A class-independent remote sensing ground object vector topology structure extraction method

Abstract

The application discloses a class-independent remote sensing ground object vector topology structure extraction method, which comprises the following steps: step 1, establishing a training set sample library; step 2, constructing a class-independent remote sensing ground object vector topology structure extraction model, TopDiG model, the TopDiG model extracts low-level topology features of ground objects through a convolution network, generates predicted nodes and visual descriptors, then adopts a dynamic label supervision strategy, dynamically generates an adjacency graph label according to the predicted nodes, and finally predicts the ground object topology structure based on a Transformer-based directed adjacency graph generator; step 3, designing a multi-task constraint loss function to train the TopDiG model in step 2; and step 4, using the trained TopDiG model to extract remote sensing ground objects from an image prediction library, so that the remote sensing image target ground object vector topology structure can be obtained. The method can directly obtain the ground object vector topology structure from the input image, and avoids information loss and complex calculation in the process of converting a grid into a vector.

Description

TECHNICAL FIELD

[0001] The present application relates to a class-independent remote sensing feature vector topology extraction method for automatic extraction of multi-class feature vector topology of satellite or aerial remote sensing images. BACKGROUND

[0002] Feature vector topology represented by directed graph structure is the basis of various remote sensing applications, such as geographic information systems, cartographic generalization, and disaster assessment. Traditional vector map production relies on manual or semi-automatic labeling, which is time-consuming and expensive. In recent years, thanks to the rapid development of deep learning, automated vector topology extraction technology has received extensive attention and research in the field of remote sensing. Current remote sensing feature vectorization extraction technology mainly falls into three categories: segmentation-based methods, contour-based methods, and graph generation methods.

[0003] Among them, segmentation-based methods need to vectorize the segmentation results to obtain vector results. Typical methods include PolyMapper, Frame Field, and ASIP, etc. Although these methods can obtain relatively stable segmentation results, details are easily missing in the vectorization process, leading to topology errors. And usually need to design a fine and complex post-processing optimization link for different features, which is computationally expensive and not robust. Contour-based methods are derived from the idea of instance segmentation, which obtains vector topology by continuously adjusting the initial contour. This kind of method does not need complex post-processing process, and can achieve good results on low-complexity planar features (such as buildings). However, this kind of method depends on the quality of the initial contour, and is difficult to apply to linear features such as roads. Representative methods include Polygon-RNN, Polygon-RNN++, Deep Snake, SharpContour, and E2EC, etc. Finally, graph generation methods simultaneously predict key nodes and their connectivity to construct feature vector topology structure. One branch of this kind of method, such as RoadTracer, VecRoad, iCurb, and RNGDet, focuses on linear features, starting from the starting point to construct the entire vector topology structure in a point-by-point prediction manner. These methods are less efficient, cannot solve the problem of error accumulation, and are not suitable for planar features. Another part of the graph generation method, such as PolyWorld and csBoundary, connects the predicted points by predicting the adjacency matrix. These methods can extract relatively regular vector topology structure, but are difficult to handle features with complex topology structure.

[0004] In summary, there is no general model method that can be applied to the topological structure extraction of various types of ground objects at present, because the topological structures of different ground objects are quite different. For example, the topological structure of linear ground objects is generally represented by the center line, which is a non-closed network structure, and the center homogenization feature needs to be focused on during extraction. The topological structure of planar ground objects is generally the boundary contour, which is a closed ring structure, and the edge inhomogeneity feature needs to be focused on during extraction. In addition, there are differences between artificial ground objects (such as buildings, etc.) and natural ground objects (such as water bodies and forests, etc.) in planar ground objects. Compared with artificial ground objects, natural ground objects have complex and variable shapes, variable areas, and fuzzy edge features. The above problems bring challenges to the unified vector topological structure extraction. SUMMARY

[0005] In order to overcome the above problems, the present application provides a class-independent remote sensing ground object vector topological structure extraction method, named TopDiG. For the topological structure of different ground objects, the convolutional neural network focusing on low-level topological information is used to extract the features of the nodes and the corresponding regions. Then, the dynamic label supervision method is used to dynamically generate the directed adjacency graph label according to the unordered prediction points. Finally, the directed adjacency graph generator based on Transformer is used to predict the adjacency matrix between nodes to obtain the vector topological structure of the ground object.

[0006] The present application provides a class-independent remote sensing ground object vector topological structure extraction method, comprising the following steps:

[0007] Step 1, establishing a training set sample library. According to the remote sensing image and historical database of the research area or test area, typical target ground object samples are selected, and corresponding labels are made. Then, according to the size of hardware computing resources, the image is cropped to a suitable slice size, such as 512x512, and the corresponding raster label and vector label are made.

