Optical satellite image pixel level change detection method

Abstract

The invention discloses an optical satellite image pixel level change detection method. The method comprises the following steps: S1, data acquisition and preprocessing; s2, designing a convolutional neural network model phi: the convolutional neural network model phi comprises a multi-scale feature fusion strategy psi and a multi-scale output structure Omega, and the multi-scale feature fusion strategy psi extracts multi-scale features through four types of nonlinear activation operations to output a multi-scale feature map, the multi-scale output structure Omega classifies the feature map through the activation layer and outputs a detection result; s3, model training; s4, optimizing the model; and S5, model prediction: fixing the convolutional neural network phi, and performing change detection on each to-be-tested image to obtain a change detection result graph CBI. The method effectively resists the influence caused by the registration error between the dual-time-phase satellite remote sensing images and the projection difference of the remote sensing images, improves the precision of pixel-level change detection, and improves the processing efficiency of the pixel-level change detection.

Description

Technical Field

[0001] The invention belongs to the field of remote sensing image processing, and in particular relates to a method for detecting pixel-level changes in optical satellite images. Background Art

[0002] Existing remote sensing image change detection models show good performance on strictly registered datasets, but poor results in practical applications. In addition to the model generalization ability factor, there are two other reasons that limit the application scope of remote sensing image change detection models:

[0003] 1. Most change detection models cannot take into account both the accuracy and efficiency of change detection tasks. Many existing change detection models focus on the localization accuracy of the detected change area, while ignoring the processing efficiency. This greatly limits the application of change detection technology in practical tasks.

[0004] 2. Due to the registration errors of dual-phase satellite remote sensing images obtained in actual production and the projection differences of satellite remote sensing images, most change detection models perform poorly in actual production. Summary of the invention

[0005] In view of the problems existing in the prior art, the present invention provides a method for detecting pixel-level changes in optical satellite images. The method is based on a multi-scale feature fusion strategy and a multi-scale output structure, and a convolutional neural network is constructed to realize pixel-level change detection through the convolutional neural network. The method can reduce the influence of remote sensing image projection difference and registration error on the detection results, so as to solve the problem of inaccuracy in the process of remote sensing image pixel-level change detection.

[0006] To this end, the present invention adopts the following technical solutions:

[0007] A method for detecting pixel-level changes in optical satellite images comprises the following steps:

[0008] S1, Data acquisition and preprocessing:

[0009] Acquire large-scale remote sensing images a and b at two different periods, preprocess the large-scale remote sensing images through strict geometric correction, orthorectification and resampling, and use the preprocessed large-scale remote sensing images a and b as input data;

[0010] S2, model design: The convolutional neural network model Φ designed by the present invention includes a multi-scale feature fusion strategy Ψ and a multi-scale output structure Ω, wherein the multi-scale feature fusion strategy Ψ includes four types of spatial feature extraction operations, the input of the multi-scale feature fusion strategy Ψ is the input data described in S1, and the output is feature maps at different scales;

[0011] Multi-scale output structure Ω={S iEach branch of} includes an activation layer, each activation layer includes a nonlinear activation operation, and the nonlinear activation operations in each activation layer are the same and are generated by the classifier Implementation, S i It represents the output feature result obtained by the branch after being processed by the i-th nonlinear activation branch, where i = 1, 2, 3, 4;

[0012] S3, model training:

[0013] The training process of the convolutional neural network model Φ includes the following sub-steps:

[0014] Step I: Change detection image pair {(a, b, CBI * ) z} as the input of the convolutional neural network model Φ, where z = 1, 2, ... Z, Z represents the total number of change detection image pairs, and the feature maps L at different scales are obtained through the multi-scale feature fusion strategy Ψ described in S2 1-0 , L 2-0 , L 3-0 and L 4-0 , CBI * is the true value, which is the manually annotated change detection patch;

[0015] Step II: L 1-0 Perform high-level feature extraction to obtain feature map L 2-1 , L 2-0 and L 2-1 According to the feature cascade operation, the feature map L is obtained 2-c , for L 2-1 Perform high-level feature extraction to obtain feature map L 3-1 ;

[0016] Step III: Through two branches, the feature map L 2-c Perform high-level feature extraction to obtain feature map L 3-2 and L 4-3 , the feature map L 3-0 , L 3-1 and L 3-2 Perform channel cascade processing to obtain L 3-c ;

