A building geographic information database updating method

Abstract

The application provides a building geographic information database updating method, and belongs to the technical field of remote sensing and geographic space information, and comprises the following steps: inputting a building remote sensing image into a stacked U-Nets model, obtaining building raster data and performing vectorization processing; performing matching on building vector data in different periods based on an evidence theory vector matching method, obtaining a newly-built or demolished building part, and updating a building geographic information database according to the newly-built or demolished building part. The method can update the building geographic information database, and can also update artificial object data in a land use / land cover database.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of remote sensing and geographic spatial information, and particularly relates to a building geographic information database updating method. BACKGROUND

[0002] The building geographic information database, as a city basic geographic information database, can provide basic data for various industry applications through construction and maintenance.

[0003] The building data is large in scale, and the buildings are removed or newly built at a high speed. The traditional construction and maintenance of the building geographic information database is usually realized through field investigation and measurement, and has problems of complicated process, high cost, long cycle and the like. SUMMARY

[0004] In order to overcome the above-mentioned problems of the prior art, the present application provides a building geographic information database updating method.

[0005] In order to achieve the above-mentioned purpose, the present application provides the following technical scheme.

[0006] A building geographic information database updating method comprises the following steps:

[0007] inputting a building remote sensing image into a stacked U-Nets model to obtain building raster data and perform vectorization processing;

[0008] performing matching on building vector data in different periods based on an evidence theory-based vector matching method to obtain a newly built or removed building part, and updating the building geographic information database according to the newly built or removed building part;

[0009] The stacked U-Nets model comprises:

[0010] a 7*7 convolution filter with a step of 2;

[0011] a residual block, an input end of which is connected to an output end of the convolution filter;

[0012] a first U-Net stacking module, an input end of which is connected to an output end of the residual block, and which comprises two series-connected U-Net modules;

[0013] a second U-Net stacking module, an input end of which is connected to the first U-Net stacking module, and which comprises seven series-connected standard U-Net modules;

[0014] a third U-Net stacking module, an input end of which is connected to the second U-Net stacking module, and which comprises seven series-connected standard U-Net modules;

[0015] a fourth U-Net stacking module, an input end of which is connected with the third U-Net stacking module, comprising one U-Net+ module, the U-Net+ module being a U-Net module containing only an encoder E1 and a decoder D1;

[0016] a 1*1 convolution + softmax layer, an input end of which is connected with an output end of the fourth U-Net stacking module.

[0017] Further, output ends of the first U-Net stacking module and the second U-Net stacking module are provided with average pooling layers, and the U-Net modules in the third U-Net stacking module and the fourth U-Net stacking module adopt extended convolution.

[0018] Further, in the first U-Net stacking module, the second U-Net stacking module, the first U-Net module of the third U-Net stacking module and the U-Net+ module of the fourth U-Net stacking module, a 1*1 convolution is used to realize a skip connection, and connection modes of all the remaining U-Net modules are realized as identity mapping connections.

[0019] Further, the U-Net module comprises 10 convolutional blocks connected in series, and a batch normalization and a ReLU activation function are used to process image vector data at an input end of each convolutional block.

[0020] 9. Further, the batch normalization comprises:

[0021] normalizing the data, and the normalization operation being:

[0022]

[0023]

[0024]

[0025] wherein m represents a sample quantity extracted by one training, μ B represents a mean value, represents a variance, and ε is a very small value close to 0 added to avoid a denominator being 0;

[0026] performing a scale transformation and a shift operation:

[0027]

[0028] wherein γ is a scale factor and β is a translation factor.

[0029] Further, the ReLU activation function is:

[0030]

[0031] Further, the vector matching method based on evidence theory matches the building vector data in different periods, comprising:

[0032] Using ArcPy to calculate the area, perimeter, spatial position, eccentricity, rectangularity and other vector characteristics F1 of each building;

[0033] The closest set of objects in the existing building geographic information database to the object to be matched F1 is used as the candidate object set {C i} to construct the recognition framework Θ F1 ={matC1,matC2,…,matC i ,…,matC N ,NM};Wherein, matC i Indicates that the candidate C i is the same real world object as the feature F1;

[0034] Judge whether C i is the same as F1;

[0035] Using Dempster rule to combine the judgment criteria of each C i , form the judgment evidence from different criteria;

[0036] Calculate the evidence quality of each candidate object, and after fusing the large amount of evidence assigned to each candidate object, use Dempster rule to fuse the candidate objects;

[0037] Find a candidate object with the highest confidence using pignistic probability, and match the candidate object with the highest confidence with the object to be matched F1.

