An optical satellite image pixel-level change detection method

Abstract

The application discloses an optical satellite image pixel-level change detection method, comprising 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, the multi-scale feature fusion strategy Psi extracts multi-scale features and outputs multi-scale feature maps through four types of nonlinear activation operations, and the multi-scale output structure Omega classifies and outputs a detection result through an activation layer on the feature map; S3, model training; S4, model optimization; S5, model prediction: fixing the convolutional neural network Phi, performing change detection on each pair of test images, and obtaining a change detection result map CBI. The method effectively resists the influence of registration errors and projection differences of remote sensing images between double-time-phase satellite remote sensing images, improves the precision of pixel-level change detection, and improves the processing efficiency of pixel-level change detection.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of remote sensing image processing, and particularly relates to an optical satellite image pixel-level change detection method. BACKGROUND

[0002] Existing remote sensing image change detection models perform well on strictly registered datasets, but do not perform well in actual applications. In addition to the model generalization ability factor, the following two reasons also limit the application range of remote sensing image change detection models:

[0003] 1. Most change detection models cannot balance the accuracy and efficiency of the change detection task. Many existing change detection models focus on the positioning accuracy of the change area, and ignore the processing efficiency. This greatly limits the application of change detection technology in actual tasks.

[0004] 2. Due to the registration error of the double-time satellite remote sensing images obtained in actual production and the projection error of the satellite remote sensing images, most change detection models perform poorly in actual production. SUMMARY

[0005] In view of the problems existing in the prior art, the application provides an optical satellite image pixel-level change detection method. The method is based on a multi-scale feature fusion strategy and a multi-scale output structure, constructs a convolutional neural network, and realizes pixel-level change detection through the convolutional neural network. The method can reduce the influence of remote sensing image projection error and registration error on the detection result, so as to solve the problem of inaccuracy in the process of remote sensing image pixel-level change detection.

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

[0007] An optical satellite image pixel-level change detection method comprises the following steps:

[0008] S1, data acquisition and preprocessing:

[0009] Obtain two large-scale remote sensing images a and b of different time periods, and preprocess the large-scale remote sensing images through strict geometric correction, orthorectification and resampling. The preprocessed large-scale remote sensing images a and b are used as input data;

[0010] S2, model design: the convolutional neural network model Φ designed by the application comprises a multi-scale feature fusion strategy Ψ and a multi-scale output structure Ω. The multi-scale feature fusion strategy Ψ comprises four types of spatial feature extraction operations. The input of the multi-scale feature fusion strategy Ψ is the input data of S1, and the output is a feature map at different scales.

[0011] The multi-scale output structure Ω = {S iEach branch of the} includes an activation layer, each activation layer includes a nonlinear activation operation, the nonlinear activation operation in each activation layer is same and is determined by the classifier Implementation, S i represents an output feature result obtained by the i-th nonlinear activation branch after processing, wherein 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: taking the change detection image pair {(a, b, CBI * ) z} as the input of the convolutional neural network model Φ, wherein z = 1, 2, … Z, Z represents the total number of change detection image pairs, and the feature maps L 1-0 , L 2-0 , L 3-0 and L 4-0 under different scales are obtained through the multi-scale feature fusion strategy Ψ described in S2; * CBI

[0015] Step II: performing high-level feature extraction on L 1-0 to obtain the feature map L 2-1 , concatenating L 2-0 and L 2-1 by feature concatenation operation to obtain the feature map L 2-c , performing high-level feature extraction on L 2-1 to obtain the feature map L 3-1 ;

[0016] Step III: performing high-level feature extraction on the feature map L 2-c through two branches to obtain the feature maps L 3-2 and L 4-3 , and performing channel concatenation processing on L 3-0 , L 3-1 and L 3-2 to obtain L 3-c ;

[0017] Step IV: performing high-level feature extraction on the feature map L 3-1 to obtain the feature map L 4-1 , performing high-level feature extraction on the feature map L 3-2 to obtain the feature map L 4-2 , performing high-level feature extraction on the feature map L 3-c to obtain the feature map L 4-4 , and performing high-level feature extraction on the feature map L 4-0 , L4-1 , L 4-2 , L 4-3 and L 4-4 perform channel concatenation processing, and obtain L 4-c ;

[0018] Step V: performing deconvolution operation and up-sampling operation on the feature map L 4-c , and concatenating with the feature map L 3-c , to obtain the feature map L 3-u , and repeating step V for the feature map groups {L 3-u , L 2-c} and {L 2-u , L 1-0}, to obtain L 2-u and L 1-u ;