[0017] Step IV: For feature map L 3-1 Perform high-level feature extraction to obtain feature map L 4-1 , for the feature map L 3-2 Perform high-level feature extraction to obtain feature map L 4-2 , the feature map L 3-c Perform high-level feature extraction to obtain feature map L 4-4 , the feature map L 4-0 , L4-1 , L 4-2 , L 4-3 and L 4-4 Perform channel cascade processing to obtain L 4-c ;

[0018] Step V: For feature map L 4-c Perform deconvolution and upsampling operations, and compare them with the feature map L 3-c Cascade to get the feature map L 3-u , for the feature map group {L 3-u ,L 2-c} and {L 2-u , L 1-0} Repeat step V to get L 2-u and L 1-u ;

[0019] Step VI: L 1-u , L 2-u , L 3-u and L 4-c Perform deconvolution to obtain the feature map of the convolution layer Among them, c represents the number of channels of the feature map, and the subscript number represents the convolutional feature map corresponding to the hth layer. After dimension transformation, c is reduced to 1 through dimensionality reduction, and we get: Where 1 means that the obtained feature map is a single-channel vector;

[0020] Step VII: Reduce the feature maps of different scales Input into the multi-scale output structure Ω described in S2, perform binary classification through the activation layer classifier in the multi-scale output structure Ω, output the pixel value change probability of the feature result, binarize the change probability results of all pixels, and obtain the prediction result CBI of the change detection;

[0021] S4, model optimization:

[0022] Through the predicted result CBI and the true value CBI * The loss function is calculated based on the similarity between them, and the convolutional neural network model Φ = {Θ, M, T} is updated by back propagation, where Θ represents the model parameters to be trained; M represents the output network feature map; and T represents the change detection classifier.

[0023] The above training and optimization process needs to be iterated, and the loss function is reduced and the network performance is improved by continuously updating the model parameters until the iteration stop condition is met;

[0024] S5, model prediction: fix the convolutional neural network Φ, perform change detection on each image to be tested, and obtain the change detection result map CBI of the corresponding size.

[0025] The multi-scale feature fusion strategy described in step S2 It includes four types of spatial feature extraction operations, among which:

[0026] It represents the output feature result obtained by the branch after being processed by r first-class spatial feature extraction branches;

[0027] It represents the output feature result obtained by the branch after being processed by j second-type spatial feature extraction branches;

[0028] It represents the output feature result obtained by the branch after being processed by k third-category spatial feature extraction branches;

[0029] It represents the output feature result obtained by the branch after being processed by l fourth-category spatial feature extraction branches;

[0030] The convolution operation used by the first type of spatial feature extraction branch is: a deformable convolution operation with a convolution kernel size of {3×3} and a step size of 1 is performed. The output result of this processing branch is

[0031] The convolution operation used by the second type of spatial feature extraction branch is: a convolution kernel size of {3×3} and a convolution step of 2 are performed. The output result of this processing branch is

[0032] The convolution operation used by the third type of spatial feature extraction branch is: first perform a convolution operation with a convolution kernel size of {3×3} and a convolution stride of 1, then perform a pooling operation; then perform a convolution operation with a convolution kernel size of {3×3} and a convolution stride of 2. The output result of this processing branch is

[0033] The convolution operation used by the fourth type of spatial feature extraction branch is: first perform three convolution operations with the size of the convolution kernels being {3×3} and the convolution stride being 1, then perform a pooling operation, and then perform a convolution operation with the size of the convolution kernel being {3×3} and the convolution stride being 1. The output result of this processing branch is

[0034] After each spatial feature extraction branch is processed, a pooling operation is performed on the output result to obtain the output feature map of the processing branch.

[0035] The classifier in step S2 According to the Sigmoid function design, represents the input feature vector, and 2 represents the input feature vector Perform binary classification;

[0036] Classifier The formula is as follows:

[0037]

[0038] Among them, f i is the input of the classifier T, exp() is the logarithmic function, F(f i ) is the result output of the classifier T. As a binary classification task, F(f i ) has an output range of [0,1], F(f i ) represents the feature map The change probability of the pixel in the mth row and nth column in .