[0038] Further, the stacked U-Nets model is trained using the cross-entropy loss function, and the cross-entropy loss function is:

[0039]

[0040] In the formula, p(X i ) represents the true probability, and q(Xi) represents the predicted probability.

[0041] The building geographic information database updating method provided by the application has the following beneficial effects:

[0042] The application constructs a stacked U-Net model, which can quickly identify and extract buildings, has high accuracy, and greatly improves work efficiency; and a vector matching method based on evidence theory is applied to match building geographic information data in different periods, obtain newly built or demolished building parts, and update the building geographic information database according to the same, thereby solving the problems of low accuracy and high cost of building update in the urban geographic database, and having high automation. BRIEF DESCRIPTION OF DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the application and the design scheme thereof, the drawings required by the embodiments will be briefly introduced as follows. The drawings in the following description are only part of the embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of the drawings.

[0044] Figure 1 It is a method flowchart of the application;

[0045] Figure 2 It is a stacked U-Net model structure diagram of the application;

[0046] Figure 3 It is a cropped version structure diagram of U-Net+ in the embodiments of the application;

[0047] Figure 4 It is a new building distribution diagram identified by the stacked U-Net model in the embodiments of the application;

[0048] Figure 5 It is a vector matching method flowchart used in the embodiments of the application;

[0049] Figure 6 It is a relationship diagram between the standard deviation and the probability mean of the detected building in the embodiments of the application; DETAILED DESCRIPTION

[0050] In order to make those skilled in the art better understand the technical scheme of the application and can be implemented, the application will be described in detail in conjunction with the drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical scheme of the application, and cannot be used to limit the protection scope of the application.

[0051] In addition, the terms "first", "second" and the like are used only for descriptive purposes and should not be construed as indicating or implying relative importance. In the description of the present application, it should be noted that, unless otherwise explicitly specified or limited, the terms "connected", "connected" should be understood broadly, for example, it can be fixedly connected, or it can be detachably connected, or it can be integrally connected; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances. In the description of the present application, unless otherwise specified, the meaning of "a plurality of" is two or more, which will not be described here.

[0052] Embodiments:

[0053] The present application provides a building geographic information database updating method, specifically as shown in the figure, comprising: inputting the building remote sensing image into the stacked U-Nets model, and extracting the building grid data image in the building remote sensing image by using the U-Nets model; vectorizing the building grid data, and performing vector matching on the vectorized building data and the existing building geographic information database, so as to obtain the change of the building, and updating the existing building geographic information database according to the change of the building; Figure 1

[0054] The stacked U-Nets model comprises: a 7*7 convolution filter with a step of 2; a residual block, the input end of which is connected with the output end of the convolution filter; a first U-Net stacking module, the input end of which is connected with the output end of the residual block, which comprises two U-Net modules connected in series; a second U-Net stacking module, connected with the first U-Net stacking module, which comprises seven standard U-Net modules connected in series; a third U-Net stacking module, connected with the second U-Net stacking module, which comprises seven standard U-Net modules connected in series; a fourth U-Net stacking module, connected with the third U-Net stacking module, which comprises one U-Net module containing only an encoder E1 and a decoder D1; a 1*1 convolution + softmax layer, the input end of which is connected with the output end of the fourth U-Net stacking module.

[0055] The following is a specific embodiment of the present application:

[0056] S1, obtaining high-resolution images, pre-processing the data, making a certain number of building samples, and constructing a stacked U-Nets network to extract urban buildings; wherein the label of the building sample is the classification result obtained by manually identifying the remote sensing image;

[0057] ​S2, the extracted building raster data is processed by using a morphological algorithm, and then raster data is vectorized to obtain building data in the same format as the existing building geographic information database, so as to facilitate subsequent building comparison;

[0058] S3, extracting building vector features, and using a vector matching method based on evidence theory to perform vector matching between the vectorized building data and the existing building geographic information library;

[0059] S4, combining auxiliary data, evaluating the building extraction result after vector matching, and updating the building database;

[0060] Specifically, the specific steps of step S1 are:

[0061] S11, obtaining high-resolution remote sensing images, identifying buildings in the images by using artificial visual interpretation, drawing building samples in ArcGIS software, and making labels in Shapefile format for subsequent model training;