[0019] Step VI: performing deconvolution operation on L 1-u , L 2-u , L 3-u and L 4-c , respectively, to obtain the feature map of the convolution layer wherein c represents the number of channels of the feature map, and the subscript number represents the convolution feature map corresponding to the h layer, and is subjected to dimension transformation, and c is reduced to 1 through dimension reduction processing, to obtain wherein 1 represents that the obtained feature map is a single-channel vector;

[0020] Step VII: inputting the feature maps of different scales after dimension reduction into the multi-scale output structure Ω of S2, performing two classification through the activation layer classifier in the multi-scale output structure Ω, outputting the pixel value change probability of the feature result, and binarizing the change probability result of all pixels to obtain the prediction result CBI of change detection;

[0021] S4, model optimization:

[0022] calculating the loss function through the similarity degree between the prediction result CBI and the true value CBI * , and updating the convolution neural network model Φ = {Θ, M, T} through back propagation, wherein 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 model parameters are updated to reduce the loss function and improve the network performance until the iteration stopping condition is met;

[0024] S5, model prediction: fixing the convolution neural network Φ, performing change detection for each pair of test images to obtain the change detection result map CBI of the corresponding size.

[0025] The multi-scale feature fusion strategy in step S2 The four types of spatial feature extraction operations include:

[0026] represents the output feature result obtained by the branch after processing via the r first-type spatial feature extraction branch;

[0027] represents the output feature result obtained by the branch after processing via the j second-type spatial feature extraction branch;

[0028] represents the output feature result obtained by the branch after processing via the k third-type spatial feature extraction branch;

[0029] represents the output feature result obtained by the branch after processing via the l fourth-type spatial feature extraction branch;

[0030] The convolution operation adopted by the first-type spatial feature extraction branch is a deformable convolution operation with a convolution kernel size of {3x3} and a stride of 1, and the output result of the processing branch is

[0031] The convolution operation adopted by the second-type spatial feature extraction branch is a convolution operation with a convolution kernel size of {3x3} and a convolution stride of 2, and the output result of the processing branch is

[0032] The convolution operation adopted by the third-type spatial feature extraction branch is first a convolution operation with a convolution kernel size of {3x3} and a convolution stride of 1, then a pooling operation, and then a convolution operation with a convolution kernel size of {3x3} and a convolution stride of 2, and the output result of the processing branch is

[0033] The convolution operation adopted by the fourth-type spatial feature extraction branch is first three convolution operations with a convolution kernel size of {3x3} and a convolution stride of 1, then a pooling operation, and then a convolution operation with a convolution kernel size of {3x3} and a convolution stride of 1, and the output result of the processing branch is

[0034] After processing by each spatial feature extraction branch, 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, wherein, represents the input feature vector, and 2 represents the binary classification of the input feature vector ​

[0036] Classifier The formula is as follows:

[0037]

[0038] wherein f i is the input of the classifier T, exp() is the logarithm function, F(f i ) is the result output of the classifier T, as a binary classification task, the output range of F(f i ) is [0, 1], and F(f i ) represents the change probability of the pixel in the mth row and the nth column of the feature map.

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

[0040] E = E bce + λE dc

[0041] wherein λ is a weight control parameter, used to regulate the proportion between E bce and E dc ;

[0042] The formula of the binary cross-entropy loss function E bce is as follows:

[0043]

[0044] wherein Q is the total number of pixels of the image a, y q represents the state of the qth pixel, y q＝1 represents the number of changed pixels in the image; y q＝0 represents the number of unchanged pixels; p q represents the change probability of the qth pixel;

[0045] The formula of the Dice coefficient loss function E dc is as follows:

[0046]

[0047] wherein Y represents the given change map true value, and represents the predicted change result map.

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

[0049] ​The four types of spatial feature extraction branches in the multi-scale feature fusion strategy have the same value, all being 1, r=j=k=l=1.

[0050] The pooling operation all adopts maximum pooling, and the step distance of the pooling operation is all 2.

[0051] The weight control parameter λ is tested multiple times with preset values, and λ is set to 0.5.

[0052] The optical satellite image pixel-level change detection method of the application 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 application has the following beneficial effects:

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

[0055] 2. The multi-scale output structure designed in the application not only can optimize the output result of the convolutional neural network model, but also can effectively resist the influence of registration errors between double-time satellite remote sensing images and projection difference existing in remote sensing images, and further improve the detection accuracy of the change detection model, which is more suitable for optical satellite image pixel-level change detection.

[0056] 3. The change detection model constructed in the application has better feature extraction capability and higher processing efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 is a multi-scale feature fusion strategy in the application;

[0058] Figure 2 is a deformable convolution structure in the application;

[0059] Figure 3 is a multi-scale output structure in the application;

[0060] Figure 4 is a convolutional neural network model in the application;

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

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

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

[0064] The detection method of the present application is described in detail below in combination with the drawings.

