Method for detecting building changes in high-consequence areas of natural gas pipelines based on remote sensing images

Abstract

The present invention discloses a method for detecting building changes in high-consequence areas of natural gas pipelines based on remote sensing images, which relates to the technical field of image processing and aims to solve the problems of high labor costs and low automation in obtaining change information. The key points of its technical solution are as follows: S1: Generate a high-consequence area strip map according to the drawn high-consequence area vector; S2: Screen multi-temporal remote sensing images around the natural gas pipeline to construct a change detection data set; S3: A method for detecting building changes in high-consequence areas of natural gas pipelines, including a backbone network and a secondary backbone network; S4: The building change patches output by the backbone network; S5: The building change attention map output by the backbone network; S6: Update the parameters of the entire model through backpropagation; S7: Train the high-consequence area building change detection model; S8: Output the building change patches; S9: Then generate a building change patch vector according to the change patches. The effects of increasing efficiency, enhancing information extraction, and reducing information loss are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a method for detecting building changes in high-consequence areas of natural gas pipelines based on remote sensing images. Background Art

[0002] A natural gas pipeline refers to a pipeline that transports natural gas (including associated gas produced in oil fields) from the production site or treatment plant to the urban gas distribution center or industrial enterprise users, also known as a gas transmission pipeline. Using natural gas pipelines to transport natural gas is a way to transport large amounts of natural gas on land. Among the total length of pipelines in the world, natural gas pipelines account for about half. According to the "Integrity Management Specification for Oil and Gas Transmission Pipelines" (GB32167―2015), "high-consequence area" is clearly defined as "an area where pipeline leakage may cause greater adverse impacts on the public and the environment". With the development of aerospace technology over the years, the spectral resolution, spatial resolution, and temporal resolution of remote sensing images have all been improved, providing a more convenient and detailed data source for building and high-consequence area change monitoring.

[0003] The existing technical solutions mentioned above have the following defects: In actual production, the method for obtaining change information still mainly relies on field surveys or manual and visual interpretation methods. This method has a high labor cost, a low degree of automation, and the production quality strongly depends on the professionalism and experience of natural gas company managers, and reduces the work efficiency in high-consequence areas. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for detecting building changes in high-consequence areas of natural gas pipelines based on remote sensing images, which can enhance the feature extraction ability of the model, accurately extract building change information, and realize the correction and update of the results of change information in high-consequence areas.

[0005] To achieve the above purpose, the present invention provides the following technical solutions:

[0006] A method for detecting building changes in high-consequence areas of natural gas pipelines based on remote sensing images, the steps of which are as follows:

[0007] S1: Superimpose the natural gas pipeline vectors on two-phase remote sensing images respectively. Taking the vector line as the center, draw the high-consequence area vector, and generate a high-consequence area strip map according to the drawn high-consequence area vector;

[0008] S2: Screen multi-phase remote sensing images around the natural gas pipeline, construct a change detection data set for training a high-consequence area building change detection model, use semantic segmentation annotation software to annotate the multi-phase remote sensing images, produce a changed building label image (Ground Truth), and increase the number of data set samples through data augmentation;

[0009] S3: The building change detection method for high-consequence areas of natural gas pipelines includes a backbone network and a secondary backbone network. The backbone network adopts an encoder-decoder structure. The encoder module has a total of seven encoder structures. The first five layers use SE_ResNet50, and the additional two layers of encoders are constructed using pooling layers and convolutional layers to extract high-level feature information. The multi-scale fusion module and attention module in the backbone network are mainly used to extract the features of changed buildings. The features of changed buildings are restored to the original size through the decoder, and the building changes in multi-temporal remote sensing images are predicted. The backbone network also outputs a building change attention map for the training of the pre-trained VGG16 model of the secondary backbone network;

[0010] S4: The building change patches output by the backbone network will calculate the loss with the Ground Truth marked in S2. The loss function uses the binary cross-entropy loss function;

[0011] S5: The building change attention map output by the backbone network, after being activated by the difference and Sigmoid function, extracts the probability map of the features of the changed building area, multiplies it with the early image and the late image respectively through matrix multiplication, and inputs it into the pre-trained VGG16 model of the secondary backbone network for model training. The loss metric of the secondary backbone network uses the mean squared error loss function;