[0039] The loss function in step S4 is composed of the binary cross entropy loss function E bce And Dice coefficient loss function E dc The formula is as follows:

[0040] E=E bce +λE dc

[0041] Among them, λ is the weight control parameter, which is used to adjust E bce With E dc The ratio between

[0042] Binary cross entropy loss function E bce The formula is:

[0043]

[0044] Where Q is the total number of pixels of image a, y q represents the state of the qth pixel, y q＝1 Indicates the number of pixels that change in the image; y q＝0 Indicates the number of pixels that have not changed; p q Indicates the probability of the qth pixel changing;

[0045] Dice coefficient loss function E dc The formula is:

[0046]

[0047] Among them, Y represents the true value of the given change graph, A graph showing the predicted change results.

[0048] The iteration stopping condition in step S4 is: at least 40 iterations.

[0049] The number of branches for extracting four types of spatial features in the multi-scale feature fusion strategy is the same, which is 1, r=j=k=l=1.

[0050] The pooling operations all adopt maximum pooling, and the step size of the pooling operations is 2.

[0051] The weight control parameter λ is tested by setting a preset value multiple times, and λ is set to 0.5.

[0052] The optical satellite image pixel level change detection method of the present invention realizes the detection of optical satellite image pixel level change by constructing a convolutional neural network model Φ.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] 1. The multi-scale feature fusion strategy proposed in this paper combines the deformable convolution operation, which can effectively resist the registration error between satellite images, effectively fuse multi-scale feature information, and thus improve the accuracy of pixel-level change detection.

[0055] 2. The multi-scale output structure designed in the present invention can not only optimize the output results of the convolutional neural network model, but also effectively resist the influence of the registration error between the dual-phase satellite remote sensing images and the projection difference of the remote sensing images themselves, thereby improving the detection accuracy of the change detection model and being more suitable for pixel-level change detection of optical satellite images.

[0056] 3. The change detection model constructed by the present invention has better feature extraction capability and more efficient processing efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a schematic diagram of the multi-scale feature fusion strategy in the present invention;

[0058] Figure 2 It is a deformable convolution structure diagram in the present invention;

[0059] Figure 3 is a schematic diagram of a multi-scale output structure in the present invention;

[0060] Figure 4 It is the convolutional neural network model in the present invention;

[0061] Figure 5 It is the test result of the existing detection method and the detection method of the present invention on the VHR data set;

[0062] Figure 6 It is the test result of the existing detection method and the detection method of the present invention on the actual production data set;

[0063] Figure 7It is the test result of the existing detection method and the detection method of the present invention on the data with registration error. DETAILED DESCRIPTION

[0064] The detection method of the present invention is described in detail below with reference to the accompanying drawings.

[0065] See also Figure 1-Figure 4 , a method for detecting pixel-level changes in optical satellite images of the present invention comprises the following steps:

[0066] S1, Data acquisition and preprocessing:

[0067] Large-scale remote sensing images a and b are collected at two different periods, and the large-scale remote sensing images are preprocessed through strict geometric correction, orthorectification and resampling, and the preprocessed large-scale remote sensing images a and b are used as input data.

[0068] S2, model design: The convolutional neural network model Φ designed in the present invention includes a multi-scale feature fusion strategy Ψ and a multi-scale output structure Ω.

[0069] like Figure 1 As shown in Figure 1, the multi-scale feature fusion strategy Ψ can extract features from the input image. This processing will output image features with multiple scales and feature levels.

[0070] Multi-scale feature fusion strategy It includes four types of image processing operations, all of which are spatial feature extraction operations. Among them:

[0071] It represents the output feature result obtained by the branch after being processed by r first-class spatial feature extraction branches;

[0072] It represents the output feature result obtained by the branch after being processed by j second-type spatial feature extraction branches;

[0073] It represents the output feature result obtained by the branch after being processed by k third-category spatial feature extraction branches;

[0074] It represents the output feature result obtained by the branch after being processed by l fourth-category spatial feature extraction branches.

[0075] Preferably, the number of branches for extracting the four types of spatial features in the multi-scale feature fusion strategy is the same, which is 1, that is, r=j=k=l=1.

