, and lack of feature fusion capabilities.

Method used

A twin network is used to extract depth features and construct correlation alignment loss, combine the channel attention mechanism to enhance the difference characteristics, and promote feature fusion through the Kullback-Leibler divergence loss function, reduce feature differences, and improve the model's ability to identify subtle changes.

Benefits of technology

It effectively reduces the difference in feature distribution of hyperspectral images in different phases, enhances the ability to mine original image features, and improves the accuracy and reliability of the model's identification of changing information.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118918458A_ABST
    Figure CN118918458A_ABST
Patent Text Reader

Abstract

The invention discloses a hyperspectral image domain adaptive change detection method, equipment and a software product, and aims to reduce image distribution difference caused by imaging time change to the greatest extent by using correlation alignment loss so as to prevent a network from excessively capturing non-critical change. Meanwhile, a channel attention mechanism is also introduced to strengthen the feature expression of the difference image, so that the network is helped to find subtle changes more accurately. And finally, the features extracted by the two branches are fused by adopting Kullback-Leibler divergence loss, so that the accuracy of the network on change detection is further enhanced.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing image processing, and relates to a hyperspectral image domain adaptive change detection method, device and software product, and in particular to a hyperspectral image domain adaptive change detection integrating correlation alignment and spectral attention features. Background Art

[0002] Change detection identifies and analyzes changes in a specific area by comparing different data sets. It is often used in many fields such as disaster assessment, urban planning, and building damage assessment (References 1, 2). Hyperspectral images contain rich spectral information and can provide more decision support for detecting changing geographic features in remote sensing images. Therefore, the application of hyperspectral images in change detection is becoming more and more extensive, which helps to better understand and analyze the changes in geographic features.

[0003] Traditional change detection methods can be divided into three categories (References 3, 4): First, image algebra-based methods, which mainly analyze the spectral differences between pixels to effectively detect significant changes in remote sensing images; second, transformation-based methods, which capture and extract key changes in images by mapping the original remote sensing images to a specific feature space; and finally, classification-based methods, which mainly compare the differences in geographic feature labels in remote sensing images to accurately identify the changed areas. However, traditional change detection methods often focus too much on the extraction of shallow features, which to some extent limits the comprehensive and in-depth understanding and utilization of remote sensing image information.

[0004] Convolutional neural networks have attracted much attention due to their powerful ability to extract deep features, and have also been well applied in hyperspectral change detection (References 5, 6, 7). Change detection in hyperspectral images often identifies change information by comparing the differences between different data sets. However, due to factors such as imaging time and sensor differences, the resulting hyperspectral images often have differences in feature distribution between images, which greatly increases the complexity of change detection. Therefore, how to effectively reduce the feature distribution differences between hyperspectral images of different phases and reduce the model's ability to over-predict change information is the first key difficulty faced by this field. At the same time, most methods mainly focus on processing the feature differences of hyperspectral images of different phases, but ignore the inherent differences between the original images. Therefore, how to deeply explore the differences between the original images themselves to enhance the model's ability to detect subtle change information is the second key difficulty faced by this field. In addition, most networks need to be further strengthened in effectively fusing features extracted from different branches, which is the third key difficulty faced by this field.

[0005] References:

[0006] [1]Sicong Liu,Daniele Marinelli,Lorenzo Bruzzone,and FrancescaBovolo,”Areview of change detection in multitemporal hyperspectral images:Currenttechniques,applications,and challenges,”IEEE Geosci.RemoteSens.Mag.,2019,

[0007] vol.7,no.2,pp:140-158.

[0008] [2]Lorenzo Bruzzone,and Francesca Bovolo,”A novel framework forthedesign ofchange-detection systems for very-high-resolution remotesensingimages,”Proc.

[0009] IEEE,2012,vol.101,no.3,pp:609-630.

[0010] [3]Xiaofeng Zhang,Shuli Cheng,Liejun Wang,and HaojinLi,”Asymmetriccross-attention hierarchical network based on CNN andtransformerforbitemporal remote sensing images change detection,”IEEE Trans.

[0011] Geosci.Remote Sens.,2021,vol.60,pp:1-12.

[0012] [4]Zengfu Hou,Wei Li,Lu Li,Ran Tao,and Qian Du,”Hyperspectralchangedetection based on multiple morphological profiles,”IEEETrans.Geosci.RemoteSens.,2022,vol.60,pp:1-12.