[0012] S6: The binary cross-entropy loss and the mean squared error loss function obtained in S3 and S4 are weighted and summed as the total loss of the building change detection network for high-consequence areas, and the parameters of the entire model are updated through backpropagation;

[0013] S7: The self-built change detection dataset in S2 is used to train the building change detection model for high-consequence areas until the building change detection model meets the actual usage requirements. At this time, the network model is the building change detection model for high-consequence areas;

[0014] S8: According to the two-phase high-consequence area strip maps obtained in S1, the two-phase high-consequence area strip maps are respectively cropped to a size of 256×256×3 and input into the building change detection network for high-consequence areas. The trained model is called to automatically detect the changed buildings in the high-consequence area images and output the building change patches;

[0015] S9: The building change patch images output by the high-consequence area change detection algorithm lack geographical reference coordinate information and cannot accurately locate the positions of the changed buildings. It is necessary to add the geographical reference information of the input temporal remote sensing images to the building change patches, and then generate building change patch vectors according to the change patches.

[0016] Further, the method for generating the high-consequence area strip map in S1 is as follows:

[0017] A1: Rasterize the input temporal remote sensing image, overlay the natural gas pipeline vector on the temporal remote sensing image, draw the high-consequence area vector with the pipeline vector line as the center within a range of 200 meters in width and 2000 meters in length on both sides, and use the defined high-consequence area vector to crop and generate the high-consequence area identification strip map;

[0018] A2: Only retain the pixel values of the high-consequence area strip region, set the pixels of the remaining regions in the image to 0, and crop the high-consequence area identification strip maps of the two-phase temporal remote sensing images.

[0019] Further, the segmentation annotation amplification method in S2 is as follows:

[0020] B1: Select the images around the natural gas pipeline, make the change detection data set, first crop the temporal remote sensing images of the same size and the same area respectively, with the cropping size of 256×256×3, to obtain the training sample sets of the two phases before and after. In addition, use the semantic segmentation annotation software to make the change detection label file;

[0021] B2: Annotate the two-phase images with building changes respectively to construct the change detection sample set. To increase the number of the sample set, the two-phase images and the annotated images are data-augmented. The specific methods include rotating 90°, 180°, 270°, as well as flipping vertically and horizontally.

[0022] Further, the data enhancement methods in A5 are image mirroring, Gaussian blur, random rotation, and randomly removing some pixels.

[0023] Further, the specific operation method of S4 is as follows:

[0024] C1: Extract the changed buildings from the temporal remote sensing image, output the building change patch image, and calculate the loss generated during the training process by using the building change patch image and the labeled label image in S2;

[0025] C2: The high-consequence area change detection mainly targets the changed buildings within the high-consequence area of the two-phase images. The buildings belong to one category, and the rest except the buildings belong to one category. For the sample (x, y), let x be the sample and y be the corresponding label value, and the prediction value set is {0, 1}. Assume that the true label of a certain building sample is y gt , the probability value of the result predicted by this building sample is y p , the loss function of this building sample is defined as follows:

[0026]

[0027] Among them, y gt is the true label value of the building sample, y p (x) i is the building sample at ygt The probability when it is equal to 1.

[0028] Furthermore, the specific operation method of S6 is as follows:

[0029] D1: After obtaining the loss values of the backbone network and the secondary backbone network, their weighted sum is obtained to get the total loss. The total loss function of the network is defined as follows:

[0030] Loss sum = Loss BCE + αLoss MSE

[0031] Among them, Loss BCE is the binary cross-entropy loss, α is the weight coefficient of the binary cross-entropy loss of the backbone network, and Loss MSE is the mean square error loss;

[0032] D2: Since the backbone network undertakes the main task in the whole network, that is, the building transformation detection task in the high-consequence area, in order to make the network focus more on building transformation detection, the loss function of the backbone network has a greater weight than that of the secondary backbone network, that is, α > 1. After obtaining the total network loss, backpropagation is carried out to update the parameters in the network.

[0033] To sum up, the beneficial technical effects of the present invention are as follows:

[0034] 1. The method for detecting building changes in the high-consequence area of natural gas pipelines based on remote sensing images uses a high-consequence area change detection algorithm based on deep learning, which improves the speed of updating building change information in the high-consequence area, improves work efficiency, and produces the effect of increasing efficiency;

[0035] 2. The method for detecting building changes in the high-consequence area of natural gas pipelines based on remote sensing images, the high-consequence area building change detection model includes a backbone network and a secondary backbone network. The backbone network adopts an encoder-decoder structure. The multi-scale feature fusion module designed in the backbone network extracts high-level feature information and low-level feature information from the encoder branch. In addition, the added channel attention module and spatial attention module improve the attention to the changed building area from the channel and spatial ranges, enhance the ability to extract changed building information, and the backbone network uses the binary cross-entropy loss function to constrain the loss generated during training, producing the effect of enhancing the information extraction ability.