[0076] The convolution operation used by the first type of spatial feature extraction branch is: a deformable convolution operation with a convolution kernel size of {3×3} and a step size of 1 is performed. The output result of this processing branch is

[0077] The convolution operation used by the second type of spatial feature extraction branch is: a convolution kernel size of {3×3} and a convolution step of 2 are performed. The output result of this processing branch is

[0078] The convolution operation used by the third type of spatial feature extraction branch is: first perform a convolution operation with a convolution kernel size of {3×3} and a convolution stride of 1, then perform a pooling operation; then perform a convolution operation with a convolution kernel size of {3×3} and a convolution stride of 2. The output result of this processing branch is

[0079] The convolution operation used by the fourth type of spatial feature extraction branch is: first perform three convolution operations with the size of the convolution kernels being {3×3} and the convolution stride being 1, then perform a pooling operation, and then perform a convolution operation with the size of the convolution kernel being {3×3} and the convolution stride being 1. The output result of this processing branch is

[0080] After each spatial feature extraction branch is processed, a pooling operation is performed on the output result to obtain the output feature map of the processing branch.

[0081] Preferably, all the above pooling operations adopt maximum pooling, and the step size of the pooling operations is 2.

[0082] Different from the output of a general convolutional neural network, the multi-scale output structure Ω in the convolutional neural network designed by the present invention processes all input image features of different scales respectively through nonlinear activation operations and outputs feature maps of different scales. The multi-scale output structure Ω can not only optimize the output features of the convolutional neural network model, but also effectively resist the registration error between dual-phase satellite remote sensing images and the projection difference of the remote sensing image itself.

[0083] like Figure 3 As shown, the multi-scale output structure Ω={S i Each branch of} includes an activation layer, each activation layer includes a nonlinear activation operation, and the nonlinear activation operations in each activation layer are the same. i It represents the output feature result obtained by the branch after being processed by the i-th nonlinear activation branch, where i = 1, 2, 3, 4.

[0084] Multi-scale output structure Ω={S i Each activation layer in} includes a nonlinear activation operation composed of the classifier Implementation, Classifier According to the Sigmoid function design, Represents the input feature map; 2 represents the input feature vector Perform binary classification;

[0085] Classifier The formula is as follows:

[0086]

[0087] Among them, f i is the input of the classifier T, that is exp() is the logarithmic function, F(f i ) is the result output of the classifier T. As a binary classification task, F(f i ) has an output range of [0,1], F(f i ) represents the feature map The change probability of the pixel in the mth row and nth column in .

[0088] Multi-scale output structure Ω={S i}The input feature map is classified into two categories, and the output is the pixel change probability of the feature map.

[0089] S3, model training:

[0090] Figure 4 The figure shows the optical satellite image pixel-level change detection model designed by the present invention based on a multi-scale fusion strategy and a multi-scale output structure.

[0091] The training process of the convolutional neural network model Φ includes the following sub-steps:

[0092] Step I: Change detection image pair {(a, b, CBI * ) z} as the input of the convolutional neural network model Φ, where z = 1, 2, ... Z, Z represents the total number of change detection image pairs, and the input feature maps L at different scales are obtained through the multi-scale feature fusion strategy Ψ 1-0 , L 2-0 , L 3-0 and L 4-0 .CBI * is the true value, which is the manually annotated change detection patch.

[0093] Step II: L 1-0 Perform high-level feature extraction to obtain feature map L 2-1 . 2-0 and L 2-1 According to the feature cascade operation, the feature map L is obtained 2-c . 2-1 Perform high-level feature extraction to obtain feature map L 3-1 .

[0094] Step III: Through two branches, the feature map L 2-c Perform high-level feature extraction to obtain feature map L 3-2 and L 4-3 . The feature map L 3-0 , L 3-1 and L 3-2 Perform channel cascade processing to obtain L 3-c .

[0095] Step IV: For feature map L 3-1 Perform high-level feature extraction to obtain feature map L 4-1 . For the feature map L 3-2 Perform high-level feature extraction to obtain feature map L 4-2 . The feature map L 3-c Perform high-level feature extraction to obtain feature map L 4-4 . The feature map L 4-0 , L 4-1 , L 4-2 , L 4-3 and L 4-4 Perform channel cascade processing to obtain L 4-c .

[0096] Step V: For feature map L 4-c Perform deconvolution and upsampling operations, and compare them with the feature map L 3-c Cascade to get the feature map L 3-u For the feature map group {L 3-u ,L 2-c} and {L 2-u , L 1-0} Repeat step V to get L 2-u and L 1-u .

