A deep learning-based optical remote sensing image building change detection method

Abstract

The application discloses a kind of based on deep learning's optical remote sensing image building change detection method, belong to remote sensing image change detection field.Firstly, optical remote sensing image data is obtained from high two and is preprocessed and change information label Label is marked;Then the image is input into the change detection model of the application, and the detection result map is output after calculation.The application effectively optimizes and improves the problems that the existing change detection method has lower detection accuracy, the detection edge is not complete, and the change detection algorithm has poor universality on different data sets.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of optical remote sensing image building change detection, and specifically is building change information extraction using deep learning technology to provide services for specific businesses such as land monitoring. BACKGROUND

[0002] With the acceleration of urbanization, buildings as an important part of urban space, its change detection has extremely important significance in urban planning, land use management, disaster monitoring and other aspects. Traditional building change detection methods often rely on manual visual interpretation, and have problems such as low efficiency and strong subjectivity. In recent years, satellite remote sensing technology has developed rapidly, especially the optical remote sensing satellite "High Resolution 2" launched by China has realized sub-meter spatial resolution, but the improvement of resolution also means the dramatic increase of data volume, which brings opportunities and challenges to the field of change detection. At the same time, computer vision technology is also progressing, and the building change detection method of optical remote sensing image based on deep learning has gradually become a research hotspot.

[0003] The change detection method for remote sensing image can be divided into remote sensing change detection based on traditional method and remote sensing change detection based on deep learning. The traditional method because of excessive dependence on expert experience and preprocessing operation, leads to a more cumbersome detection process, and its detection accuracy is not high, manual intervention is more, the degree of automation is low, and the time consumption is large. And the remote sensing change detection based on deep learning can automatically extract the deep features of remote sensing image data, abstract more suitable features for the current task, and get more accurate results. And it can realize end-to-end learning without human intervention. And with the increase of data volume, the performance of the model will be improved, these characteristics are very good to adapt to the trend of high resolution of remote sensing image.

[0004] The research of change detection based on deep learning can be divided into two directions, one is to fuse the two or more remote sensing images in the input stage, and then send them into the convolutional neural network, and finally get the change binary graph. The other detection method is to use the twin network structure to extract the features of the two input images respectively, and then compare the extracted features to get the change detection result graph.

[0005] At present, there are many remote sensing image building change detection methods based on deep learning at home and abroad, such as LSTM model based on recurrent neural network, which can well process remote sensing image sequence with time sequence, but the image feature extraction capability is relatively poor; BIT model based on Transformer mechanism, which can well obtain the global attention of the image and capture long-term dependence, but it needs a large amount of computing resources and a large-scale data set; U-Net model based on convolutional neural network, which can well extract the feature information of the image, and the requirement for data volume is not high, but the global attention is insufficient; at present, the research on U-Net model based on convolutional neural network is very popular, and a large number of researchers improve the performance of the network by adding skip connection and designing attention module. SUMMARY

[0006] The purpose of the present application is to provide an optical remote sensing image building change detection method based on deep learning, which describes the whole process from obtaining original data from optical remote sensing satellite to finally obtaining change detection prediction map through the designed model. The present application improves the accuracy of change feature modeling and extraction, solves the problems of low detection accuracy of existing building change detection methods for optical remote sensing images, incomplete detection edge, and poor universality of change detection algorithm on different data sets.

[0007] The technical scheme adopted by the present application is as follows:

[0008] An optical remote sensing image building change detection method based on deep learning

[0009] Step 1: Obtain two different time period multi-spectral images and panchromatic images of the target research area, and use ENVI software to perform radiation calibration, atmospheric correction and orthorectification on the multi-spectral images respectively, and perform radiation calibration and orthorectification on the panchromatic images, then register and image fuse the processed multi-spectral images and panchromatic images to obtain the final remote sensing image;

[0010] Step 2: manually label the changed part of the building in the two different time remote sensing images by using LabCD software, obtain the corresponding change Label image, and then process the corresponding change area edge label by using sobel operator;