[0013] [5]Meiqi Hu, Chen Wu, and Liangpei Zhang,”HyperNet:

[0014] Self-supervised hyperspectral spatial spectral feature understanding network for hyperspectral change detection,” IEEE Trans.Geosci.Remote Sens., 2022, vol.60, pp: 1-17.

[0016] [6]Jigang Ding, Xiaorun Li, Shu Xiang, and Shuhan Chen, "Multilevelfeatures fused and change information enhanced neural network for hyperspectral image change detection," IEEE Trans.Geosci.RemoteSens., 2023.

[0017] [7] Xianghai Wang, Keyun Zhao, Xiaoyang Zhao, and Siyao Li, "TriTF: Atriplettransformer framework based on parents and brother attention for hyperspectral image change detection." IEEE Trans.Geosci.RemoteSens., 2023, vol.61, pp: 1-13. Summary of the invention

[0019] In order to alleviate the deficiencies of the prior art, the present invention provides a high-precision and high-efficiency hyperspectral image domain adaptive change detection method, device and software product that integrates correlation alignment and spectral attention features.

[0020] In a first aspect, the present invention provides a hyperspectral image domain adaptive change detection method, comprising the following steps:

[0021] Step 1: The hyperspectral remote sensing image acquired at the first time point in the same area Hyperspectral remote sensing images acquired at different times and at the second time point Among them, C, H, and W refer to the spectral dimension, length, and width of the hyperspectral remote sensing image, respectively. Based on the twin network, the difference depth feature F between T1 and T2 is extracted. DF ;

[0022] Step 2: Extract deep features of T1 and T2 based on the Siamese network and And construct correlation alignment loss based on deep features

[0023] Step 3: Convolutional neural network based on channel attention mechanism to extract enhanced difference features F of T1 and T2 DI ;

[0024] Step 4: Transform the deep features F DF and deep features F DI Directly add to obtain the fusion feature FF;

[0025] Step 5: Based on the Kullback-Leibler divergence, construct F DF and F DI of The loss function promotes feature fusion, and the fused feature FF is input into the Softmax layer to obtain the change detection result.

[0026] Preferably, in step 1, the deep features of T1 and T2 are extracted based on the twin network. and And construct the difference deep feature F through the deep feature DF ;

[0027] The twin network first processes T1 or T2 through a two-dimensional convolutional neural network and then outputs in, is the feature output of phase T1 or T2 after being processed by a two-dimensional convolutional neural network, W 1 / 2 is the convolution weight, B 1 / 2 is the bias weight, σ(·) is the activation function; then and Subtract and take the absolute value to get the difference feature Where |·| represents the operation of taking the absolute value; the difference feature is then input into a two-dimensional convolutional neural network to extract the deep difference feature, and finally the deep difference feature T DF Input into the fully connected layer to obtain the difference depth feature F DF .

[0028] Preferably, in step 2, the correlation alignment loss extracts the deep features of T1 and T2 based on the twin network and During the training process, one of the loss functions used is the domain adaptation correlation alignment loss function

[0029]

[0030]

[0031] in, represents the square of the matrix Frobenius criterion, 1 represents a column vector of all 1s, and n1 and n2 represent the number of samples in the two phases T1 and T2 respectively.

[0032] Preferably, in step 3, the enhanced difference feature F is extracted based on the convolutional neural network. DI ; The twin network includes a channel attention mechanism, a two-dimensional convolutional neural network and a fully connected layer;

[0033] The channel attention mechanism directly subtracts the original images of the two phases T1 and T2, and performs an absolute value operation to obtain a difference image T DI , and enhance the difference image T based on the channel attention mechanism DI Dimensional features to obtain enhanced features Then use a two-dimensional convolutional neural network to extract Deep features of Finally, the deep features of the difference image will be enhanced Input into the fully connected layer to obtain the enhanced difference feature F based on the channel attention mechanism DI ;

[0034] in, Depend on Reshape to get; Depend on and Transpose and do matrix multiplication to get; is the scaling parameter;

[0035] in, After being processed by a 2D convolutional neural network The feature output of , W is the convolution weight, B is the bias weight, and σ(·) is the activation function.