[0097] Step VI: L 1-u , L 2-u , L 3-u and L 4-c Perform deconvolution to obtain the feature map of the convolution layer Where c represents the number of channels of the feature map, and the subscript h represents the convolutional feature map corresponding to the hth layer. After dimension transformation, c is reduced to 1 through dimensionality reduction, and we get: Where 1 means that the obtained feature map is a single-channel vector.

[0098] Step VII: Reduce the feature maps of different scales The input is input into the multi-scale output structure Ω, and the activation layer classifier in the multi-scale output structure Ω performs binary classification, outputs the probability of pixel value change of the feature map, and binarizes the change probability results of all pixels to obtain the prediction result CBI of change detection.

[0099] S4, model optimization:

[0100] Through the predicted result CBI and the true value CBI * The loss function is calculated based on the similarity between them, and the convolutional neural network model Φ = {Θ, M, T} is updated using back propagation. Among them, Θ represents the model parameters to be trained; M represents the output network feature map; and T represents the change detection classifier.

[0101] As a preferred method, the loss function used in the model training process is the binary cross entropy loss function E bce And Dice coefficient loss function E dc Combination of E bce Defined as:

[0102]

[0103] Where Q is the total number of pixels in image a. q represents the state of the qth pixel, y q＝1 Indicates the number of pixels that change in the image;

[0104] y q＝0 Indicates the number of pixels that have not changed; p q Indicates the probability of the qth pixel changing;

[0105] E dc Defined as:

[0106]

[0107] Among them, Y represents the true value of the given change graph, A graph showing the predicted change results.

[0108] Preferably, the loss function used in the model training process can be defined as:

[0109] E=E bce +λE dc

[0110] Among them, λ is the weight control parameter, which is used to adjust E bce With E dc The ratio between them, the λ weight control parameter is tested multiple times with preset values, and as a preferred setting, is set to 0.5.

[0111] The above training and optimization process needs to be iterated, by continuously updating the model parameters to reduce the loss function and improve the network performance until the iteration stop condition is met (at least 40 iterations).

[0112] S5, model prediction: fix the convolutional neural network Φ, perform change detection on each image to be tested, and obtain the change detection result map CBI of its corresponding size.

[0113] The detection is performed by the existing detection method and the detection method of the present invention, wherein the input data of the convolutional neural network model Φ is two large-scale remote sensing images a and b at different periods after preprocessing; the output data is a binary change detection result map, i.e., a binary image CBI. The input data a and b and the output data CBI have completely consistent image size, ground resolution, and geographic coverage. In the output data, if the pixel value CBI(m,n)=0 in the mth column and nth row of the output image CBI, it means that no ground object change has occurred at the location; if CBI(m,n)=1, it means that a ground object change has occurred at the location.

[0114] The VHR data set is tested by the existing detection method and the detection method of the present invention, and the obtained change detection results are as follows: Figure 5 The actual production data set is tested by the existing detection method and the detection method of the present invention, and the change detection results obtained are as follows: Figure 6 As shown, where:

[0115] a is the image before the change;

[0116] b is the image after the change;

[0117] c is the true value of the manually labeled change;

[0118] d is the prediction result of FC-EF (Daudt, Rodrigo Caye, Bertr Le Saux, and Alexandre Boulch. "Fully convolutional siamese networks for change detection." 2018 25th IEEE International Conference on Image Processing (ICIP). IEEE, 2018.) method;

[0119] e is the prediction result of SNUNet-CD / 16 (Fang, Sheng, et al. "SNUNet-CD: A densely connected Siamese network for change detection of VHR images." IEEE Geoscience and Remote Sensing Letters 19(2021): 1-5.) method;

[0120] f is the prediction result of Unet++_MSOF (Fang, Sheng, et al. "SNUNet-CD: A densely connected Siamese network for change detection of VHR images." IEEE Geoscience and Remote Sensing Letters 19(2021): 1-5.) method;

[0121] g is the prediction result of the detection method of the present invention.

[0122] from Figure 5 and Figure 6 It can be seen that compared with the FC-EF, SNUNet-CD / 16, and Unet++_MSOF methods, the detection method of the present invention has a stronger ability to distinguish small objects, a larger range of correct detection, and fewer false detections.

