Self-supervised hyperspectral change detection method and system

Abstract

The invention relates to a self-supervised hyperspectral change detection method and system, and belongs to the technical field of deep learning. The method comprises the following steps: S1, preprocessing a hyperspectral image; s2, performing multi-scale mask change simulation; s3, inputting the single time phase sample and a pseudo second time phase sample obtained after mask simulation change into a global local feature aggregation codec for training; and S4, after training is completed, performing change detection on the complete image to obtain a result. The performance of the method is superior to that of other hyperspectral image farmland change detection methods, and the model can be efficiently trained without sample marking and downstream fine tuning.

Description

Technical Field

[0001] The present invention belongs to the technical field of deep learning and relates to a self-supervised hyperspectral change detection method and system. Background Art

[0002] Remote sensing image change detection detects changes by observing remote sensing images of the same area at different times, and is widely used in urban planning, environmental monitoring, agricultural surveys, and disaster assessments. Compared with multispectral remote sensing images that only contain part of the band, hyperspectral data has the advantage of providing continuous and detailed characteristic spectral features over a large spectral range, and has the characteristics of "image-spectrum integration", which can better distinguish the changed parts of the two images. With the development of imaging spectroscopy technology, many satellites have provided rich hyperspectral images for studying surface changes, but they also face huge labeling challenges.

[0003] Classic change detection methods mainly use spectral information to construct algebraic operations, or project hyperspectral images into a low-dimensional feature space to reveal the properties of changes, represented by methods such as change vector analysis, iterative multivariate change detection, slow feature analysis, and principal component analysis. Traditional change detection methods are often based on spectral differences between images of different phases, which cannot fully utilize the inherent characteristics of complex hyperspectral images. In recent years, the development of deep learning technology has brought new solutions to improve the efficiency and accuracy of change detection. Initially, deep learning methods were highly dependent on supervised information for learning. Some supervised methods based on labeled data, such as BCNNS, MLEDAN, MSDFFN, etc., have good accuracy performance. However, the acquisition and labeling of samples are time-consuming and laborious. Therefore, some work began to introduce self-supervised strategies into change detection to reduce the dependence on labeled data. Existing self-supervised hyperspectral change detection methods can be roughly divided into two categories. One category generates high-confidence pseudo-labels as supervisory information, and representative ones include BCG Net and S3Net. The other category constructs supervisory information between features based on contrastive learning, and representative ones include DSConvAEs, HyperNet, UA-GSSL, etc. However, the above methods usually require preprocessing of paired images from the same geographical location during the training phase, and some methods also require a small amount of labels for fine-tuning. Compared with paired labeled multi-temporal hyperspectral images, unlabeled single-temporal hyperspectral images are easier to obtain and cheaper.

[0004] In order to overcome the challenges brought by limited labeled data and expand the generalization ability of the model, the present invention introduces self-supervised learning to make full use of the information of unlabeled samples, based on single-phase images, and improves the accuracy of current change detection from the perspective of masks. Summary of the invention

[0005] In view of this, the primary purpose of the present invention is to provide a self-supervised hyperspectral change detection method, which uses a multi-scale mask to simulate the change area, learns effective change features from a single-phase hyperspectral image, and obtains an efficient change detector. First, a multi-scale mask change simulation strategy based on the original spectrum is designed to simulate the image of the pseudo second phase change. Then, a global local multi-scale feature extraction aggregation network is designed to detect the changes between the single-phase hyperspectral image and the masked hyperspectral image. After that, the training of the self-supervised network is carried out through mask supervision loss and reconstruction loss. This method can efficiently train the model without the need for labeled samples and downstream fine-tuning.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] A self-supervised hyperspectral change detection method comprises the following steps: S1: hyperspectral image preprocessing; S2: multi-scale mask change simulation; S3: inputting single-phase samples and pseudo second-phase samples obtained after mask simulation change into a global local feature aggregation codec for training; S4: after the training is completed, performing change detection on the complete image to obtain the result.

[0008] Furthermore, in step S1, the single-phase hyperspectral image is preprocessed, mainly including geometric distortion correction, spectral image vignetting correction, radiation correction, etc.; the preprocessed data set is subjected to a block operation. In this method, when selecting samples, the patches centered on the edge pixels of the image are taken into consideration. The edges of the image are first expanded in a mapping manner, and samples are generated using 32×32 blocks as sliding windows.