[0011] Step 3: pre-process the obtained two time optical remote sensing image data;

[0012] Step 4: input the read image T1 before change and the image T2 after change into the weight shared UHRNet encoder respectively, and extract the feature maps of different levels from stage1 to stage5 of each image;

[0013] Step 5: The feature maps of the same stage of T1 and T2 are respectively input into the difference feature extraction module to obtain the difference feature maps of the remote sensing feature images of each stage, and the calculation formula is as follows:

[0014] D i = abs(X i1 ,X i2 )

[0015] Wherein X i1 ,X i2 represent the feature maps of T1 and T2 of the i-th stage respectively, and D i represents the output of the i-th stage.

[0016] Step 6: The difference feature maps of each stage are respectively input into the channel and spatial attention module. The channel attention module can calculate the importance size of each channel to give different attention to each channel; the spatial attention module can give different weights to different positions of the input data, so that the model pays more attention to important information, thereby obtaining the enhanced difference feature map.

[0017] Step 7: The minimum resolution difference feature map is input into the UHRNet decoder, and a jump connection is added in the decoder.

[0018] Step 8: The output of the UHRNet decoder is input into the classifier through the edge guiding module, and then spliced with itself to obtain the final change detection binary image, and deep supervision is added to accelerate the model convergence.

[0019] Further, the specific steps of step 1 are as follows:

[0020] Step 1.1: The multispectral image and the panchromatic image are first radiometrically calibrated, which can convert the digital quantization value DN into a radiometric brightness value with physical meaning, so as to more accurately describe the reflection or emission characteristics of the ground object. The formula is as follows:

[0021] Rad = gain * DN + Bias

[0022] Wherein grain is the gain coefficient, DN is the digital quantization value in the image, and Bias is the offset.

[0023] Step 1.2: The FLAASH atmospheric correction tool is selected to perform remote sensing image atmospheric correction on the multispectral image.

[0024] Step 1.3: The RPC information is used to orthorectify the multispectral image and the panchromatic image, and the three convolution method is selected as the resampling method.

[0025] Step 1.4: Fuse the multispectral image and the panchromatic image using the NNDiffuse Pan Sharpening method in ENVI to obtain a remote sensing image with both high spatial resolution and spectral resolution.

[0026] Further, the specific steps of step 3 are as follows:

[0027] Step 3.1: After obtaining the multi-temporal optical remote sensing image data, it is segmented into non-overlapping image blocks with a resolution of 256*256, and three data sets of training set, validation set and test set are obtained according to the data ratio of 7:1:2;

[0028] Step 3.2: Read the remote sensing image after segmentation;

[0029] Step 3.3: Perform data augmentation operations on the read training set, including horizontal random flip, vertical random flip, image random cropping, and image Gaussian blur, to increase the data complexity in the training process and enhance the model robustness.

[0030] Further, the specific steps of step 4 are as follows:

[0031] Step 4.1: For the input feature map X, pass it through two two-dimensional convolutions with a step size of 2 to obtain L feature maps A with a resolution size of 1 / 4 of the input, the calculation formula is as follows:

[0032] A = ReLU(BN(Conv(ReLU(BN(Conv(x)))))

[0033] Step 4.2: For each stage of the UHRNet encoder, only two parallel HR modules are used, and convolution is performed to extract features and fuse features of two different scales within each stage. Finally, each stage outputs two feature maps with a resolution difference of one.

[0034] Step 4.3: Between every two adjacent stages of the UHRNet encoder, the output feature map with smaller resolution in the previous stage is downsampled to obtain a feature map with half the resolution but twice the number of channels, and it is used as the input of the two HR modules in the next stage along with the input with the original resolution before downsampling. Finally, five feature maps with different resolution sizes are obtained.