[0062] S12, using Python to write random cropping code to uniformly crop the obtained high-resolution images and building label vector data into 512x512 pictures, and sequentially name them 1, 2, 3, 4, …; wherein the high-resolution remote sensing image data has R, G, and B channels, and the pixel value range is 0-255; the label vector data needs to be rasterized and converted into a binary raster format with pixel values of 0 (background value) and 255 (building); the above cropped high-resolution images and building labels are in JPG and PNG formats respectively;

[0063] S13, dividing the processed image samples and label data into a sample set and a training set, in the embodiment of the present application, 80% is randomly selected for network model training, and the remaining 20% is used to verify the performance of the trained network model, and the optimal network model is obtained through multiple training and verification, in the present example, a total of 4000 training images and corresponding label images are used, and 1000 verification images and corresponding label images are used.

[0064] S14, constructing a stacked U-Nets model, normalizing the cropped remote sensing image pictures and label picture training data, and inputting them into the network in batches for training to obtain the trained neural network;

[0065] The U-Net architecture was originally proposed for segmenting biomedical images. U-Net captures context information at multiple scales through a compression (encoder) and expansion (decoder) path, which can achieve good segmentation results without requiring a large amount of training data, and is widely used in the field of biomedical image segmentation. However, when facing complex color profiles, multiple bands and occlusions in remote sensing images, the segmentation performance of the U-Net network is not good.

[0066] The present application expands the capabilities of U-Net by stacking U-Net blocks into a deep architecture. When data passes through multiple U-Net modules, high-resolution features and low-resolution context information can be mixed together, and after many layers of processing, information-rich high-resolution features are generated. In addition, compared with traditional architectures, stacked U-Net models require fewer feature maps per layer, so they can achieve higher performance with fewer parameters.

[0067] In a specific implementation, as shown in Figure 2 The stacked U-Net model for building extraction includes 4 blocks of stacked U-Net, each containing 2, 7, 7 and 1 standard U-Net modules, respectively. Each U-Net module is composed of 10 convolutional blocks, each of which uses batch normalization and ReLU activation function. Low-resolution and high-resolution feature maps are generated using transposed convolution and deconvolution, respectively. The feature map resolution at the input and output of each U-Net module is the same, and the convolution process uses a 3x3 convolution kernel. In addition, the jump connection within the module helps to avoid the gradient vanishing phenomenon.

[0068] Batch normalization (Batch Normalization) is a data processing specification method, which essentially pulls the data output by the network structure back to the standard normal distribution, i.e. the mean of the data is 0 and the variance is 1. When the data is distributed in the standard normal distribution, the non-linear transformation generated after the activation function is more sensitive, so that the activation function can exert greater non-linear capability, which can effectively avoid gradient disappearance and speed up network convergence and training speed. Using batch normalization first normalizes the data obtained, as shown in the following formula:

[0069]

[0070]

[0071]

[0072] In the formula, m represents the number of samples extracted for training once, μ B represents the mean, denotes the variance, and ε is a small value close to 0 added to avoid the denominator being 0. Then, a scaling and shifting operation is performed, as shown in the following formula:

[0073]

[0074] where γ is the scale factor and β is the shift factor. The scaling and shifting operation is to make it possible for the batch normalization step added due to the training requirement to restore the original input, so as to ensure the capacity of the entire network. The γ and β can be learned during the network training, so that the neural network can select a most suitable distribution during the training process, which makes the network have a better performance.

[0075] The activation function provides the nonlinear modeling capability of the network, defines the mapping relationship between the input and the output, and to a large extent determines the ability of the neural network. In the example of the present application, the ReLU function is used. The ReLU function outputs 0 when x≤0, which will cause the death of neurons. The activation function introduces a nonlinear factor to the neuron, so that the neural network can arbitrarily approximate any nonlinear function, and thus can be applied to many nonlinear models. The calculation formula of the ReLU function is as follows:

[0076]

[0077] The input image passes through a 7x7 convolutional filter (with a stride of 2) and a residual block. In this stage, the traditional max-pooling layer is replaced by the cross-convolution layer inside the residual block. Then, four stacked U-Net modules are used. These modules combine high-resolution captured details with low-resolution captured long-range contextual information to generate a segmentation map of the scene.