[0009] Furthermore, in step S2, in order to obtain a hyperspectral image of the second phase that is close to the real one, we use a multi-scale mask to simulate the change area; in order to adapt to the change characteristics of the ground objects in different situations, we set different mask module sizes and different mask ratios, and randomly mix these independent multi-scale masks with equal probability to obtain the final multi-scale mask M; the spectrum of the central pixel of the original patch is horizontally flipped as the change spectrum, the spectrum of the masked area is replaced by the change spectrum, and random noise is added to the remaining areas to simulate the imaging differences in the real scene. The formula of the mask change simulation strategy is as follows:

[0010]

[0011] Among them, (1-M) represents the unmasked area, represents the hyperspectral image after adding the mask, represents the hyperspectral image after adding Gaussian noise, Represents the spectrum value after flipping.

[0012] Further, in step S3, the single-phase hyperspectral image and the pseudo second-phase hyperspectral image obtained after mask simulation change are input into the global-local feature aggregation codec; first, rich multi-scale features are obtained in the encoder stage; then, the feature map of the same scale is subtracted from the dual-time feature map to obtain a differential feature map; then, these multi-scale differential features are connected to the decoder through jump connections to achieve cross-layer connection of features and reduce the loss of details.

[0013] Furthermore, in the global-local feature aggregation codec, the encoding stage consists of three downsampling units, each of which contains a global fusion module, a local fusion module, and a downsampling operation. The encoder stage can be summarized as follows:

[0014]

[0015] in, represents the feature map generated by the i-th convolutional layer of the input patch, f Conv 3×3 represents a convolution operation with a kernel size of 3×3, LMM(·) and GMM(·) represent local fusion modules and global fusion modules, and all convolutional layers are followed by a batch normalization layer and a LeakyReLU activation function layer.

[0016] Furthermore, in the global local feature aggregation codec, the decoding stage includes a detection decoder and a reconstruction decoder; the feature maps of the two branches of the obtained encoder are differentially operated to obtain a differential feature map; the differential feature map is input into the multi-layer upsampling detection decoder to restore the feature map to the original input size. The process of the detection decoder can be summarized as follows:

[0017]

[0018] in, represents the output feature map after the i-th deconvolution operation, f DConv 3×3 represents the deconvolution operation with a kernel size of 3×3, f Conv 1×1 represents a convolution operation with a kernel size of 1×1, [;] represents a stacking operation along the channel dimension, which realizes the skip connection of the feature map;

[0019] At the same time, we input the output of the mask branch into the reconstruction decoder to obtain the reconstructed feature map of the original single-phase hyperspectral image. The process of the reconstruction decoder can be summarized as follows:

[0020]

[0021] in, Represents the reconstructed feature map after the i-th deconvolution operation; then, we use a multi-layer perceptron that does not share weights to project the obtained feature map into another dimensional space. The formula for the projection process is as follows:

[0022]

[0023] Finally, the reconstructed feature map X is obtained rec And the detection feature map X det .

[0024] Furthermore, in the encoding stage, in order to achieve effective aggregation of spatial spectral information and obtain discriminative differential features, we embed the global fusion module and the local fusion module to realize long-range information interaction and local information aggregation respectively; in the global fusion module, first, given an input feature Divide the features into P blocks along the channel dimension and connect them along the columns to obtain A 3×3 convolution is applied to this feature map, which is then reduced to its original dimension. Next, the reduced feature map is fused with the initial feature map through a fully connected layer and a GELU function. The feature is then divided into P blocks along the channel dimension and connected along the rows to obtain The same 3×3 convolution, fully connected layer and GELU function are performed and fused with the initial feature map; the global fusion module slices and reorganizes the feature maps in different directions of rows and columns to make the spatial distances of pixels close to each other, thereby improving the interaction ability between long-range information.

[0025] Furthermore, the output of the global fusion module is input into the local fusion module for further feature extraction. First, 3×7 and 7×3 convolution operations are applied to the input feature map to extract local features in different directions. The formula is as follows:

[0026] x add =x+f Conv 3×7 (x)+f Conv 7×3 (x)

[0027] Next, global average pooling is used to compress the spatial dimensions of the input feature map, and then a multi-layer perceptron is used to compress and extract features, and the Swish function is used for activation. The formula for the attention process is as follows:

[0028] Att(x) = Swish(GAP(MLP(x) add )))

[0029] Where Swish(x) = xsigmoid(x), the final output of the local fusion module can be described as follows:

[0030] LMM(x)=Att(x)x add

[0031] The feature map is multiplied by the shared channel attention weight, and the obtained feature map highlights the useful areas and suppresses the useless areas. Through the designed local fusion module, the local features can be effectively aggregated to achieve the fusion of spatial and spectral attention.