[0035] Further, the specific steps of step 6 are as follows:

[0036] Step 6.1: Obtain the difference feature map D iThe channel attention module CAM is applied to maximum pooling, and then the attention weight is calculated through two fully connected layers and a sigmoid activation function. The reconstructed attention weight vector is multiplied by the input feature map to obtain the output of the CAM, and the calculation formula is as follows:

[0037] C i = σ(C2(LRe(C1(max(D i ))))

[0038]

[0039] Wherein C1, C2 represents a fully connected layer, LRe represents a Leaky-ReLU activation function, σ represents a sigmoid activation function, C i represents the channel attention weight of the i-th stage, D i represents the difference feature map of the i-th stage, D ic The difference feature map of the i-th layer is added with channel attention;

[0040] Step 6.2: At the same time, the difference feature map D i is also sent to the spatial attention module SAM. The SAM condenses the spatial information of the feature map through max pooling and average pooling, and then calculates the spatial attention weight using a convolution layer and a sigmoid activation function. The output of the SAM is obtained by multiplying the obtained spatial attention weight with the input feature map, and the calculation formula is as follows:

[0041] S i = σ(Conv([max(D i )+avg(D i )]))

[0042]

[0043] Wherein S i represents the channel attention weight of the i-th stage, D i represents the difference feature map of the i-th stage, D is The difference feature map of the i-th layer is added with spatial attention;

[0044] Step 6.3: Finally, the outputs of the channel and spatial attention modules are added to obtain the feature output D i ', and the calculation formula is as follows:

[0045]

[0046] Further, the specific steps of step 7 are as follows:

[0047] Step 7.1: Up-sample the minimum resolution difference feature map D5' and concatenate it with the difference feature map D4' of the 4th stage of the encoder to obtain D4", and input D4" and D5' into the UHRNet decoder;

[0048] Step 7.2: For each stage of the UHRNet decoder, only two parallel HRNet branches are composed, and convolution and fusion of two-scale resolution features are also performed within each stage, and finally two feature maps with a difference of one in resolution are output for each stage;

[0049] Step 7.3: Between every two adjacent stages of the UHRNet decoder, the resolution of the output feature map in the previous stage is increased by half by up-sampling, and the feature map with reduced one-time channel is concatenated with the feature map of the same resolution size of the encoder, and then the resolution input before up-sampling is used as the input of the two HR modules of the next stage, and finally five feature maps with different resolution sizes are obtained.

[0050] Further, the specific steps of step 8 are as follows:

[0051] Step 8.1: For the output maximum resolution feature map D1", pass it through a 1*1 convolution module to obtain an edge prediction map, and obtain output D 1edge by a softmax function, and then up-sample to the original image size and compare with the edge label to calculate the edge loss, and the corresponding loss function is:

[0052]

[0053] Step 8.2: Concatenate the output D1'" and D1" in the channel dimension, and then input it into another 1*1 convolution module, and obtain output D 1cd by a softmax function, and then up-sample to the original image size and compare with the change detection label to calculate the change detection loss, and the corresponding loss function is:

[0054] Loss cd =-α t (1-p t ) γ log(p t )

[0055] Step 8.3: Use deep supervision, directly up-sample other resolution outputs to the original image size and compare them with the change detection label to calculate the change detection loss Loss ds . The corresponding loss function calculation formula is the same as above; the final loss function calculation formula is as follows:

[0056] Loss=Lossedge +Loss cd +Loss ds

[0057] In summary, by adopting the technical scheme, the application has the beneficial effects of:

[0058] 1. A deep learning-based optical remote sensing image building change detection method adopts an encoder-decoder network structure, combines an HRNet network structure with a U-Net network to build a U-HRNet model, extracts feature information by using an encoder part of the U-HRNet model, and reconstructs and recovers change feature information by using a decoder part of the U-HRNet model.

[0059] 2. In the application, step 4 can maintain multi-scale parallelism at any stage of the encoder, perform multi-scale fusion, well fuse low-level detail information and high-level semantic information, make the obtained feature map contain more information, and compared with the HRNet network, reduce the calculation of the high-resolution feature map, save the computing resources, and reduce the running time of the network.

[0060] 3. In the application, the channel and spatial attention module is introduced in step 6, which can make the model pay more attention to important information and enhance the feature information of the difference feature map.