[0123] On the data with registration errors, the method (EMS-CDNet) that also adopts the multi-scale feature fusion strategy and the detection method of the present invention are used to perform detection respectively, and the obtained change detection results are as follows: Figure 7 As shown, where:

[0124] a is the image before the change;

[0125] b is the image after the change;

[0126] c is the true value of the manually labeled change;

[0127] d is the prediction result of the EMS-CDNet (Zheng, Zhi, et al. "EMS-CDNet: an efficient multi-scale-fusion change detection network for very high-resolution remote sensing images." International Journal of Remote Sensing 43.14(2022):5252-5279.) method;

[0128] e is the prediction result of the detection method of the present invention.

[0129] To highlight the robustness of different detection methods to registration errors, Figure 7 The prediction results in are superimposed with the changed image. It can be seen from the circled parts in d and e that, compared with the EMS-CDNet method which also uses the multi-scale feature fusion strategy, the detection method of the present invention is less affected by the registration error, and the edge of the detected changed area is more closely aligned with the actual edge of the object in the image. This also verifies that the present invention combines deformable convolution, multi-scale fusion strategy and multi-scale output structure to effectively reduce the impact of the registration error on the detection results.

Claims

1. A method for detecting pixel-level changes in optical satellite images, characterized in that: The following steps are involved: S1, Data acquisition and preprocessing: Acquire large-scale remote sensing images a and b at two different periods, preprocess the large-scale remote sensing images through strict geometric correction, orthorectification and resampling, and use the preprocessed large-scale remote sensing images a and b as input data; S2, model design: The convolutional neural network model Φ designed by the present invention includes a multi-scale feature fusion strategy Ψ and a multi-scale output structure Ω, wherein the multi-scale feature fusion strategy Ψ includes four types of spatial feature extraction operations, the input of the multi-scale feature fusion strategy Ψ is the input data described in S1, and the output is feature maps at different scales; Multi-scale output structure Ω={S i Each branch of} includes an activation layer, each activation layer includes a nonlinear activation operation, and the nonlinear activation operations in each activation layer are the same and are generated by the classifier Implementation, S i It represents the output feature result obtained by the branch after being processed by the i-th nonlinear activation branch, where i = 1, 2, 3, 4; S3, model training: The training process of the convolutional neural network model Φ includes the following sub-steps: Step I: Change detection image pair {(a, b, CBI * ) z } as the input of the convolutional neural network model φ, where z = 1, 2, ... Z, Z represents the total number of change detection image pairs, and the feature maps L at different scales are obtained through the multi-scale feature fusion strategy ψ described in S2 1-0 , L 2-0 , L 3-0 and L 4-0 , CBI * is the true value, which is the manually annotated change detection patch; Step II: L 1-0 Perform high-level feature extraction to obtain feature map L 2-1 , L 2-0 and L 2-1 According to the feature cascade operation, the feature map L is obtained 2-c , for L 2-1 Perform high-level feature extraction to obtain feature map L 3-1 ; Step III: Through two branches, the feature map L is 2-c Perform high-level feature extraction and obtain feature maps L 3-2 and L 4-3 , the feature map L 3-0 , L 3-1 and L 3-2 Perform channel cascade processing to obtain L 3-c ; Step IV: For feature map L 3-1 Perform high-level feature extraction to obtain feature map L 4-1 , for the feature map L 3-2 Perform high-level feature extraction to obtain feature map L 4-2 , the feature map L 3-c Perform high-level feature extraction to obtain feature map L 4-4 , the feature map L 4-0 , L 4-1 , L 4-2 , L 4-3 and L 4-4 Perform channel cascade processing to obtain L 4-c ; Step V: For feature map L 4-c Perform deconvolution and upsampling operations, and compare them with the feature map L 3-c Cascade to get the feature map L 3-u , for the feature map group {L 3-u ,L 2-c } and {L 2-u , L 1-0 } Repeat step V to get L 2-u and L 1-u ; Step VI: L 1-u , L 2-u , L 3-u and L 4-c Perform deconvolution to obtain the feature map of the convolution layer Among them, c represents the number of channels of the feature map, and the subscript h represents the feature map corresponding to the hth layer. After dimension transformation, c is reduced to 1 through dimensionality reduction, and we get: Where 1 means that the obtained feature map is a single-channel vector; Step VII: Reduce the feature map Input into the multi-scale output structure Ω described in S2, perform binary classification through the activation layer classifier in the multi-scale output structure Ω, output the pixel value change probability of the feature result, binarize the change probability results of all pixels, and obtain the prediction result CBI of the change detection; S4, model optimization: Through the predicted result CBI and the true value CBI * The loss function is calculated based on the similarity between them, and the convolutional neural network model φ={Θ,M,T} is updated by back propagation, where Θ represents the model parameters to be trained; M represents the output network feature map; T represents the change detection classifier; The above training and optimization process needs to be iterated, and the loss function is reduced and the network performance is improved by continuously updating the model parameters until the iteration stop condition is met; S5, model prediction: fix the convolutional neural network Φ, perform change detection on each image to be tested, and obtain the change detection result map CBI of the corresponding size.