[0032] Further, the output feature maps of the detection decoder and the corresponding projection head are selected for final detection, and the final change detection result can be described as follows:

[0033]

[0034] Among them, y i To predict the probability result, fc(·) represents the fully connected layer for extracting features and reducing dimensionality; multi-scale masks are used to supervise the change detection results, and the formula for using the cross entropy function as the loss function is as follows:

[0035]

[0036] Where N is the number of samples, m is i is the mask label of a given sample; at the same time, a reconstruction loss is designed to reconstruct the output feature map of the decoder and the corresponding projection head. The formula of the reconstruction loss function is as follows:

[0037]

[0038] The total loss function is composed of the above loss functions as follows and is calculated as follows:

[0039] L total =L sup +L rec

[0040] Calculate the loss function, optimize the model parameters of the change detection framework according to the loss function and the back-propagation process, and obtain a trained change detection framework after the training is completed; use the trained change detection framework to discriminate the input samples and output a change detection result graph.

[0041] In addition, the present invention also provides a self-supervised hyperspectral change detection system, which is configured with a control program for implementing the method described above.

[0042] The beneficial effects of the present invention are:

[0043] The self-supervised hyperspectral change detection method and system proposed in the present invention can efficiently train the model without the need for labeled samples and downstream fine-tuning, solving the problems of difficult pairing and labeling of multi-phase samples. The multi-scale mask simulation strategy proposed in the present invention utilizes the original spectral information with multi-scale masks to achieve adaptive learning of features across different scales. The global local multi-scale feature extraction aggregation network proposed in the present invention can realize remote feature interaction and aggregation of local spatial spectral features. Experimental results on a public hyperspectral image dataset show that the performance of the proposed STMNet is better than the most advanced hyperspectral change detection method.

[0044] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0046] Figure 1 is a flow chart of the method of the present invention;

[0047] Figure 2 Figure 2. Single Temporal Mask-Based Change Detection Network (STMNet) for Self-Supervised Hyperspectral Change Detection.

[0048] Figure 3 A schematic diagram of the multi-scale variation simulation of the present invention;

[0049] Figure 4 It is a structural diagram of the global-local fusion module of the present invention;

[0050] Figure 5 is a land dataset, wherein (a) is an image before the change, (b) is an image after the change, and (c) is a change detection result of the method of the present invention;

[0051] Figure 6 Visualization results of different methods on the farmland dataset, including (a) CVA, (b) PCAKM, (c) MSCD, (d) HyperNet, (e) BCGNet, (f) UAGSSL, (g) STMNet, and (h) ground truth map. DETAILED DESCRIPTION

[0052] The technical solution of the present invention is described in detail below with reference to the accompanying drawings.

[0053] Figure 1 The flowchart of the method of the present invention provides a self-supervised hyperspectral change detection method and system. As shown in the figure, in the image acquisition stage, geometric distortion correction, radiation correction and other means are used to quickly achieve image registration. The deep learning network used for change detection is as follows Figure 2 As shown in the figure, it uses multi-scale masks to simulate the change area, learns effective change features from single-phase hyperspectral images, and obtains an efficient change detector. The network consists of three parts: multi-scale mask change simulation, global-local feature aggregation encoding and decoding, and self-supervised training. First, a multi-scale mask strategy based on the original spectrum is designed to simulate the image of pseudo second-phase change, which utilizes the original spectral information with multi-scale masks to achieve adaptive learning of features across different scales. Secondly, a global-local multi-scale feature extraction aggregation network is designed to detect the changes between single-phase hyperspectral images and masked hyperspectral images, realizing long-range feature interaction and aggregation of local spatial spectral features. The end-to-end network proposed in the present invention can efficiently train the model without the need for labeled samples and downstream fine-tuning.

[0054] Specifically, the technical solution of the present invention includes the following contents:

[0055] 1. Hyperspectral image preprocessing: Preprocess the single-phase hyperspectral image, mainly including geometric distortion correction, spectral image vignetting correction, radiation correction, etc. The preprocessed data set is subjected to block extraction. In this method, when selecting samples, the patches centered on the edge pixels of the image are considered. The edges of the image are first expanded in a mapping manner, and samples are generated using 32×32 blocks as sliding windows.