[0065] Referring to Figures 1-4 , the optical satellite image pixel-level change detection method of the present application comprises the following steps:

[0066] S1, data acquisition and preprocessing:

[0067] Acquire two large-scale remote sensing images a and b at different time periods, and preprocess the large-scale remote sensing images by strict geometric correction, orthorectification and resampling, taking the preprocessed large-scale remote sensing images a and b as input data.

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

[0069] As shown in Figure 1 , the multi-scale feature fusion strategy Ψ can extract features from the input image, and this processing process outputs multiple image features with different scales and feature levels.

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

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

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

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

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

[0075] As a preferred, the number of four types of spatial feature extraction branches in the multi-scale feature fusion strategy is the same, all of which are 1, i.e. r = j = k = l = 1.

[0076] The convolution operation adopted by the first-type spatial feature extraction branch is: a deformable convolution operation with a convolution kernel size of {3x3} and a step of 1, and the output result of this processing branch is

[0077] The convolution operation adopted by the second type of spatial feature extraction branch is: a convolution kernel size of {3x3} and a convolution stride of 2 are performed, and the output result of the processing branch is

[0078] The convolution operation adopted by the third type of spatial feature extraction branch is: first, a convolution kernel size of {3x3} and a convolution stride of 1 are performed, and then a pooling operation is performed; and then, a convolution kernel size of {3x3} and a convolution stride of 2 are performed, and the output result of the processing branch is

[0079] The convolution operation adopted by the fourth type of spatial feature extraction branch is: first, three convolution kernel sizes of {3x3} and a convolution stride of 1 are performed, and then a pooling operation is performed, and then a convolution kernel size of {3x3} and a convolution stride of 1 are performed, and the output result of the processing branch is

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

[0081] As preferred, all the above pooling operations adopt maximum pooling, and the stride of the pooling operation is 2.

[0082] Unlike the output of a general convolutional neural network, the multi-scale output structure Ω in the convolutional neural network designed by the present application processes all input image features of different scales through nonlinear activation operations respectively, and outputs feature maps of different scales. The multi-scale output structure Ω not only optimizes the output features of the convolutional neural network model, but also effectively resists the registration error between the two-phase satellite remote sensing images and the projection difference of the remote sensing images themselves.

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

[0084] The nonlinear activation operation included in each activation layer in the multi-scale output structure Ω={S i} is implemented by a classifier , which is designed according to a Sigmoid function, where ​represents an input feature map; 2 represents an input feature vector performing binary classification;

[0085] classifier The formula is as follows:

[0086]

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

[0088] The multi-scale output structure Ω = {S i} performs binary classification on the input feature map, and the output is the change probability of the pixel of the feature map.

[0089] S3, model training:

[0090] Figure 4 The optical satellite image pixel-level change detection model based on the multi-scale fusion strategy and the multi-scale output structure designed by the application is shown.

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

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

[0093] Step II: performing high-level feature extraction on L 1-0 to obtain the feature map L 2-1 . L 2-0 and L 2-1 are concatenated according to the feature concatenation operation to obtain the feature map L 2-c . High-level feature extraction is performed on L 2-1 to obtain the feature map L 3-1 .

[0094] Step III: High-level feature extraction is performed on the feature maps L 2-c to obtain feature maps L 3-2 and L 4-3 . Channel concatenation is performed on the feature maps L 3-0 , L 3-1 and L 3-2 to obtain L 3-c .

[0095] Step IV: High-level feature extraction is performed on the feature map L 3-1 to obtain feature map L 4-1 . High-level feature extraction is performed on the feature map L 3-2 to obtain feature map L 4-2 . High-level feature extraction is performed on the feature map L 3-c to obtain feature map L 4-4 . Channel concatenation is performed on the feature maps L 4-0 , L 4-1 , L 4-2 , L 4-3 and L 4-4 to obtain L 4-c .

[0096] Step V: Deconvolution and up-sampling operations are performed on the feature map L 4-c , and the feature map L 3-c is concatenated to obtain the feature map L 3-u . Steps V are repeated for the feature map groups {L 3-u , L 2-c} and {L 2-u , L 1-0} to obtain L 2-u and L 1-u , respectively.

[0097] Step VI: Deconvolution is performed on L 1-u , L 2-u , L 3-u and L 4-c to obtain the feature maps of the convolutional layers where c represents the number of channels of the feature map, and subscript h represents the convolutional feature map corresponding to the h layer. Dimension transformation is performed on to reduce c to 1 through dimension reduction to obtain where 1 indicates that the obtained feature map is a single-channel vector.