2. The optical satellite image pixel-level change detection method according to claim 1, characterized in that: The multi-scale feature fusion strategy described in step S2 It includes four types of spatial feature extraction operations, among which: It represents the output feature result obtained by the branch after being processed by r first-class spatial feature extraction branches; It represents the output feature result obtained by the branch after being processed by j second-type spatial feature extraction branches; It represents the output feature result obtained by the branch after being processed by k third-category spatial feature extraction branches; It represents the output feature result obtained by the branch after being processed by l fourth-category spatial feature extraction branches; The convolution operation used by the first type of spatial feature extraction branch is: a deformable convolution operation with a convolution kernel size of {3×3} and a step size of 1 is performed. The output result of this processing branch is The convolution operation used by the second type of spatial feature extraction branch is: a convolution kernel size of {3×3} and a convolution step of 2 are performed. The output result of this processing branch is The convolution operation used by the third type of spatial feature extraction branch is: first perform a convolution operation with a convolution kernel size of {3×3} and a convolution stride of 1, then perform a pooling operation; then perform a convolution operation with a convolution kernel size of {3×3} and a convolution stride of 2. The output result of this processing branch is The convolution operation used by the fourth type of spatial feature extraction branch is: first perform three convolution operations with the size of the convolution kernels being {3×3} and the convolution stride being 1, then perform a pooling operation, and then perform a convolution operation with the size of the convolution kernel being {3×3} and the convolution stride being 1. The output result of this processing branch is After each spatial feature extraction branch is processed, a pooling operation is performed on the output result to obtain the output feature map of the processing branch.

3. The optical satellite image pixel-level change detection method according to claim 1, characterized in that: The classifier in step S2 According to the Sigmoid function design, Represents the input feature map, 2 represents the input feature vector Perform binary classification; Classifier The formula is as follows: Among them, f i is the input of the classifier T, exp() is the logarithmic function, F(f i ) is the result output of the classifier T. As a binary classification task, F(f i ) has an output range of [0,1], F(f i ) represents the feature map The change probability of the pixel in the mth row and nth column in .

4. The optical satellite image pixel-level change detection method according to claim 1, characterized in that: The loss function in step S4 is composed of the binary cross entropy loss function E bce And Dice coefficient loss function E dc The formula is as follows: And=And bce +λE dc Among them, λ is the weight control parameter, which is used to adjust E bce With E dc The ratio between Binary cross entropy loss function E bce The formula is: Where Q is the total number of pixels of image a, y q represents the state of the qth pixel, y q＝1 Indicates the number of pixels that change in the image; y q＝0 Indicates the number of pixels that have not changed; p q Indicates the probability of the qth pixel changing; Dice coefficient loss function E dc The formula is: Among them, Y represents the true value of the given change graph, A graph showing the predicted change results.

5. The optical satellite image pixel-level change detection method according to claim 1, characterized in that: The iteration stopping condition in step S4 is: at least 40 iterations.

6. The optical satellite image pixel-level change detection method according to claim 2, characterized in that: The number of branches for extracting four types of spatial features in the multi-scale feature fusion strategy is the same, which is 1, r=j=k=l=1.

7. The optical satellite image pixel level change detection method according to claim 2, characterized in that: The pooling operations all adopt maximum pooling, and the step size of the pooling operations is 2.

8. The optical satellite image pixel-level change detection method according to claim 4, characterized in that: The weight control parameter λ is tested by setting a preset value multiple times, and λ is set to 0.5.

Images