[0056] 2. Multi-scale mask change simulation: In order to obtain hyperspectral images close to the real second phase, we use multi-scale masks to simulate the change area, such as Figure 3 As shown. In order to adapt to the changing characteristics of ground objects in different situations, we set different mask module sizes and different mask ratios, and randomly mix these independent multi-scale masks with equal probability to obtain the final multi-scale mask M. The spectrum of the central pixel of the original patch is horizontally flipped as the change spectrum, the spectrum of the masked area is replaced with the change spectrum, and random noise is added to the remaining areas to simulate the imaging differences in real scenes. The formula of the mask change simulation strategy is as follows:

[0057]

[0058] Among them, (1-M) represents the unmasked area, represents the hyperspectral image after adding the mask, represents the hyperspectral image after adding Gaussian noise, Represents the spectrum value after flipping.

[0059] 3. The single-phase hyperspectral image and the pseudo second-phase hyperspectral image obtained after the mask simulation change are input into the global local feature aggregation codec. First, rich multi-scale features are obtained in the encoder stage. Then, the feature map of the same scale is subtracted from the dual-time feature map to obtain a differential feature map. Then, these multi-scale differential features are connected to the decoder through jump connections to achieve cross-layer connection of features and reduce the loss of details.

[0060] 4. In the global-local feature aggregation codec, the encoding stage consists of three downsampling units, each of which contains a global fusion module, a local fusion module, and a downsampling operation. The encoder stage can be summarized as follows:

[0061]

[0062] in, represents the feature map generated by the i-th convolutional layer of the input patch, f Conv 3×3 represents a convolution operation with a kernel size of 3×3, LMM(·) and GMM(·) represent local fusion modules and global fusion modules. All convolutional layers are followed by a batch normalization layer and a LeakyReLU activation function layer.

[0063] 5. In the global local feature aggregation codec, the decoding stage includes a detection decoder and a reconstruction decoder. The feature maps of the two branches of the encoder are differentiated to obtain a differential feature map. The differential feature map is input to the multi-layer upsampling detection decoder to restore the feature map to the original input size. The process of the detection decoder can be summarized as follows:

[0064]

[0065] in, represents the output feature map after the i-th deconvolution operation, f DConv 3×3 represents the deconvolution operation with a kernel size of 3×3, f Conv 1×1 Represents a convolution operation with a kernel size of 1×1. [;] represents a stacking operation along the channel dimension, which implements the skip connection of the feature map.

[0066] At the same time, we input the output of the mask branch into the reconstruction decoder to obtain the reconstructed feature map of the original single-phase hyperspectral image. The process of the reconstruction decoder can be summarized as follows:

[0067]

[0068] in, Represents the reconstructed feature map after the i-th deconvolution operation. Then, we use a multi-layer perceptron that does not share weights to project the obtained feature map into another dimensional space. The formula for the projection process is as follows:

[0069]

[0070] Finally, the reconstructed feature map X is obtained rec And the detection feature map X det .

[0071] 6. In the encoding stage, in order to achieve effective aggregation of spatial spectral information and obtain discriminative differential features, we embed the global fusion module and the local fusion module to realize remote information interaction and local information aggregation, respectively. Figure 4 As shown. In the global fusion module, first, given an input feature Divide the features into P blocks along the channel dimension and connect them along the columns to obtain A 3×3 convolution is applied to this feature map, which is then reduced to its original dimension. Next, the reduced feature map is fused with the initial feature map through a fully connected layer and a GELU function, and then the feature is divided into P blocks along the channel dimension and connected along the rows to obtain The same 3×3 convolution, fully connected layer and GELU function are performed and fused with the initial feature map. The global fusion module slices and reorganizes the feature maps in different directions of rows and columns to make the spatial distances of pixels close to each other, thereby improving the interaction ability between long-range information.

[0072] 7. Next, the output of the global fusion module is input into the local fusion module for further feature extraction. First, 3×7 and 7×3 convolution operations are applied to the input feature map to extract local features in different directions. The formula is as follows:

[0073] x add =x+f Conv 3×7 (x)+f Conv 7×3 (x)

[0074] Next, global average pooling is used to compress the spatial dimensions of the input feature map, and then a multi-layer perceptron is used to compress and extract features, and the Swish function is used for activation. The formula for the attention process is as follows:

[0075] Att(x) = Swish(GAP(MLP(x) add )))

[0076] Where Swish(x) = xsigmoid(x), the final output of the local fusion module can be described as follows:

[0077] LMM(x)=Att(x)x add

[0078] The feature map is multiplied by the shared channel attention weight, and the obtained feature map highlights the useful areas and suppresses the useless areas. Through the designed local fusion module, local features can be effectively aggregated to achieve the fusion of spatial and spectral attention.