[0098] Step VII: Different scale feature maps after dimension reduction The input is input into the multi-scale output structure Ω, and the pixel value change probability of the feature map is output through the activation layer classifier in the multi-scale output structure Ω. The change detection prediction result CBI can be obtained by binarizing the change probability results of all pixels.

[0099] S4, model optimization:

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

[0101] As a preferred, the loss function used in the model training process is composed of the binary cross-entropy loss function E bce and the Dice coefficient loss function E dc , wherein E bce is defined as:

[0102]

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

[0104] y q＝0 represents the number of unchanged pixels; p q represents the change probability of the qth pixel;

[0105] E dc is defined as:

[0106]

[0107] Wherein, Y represents the given change map true value, represents the predicted change result map.

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

[0109] E=E bce +λE dc

[0110] Wherein, λ is a weight control parameter, which is used to control the proportion between E bce and E dc . The λ weight control parameter is tested multiple times, and as a preferred, it is set to 0.5.

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

[0112] S5, model prediction: the fixed convolutional neural network Φ can perform change detection on each pair of test images to obtain the corresponding size change detection result map CBI.

[0113] The existing detection method and the detection method of the present application are used for detection, wherein the input data of the convolutional neural network model Φ is two different time period large-scale remote sensing images a and b after preprocessing; the output data is the binary change detection result map, i.e. the binary image CBI. The input data a and b and the output data CBI have completely consistent image size, ground resolution and geographical coverage. In the output data, if the pixel value CBI(m,n) of the mth column and the nth row in the output image CBI is 0, it means that there is no change of ground objects at this position, and if CBI(m,n) = 1, it means that there is a change of ground objects at this position.

[0114] The existing detection method and the detection method of the present application are used for detection on the VHR data set, and the obtained change detection results are as shown in Figure 5 The existing detection method and the detection method of the present application are used for detection on the actual production data set, and the obtained change detection results are as shown in Figure 6 , wherein:

[0115] a is the pre-change image;

[0116] b is the post-change image;

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

[0118] d is the prediction result of the 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 the 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 the 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 application.

[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 application has stronger discrimination ability for small block ground objects, larger correct detection range and fewer false detection cases.

[0123] On the data with registration error, the methods (EMS-CDNet) also using the multi-scale feature fusion strategy and the detection method of the present application are used for detection respectively, and the change detection results obtained are as shown in Figure 7 , wherein:

[0124] a is the image before change;

[0125] b is the image after change;

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

[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 application.

[0129] To highlight the robustness of different detection methods to registration errors, Figure 7 The prediction results in d and e are superimposed with the post-change image. As can be seen from the circled parts in d and e, compared with the EMS-CDNet method which also uses a multi-scale feature fusion strategy, the detection method of the application is less affected by registration errors, and the edges of the changed areas detected are more consistent with the actual edges of the ground objects in the image, which also verifies that the combination of deformable convolution, multi-scale fusion strategy and multi-scale output structure in the application can effectively reduce the influence of registration errors on the detection results.