[0061] 4. In the application, the edge guiding module is introduced in step 7.4, the edge label obtained by using the change label processing is used to guide the model to learn the edge feature, and the model can obtain a more complete change detection prediction map of the edge information.

[0062] 5. The application respectively trains and tests on the change detection data set of Chengdu High-tech Zone and the public data set LEVIR-CD data set, and the experimental results show that the proposed method effectively optimizes and improves the problems that the existing building change detection method has low detection accuracy, incomplete edge detection on the optical remote sensing image with high resolution, and poor universality of the change detection algorithm on different data sets. BRIEF DESCRIPTION OF DRAWINGS

[0063] The application will be described by examples and with reference to the accompanying drawings, in which:

[0064] Figure 1 is a flowchart of a deep learning-based optical remote sensing image building change detection method;

[0065] Figure 2 is a network overall architecture diagram adopted in the embodiment 1 of the application;

[0066] Figure 3is the encoder structure adopted in embodiment 1 of the present application;

[0067] Figure 4 is the decoder structure adopted in embodiment 1 of the present application;

[0068] Figure 5 is the channel and spatial attention module structure diagram adopted in embodiment 1 of the present application;

[0069] Figure 6 is the edge guiding module structure diagram adopted in embodiment 1 of the present application;

[0070] Figure 7 is the change detection result display of embodiment 1 of the present application on the LEVIR-CD test set, wherein (a) is the pre-change image, (b) is the post-change image, (c) is the change reference image, and (d) is the obtained detection result image;

[0071] Figure 8 is the change detection result display of embodiment 1 of the present application on the self-made change detection data set of the target research area of Chengdu High-tech Zone, wherein (a) is the pre-change image, (b) is the post-change image, (c) is the change reference image, and (d) is the obtained detection result image. DETAILED DESCRIPTION

[0072] All features disclosed in this specification, and / or all steps of any methods or processes disclosed in this specification, may be combined in any combination, except combinations where at least some of such features and / or steps are mutually exclusive.

[0073] The present application will be described in detail below with reference to the following Figures 1 to 8 The present application will be described in detail below with reference to the following

[0074] Embodiment 1:

[0075] Step 1: Obtain two multispectral images and panchromatic images of different time periods of the target research area of Chengdu High-tech Zone from Gao Fen II, and use ENVI software to perform radiation calibration, atmospheric correction, and orthorectification on the multispectral images, and perform radiation calibration and orthorectification on the panchromatic image, and then register and image fuse the processed multispectral images and panchromatic image to obtain the final remote sensing image;

[0076] Step 1.1: Perform radiation calibration on the multispectral images and panchromatic images first, which can convert digital quantization values (DN values) into radiation brightness values with physical meaning, so as to more accurately describe the reflection or emission characteristics of ground objects. The formula is as follows:

[0077] Rad = gain * DN + Bias

[0078] wherein grain is the gain coefficient, DN is the digital quantization value in the image, and Bias is the offset.

[0079] Step 1.2: Select FLAASH atmospheric correction tool to perform remote sensing image atmospheric correction on the multispectral image;

[0080] Step 1.3: Perform orthorectification on the multispectral image and the panchromatic image using RPC information, and select cubic convolution method for resampling;

[0081] Step 1.4: Use the NNDiffuse Pan Sharpening method in ENVI to fuse the multispectral image and the panchromatic image to obtain a remote sensing image with high spatial resolution and spectral resolution.

[0082] Step 2: Use LabCD software to manually label the changed part of the building in the two different time remote sensing images, obtain the corresponding change Label image, and then process the corresponding change area edge label through the sobel operator;

[0083] Step 3: Preprocess the two time optical remote sensing image data obtained;

[0084] Step 3.1: After obtaining the multi-temporal optical remote sensing image data, divide it into non-overlapping image blocks with a resolution of 256*256, and divide it into training set, validation set and test set according to the data ratio of 7:1:2;

[0085] Step 3.2: Read the remote sensing image after segmentation;

[0086] Step 3.3: Perform data enhancement operations on the read training set, including horizontal random flip, vertical random flip, image random cropping, and image Gaussian blur, to increase the data complexity in the training process and enhance the model robustness.