[0079] 8. Select the output feature map of the detection decoder and the corresponding projection head for final detection. The final change detection result can be described as follows:

[0080]

[0081] Among them, y i To predict the probability results, fc(·) represents the fully connected layer for feature extraction and dimensionality reduction.

[0082] Multi-scale masks are used to supervise the change detection results, and the formula using the cross entropy function as the loss function is as follows:

[0083]

[0084] Where N is the number of samples, m is i is the mask label for a given sample.

[0085] At the same time, a reconstruction loss is designed to reconstruct the output feature maps of the decoder and the corresponding projection head. The formula of the reconstruction loss function is as follows:

[0086]

[0087] The total loss function is composed of the above loss functions as follows and is calculated as follows:

[0088] L total =L sup +L rec

[0089] Calculate the loss function, optimize the model parameters of the change detection framework according to the loss function and the back-propagation process, and obtain a trained change detection framework after the training is completed; use the trained change detection framework to discriminate the input samples and output a change detection result graph.

[0090] like Figure 5This is the experimental result of the STMNet change detection network described in the present invention on an open source hyperspectral farmland dataset. It can be seen that the changed area is well detected. The detection effect of the present invention can be further illustrated by comparative experiments. The method of the present invention is compared with other existing methods CVA, PCAKM, MSCD, HyperNet, BCGNet, and UAGSSL methods on the farmland dataset, and the overall accuracy, Kappa coefficient, accuracy, recall, and F1-score are calculated respectively. The larger the accuracy, the higher the proportion of all results predicted as positive; the larger the Kappa coefficient, the higher the consistency of the detection results; the larger the recall, the higher the proportion of all positive results predicted correctly; the larger the F1-score, the better the comprehensive evaluation of the results. Table 1 gives the values ​​of various indicators of the detection results of different methods:

[0091] Table 1 Comparison of STMNet and various methods on farmland dataset

[0092]

[0093] It can be seen that the method of the present invention achieves the best accuracy on this dataset without labeled samples. Figure 6 The visual detection results of the above various methods are given, and it can be seen that the performance of the method described in the present invention is better than other self-supervised hyperspectral image farmland change detection methods.

[0094] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A self-supervised hyperspectral change detection method, characterized in that: The following steps are involved: S1: Hyperspectral image preprocessing; S2: Multi-scale mask variation simulation; S3: Input the single-phase samples and the pseudo second-phase samples obtained after the mask simulation change into the global local feature aggregation codec for training; S4: After training is completed, change detection is performed on the complete image to obtain the results.

2. A self-supervised hyperspectral change detection method according to claim 1, characterized in that: In step S1, the single-phase hyperspectral image is preprocessed, including geometric distortion correction, spectral image vignetting correction, and radiation correction; the preprocessed data set is subjected to a block operation, and when selecting samples, the patches centered on the edge pixels of the image are considered. The edges of the image are first expanded in a mapping manner, and samples are generated using 32×32 blocks as sliding windows.

3. The self-supervised hyperspectral change detection method according to claim 1, characterized in that: In step S2, in order to obtain a hyperspectral image of the second phase that is close to the real one, a multi-scale mask is used to simulate the change area; in order to adapt to the change characteristics of the ground objects in different situations, different mask module sizes and different mask ratios are set, and these independent multi-scale masks are randomly mixed with equal probability to obtain the final multi-scale mask M; The spectrum of the central pixel of the original patch is horizontally flipped as the change spectrum, the spectrum of the masked area is replaced by the change spectrum, and random noise is added to the remaining areas to simulate the imaging difference in the real scene. The formula of the mask change simulation strategy is as follows: Among them, (1-M) represents the unmasked area, represents the hyperspectral image after adding the mask, represents the hyperspectral image after adding Gaussian noise, Represents the spectrum value after flipping.

4. The self-supervised hyperspectral change detection method according to claim 1, characterized in that: In step S3, the single-phase hyperspectral image and the pseudo second-phase hyperspectral image obtained after mask simulation change are input into the global-local feature aggregation codec; first, rich multi-scale features are obtained in the encoder stage; then, the feature map of the same scale is subtracted from the dual-time feature map to obtain a differential feature map; then, these multi-scale differential features are connected to the decoder through jump connections to achieve cross-layer connection of features and reduce the loss of details.