[0036] 3. In the building change detection method for high-consequence areas of natural gas pipelines based on remote sensing images, the building change attention map derived from the backbone network is used as the input of the secondary network. After the building change attention map is activated by differential and sigmoid functions, it is multiplied by the matrix of the early image and the late image respectively, and then input into the VGG16 pre-trained model respectively to guide the training process of the secondary network. The secondary network uses the mean square error loss function to reduce information loss and improve the segmentation performance of the model, resulting in the effect of reducing information loss. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a schematic diagram of the 7-layer encoder structure of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0038] The method of the present invention will be further described in detail below with reference to the accompanying drawings.

[0039] Refer to Figure 1 , and the building change detection method for high-consequence areas of natural gas pipelines based on remote sensing images is as follows:

[0040] S1: Model establishment. Use the natural gas pipeline vector, overlay it on multi-temporal remote sensing images, take the natural gas pipeline vector line as the center, automatically draw high-consequence area vectors on both sides, and generate a high-consequence area identification strip map using the high-consequence area vectors. The specific implementation process is described as follows:

[0041] Rasterize multi-temporal remote sensing images, overlay the natural gas pipeline vector on multi-temporal remote sensing images, take the pipeline vector line as the center, draw high-consequence area vectors within a range of 200 meters in width and 2000 meters in length on both sides, only retain the pixel values within the high-consequence area strip region, and set the pixels in the remaining regions of the image to 0 to obtain the high-consequence area identification strip maps of two periods of multi-temporal remote sensing images;

[0042] S2: Select the images around the natural gas pipeline, and make a change detection data set. First, use multi-temporal remote sensing images of the same size and the same area for cropping respectively, and the cropping size is 256×256×3 to obtain the training sample sets of the previous and later periods. In addition, make a change detection label file, label the images of the previous and later periods with building changes respectively, and construct a change detection sample set. In order to increase the number of the sample set, the images of the previous and later periods and the labeled images are expanded, and the specific methods include rotating 90°, 180°, 270°, and flipping up and down and left and right for data expansion;

[0043] S3: The building change detection network framework for high-consequence areas includes a backbone network and a secondary network. The backbone network adopts an encoder-decoder structure. The encoder has a total of seven layers. Specifically, the first five layers adopt the SE_ResNet50 network, and the additional two-layer encoder structure is constructed using a pooling layer and a convolutional layer. The encoder structure parameters are asFigure 1 As shown in the figure. The output sizes of the feature maps of the first three encoder branches are 64×64×64, 256×64×64, and 512×32×32 respectively. The feature fusion module first upsamples the output branch of the third-layer encoder by a factor of two, and then reduces the dimensions of the three encoder branches. The output size of the feature map of each encoder branch after processing is 64×64×64, and then the three encoder branches are merged. The output sizes of the feature maps of the last four encoder branches are 1024×32×32, 2048×16×16, 2048×8×8, and 2048×4×4 respectively. The feature fusion module upsamples the fifth, sixth, and seventh-layer encoder branches by a factor of two, four, and eight respectively, and then performs dimensionality reduction on the last four encoder branches. The output size of the feature map of each encoder branch after processing is 256×32×32, and then the outputs of the four encoder branches are merged. After the features merged from the first three encoder branches are upsampled by a factor of two, they are fused with the features merged from the last four encoder branches. The fused features preserve more high-level and low-level change feature information. For the operations of feature map upsampling, dimensionality reduction, and merging in the feature fusion module, the attention module is set after the feature fusion module. The attention module includes an attention module and a spatial attention module, which are connected in parallel. Information on building changes is extracted in the channel and spatial ranges to increase the attention to the changed building areas. The decoder uses sub-pixel convolution to restore the feature map output by the attention module. Finally, the network outputs the results of the input building change patches, whose size is the same as the labeled image. The building change attention map derived from the backbone network only retains the information of the building change areas, and the irrelevant information is filtered. The size of the building change attention feature map is 256×256×2, and its main function is to be input into the pre-trained model of the secondary backbone network VGG16 to guide the training of the secondary backbone network;