[0087] Step 4: Input the image T1 before the change and the image T2 after the change into the weight-shared UHRNet encoder respectively, and the structure of a single UHRNet encoder is as shown in Figure 3 , which can extract feature maps of different levels from stage1 to stage5 of the input image;

[0088] Step 4.1: For the input feature map X, obtain L feature maps A with a resolution of 1 / 4 of the input through two-dimensional convolution with a step of 2, and the calculation formula is as follows:

[0089] A = ReLU(BN(Conv(ReLU(BN(Conv(x)))))

[0090] Step 4.2: For each stage of the UHRNet encoder, only two parallel HR modules are composed, and within each stage, convolution is performed to extract features and fuse two scale resolution features, and finally two feature maps with one-fold difference in resolution are output for each stage;

[0091] Step 4.3: Between every two adjacent stages of the UHRNet encoder, the feature map with half resolution but doubled channels is obtained by downsampling the output feature map with smaller resolution in the previous stage, and is used as the input of the two HR modules of the next stage together with the resolution input before downsampling, and finally five feature maps with different resolution sizes are obtained.

[0092] Step 5: The feature maps of the same stage of T1 and T2 are respectively sent into the difference feature extraction module to obtain the difference feature maps of the remote sensing feature maps of each stage, as shown in Figure 2 i1 i2 i respectively represent the feature maps of T1 and T2 of the i-th stage, and D i represents the output of the i-th stage, and the calculation formula is as follows:

[0093] D i1 = abs(X i2 )

[0094] Step 6: The difference feature maps of each stage are respectively sent into the channel and spatial attention module, and the channel and spatial attention module is as shown in Figure 5 . The channel attention module can calculate the importance size of each channel and give different attention to each channel; the spatial attention module can give different weights to different positions of the input data, so that the model pays more attention to important information, thereby obtaining enhanced difference feature maps;

[0095] Step 6.1: The difference feature maps D i of each stage are sent into the channel attention module CAM, and maximum pooling is applied first, then the attention weight is calculated through two fully connected layers and a sigmoid activation function, and the reconstructed attention weight vector is multiplied with the input feature map element by element to obtain the output of CAM, and the calculation formula is as follows:

[0096] C i = σ(C2(LRe(C1(max(D i ))))

[0097]

[0098] ​​​where C1, C2 represent fully connected layers, LRe represents a Leaky-ReLU activation function, σ represents a sigmoid activation function, C i represents the channel attention weight of the i-th stage, D i represents the difference feature map of the i-th stage, D ic is the difference feature map of the i-th layer with added channel attention;

[0099] Step 6.2: At the same time, the difference feature maps D i of each stage are also sent into the spatial attention module SAM, which condenses the spatial information of the feature maps through max pooling and average pooling, and then calculates the spatial attention weight using a convolution layer and a sigmoid activation function. The output of SAM is obtained by multiplying the obtained spatial attention weight with the input feature map element by element, and the calculation formula is as follows:

[0100] S i = σ(Conv([max(D i )+avg(D i )]))

[0101]

[0102] where S i represents the channel attention weight of the i-th stage, D i represents the difference feature map of the i-th stage, D is is the difference feature map of the i-th layer with added spatial attention;

[0103] Step 6.3: Finally, the outputs of the channel and spatial attention modules are added to obtain the feature output D i ', and the calculation formula is as follows:

[0104]

[0105] Step 7: The minimum resolution difference feature map is sent into the UHRNet decoder, the decoder structure is as shown in Figure 4 , and the input of each layer of the decoder is connected with the output of each layer of the encoder through a skip connection, as shown in Figure 2 which shows the connection mode of the encoder and the decoder;

[0106] Step 7.1: The minimum resolution difference feature map D5' is up-sampled, and is spliced with the difference feature map D4' of the fourth stage of the encoder to obtain D4", and D4" and D5' are taken as inputs and input into the UHRNet decoder;