Claims

1. A method for detecting pixel-level changes in optical satellite images, characterized in that, Includes the following steps: S1, Data Acquisition and Preprocessing: Acquire large-scale remote sensing images from two different periods and The large-scale remote sensing image is preprocessed through rigorous geometric correction, orthorectification, and resampling. and As input data; S2, Model Design: Convolutional Neural Network Model Including multi-scale feature fusion strategies and multi-scale output structure Among them, multi-scale feature fusion strategy Includes four types of spatial feature extraction operations and multi-scale feature fusion strategies. The input is the input data described in S1, and the output is feature maps at different scales; Multi-scale output structure Each branch includes an activation layer, and each activation layer includes a non-linear activation operation. The non-linear activation operations in each activation layer are identical and are implemented by a classifier. Indicates via the first After processing by a nonlinear activation branch, the output feature result obtained by this branch is, where ; S3, Model Training: Convolutional Neural Network Model The training process includes the following sub-steps: Step 1: Pair the change detection images As a convolutional neural network model The input, where Z represents the total number of change detection image pairs, which are determined by the multi-scale feature fusion strategy described in S2. Feature maps at different scales are obtained. , , and , The true value represents a manually labeled change detection patch. Step II: For High-level feature extraction is performed to obtain feature maps. ,Will and By performing feature cascading operations, we obtain the feature map. ,right High-level feature extraction is performed to obtain feature maps. ; Step III: Process the feature map through two branches respectively. High-level feature extraction is performed to obtain feature maps. and , feature map and Perform channel cascading processing to obtain ; Step IV: Process the feature map High-level feature extraction is performed to obtain feature maps. For feature maps High-level feature extraction is performed to obtain feature maps. , feature map High-level feature extraction is performed to obtain feature maps. , feature map , , , and Perform channel cascading processing to obtain ; Step V: Process the feature map Perform deconvolution and upsampling operations, and then compare them with the feature map. Cascade to obtain feature maps For feature map group { }and{ , Repeat step V to obtain the following results sequentially. and ; Step VI: For each , , and Perform deconvolution to obtain the feature map of the convolutional layer. Where c represents the number of channels in the feature map, and the subscript... Indicates the first The feature map corresponding to the layer, for By performing a dimensionality transformation and reducing c to 1, we obtain... , where 1 indicates that the obtained feature map is a single-channel vector; Step VII: Convert the dimensionality-reduced feature map Input to the multi-scale output structure described in S2 In the middle, through multi-scale output structure The activation layer classifier performs binary classification, outputting the pixel value change probability of the feature results. The change probability results of all pixels are binarized to obtain the prediction result of change detection. ; S4, Model Optimization: Based on the prediction results truth value The loss function is calculated based on the similarity between the components, and the convolutional neural network model is updated using backpropagation. ,in, Indicates the parameters of the model to be trained; This represents the output network feature map; This represents a change detection classifier; The above training and optimization process needs to be carried out iteratively, by continuously updating the model parameters to reduce the loss function and improve network performance until the iteration stopping condition is met; S5, Model Prediction: Fixed Convolutional Neural Network For each image to be tested, change detection is performed to obtain the change detection result image of the corresponding size. ; The multi-scale feature fusion strategy It includes four types of spatial feature extraction operations, among which: , , and They represent respectively via After processing by the first type of spatial feature extraction branch, via After processing by the second type of spatial feature extraction branch, via After processing by the third type of spatial feature extraction branch and via The output feature results are obtained after processing by the fourth type of spatial feature extraction branch; The first type of spatial feature extraction branch uses deformable convolution operation with a kernel size of {3×3} and a stride of 1. The output result is... ; In the convolution operation used in the second type of spatial feature extraction branch, the kernel size is {3×3} and the stride is 2. The output result is... ; The third type of spatial feature extraction branch uses the following convolution operations in sequence: a convolution operation with a kernel size of {3×3} and a stride of 1, a pooling operation, and a convolution operation with a kernel size of {3×3} and a stride of 2. The output of this branch is: ; The fourth type of spatial feature extraction branch uses the following convolution operations in sequence: three convolution operations with kernel size {3×3} and stride of 1, a pooling operation, and a convolution operation with kernel size {3×3} and stride of 1. The output of this branch is: ; After processing each spatial feature extraction branch, the output result is pooled to obtain the output feature map of that processing branch.

2. The method for detecting pixel-level changes in optical satellite images according to claim 1, characterized in that: The classifier described in step S2 Based on the design of the Sigmoid function, where, 2 represents the input feature map, and 2 represents the input feature vector. Perform binary classification; Classifier The formula is as follows: , in, For classifier The input is exp(), which is the logarithmic function. For classifier The output result is used as a binary classification task. The output range is [0,1]. Representation of feature map The Middle Line number The probability of change of column pixels.

3. The method for detecting pixel-level changes in optical satellite images according to claim 1, characterized in that: The loss function described in step S4 is the binary classification cross-entropy loss function. With Dice coefficient loss function Combining the above, the formula is as follows: ， in, These are weight control parameters used to adjust... and The ratio between them; Binary cross-entropy loss function The formula is: ， Where Q represents the image. The total number of pixels, Indicates the first The state of each pixel Indicates the number of pixels that have changed in the image; Indicates the number of pixels that remained unchanged; Indicates the first The probability of a pixel changing; Dice coefficient loss function The formula is: ， in, Represents the true value of a given change graph. A graph showing the predicted changes.

4. The method for detecting pixel-level changes in optical satellite images according to claim 1, characterized in that: The iteration stopping condition in step S4 is: at least 40 iterations.

5. The method for detecting pixel-level changes in optical satellite images according to claim 1, characterized in that: The four types of spatial feature extraction branches in the multi-scale feature fusion strategy have the same number of values. .

6. The method for detecting pixel-level changes in optical satellite images according to claim 1, characterized in that: All pooling operations employ maximum pooling, and the step size for each pooling operation is 2.

7. The method for detecting pixel-level changes in optical satellite images according to claim 3, characterized in that: The weight control parameters This was obtained after multiple preset values ​​and tests. Set it to 0.5.

Images