[0036] Preferably, the twin network and the convolutional neural network constitute a hyperspectral image domain adaptation change detection network; the hyperspectral image domain adaptation change detection network is a trained network, and the loss function used in the training is for:

[0037]

[0038] in, is the cross entropy loss function for binary classification, is the correlation alignment loss function for domain adaptation, is the Kullback-Leibler divergence loss function of feature fusion; x represents the specific value of each feature; n represents the number of samples, Y i and Represent the true label and predicted label of sample i respectively; λ1 and λ2 are and Hyperparameters of .

[0039] In a second aspect, the present invention provides a hyperspectral image domain adaptive change detection device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the hyperspectral image domain adaptive change detection method is implemented when the processor executes the program.

[0040] In a third aspect, the present invention provides a software product, comprising a computer program, which implements the hyperspectral image domain adaptive change detection method when executed by a processor.

[0041] In a fourth aspect, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the hyperspectral image domain adaptive change detection method.

[0042] The advantages and beneficial effects of the present invention include:

[0043] (1) In order to reduce the distribution differences of hyperspectral images at different phases, this paper proposes domain adaptation through correlation alignment loss, aiming to minimize the distribution differences by aligning the correlation features extracted from the hyperspectral images, so as to reduce the model's ability to over-predict change information.

[0044] (2) In order to enhance the network's ability to mine the features of the original image, the present invention enhances the features of the difference image channel dimension through a channel attention mechanism, aiming to explore the correlation between the difference image channels and further mine semantic features to improve the model's ability to recognize subtle changes.

[0045] (3) In order to better integrate the different features extracted by the network, the present invention reduces the differences in features of different branches through Kullback-Leibler divergence loss, aiming to promote the fusion of features extracted by two branches, so as to improve the accuracy and reliability of the model in identifying change information. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The technical solution of the present invention is further described below using embodiments and specific implementation methods. In addition, in the process of describing the technical solution, some drawings are also used. For those skilled in the art, other drawings and the intention of the present invention can also be obtained based on these drawings without paying creative work.

[0047] Figure 1 It is a schematic diagram of the method principle of an embodiment of the present invention. DETAILED DESCRIPTION

[0048] Combine the following Figure 1 , the specific implementation methods of the present invention are further described in detail.

[0049] In order to facilitate ordinary technicians in the field to understand and implement the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the implementation examples described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0050] Change detection is an important task in geospatial analysis, which aims to identify obvious changes in geographic features between images of different data sets. However, existing work often ignores the distribution differences between images caused by changes in imaging time. At the same time, the features of hyperspectral images still have great development potential in describing and detecting changes. In order to alleviate this situation, the present invention proposes a hyperspectral image domain adaptation change detection method called fusion correlation alignment and spectral attention features, which combines domain adaptation and attention mechanisms to make up for the shortcomings of existing methods, thereby completing the hyperspectral image change detection task more accurately. Through this method, the present invention aims to improve the accuracy, reliability and efficiency of hyperspectral image change detection, and provide strong support for the research and application of change detection.

[0051] Please see Figure 1 , this embodiment provides a hyperspectral image domain adaptation change detection method, comprising the following steps:

[0052] Step 1: The hyperspectral remote sensing image acquired at the first time point in the same area Hyperspectral remote sensing images acquired at different times and at the second time point Among them, C, H, and W refer to the spectral dimension, length, and width of the hyperspectral remote sensing image, respectively. The deep features of T1 and T2 are extracted based on the twin network. and Finally, the difference depth feature F is constructed through the depth feature DF ;

[0053] In one embodiment, the deep features of T1 and T2 are extracted based on the Siamese network. and Finally, the difference depth feature F is constructed through the depth feature DF ;

[0054] The twin network first processes T1 or T2 through a two-dimensional convolutional neural network and then outputs in, is the feature output of phase T1 or T2 after being processed by a two-dimensional convolutional neural network, W 1 / 2 is the convolution weight, B 1 / 2 is the bias weight, σ(·) is the activation function, and ReLU is used as the activation function here; then and Subtract and take the absolute value to get the difference feature Where |·| represents the operation of taking the absolute value; the difference feature is then input into a two-dimensional convolutional neural network to extract the deep difference feature, and finally the deep difference feature T DF Input into the fully connected layer to obtain the difference depth feature F DF .

[0055] Step 2: In one embodiment, the correlation alignment (CORAL) loss extracts deep features of T1 and T2 based on the Siamese network and During the training process, one of the loss functions used is the domain adaptation correlation alignment loss function

[0056]

[0057] in, represents the square of the matrix Frobenius criterion, 1 represents a column vector of all 1s, and n1 and n2 represent the number of samples in the two phases T1 and T2 respectively.