[0107] Step 7.2: For each stage of the UHRNet decoder, only two parallel HRNet branches are composed, and convolution and fusion of two-scale resolution features are also performed within each stage, and finally two feature maps with one-fold difference in resolution are outputted by each stage;

[0108] Step 7.3: Between every two adjacent stages of the UHRNet decoder, the output feature map with larger resolution in the previous stage is up-sampled to obtain a feature map with half-increased resolution and halved channel number, and then it is connected with the feature map of the same resolution size of the encoder by skip connection, and finally it is inputted into the two HR modules of the next stage together with the resolution input before up-sampling, and finally five feature maps with different resolution sizes are obtained.

[0109] Step 8: The maximum resolution output of the above decoder is sent into the classifier to obtain the final change detection binary image through the edge guiding module as shown in Figure 6 , and the remaining small resolution outputs are added with deep supervision to accelerate the model convergence;

[0110] Step 8.1: For the output maximum resolution feature map D1", it is passed through a 1*1 convolution module to obtain an edge prediction map, and the output D 1edge is obtained through the softmax function, and then it is up-sampled to the original image size and compared with the edge label to calculate the edge loss, and the corresponding loss function is:

[0111]

[0112] Step 8.2: The output D1'" and D1" are spliced in the channel dimension, and then sent into another 1*1 convolution module, and the output D 1cd is obtained through the softmax function, and then it is up-sampled to the original image size and compared with the change detection label to calculate the change detection loss, and the corresponding loss function is:

[0113] Loss cd = -α t (1-p t ) γ log(p t )

[0114] Step 8.3: Using deep supervision, the other resolution outputs are directly up-sampled to the original image size and compared with the change detection label to calculate the change detection loss Loss ds . The corresponding loss function calculation formula is the same as above; the final loss function calculation formula is as follows:

[0115] Loss = Loss edge + Loss cd+Loss ds

[0116] Step 8.4: The results on the LEVIR-CD test set are shown in FIG. 8, where (a) is the pre-change image, (b) is the post-change image, (c) is the actual change reference image, and (d) is the obtained detection result image, and the white part in FIG. (c) and FIG. (d) represents the changed area, and the black part represents the unchanged area. The results on the self-made Chengdu High-tech Zone target research area test set are shown in FIG. 9, and similarly, where (a) is the pre-change image, (b) is the post-change image, (c) is the change reference image, and (d) is the obtained detection result image, and the white part in FIG. (c) and FIG. (d) represents the changed area, and the black part represents the unchanged area. Figure 7 Figure 8

[0117] The present application adopts an encoder-decoder network structure, combines the HRNet network structure with the U-Net network to build a U-HRNet model. The feature information is extracted by the encoder part of the U-HRNet model, and the changed feature information is recovered by the decoder part of the U-HRNet model. In the present application, the U-HRNet structure ensures that the encoder and the decoder can maintain multi-scale parallelism at each stage, perform multi-scale fusion, and well fuse the low-level detailed information and the high-level semantic information, so that the obtained feature map contains more rich information. At the same time, the channel and spatial attention module introduced in the present application can make the model pay more attention to important information and enhance the feature information of the difference feature map. In addition, the edge guiding module is introduced in the present application, which can make the model obtain more complete edge information by using the edge label obtained by processing the change label to guide the model to learn the edge features. Through the above method, the present application effectively improves the problems of the existing change detection method, such as low detection accuracy on optical remote sensing images with high resolution and incomplete edge of the output change detection map. Furthermore, the present application has carried out training and testing on two different data sets including a public data set LEVIR-CD and a self-made data set, which also confirms the universality of the proposed method on different data sets.