[0078] Two strided convolutions and two average pooling are used to increase the receptive field, which helps to capture long-range information. For an input size of 512x512, the output map size is 32x32. Average pooling is applied to the output of the first and second stacked U-Net blocks. In the third and fourth blocks, the U-Net modules perform dilated convolution to keep the feature map resolution unchanged. The input size of the feature map of the fourth U-Net module is 7x7, which is further reduced to 2x2 at the input of the encoder E2. At this resolution, it is not possible to have E2 and D2 layers, so a cropped version of U-Net (U-Net+) is used in the fourth block, i.e., only the encoder E1 and the decoder D1, as shown in Figure 3 .

[0079] In the first U-Net of each block, a 1x1 convolution is used to implement the skip connection. The remaining connections of all U-Net in each block except the first U-Net are implemented as identity mapping. The number of output feature maps of each block is roughly the same as the total number of feature maps generated by all previous U-Net modules, which enables the architecture to flexibly retain all these feature maps.

[0080] The output is rescaled to the original size using bilinear interpolation. The model trains the network using a cross-entropy loss function, the formula of which is as follows:

[0081]

[0082] In the formula, p(X i ) represents the true probability, and q(Xi) represents the predicted probability.

[0083] The Adam optimizer is used with a starting learning rate of 1.0e-4, weight decay and momentum values of 1.5e-4 and 0.95 respectively, the pooling step at the end of the third and fourth blocks is removed, the output is performed with a stride of 16, and the training and testing processes are implemented using the PyTorch framework. The model is trained on an RTX3060 GPU with a batch size of 8.

[0084] S15, call the trained model to extract buildings in a certain city in 2018; use the above training data preprocessing method to perform data preprocessing operations such as cropping on the images of the study area, and the image size is also 512x512. Use the cv2.imread() function in OpenCV to read the image and perform normalization operation, call the trained model to recognize the building, and in the study area of the present embodiment, a total of 3242 buildings are recognized by the stacked U-Net model used, as shown in Figure 4 .

[0085] Step S2 in this embodiment is to process the extracted building raster data and then vectorize the raster data, specifically:

[0086] S21, use the dilation algorithm and opening algorithm in mathematical morphology to preprocess the building raster data to reduce the "salt and pepper noise" in the raster image and smooth the boundary. The dilation operation and opening operation are implemented using the cv2.dilate() and cv2.morphologyEx() functions in OpenCV.

[0087] S22, import GDAL library, use gdal.Polygonize() function in it to vectorize, set six parameters of input raster image band, vector layer after vectorization, index of DN value to be written into vector field, algorithm option, progress bar callback function, progress bar parameter respectively to call the method, get the shape and value of the connected region in the image, and the polygon representing the background will be deleted from the result.

[0088] Step S3 in the embodiment performs vector matching on the vectorized building data and the existing building database. Vector matching is a process of identifying the same ground object in different source maps by analyzing the differences and similarities of vector elements. The matching basis includes distance measure, geometric shape, topological relationship, graph structure and other similarity indexes. Therefore, it is necessary to extract vector features before matching. In the embodiment, ArcPy is used to calculate the vector features of the building data, including the area, perimeter, spatial position and shape features such as eccentricity and rectangularity of each building vector, to prepare for the subsequent vector matching work.

[0089] The vector features of the building data are calculated by using ArcPy, and the vector features include:

[0090] The area, perimeter, spatial position, eccentricity and rectangularity of the building are calculated as the basis for vector matching. The eccentricity is used to reflect the flatness of the polygon shape relative to the circle, and the aspect ratio of the similar shape building should be approximate. The calculation formula is as follows:

[0091]

[0092] In the formula, L represents the length of the building, and W represents the width of the building.

[0093] The rectangularity represents the filling degree of the shape in the minimum circumscribed rectangle. The calculation formula is as follows:

[0094]

[0095] In the formula, S0 represents the area of the building, and S R represents the area of the minimum circumscribed rectangle. The minimum circumscribed rectangle of the building is generated by using the function of arcpy.management.MinimumBoundingGeometry(), and then the rectangularity is calculated.