[0058] Step 3: Extract the deep features F of T1 and T2 based on the convolutional neural network DI ;

[0059] In one embodiment, a Siamese network is used to extract deep features F DI ; The twin network includes a channel attention mechanism, a two-dimensional convolutional neural network and a fully connected layer;

[0060] The channel attention mechanism combines the two time phases T 1 / 2 The images are directly subtracted and the absolute value operation is performed to obtain the difference image T DI :

[0061] T DI =|T1-T2|;

[0062] In order to more deeply reveal the interdependence between channels of the difference image and effectively extract the semantic features contained therein, the channel attention mechanism is used here to further enhance the feature expression of the difference image:

[0063]

[0064] in, Depend on Reshape to get; Depend on and Do matrix multiplication to get; is the scaling parameter;

[0065] Then, a 2D convolutional neural network is used to extract The deep features of

[0066]

[0067] in, After being processed by a two-dimensional convolutional neural network The feature output of , W is the convolution weight, B is the bias weight, σ(·) is the activation function, and ReLU is used here as the activation function.

[0068] Finally, the deep features of the difference image are enhanced Input into the fully connected layer to obtain the enhanced difference feature F based on the channel attention mechanism DI .

[0069] Step 4: Difference depth feature F DF and enhanced difference feature F DI Directly add to obtain the fusion feature FF;

[0070] In one embodiment, in order to obtain a more comprehensive and rich feature representation, the features F extracted from the two branch fully connected layers are DF and F DI Perform an addition operation to fuse their respective information. This step aims to combine the features extracted by the two branches to form the final feature extracted by the entire network, which is named the fusion feature FF:

[0071] FF=F DF +F DI ;

[0072] Step 5: Based on the Kullback-Leibler divergence, construct F DF and F DI of The loss function promotes feature fusion, and the fused feature FF is input into the Softmax layer to obtain the change detection result.

[0073] In one embodiment, the twin network and the convolutional neural network constitute a hyperspectral image domain adaptation change detection network; the hyperspectral image domain adaptation change detection network is a trained network, and the loss function used in the training is for:

[0074]

[0075] in, is the cross entropy loss function for binary classification, is the correlation alignment loss function for domain adaptation, is the Kullback-Leibler divergence loss function of feature fusion; x represents the specific value of each feature; n represents the number of samples, Y i and Represent the true label and predicted label of sample i respectively; λ1 and λ2 are and Hyperparameters of .

[0076] This embodiment also provides a hyperspectral image domain adaptation change detection device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the hyperspectral image domain adaptation change detection method is implemented when the processor executes the program.

[0077] This embodiment also provides a software product, including a computer program, which implements the hyperspectral image domain adaptation change detection method when executed by a processor.

[0078] This embodiment also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the hyperspectral image domain adaptive change detection method is implemented.

[0079] The present invention is further described below through experiments.

[0080] This experiment is written in PYTHON language and implemented using the classic deep learning framework PYTORCH. The PYTHON remote sensing image reading and writing functions are the implementation basis, aiming to discover the change information of two C×H×W-sized hyperspectral images of different phases, where C, H and W refer to the spectral dimension, length and width of the hyperspectral image respectively. The PYTHON remote sensing image reading and writing functions are public technologies in this field and will not be described in detail here.

[0081] In this experiment, two hyperspectral datasets named "China" and "Rivers" were selected as datasets to verify the effectiveness of the model, in order to verify the accuracy and reliability of the present invention. The China dataset was acquired over a farmland in Yancheng City, Jiangsu Province, China on May 3, 2006 and April 23, 2007. The dataset was taken using the Hyperion sensor of the EO-1 satellite, with a wavelength range of 0.4 to 2.5 microns, a spatial resolution of about 30 meters, and a spectral resolution of about 10 nanometers. Each image has a pixel of 420×140 and includes 154 spectral bands. The changes observed in this dataset are mainly due to crop rotation. The Rivers dataset was collected in the river area of ​​Jiangsu Province, China on May 3, 2013 and December 31, 2013. The dataset was taken using the Hyperion sensor of the EO-1 satellite, with a wavelength range of 0.4 to 2.5 microns, a spatial resolution of about 30 meters, and a spectral resolution of about 10 nanometers. Each image consists of 463×241 pixels and includes 198 spectral bands. The changes observed in this dataset are mainly caused by the disappearance of materials in the river. In order to comprehensively evaluate the performance of the method of the present invention, the classic difference principal component analysis (method 1) and support vector machine (method 2), deep learning multi-layer encoder-decoder attention network (method 3), self-supervised hyperspectral spatial-spectral feature understanding network (method 4), two-branch differential amplification graph convolution (method 5), and multi-scale difference change feature fusion network (method 6) are used to compare with the method of the present invention.