[0118] The above merely provides the preferred embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can think of changes or replacements within the technical range disclosed by the present application without creative labor, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be limited by the protection scope defined in the claims.​​

Claims

1. A deep learning-based optical remote sensing image building change detection method, comprising the following steps: Step 1: Obtain two different time period multi-spectral images and panchromatic images of the target research area, and use ENVI software to perform radiation calibration, atmospheric correction and orthorectification on the multi-spectral images respectively, and perform radiation calibration and orthorectification on the panchromatic image, register and image fusion the processed multi-spectral image and panchromatic image to obtain the final remote sensing image; Step 2: Use LabCD software to manually label the changed part of the building in the two different time remote sensing images, obtain the corresponding change Label image, and then process it by sobel operator to obtain the edge label of the corresponding change area; Step 3: Preprocess the obtained two different time optical remote sensing image data; Step 4: Input the read image T1 before the change and the image T2 after the change into the weight-shared UHRNet encoder respectively, and extract the feature maps of image T1 and image T2 at different levels from stage1 to stage5; Step 5: Send the feature maps of image T1 and image T2 at the same stage into the difference feature extraction module respectively to obtain the difference feature maps of each stage remote sensing feature image, the calculation formula is as follows: D i =abs(X i1 ,X i2 ) wherein , respectively represent the feature maps of image T1 and image T2 of the i-th stage, D i represents the output of the i-th stage; Step 6: Send each stage difference feature map into the channel and spatial attention module, the channel attention module obtains the channel coefficient by maximum pooling, full connection and average pooling, which represents the importance of each channel, and gives different attention to each channel, the calculation formula is as follows: C i =σ(avg(C2(LRe(C1(max(D i )))))) ； wherein , denotes a fully connected layer, denotes a Leaky-ReLU activation function, denotes a sigmoid activation function, denotes the channel attention weight of the i-th stage, denotes the difference feature map of the i-th stage, the difference feature map with added channel attention of the i-th layer; Then enter the spatial attention module, and obtain the weight coefficient by maximum pooling and average pooling pixel by pixel addition and convolution, give different weights on different spatial positions of the input data, and make the model selectively pay more attention to important information, the formula is as follows: ； ； wherein denotes the channel attention weight of the i-th stage, denotes the difference feature map of the i-th stage, the difference feature map added with spatial attention of the i-th layer; Finally, the enhanced difference feature map is obtained: ； Step 7: Send the minimum resolution difference feature map into the UHRNet decoder, and add a jump connection in the decoder; Step 8: Output the UHRNet decoder through the edge guide module, and then splice it with itself and send it into the classifier to obtain the final change detection binary image. 2.The deep learning-based optical remote sensing image building change detection method of claim 1, wherein, The specific steps of step 1 are as follows: Step 1.1: Radiometric calibration is performed on the multi-spectral image and panchromatic image first, and the digital quantization value DN is converted into the radiometric brightness value, which describes the reflection or emission characteristics of the ground object, and is expressed as: ； wherein is a gain coefficient, is a digital quantized value in the image, is an offset; Step 1.2: Select FLAASH atmospheric correction tool to perform remote sensing image atmospheric correction on the multi-spectral image; Step 1.3: Use RPC information to perform orthorectification on the multi-spectral image and panchromatic image, and select cubic convolution method for resampling; Step 1.4: Use NNDiffuse Pan Sharpening method in ENVI to fuse the multi-spectral image and panchromatic image to obtain a remote sensing image with high spatial resolution and spectral resolution. 3.The deep learning-based optical remote sensing image building change detection method of claim 1, wherein, The specific steps of step 3 are as follows: Step 3.1: After obtaining the multi-temporal optical remote sensing image data, it is segmented into non-overlapping image blocks with a resolution of 256*256, and divided into a training set, a validation set and a test set according to a data ratio of 7:1:2; Step 3.2: reading the remote sensing image after segmentation; Step 3.3: performing data enhancement operation on the read training set, and the data enhancement operation includes horizontal random flip, vertical random flip, image random cropping and image Gaussian blur. 4.The deep learning-based optical remote sensing image building change detection method of claim 1, wherein, The specific steps of the step 4 are as follows: Step 4.1: for the input feature map X, L