[0096] In the embodiment, the knowledge fusion data matching method based on evidence theory is used for vector data matching, and the method flow is as shown in Figure 5As shown, for two vector data sets to be matched, in the matching process, first, according to each feature belonging to a data set 1, find the potentially same features in another data set 2, and then analyze these candidate objects to determine the final matching link, the matching process used in the examples of the present application follows this method, including the following five steps, specifically:

[0097] S31, construct the to-be-matched vectorized building data and the existing building geographic information database identification framework. For the vectorized feature set F1 in the existing building geographic information database, find the close feature set in the to-be-matched vectorized building data according to the distance criterion, these feature sets are candidate {Ci}, i = 1, …, N, and the candidate features are used for matching F1. The combination of features describing the same building object with different features is regarded as a feature set. The feature sets in the two data sources are stored using np.array() respectively. The general identification framework is defined as follows:

[0098] Θ = {matC1, matC2, …, matC i ,…,matC N ,NM}

[0099] Where matCi is the hypothesis matched with the Ci-th vector feature set, and NM represents the hypothesis that "F1 does not match any feature in the to-be-matched vectorized building data".

[0100] S32, analyze each building object independently. Construct a Euclidean distance matrix with 2 Θ as the column and 2 Θ as the row of the to-be-matched building object, and regard each row as a piece of evidence source. (The matrix is a feature matrix)

[0101] S33, fusion of matching criteria. The matching criterion is a piece of evidence source with a length of 2 Θ . Adopt the Euclidean distance criterion, use the numpy library to calculate the Euclidean distance matrix of the to-be-matched vectorized building data relative to the vectorized feature set of the existing building geographic information database object; according to the linear normalization criterion, map each row of the Euclidean distance matrix to [0, 1] and ensure that the sum is 1; consider the buildings of the existing building geographic information database as labels, which are represented as diagonal 1 and other positions 0 in the Euclidean distance matrix. Arrange each piece of evidence source in the two data sources as evidence data m A (c i ), A = 1, …, N and m B (c i ), B = 1, …, N, and the arrangement process is basic probability assignment determination, the formula is:

[0102]

[0103] where |z i represents a single hypothesis; x and z i both represent 2 Θ a hypothesis in 2 i ; a certain hypothesis z i , x is all other hypotheses in 2 Θ except z i | = 1 represents a single hypothesis.

[0104] S34, fusion of evidence data. The same evidence data in different data sources is fused using the Dempster rule. The Dempster rule fusion formula is:

[0105]

[0106] where P represents one of the Θ hypotheses, m(P) represents the evidence combination quality allocated to a certain hypothesis, and A and B represent different information sources. At the end of this step, a combination quality of evidence is allocated to each of the Θ hypotheses.

[0107] S35, final decision of data matching. The standard used for decision is the maximum value of the pignistic probability. According to the pignistic probability formula, the final decision result of the existing building geographic information database and the vectorized building data to be matched is calculated. In this process, the evidence with the highest confidence is selected. For each selected hypothesis, a confidence equal to the difference between the first maximum value and the second maximum value is calculated, and the suspicious result (confidence less than a threshold or conflict between standards) is highlighted for subsequent checking.

[0108] Step S4 in this embodiment evaluates the result of vector matching, and further updates the building database, specifically:

[0109] The buildings in the building geographic information database of a city in 2016 are regarded as ground truth data, high-resolution orthophotos in 2018 are adopted, if the building polygon detected by the neural network is matched with the vector building in the database, it is considered that the classification is correct and does not need to be updated, otherwise, it is determined whether the update type is a new building or a removed building, if the building detected by the neural network is not successfully matched with the vector building in the database, it is determined as a new building; if the vector building in the database is not successfully matched with the building vector detected by the neural network, it is determined as a removed building, and the building geographic information database is updated.

[0110] The average value and the standard deviation value of the probability output by the stacked U-Nets network are comprehensively used to evaluate the building detection situation, such as Figure 6The buildings with false classification tend to cluster in the lower left corner with lower mean and standard deviation values of building probability. The mean value of building probability has a greater impact on distinguishing the buildings with correct and false detection. When the mean value is greater than 0.8, there are few buildings with incorrect detection. In order to maximize the results obtained by the embodiments of the present application, the threshold values of the mean and standard deviation values are set to 0.7 and 0.1, and the results retain 1471 objects, of which 1139 objects (77.4%) are the buildings with correct detection to be updated.

[0111] The above embodiments are only the preferred specific embodiments of the present application, and the protection scope of the present application is not limited thereto, and any simple change or equivalent replacement of the technical solutions within the technical scope disclosed by the present application can be obviously obtained by any person skilled in the art, and all of them belong to the protection scope of the present application.