5. A self-supervised hyperspectral change detection method according to claim 4, characterized in that: In the global-local feature aggregation codec, the encoding stage consists of three downsampling units, each of which contains a global fusion module, a local fusion module, and a downsampling operation. The encoder stage can be summarized as follows: in, represents the feature map generated by the i-th convolutional layer of the input patch, f Conv 3×3 represents a convolution operation with a kernel size of 3×3, LMM(·) and GMM(·) represent local fusion modules and global fusion modules, and all convolutional layers are followed by a batch normalization layer and a LeakyReLU activation function layer.

6. The self-supervised hyperspectral change detection method according to claim 4, characterized in that: In the global local feature aggregation codec, the decoding stage includes a detection decoder and a reconstruction decoder; the feature maps of the two branches of the encoder are differentiated to obtain a differential feature map; the differential feature map is input to the multi-layer upsampling detection decoder to restore the feature map to the original input size. The process of the detection decoder can be summarized as follows: in, represents the output feature map after the i-th deconvolution operation, f DConv 3×3 represents the deconvolution operation with a kernel size of 3×3, f Conv 1×1 represents a convolution operation with a kernel size of 1×1, [;] represents a stacking operation along the channel dimension, which realizes the skip connection of the feature map; At the same time, the output of the mask branch is input into the reconstruction decoder to obtain the reconstructed feature map of the original single-phase hyperspectral image. The process of the reconstruction decoder can be summarized as follows: in, Represents the reconstructed feature map after the i-th deconvolution operation; then, the obtained feature map is projected into another dimensional space using a multilayer perceptron that does not share weights. The formula for the projection process is as follows: Finally, the reconstructed feature map X is obtained rec And the detection feature map X det .

7. The self-supervised hyperspectral change detection method according to claim 5, characterized in that: In the encoding stage, in order to achieve effective aggregation of spatial spectral information and obtain discriminative differential features, the global fusion module and the local fusion module are embedded to realize remote information interaction and local information aggregation. In the global fusion module, first, given an input feature Divide the features into P blocks along the channel dimension and connect them along the columns to obtain A 3×3 convolution is applied to this feature map, which is then reduced to its original dimension. Next, the reduced feature map is fused with the initial feature map through a fully connected layer and a GELU function. The feature is then divided into P blocks along the channel dimension and connected along the rows to obtain Perform the same 3×3 convolution, fully connected layer and GELU function, and fuse with the initial feature map; The global fusion module slices and reorganizes the feature maps in different directions of rows and columns to make the spatial distances of pixels close to each other, thereby improving the interaction ability between remote information.

8. The self-supervised hyperspectral change detection method according to claim 5, characterized in that: The output of the global fusion module is input into the local fusion module for further feature extraction. First, 3×7 and 7×3 convolution operations are applied to the input feature map to extract local features in different directions. The formula is as follows: x add =x+f Conv 3×7 (x)+f Conv 7×3 (x) Next, global average pooling is used to compress the spatial dimensions of the input feature map, and then a multi-layer perceptron is used to compress and extract features, and the Swish function is used for activation. The formula for the attention process is as follows: Att(x)=Swish(GAP(MLP(x add ))) Where Swish(x) = xsigmoid(x), the final output of the local fusion module can be described as follows: LMM(x)=At(x)x add The feature map is multiplied by the shared channel attention weight, and the obtained feature map highlights the useful areas and suppresses the useless areas. Through the designed local fusion module, the local features are effectively aggregated to achieve the fusion of spatial and spectral attention.

9. The self-supervised hyperspectral change detection method according to claim 4, characterized in that: The output feature maps of the detection decoder and the corresponding projection head are selected for final detection. The final change detection result can be described as follows: Among them, y i To predict the probability result, fc(·) represents the fully connected layer for extracting features and reducing dimensionality; multi-scale masks are used to supervise the change detection results, and the formula for using the cross entropy function as the loss function is as follows: Where N is the number of samples, m is i is the mask label of a given sample; at the same time, a reconstruction loss is designed to reconstruct the output feature map of the decoder and the corresponding projection head. The formula of the reconstruction loss function is as follows: The total loss function is composed of the above loss functions as follows and is calculated as follows: L total =L sup +L rec Calculate the loss function, optimize the model parameters of the change detection framework according to the loss function and the back-propagation process, and obtain a trained change detection framework after the training is completed; use the trained change detection framework to discriminate the input samples and output a change detection result graph.

10. A self-supervised hyperspectral change detection system, characterized in that: The system is configured with a control program for implementing the method described in any one of claims 1 to 9.

Images