feature maps A with a resolution of 1 / 4 of the input are obtained by two-dimensional convolution with a step of 2, and the calculation formula is as follows: ； Step 4.2: for each stage of the UHRNet encoder, only two parallel HR modules are used, and convolution is performed to extract features and fuse features with two scales of resolution in each stage, and finally two feature maps with a resolution difference of one time are output in each stage; Step 4.3: between every two adjacent stages of the UHRNet encoder, a feature map with a resolution reduced by half but a channel number increased by one time is obtained by downsampling the output feature map with a smaller resolution in the previous stage, and is used as the input of the two HR modules of the next stage together with the input with the resolution before downsampling, and finally five feature maps with different resolution sizes are obtained. 5.The deep learning-based optical remote sensing image building change detection method of claim 1, wherein, The specific steps of the step 6 are as follows: Step 6.1: The feature maps of each stage difference are sent into the channel attention module CAM, first apply max pooling, then pass through two fully connected layers and a sigmoid activation function to calculate the attention weight, the reconstructed attention weight vector is multiplied element by element with the input feature map to get the output of CAM, the calculation formula is as follows: Step 6.1: The feature maps of each stage difference are sent into the channel attention module CAM, first apply max pooling, then pass through two fully connected layers and a sigmoid activation function to calculate the attention weight, the reconstructed attention weight vector is multiplied element by element with the input feature map to get the output of CAM, the calculation formula is as follows: C i =σ(avg(C2(LRe(C1(max(D i )))))) ； wherein , denotes a fully connected layer, denotes a Leaky-ReLU activation function, denotes a sigmoid activation function, denotes the channel attention weight of the i-th stage, denotes the difference feature map of the i-th stage, the difference feature map with added channel attention of the i-th layer; Step 6.2: The difference feature map D i Also sent into the spatial attention module SAM, SAM condenses the spatial information of the feature map through max pooling and average pooling, and then calculates the spatial attention weight using the convolution layer and the sigmoid activation function; the output of SAM is obtained by multiplying the obtained spatial attention weight with the input feature map element by element, and the calculation formula is as follows: ； ； wherein denotes the channel attention weight of the i-th stage, denotes the difference feature map of the i-th stage, the difference feature map added with spatial attention of the i-th layer; Step 6.3: Add the outputs of the channel and spatial attention modules to get the feature output The calculation formula is as follows: 。 6.The deep learning-based optical remote sensing image building change detection method of claim 1, wherein, The specific steps of the step 7 are as follows: Step 7.1: The minimum resolution difference feature map is upsampled and concatenated with the difference feature map of the 4th stage of the encoder is obtained by concatenation , which is input into the UHRNet decoder as ​​ Step 7.2: for each stage of the UHRNet decoder, only two parallel HRNet branches are used, and convolution and fusion of features with two scales of resolution are performed in each stage, and finally two feature maps with a resolution difference of one time are output in each stage; Step 7.3: between every two adjacent stages of the UHRNet decoder, a feature map with a resolution increased by half but a channel number reduced by one time is obtained by upsampling the output feature map with a larger resolution in the previous stage, and is connected with the feature map with the same resolution of the encoder by jumping, and is used as the input of the two HR modules of the next stage together with the input with the resolution before upsampling, and finally five feature maps with different resolution sizes are obtained.

7. The deep learning-based optical remote sensing image building change detection method according to claim 1, wherein, The specific steps of the step 8 are as follows: Step 8.1: For the maximum resolution feature map of the output The edge prediction map is obtained by a 1*1 convolution module, and the output D is obtained by a softmax function 1edge The original image size is up-sampled again and compared with the edge label to calculate the edge loss, and the corresponding loss function is: ； Step 8.2: The output is concatenated with the channel dimension, and then sent to another 1*1 convolution module to obtain the output D 1cd through a softmax function. The output D is up-sampled to the original image size and compared with the change detection label to calculate the change detection loss. The corresponding loss function is: ； Step 8.3: Using deep supervision, the other resolution outputs are directly re-upsampled to the original image size and compared with the change detection label to compute the change detection loss ; the corresponding loss function calculation formula is the same as above; the final loss function calculation formula is as follows: 。

Images