[0082] Experimental settings: Both datasets use 5% of samples for training and 95% of samples for testing. The China dataset selects 3150 samples as training samples and 59850 samples as testing samples, and the River dataset selects 5579 samples as training samples and 106004 samples as testing samples.

[0083] Evaluation indicators: Overall Accuracy (OA), Kappa coefficient, Precision, Recall and F1 score (F1). These indicators are commonly used quantitative evaluation indicators in hyperspectral image change detection and can be calculated through confusion matrix.

[0084] Table 1 Confusion matrix

[0085]

[0086]

[0087] In Table 1, TP represents the number of pixels that have changed in the true label and are correctly predicted to have changed. FP represents the number of pixels that have not changed in the true label and are incorrectly predicted to have changed. FN represents the number of pixels that have changed in the true label and are incorrectly predicted to have not changed. TN represents the number of pixels that have not changed in the true label and are correctly predicted to have not changed.

[0088] The calculation formulas for these five evaluation indicators are as follows:

[0089]

[0090] Table 2 Comparative test results

[0091]

[0092] From the data comparison in Table 2, it can be clearly observed that the method of the present invention shows excellent performance compared with other methods. In terms of the three key indicators of overall accuracy, Kappa coefficient and F1 score, the present invention has achieved higher values. As a commonly used indicator for measuring classification accuracy, the improvement of Kappa coefficient means that the present invention has higher accuracy and stability when distinguishing changes from non-changes. The improvement of F1 score further proves the outstanding performance of the present invention in balancing accuracy and recall. Therefore, the evaluation of Kappa coefficient and F1 score is more objective and comprehensive.

[0093] Compared with traditional change detection methods, the method of the present invention not only achieves a better balance between accuracy and recall. Compared with the change detection methods of the same depth, it can be found that the method of the present invention has significant advantages in improving the ability to identify change information. In summary, the method of the present invention has shown significant advantages and unique performance in change detection. Whether in overall accuracy, Kappa coefficient or F1 score, high scores can be achieved. The present invention aims to improve the performance of change detection in hyperspectral images and provide strong support for the research and application of change detection.

[0094] The key point of the present invention is to propose a method for domain-adaptive change detection of hyperspectral images that integrates correlation alignment and spectral attention features. The method completes the task of change detection based on domain-adaptive correlation alignment, channel attention mechanism and Kullback-Leibler divergence. The present invention is based on a convolutional neural network, and reduces the difference in feature distribution of images at different times through correlation alignment, thereby reducing the network's ability to over-predict change information; in addition, the present invention increases the correlation of the channel dimension of the difference image through the attention mechanism, and further explores semantic information to improve the network's ability to recognize subtle changes. Finally, the present invention uses Kullback-Leibler divergence loss to fuse the features extracted by the two branches, thereby further enhancing the network's accuracy in change detection.

[0095] It should be understood that the embodiments described above are part of the embodiments of the present invention, rather than all of the embodiments. In addition, the technical features in the various embodiments or single embodiments provided by the present invention can be combined with each other arbitrarily to form a feasible technical solution. Such combination is not restricted by the sequence of steps and / or the structural composition mode, but must be based on the ability of ordinary technicians in the field to implement. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the protection scope required by the present invention.

[0096] It should be understood that the above description of the preferred embodiment is relatively detailed and cannot be regarded as limiting the scope of patent protection of the present invention. Under the enlightenment of the present invention, ordinary technicians in this field can also make substitutions or modifications without departing from the scope of protection of the claims of the present invention, which all fall within the scope of protection of the present invention. The scope of protection requested for the present invention shall be based on the attached claims.