[0044] S4: The loss between the change patches of the two-phase images output by the last layer of the backbone network and the Ground Truth labeled in Step 2 is statistically calculated. The loss function is the binary cross-entropy loss function. The specific implementation process is described as follows:

[0045] For the building change detection algorithm based on multi-temporal remote sensing images, changed buildings are extracted from multi-temporal remote sensing images, and the building change patch images are output. The loss is calculated using the output building change patch images and the labeled images in Step 2. The backbone network for high-consequence area change detection uses the cross-entropy loss function for binary classification. For the training samples (x, y), let x be the sample and y be the corresponding label value, and the set of predicted values is {0, 1}. Suppose the true label of a certain building sample is y gt , and the probability value of the result predicted for this building sample is y p , and the loss function of this building sample is defined as follows:

[0046]

[0047] Among them, y gt is the true label value of the building sample, and y p (x) i is the probability of the building sample when y gt = 1;

[0048] S5: For the building change detection algorithm in the high-consequence area of the pipeline based on multi-temporal remote sensing images, the building change attention map derived from the backbone network is mainly used for the training of the sub-backbone network. After the building change attention map is activated by the difference and Sigmoid functions, it is multiplied by the matrix of the early image and the late image respectively, and then input into the sub-backbone network using the VGG16 pre-trained model for the training of the sub-backbone network. The sub-backbone network uses the mean square error loss function, and the specific implementation process is described as follows:

[0049] The building change attention is derived from the layer before the output prediction result of the backbone network. The size of the change attention map feature map is 256×256×2. After being activated by the difference and Sigmoid functions and multiplied by the matrix of the early image and the late image respectively, it is input into the VGG16 pre-trained model. Different from other change detection algorithms, the building change detection algorithm in the high-consequence area of the pipeline uses the building change attention map generated by the backbone network to not only constrain the training process of the sub-backbone network, but also constrain the training process of the entire network. The sub-backbone network uses the mean square error loss function to reduce the loss of building information and improve the accuracy of building target segmentation. The mean square error loss function used by the sub-backbone network is defined as follows:

[0050] y' p (x) i = y P (x)[0] i - y P (x)[1] i

[0051]

[0052]

[0053]

[0054] Among them, the size of the change attention feature map is 256×256×2, and y p (x) i represents the change attention map obtained by the prediction of the i-th sample, and y p (x)[0] i represents the 0-channel feature of the change attention map of the i-th sample, and y p (x)[1] iIndicates the change attention of the i-th sample Figure 1 Channel feature, y' p (x) i Is the difference result, y A (x) i And y B (x) i Are the i-th set of early-stage image samples and late-stage image samples, And Are the results after the early-stage image and the late-stage image are differentiated and activated by the sigmoid function, Loss BCE Is And The result after mean square error;

[0055] S6: Weighted sum the losses obtained in S3 and S4 as the loss function of the high-consequence area building change detection network, and then perform backpropagation. The specific implementation process is as follows:

[0056] After obtaining the loss values of the backbone network and the secondary backbone network, weighted sum them to obtain the total loss. The total loss function of the network is defined as follows:

[0057] Loss sum = Loss BCE + αLoss MSE

[0058] Where, Loss BCE Is the binary cross-entropy loss, α is the weight coefficient of the binary cross-entropy loss of the backbone network, Loss MSE Is the mean square error loss. Since the backbone network undertakes the main task in the whole network, that is, the high-consequence area building transformation detection task, in order to make the network focus more on building transformation detection, the loss function of the backbone network should have a greater weight than the loss function of the secondary backbone network, that is, α > 1. After obtaining the total network loss, perform backpropagation to update the parameters in the network;

[0059] S7: Train the high-consequence area building change detection network until the building change detection network meets the actual use requirements. At this time, the backbone network model is the required high-consequence area building change detection model. The specific implementation process is as follows:

[0060] Set parameters such as the optimizer, number of iterations, batch size, etc. of the network to make the model adaptively learn the optimal high-consequence area building change detection scheme. During the training process, the network updates the weight parameters through backpropagation. After multiple rounds of iteration, the loss function converges, and then test whether the building change detection network meets the actual use requirements. If not, continue training by increasing the data volume. If it meets the requirements, the backbone network model at this time is the required high-consequence area building change detection model;