Claims

1. A method for updating a building geographic information database, characterized in that, include: The remote sensing images of buildings are input into the stacked U-Nets model to obtain the building raster data and then vectorize it. The vector matching method based on evidence theory matches building vector data from different periods to obtain newly built or demolished building parts, and updates the building geographic information database based on the newly built or demolished building parts. The stacked U-Nets model includes: A 7x7 convolutional filter with a stride of 2; The residual block has its input terminal connected to the output terminal of the convolutional filter; The first U-Net stacking module has its input connected to the output of the residual block, and it includes two U-Net modules connected in series. The second U-Net stacking module has its input connected to the first U-Net stacking module and includes seven standard U-Net modules connected in series. The third U-Net stacking module has its input connected to the second U-Net stacking module and includes seven standard U-Net modules connected in series. The fourth U-Net stacking module, whose input is connected to the third U-Net stacking module, includes one... Module, the The module is a U-Net module containing only encoder E1 and decoder D1; A 1*1 convolutional layer with softmax is connected to the output of the fourth U-Net stacked module. The vector matching method based on evidence theory matches building vector data from different periods, including: The vectorized building data is used as the first dataset, and the existing building geographic information database is used as the second dataset. Extract the vector feature set F1 of buildings in the first dataset; Based on the distance criterion, find the feature set in the second dataset that is close to the vector feature set F1. }, construct a recognition framework ;in, express The hypothesis that the vector feature set F1 matches is NM, which represents the hypothesis that the vector feature set F1 does not match any feature in the existing building geographic information database. For each building in the first and second datasets, construct a... For example, to match building objects The Euclidean distance matrix is ​​a set of rows, and each row of the Euclidean distance matrix is ​​considered as a source of evidence. Using the Euclidean distance criterion, the Euclidean distance matrix of the second dataset relative to the vectorized feature set of the first dataset is calculated using the NumPy library; according to the linear normalization criterion, each row of the Euclidean distance matrix is ​​mapped to the interval [0,1], and the sum of each row is ensured to be 1; Organize each source of evidence in the first and second datasets into evidence data. A=1, ..., N and evidence data B = 1, ..., N; The Dempster rule is used to fuse the same piece of evidence data in the first and second datasets to obtain the fused evidence. Each feature set in the identification framework is calculated based on the fused evidence. The quality of evidence; according to The quality of evidence is calculated by determining the confidence level of the evidence, and the evidence with the highest confidence level is selected. As a set of features that match F1.

2. The method for updating a building geographic information database according to claim 1, characterized in that, The outputs of the first and second U-Net stacked modules are provided with average pooling layers, and the U-Net modules in the third and fourth U-Net stacked modules employ extended convolution.

3. The method for updating a building geographic information database according to claim 1, characterized in that, The first U-Net module, the second U-Net stacking module, the first U-Net module of the third U-Net stacking module, and the fourth U-Net stacking module In this module, 1×1 convolutions are used to implement skip connections, while all other U-Net modules implement identity mapping connections.

4. The method for updating a building geographic information database according to claim 1, characterized in that, The U-Net module comprises 10 concatenated convolutional blocks, with batch normalization and ReLU activation functions applied to the input of each block to process the image vector data.

5. The method for updating a building geographic information database according to claim 4, characterized in that, The batch processing normalization includes: The data is normalized. The normalization operation is as follows: In the formula, m This indicates the number of samples drawn in one training session. This represents the mean. Represents variance. To avoid the denominator being 0, a minimum value close to 0 is added; Perform scaling and offset operations: In the formula, It is a scale factor. It is the translation factor.

6. The method for updating a building geographic information database according to claim 1, characterized in that, The process involves organizing each source of evidence in the first and second datasets into evidence data. A=1, ..., N and evidence data The formula for B=1, ...,N is: In the formula, Indicates a single hypothesis; and express One of the assumptions, in which For the current calculation item, In addition to outside All other assumptions in it, This indicates that the hypothesis is a single hypothesis.

7. The method for updating a building geographic information database according to claim 1, characterized in that, The fusion formula for fusing the same piece of evidence data from the first and second datasets using the Dempster rule is as follows: in, express One of the hypotheses, This represents the quality of the evidence combination assigned to a certain hypothesis. ,B They represent different information sources.

8. The method for updating a building geographic information database according to claim 1, characterized in that, The stacked U-Nets model is trained using the cross-entropy loss function, which is: In the formula, Represents the true probability. This represents the predicted probability.

Images