[0008] Step 2, TopDiG model training and testing are performed to realize the extraction of the vector topological structure of various types of ground objects; the TopDiG represents a class-independent remote sensing ground object vector topological structure extraction method.

[0009] The TopDiG model comprises a node detector focusing on topology, a dynamic label supervision strategy, and a directed adjacency graph generator.

[0010] Step 3, designing a multi-task constrained loss function to train the TopDiG model in step 2;

[0011] Step 4, after training, the trained TopDiG model is used to predict the remote sensing ground object extraction of the image prediction library, and the remote sensing image target ground object vector topological structure can be obtained.

[0012] Further, the Topological-concentrated node detector (TCND) of the focused topology includes a feature encoder, a side branch outputter, a feature learning block, a weighting block, a merging block and a feature sampling block; the feature encoder is ResNet50. The side branch outputter mainly includes a convolution layer (Convolution Layer), a batch normalization layer (Batch Normalization, BN) and a transpose convolution layer (Transpose Concolution Layer). The feature learning block includes two convolution blocks, which enhance the low-level topology features and output feature maps. The weighting block includes three convolution blocks, which are used to balance the semantic information. The merging block fuses the outputs of the feature block and the weighting learning block, includes a node detection head (detection head) and a weighted summation operation, and then obtains a fixed number of prediction points according to the Non Maximum Suppression (NMS) algorithm. The feature sampling module further extracts the region features corresponding to the obtained prediction points, i.e. visual descriptors.

[0013] Further, the dynamic label supervision strategy (Directional graph supervision, DGS) is to generate labels in real time during the training process, which is used to supervise the learning of the adjacency matrix in the directed adjacency graph generator. This strategy takes the ordered ground truth nodes as reference, generates the adjacency matrix label represented by the directed adjacency graph based on the unordered predicted nodes, and generates the adjacency matrix label represented by the directed adjacency graph in real time and dynamically in each training period.

[0014] The dynamic label supervision strategy module generates the adjacency graph label according to the predicted node to supervise the learning of the directed adjacency graph. This module receives the unordered predicted nodes v i |i=1,2,…,k},v i ={(x i ,y i )} and the ordered ground truth nodes x i ,y i respectively represent the horizontal and vertical coordinates of node i. First, each ground truth node is traversed to find its nearest predicted node v i , which has a one-to-one matching relationship with v i , that is, each or v i corresponds to v i or The number is not more than 1, and after the matching is completed, the directed adjacency graph label is constructed, the first step is to create a k x k zero matrix, and the row number and column number of the matrix are the sequence numbers of the k prediction nodes; the second step is to traverse all the prediction nodes, if the prediction node v i has a corresponding true value node , all the adjacent nodes of the true value node are found, if the adjacent nodes have corresponding prediction nodes, the prediction nodes are the adjacent nodes of the v i being traversed; the adjacent relationship of the prediction nodes is recorded in the zero matrix constructed in the first step according to the process of the first step and the second step, and the recording method is as follows: assuming that the sequence number of the prediction node v i being traversed in the prediction node list v i is m, and the sequence number of an adjacent prediction node of v i found is n, the value of the element of the coordinate (m, n) in the zero matrix is modified to 1; if there is no adjacent node, the value of the element of the coordinate (m, m) in the zero matrix is modified to 1; after all the prediction nodes are traversed according to the above process, the directed adjacency graph label is constructed.

[0015] Further, the directional graph generator (DiG generator) predicts the adjacency matrix between the extracted nodes, which mainly includes a position encoding module, a multi-layer perception (MLP) module, a connection network and an adjacency graph prediction module. The position encoding module stacks the node coordinates and the visual descriptors in the feature dimension. The multi-layer perception module mainly includes a fully connected layer, a layer normalization (LN) layer and a rectified linear layer (ReLU). The connection network mainly includes 6 Transformer encoder modules, which are composed of a multi-head self-attention layer, a feed forward layer, a multi-layer perception layer and an LN layer. The adjacency graph prediction module includes two parallel adjacency graph prediction heads, which mainly include a convolution layer, a BN layer and a ReLU layer. The results of the two predictions are added to obtain the final adjacency graph. When extracting planar features, the adjacency graph prediction module further includes an optimization module to optimize the predicted adjacency graph.

[0016] Further, the multi-task loss function includes two kinds, a node detection loss function and a directed adjacency graph prediction loss function.