[0061] S8: Input the high-consequence area strip map obtained in S1 into the high-consequence area building change detection model. The model automatically detects the building changes in the high-consequence area to obtain change patches. The specific implementation process is described as follows:

[0062] In S1, using the natural gas pipeline vector, the high-consequence area vector is drawn according to the natural gas pipeline vector line. Then, the multi-temporal remote sensing images are respectively cropped using the high-consequence area vector to obtain the high-consequence area identification strip map. The two-phase high-consequence area identification strip maps are cropped into tiled images. The cropping size of the tiled images is 256×256×3. The obtained tiled images are input into the high-consequence area building change detection algorithm, and the high change detection algorithm model is called to predict the tiled building change patches. In order to obtain the whole building change patch result, a full-zero mask image of the same size as the test image is generated. The large image is cut with a step size of 256. The predicted tiled building change patch results are restored to the corresponding positions of the mask image according to the slice coordinate positions, and are stitched in sequence according to the coordinate serial numbers at the time of cropping to obtain the whole building change patch image, completing the entire prediction process;

[0063] S9: The whole building change patch image result stitched in S8 lacks georeference coordinate information and cannot locate the specific position where the building changes. It is necessary to add georeference information to the stitched whole change patch image. The two-phase high-consequence area images input carry georeference information. Only the geocoordinate information of the input images needs to be copied to the whole building change patch image to generate the whole change patch image with georeference coordinate information. In order to obtain the vector of the whole change patch image, a building change patch vector is generated according to the building change patches. This vector records the specific positions of the changed buildings. The change patch vectors are respectively superimposed on the previous image and the later image to visually display the building changes in the multi-temporal remote sensing images.

[0064] The embodiments of this specific implementation manner are all preferred embodiments of the present invention, and do not limit the protection scope of the present invention accordingly. Therefore, all equivalent changes made according to the structure, shape, and principle of the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for detecting building changes in high-consequence areas of natural gas pipelines based on remote sensing images, characterized in that, The steps are as follows: S1: Superimpose the natural gas pipeline vectors onto two multi-temporal remote sensing images respectively. With the vector lines as the center, draw the high-consequence area vectors, and generate a high-consequence area strip map based on the drawn high-consequence area vectors; S2: Screen the multi-temporal remote sensing images around the natural gas pipeline, construct a change detection dataset for training the high-consequence area building change detection model. Use semantic segmentation annotation software to annotate the multi-temporal remote sensing images, produce the change building label image Ground Truth, and increase the number of dataset samples through data augmentation; S3: The natural gas pipeline high-consequence area building change detection method includes a backbone network and a secondary network. The backbone network adopts an encoder-decoder structure. The encoder module has a total of seven encoder structures. The first five layers use SE_ResNet50, and the additional two layers of encoders are constructed using pooling layers and convolutional layers for extracting high-level feature information. The multi-scale fusion module and the attention module in the backbone network are mainly used for extracting change building features. The change building features are restored to the original size through the decoder, and the building changes in the multi-temporal remote sensing images are predicted. The backbone network also outputs a building change attention map for the training of the VGG16 pre-trained model of the secondary network; S4: Calculate the loss between the building change patches output by the backbone network and the Ground Truth annotated in S2. The loss function uses the binary cross-entropy loss function; C1: Extract the change buildings from the multi-temporal remote sensing images, output the building change patch image, and calculate using the building change patch image and the label image annotated in S2 to count the loss generated during the training process; C2: High-consequence area change detection mainly targets the changed buildings within the high-consequence areas in two-phase images. Buildings belong to one category, and those other than buildings belong to another category. For a sample (x, y), let x be the sample and y be the corresponding label value, and the set of predicted values is {0, 1}. Suppose the true label of a building sample is y gt , and the probability value of the result predicted for this building sample is y p , and the loss function of this building sample is defined as follows: where y gt is the true label value of the building sample, and y p (x) i is the probability of the building sample when y gt = 1; S5: After the building change attention map output by the backbone network is activated by the difference and the Sigmoid function, the change building area feature probability map is extracted, and it is multiplied by the matrix of the early image and the late image respectively, and input into the VGG16 pre-trained model of the secondary network for model training. The loss metric of the secondary network uses the mean squared error loss function; D1: The backbone network predicts the building change patches, and the secondary network outputs the building change attention map, which is mainly used for the training of the secondary network. After the building change attention is activated by the difference and the Sigmoid function, it is multiplied by the matrix of the early and late images in the training set respectively, and then input into the VGG16 pre-trained model, and the building change attention map is used to guide the training of the secondary network; D2: The secondary network uses the mean squared error loss function to optimize the training process of the model and improve the effect of building target segmentation. The mean squared error loss function adopted by the secondary network is defined as follows: y' p (x) i =y p (x)[0] i -y p (x)[1] i Among them, the size of the change attention feature map is 256×256×2, y p (x) i represents the change attention map obtained by the prediction of the i-th sample, y p (x)[0] i represents the 0-channel feature of the change attention map of the i-th sample, y p (x)[1] i represents the 1-channel feature of the change attention map of the i-th sample, y' p (x) i is the difference result, y A (x) i and y B (x) i respectively represent the i-th set of early and late image samples, and are the results after the early and late images are differentiated and activated by the Sigmoid function, LOSS BCE is the result of the mean squared error sum of all samples; S6: Weightedly sum the binary cross-entropy loss and the mean squared error loss function obtained in S4 and S5 as the total loss of the high-consequence area building change detection network, and update the parameters of the entire model through backpropagation; S7: Use the self-built change detection dataset in S2 to train the high-consequence area building change detection model until the building change detection model meets the actual usage requirements. The network model at this time is the high-consequence area building change detection model; S8: According to the two-phase high-consequence area strip maps obtained in S1, the two-phase high-consequence area strip maps are respectively cropped to a size of 256×256×3 and input into the high-consequence area building change detection network. The trained model is called to automatically detect the changed buildings in the high-consequence area images, and the building change patches are output; S9: The building change patch images output by the high-consequence area change detection algorithm lack georeference coordinate information and cannot accurately locate the positions of the changed buildings. It is necessary to add the georeference information of the input temporal remote sensing images to the building change patches, and then generate the building change patch vectors according to the change patches.