Claims

1. A domain-adaptive change detection method for hyperspectral images, It is characterized in that The following steps are involved: Step 1: The hyperspectral remote sensing image acquired at the first time point in the same area and the hyperspectral remote sensing image acquired at the second time point Among them, C, H, and W refer to the spectral dimension, length, and width of the hyperspectral remote sensing image, respectively. 1 and T 2 The deep features T 1 fea and T 2 fea , and construct the difference depth feature F through the depth feature DF ; Step 2: Extract T based on the Siamese network 1 and T 2 The deep features T 1 fea and T 2 fea , and construct correlation alignment loss based on deep features Step 3: Convolutional neural network based on channel attention mechanism to extract T 1 and T 2 The enhanced difference feature F DI ; Step 4: Difference depth feature F DF and enhanced difference feature F DI Directly add to obtain the fusion feature FF; Step 5: Based on the Kullback-Leibler divergence, construct F DF and F DI The loss function Promote feature fusion and input the fused feature FF into the Softmax layer to obtain the change detection result.

2. The hyperspectral image domain adaptation change detection method according to claim 1, Features: In step 1, the twin network first transforms T 1 or T 2 The output is processed by two-dimensional convolutional neural network respectively. in, is the time phase T after being processed by the two-dimensional convolutional neural network 1 or T 2 The characteristic output, W 1 / 2 is the convolution weight, B 1 / 2 is the bias weight, σ(·) is the activation function; then and Subtract and take the absolute value to get the difference feature Where |·| represents the operation of taking the absolute value; the difference feature is then input into a two-dimensional convolutional neural network to extract the deep difference feature, and finally the deep difference feature T DF Input into the fully connected layer to obtain the difference depth feature F DF .

3. The hyperspectral image domain adaptation change detection method according to claim 1, Features: In step 2, the correlation alignment loss is based on the twin network to extract T 1 and T 2 The deep features T 1 fea and T 2 fea ; During the training process, one of the loss functions used is the domain adaptation correlation alignment loss function in, represents the square of the matrix Frobenius criterion, 1 represents a column vector of all 1s, n 1 and n 2 Represents two time phases T 1 and T 2 The number of samples.

4. The hyperspectral image domain adaptation change detection method according to claim 1, Features: In step 3, the enhanced difference feature F is extracted based on the convolutional neural network DI ; The network includes a channel attention mechanism, a two-dimensional convolutional neural network and a fully connected layer; The channel attention mechanism combines the two phases T 1 and T 2 The original image is directly subtracted and the absolute value operation is performed to obtain the difference image T DI , and then enhance the difference image T based on the channel attention mechanism DI Dimensional features to obtain enhanced features Then use a two-dimensional convolutional neural network to extract Deep features of Finally, the deep features of the difference image will be enhanced Input into the fully connected layer to obtain the enhanced difference feature F based on the channel attention mechanism DI ; in, Depend on Reshape to get; Depend on and Do matrix multiplication to get; is the scaling parameter; in, After being processed by a 2D convolutional neural network The feature output of , W is the convolution weight, B is the bias weight, and σ(·) is the activation function.

5. The hyperspectral image domain adaptation change detection method according to any one of claims 1 to 4, Features: The twin network and the convolutional neural network constitute a hyperspectral image domain adaptation change detection network; the hyperspectral image domain adaptation change detection network is a trained network, and the loss function used in the training is for: in, is the cross entropy loss function for binary classification, is the correlation alignment loss function for domain adaptation, is the Kullback-Leibler divergence loss function of feature fusion; x represents the specific value of each feature; n represents the number of samples, Y i and Represent the true label and predicted label of sample i respectively; λ 1 and λ 2 yes and Hyperparameters of .

6. A hyperspectral image domain adaptive change detection device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, It is characterized in that When the processor executes the program, the hyperspectral image domain adaptive change detection method as claimed in any one of claims 1 to 5 is implemented.

7. A software product comprising a computer program, It is characterized in that When the computer program is executed by a processor, the hyperspectral image domain adaptive change detection method according to any one of claims 1 to 5 is implemented.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, It is characterized in that When the computer program is executed by a processor, the hyperspectral image domain adaptive change detection method according to any one of claims 1 to 5 is implemented.

Citation Information

Patent Citations

  • High-resolution remote sensing image change detection method

    CN113706482A

  • Hyperspectral image change detection method and device based on multilevel feature fusion and change information enhancement

    CN117392538A

  • Hyperspectral image multi-classification change detection method and system based on spectrum guided Transform network model

    CN117710805A