[0017] The node detection loss function is as follows:

[0018]

[0019] where M(.) denotes the operation of averaging, denotes the ground truth heat map of the node, h denotes the heat map of the predicted node.

[0020] The directed adjacency graph prediction loss function is as follows:

[0021]

[0022] where p denotes the predicted adjacency graph, denotes the adjacency graph label.

[0023] Moreover, the final expression of the loss function of the integrated multiple tasks is as follows,

[0024] L total = node + graph .

[0025] Compared with the prior art, the advantages and beneficial effects of the present application are: the present application provides a class-independent remote sensing ground object vector topology structure extraction method, on the one hand, the present application can be applied to the vector topology structure extraction of multiple types of remote sensing ground objects by the same method process, and is a universal model; on the other hand, the present method does not need to pass through the grid intermediate result, and can directly obtain the ground object vector topology structure from the input image. This avoids the information loss and complex calculation in the process of grid to vector conversion. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 is a schematic diagram of TopDiG;

[0027] Figure 2 is a schematic diagram of dynamic production of directed adjacency graph labels. DETAILED DESCRIPTION

[0028] In order to better understand the technical solutions of the present application, the technical solutions of the present application will be described in detail below in combination with the drawings and examples.

[0029] The present application provides a class-independent remote sensing ground object vector topology structure extraction method. For the topology structure of different ground objects, the convolutional neural network focusing on low-level topology information is first used to extract the predicted node and the corresponding region feature. Then, the dynamic label supervision method is used to dynamically generate the adjacency matrix label according to the unordered predicted node. Finally, the directed adjacency graph generator based on the Transformer is used to predict the topology structure of the ground object.

[0030] Based on the structure of the model, the application provides a class-independent remote sensing ground object vector topology extraction method. The specific implementation steps are as follows:

[0031] Step 1. Establishing training set and validation set sample library

[0032] First, the orthographic or corrected satellite or aerial remote sensing image and the corresponding typical ground object labeled vector data need to be prepared. Then, the data is preprocessed, that is, the target ground object labeled vector data is rasterized, the image and the labeled raster data are resampled and cropped to obtain an image and a labeled raster pair with appropriate resolution and consistent size. Finally, combined with computer memory resources, the characteristics of the ground object and other factors, a training sample library with appropriate slice size (such as 512x512 or 300x300) is prepared. In addition, the same processing is performed on the to-be-predicted image using data preprocessing to establish an image prediction library for subsequent model prediction.

[0033] Step 2. TopDiG model training and testing

[0034] The low-level topological features of the ground object are extracted by designing a convolutional network and a feature sampler to generate predicted nodes and visual descriptors. Then, a dynamic label supervision strategy is adopted to dynamically generate an adjacency graph label according to the predicted point. Finally, a directed adjacency graph generator based on the Transformer predicts the topological structure of the ground object. A reasonable loss function is designed as the driving force. The specific modules and steps in the embodiment are as follows:

[0035] 2.1 Node detector focusing on topology

[0036] Figure 1(a) shows a schematic diagram of the structure involved, which includes in turn a feature encoder, a side branch outputter, a feature learning module, a weight learning module, a fusion module and a feature sampling module. The feature encoder selects ResNet50, and 1 heat map with the same size as the original picture and a feature dimension of 1 is outputted by the side branch outputter in each of the 4 stages in the ResNet50 network. The side branch outputter has a total of 4, of which the first one is composed of 1 convolutional layer and 1 BN layer, and the remaining three are composed of a convolutional layer, a BN layer, a transposed convolutional layer and a bilinear upsampling layer. The convolutional layer has a convolution kernel of 1x1 and a step of 1; the transposed convolutional layer has a convolution kernel size of 4, 8 and 16 respectively, and a step of 1, 2 and 8 respectively; the BN layer has a dimension of 1. The 4 heat maps outputted by the 4 side branch outputters are stacked along the feature dimension and sent to the feature learning module. The feature learning module contains 2 convolutional blocks, each of which contains 1 convolutional layer with a convolution kernel size of 3x3, a step of 1 and a padding of 1, 1 BN layer with a dimension of 64 and 1 ReLU layer. The input and output channel numbers of the convolutional layers are (4, 64) and (64, 64) respectively. The weight learning module contains 3 convolutional blocks, the first two of which are composed of 1 convolutional layer with a convolution kernel of 1x1 and a step of 1, 1 BN layer with a dimension of 4 and 1 ReLU layer, and the last convolutional block does not contain a ReLU layer. The input and output feature channel numbers of the 3 convolutional layers are all 4. The fusion module mainly includes a node detection head and a weighted summation operation. The node detection head decodes the feature map outputted by the feature learning module, which consists of 2 convolutional layers with a convolution kernel of 1x1 and a step of 1, 1 BN layer with a dimension of 64 and 1 ReLU layer. The input and output channel numbers of the two convolutional layers are (64, 64) and (64, 1) respectively. Then the heat map outputted by the node detection head and the weight map outputted by the weight learning module are fused through the weighted summation operation. The obtained 1xHxW heat map obtains 320 predicted nodes through the non-maximum suppression algorithm. The feature sampling module obtains the features of the corresponding region from the feature map outputted by the feature learning module according to the coordinates of the predicted points, that is, the visual descriptor, by using the Grid sampling method.