2. The method for detecting building changes in high-consequence areas of natural gas pipelines based on remote sensing images according to claim 1, wherein: The method for generating the high-consequence area strip map in S1 is as follows: A1: Rasterize the input temporal remote sensing image, overlay the natural gas pipeline vector on the temporal remote sensing image. With the pipeline vector line as the center, draw the high-consequence area vector within a range of 200 meters in width and 2000 meters in length on both sides, and use the defined high-consequence area vector to crop and generate the high-consequence area identification strip map; A2: Only retain the pixel values in the high-consequence area strip region, set the pixel values in the remaining regions of the image to 0, and crop the high-consequence area identification strip maps of the two-phase temporal remote sensing images.

3. The method for detecting building changes in high-consequence areas of natural gas pipelines based on remote sensing images according to claim 1, wherein: The segmentation annotation amplification method in S2 is as follows: B1: Select the images around the natural gas pipeline, make the change detection data set. First, crop the same-size and same-region temporal remote sensing images respectively, with the cropping size of 256×256×3, to obtain the training sample sets of the two periods before and after. In addition, use the semantic segmentation annotation software to make the change detection label files; B2: Annotate the two-phase images with building changes respectively to construct the change detection sample set. To increase the number of the sample set, the two-phase images and the annotated images are data-augmented. The specific methods include rotating 90°, 180°, 270°, and flipping vertically and horizontally.

4. The method for detecting building changes in high-consequence areas of natural gas pipelines based on remote sensing images according to claim 1, wherein: The specific operation method of S6 is as follows: D1: After obtaining the loss values of the backbone network and the secondary backbone network, their weighted sum is obtained to get the total loss. The total loss function of the network is defined as follows: LOSS sum = LOSS BCE + αLOSS MSE Among them, LOSS BCE is the binary cross-entropy loss, α is the weight coefficient of the binary cross-entropy loss of the backbone network, and LOSS MSE is the mean squared error loss; D2: Since the backbone network undertakes the main task in the whole network, that is, the high-consequence area building transformation detection task, in order to make the network focus more on the building transformation detection, the loss function of the backbone network has a greater weight than that of the secondary backbone network. That is, the weight coefficient α of the binary cross-entropy loss of the backbone network is >1. After obtaining the total network loss, backpropagation is performed to update the parameters in the network.

Images