[0037] 2.2 Dynamic label supervision.

[0038] This module dynamically generates an adjacent graph label to supervise the learning of the directed adjacent graph, as shown in the schematic diagram of Figure 2 (b). This module receives unordered predicted nodes v = {v i |i = 1, 2, …, 320}, v i = {(x i , y i )} and ordered true value nodes x i , y irespectively represent the horizontal and vertical coordinates of node i. First, traverse each true value node to find its nearest neighbor prediction node v i . This matching relationship between v i is one-to-one, that is, each or v i corresponds to v i or the number of which does not exceed 1. After matching, start constructing the directed adjacency graph label. The first step is to create a 320x320 zero matrix, and the row and column numbers of the matrix are the sequence numbers of the 320 prediction nodes. The second step is to traverse all prediction nodes. If the prediction node v i has a corresponding true value node , find all adjacent nodes of the true value node. If these adjacent nodes have corresponding prediction nodes, these prediction nodes are the adjacent nodes of the traversed v i . The adjacent relationship of the prediction nodes found by the above process needs to be recorded in the zero matrix constructed in the first step. The recording method is as follows: assuming that the traversed prediction node v i is in the prediction node list v with the sequence number m, and the sequence number of a found adjacent prediction node of v i is n, then the value of the element at the coordinates (m, n) in the zero matrix is modified to 1; if v i has no adjacent nodes, then the value of the element at the coordinates (m, m) in the zero matrix is modified to 1. After traversing all prediction nodes according to the above process, the directed adjacency graph label is constructed. This dynamically generated adjacency graph based on prediction nodes is used to supervise the learning of the adjacency matrix in the directed adjacency graph generator.

[0039] 2.3 Directed Adjacency Graph Generator

[0040] The module is composed of a position encoding module, a multi-layer perception module, a connectivity network, and an adjacency graph prediction module in sequence. The position encoding module stacks the normalized node image coordinates (x, y) = {(x i , y i )|i = 1, 2, …, 320} and the visual descriptor d = {d i |i = 1, 2, …, 320} along the feature dimension to form an integrated descriptor d 320×64 ∈ R i (d 64 ∈ R emb ) along the feature dimension to form an integrated descriptor d 320 ×(64+2)(R represents all real numbers). The multi-layer perception module comprises two sub-modules, each of which is composed of one input, one fully connected layer with an output channel number of (66, 768), one LN layer with a dimension of 64, and one ReLU layer. The output of the multi-layer perception module is a feature descriptor d' e R 320×768 . The connectivity network comprises six Transformer encoder modules, each of which comprises a multi-head attention layer, a feedforward layer, an MLP layer, and an LN layer. The multi-head attention layer is subject to the formula:

[0041]

[0042] In the formula, Q, K, and V represent query, index, and content vectors, respectively. In the present method, Q, K, and V are represented by d' processed by three independent input-output fully connected layers with a feature dimension of 768. d k is the feature dimension of the feature descriptor d', softmax represents an activation function, and T represents the transpose of a matrix. The multi-head attention layer sets a 10% proportion of random neurons to be inactivated for the result of the softmax operation; the feedforward layer comprises, in sequence, one fully connected layer with an input and an output channel number of (768, 3072), a Gaussian error linear unit (GELU), and one fully connected layer with an input and an output channel number of (3072, 768). The MLP layer comprises, in sequence, one fully connected layer with an input and an output channel number of (1536, 1536), one LN layer with a dimension of 1536, one ReLU layer, and one input and output channel number of (1536, 768). The adjacency graph prediction module comprises two parallel adjacency graph prediction heads and an optimization module, each of which comprises four convolutional blocks, the first three of which are composed of a convolutional layer with a convolution kernel size of 1x1 and a step of 1, a BN layer, and a ReLU layer, and the last convolutional block does not comprise a ReLU layer. The input and output channel numbers of the four convolutional blocks are (768, 256), (256, 128), (128, 64), and (64, 1), respectively. Finally, the prediction result of the first prediction head and the transpose of the prediction result of the second prediction head are added to obtain a predicted directed adjacency graph representing the connection relationship between nodes. If the extracted feature is a planar feature, an optimization module based on the Sinkhorn algorithm is also used to optimize the predicted adjacency graph. The Sinkhorn algorithm iteratively normalizes each row and column of the adjacency matrix, and the number of iterations is 100. The final vector topology structure of the target feature can be generated from the predicted nodes and the directed adjacency graph. The specific method is as follows: the non-zero values on the non-diagonal lines in the directed adjacency graph represent a connected predicted node pair. The corresponding predicted nodes are taken out according to the row and column numbers, respectively, and connected. The zero values and non-zero values on the diagonal lines in the adjacency graph are ignored.

[0043] 2.4 Loss function design with multi-task constraints

[0044] The loss function used in the present application includes two kinds, a node detection loss function and an adjacency graph prediction loss function.

[0045] The node detection loss function is as follows:

[0046]

[0047] wherein M(·) represents an operation of averaging, represents the true value node heat map, and h represents the predicted node heat map.

[0048] The directed adjacency graph prediction loss function is as follows:

[0049]

[0050] wherein p represents the predicted adjacency graph, represents the adjacency graph label.

[0051] Moreover, the final expression of the loss function integrating multiple tasks is as follows,

[0052] L total = node + graph (8

[0053] 2.5 Test of the TopDiG model

[0054] The TopDiG model designed in steps 2.1-2.3 can be effectively trained by using the multi-task constrained loss function in step 2.4 and the Adam optimizer. After the training is completed, the trained model is used to extract remote sensing features from the image prediction library, and the remote sensing image target feature vector topology can be obtained.

[0055] In the implementation, the present application can be realized by computer software technology to automatically run the process, and the device running the process of the present application should also be within the protection scope.

[0056] The specific embodiments described herein are merely illustrative of the spirit of the present application. Those skilled in the art of the present application can make various modifications or supplements to the described specific embodiments or use similar ways to replace them, without deviating from the spirit of the present application or exceeding the scope defined by the appended claims.

Claims

1. A method for extracting the vector topological structure of remote sensing objects that is independent of category, characterized in that: The steps include: Step 1: Establish a training set sample library. Based on the remote sensing images of the study area or test area and the historical database, select typical target ground feature samples and create corresponding labels. Then, based on the hardware computing resource capacity, crop the images into slices of a certain size and create corresponding raster labels and vector labels. Step 2: Build a category-independent model for extracting the vector topology of remote sensing objects, denoted as the TopDiG model. The TopDiG model extracts low-level topological features of objects through a convolutional network, generates predicted nodes and visual descriptors, and then uses a dynamic label supervision strategy to dynamically generate adjacency graph labels based on the predicted nodes. Finally, a Transformer-based directed adjacency graph generator predicts the object topology. The TopDiG model includes a topology-focused node detector, a dynamic label supervision strategy module, and a directed adjacency graph generator; the topology-focused node detector sequentially includes a feature encoder, a branch outputter, a feature learning module, a weight learning module, a fusion module, and a feature sampling module; The feature encoder uses ResNet50. In each of the four stages of ResNet50, a side branch outputter is used to output a heat map with a feature dimension of 1 and the same size as the original image. There are four side branch outputters in total. The first one consists of a convolutional layer and a batch normalization layer, and the remaining three consist of a convolutional layer, a batch normalization layer, a transposed convolutional layer, and a bilinear upsampling layer. The four heat maps output by the four side branch outputters are stacked along the feature dimension and fed into the feature learning module. The feature learning module consists of two convolutional blocks. Each convolutional block contains a convolutional layer with a kernel size of 3×3, a stride of 1, and a padding of 1, a batch normalization layer with a dimension of 64, and a ReLU layer. The number of input and output channels of the convolutional layer are (4, 64) and (64, 64), respectively. The weight learning module contains three convolution blocks. The first two convolution blocks consist of a convolution layer with a convolution kernel of 1×1 and a stride of 1, a batch normalization layer with a dimension of 4, and a ReLU layer. The last convolution block does not contain a ReLU layer. The number of input and output feature channels of the three convolution layers is 4. The fusion module consists of a node detection head and a weighted summation operation. The node detection head decodes the feature map output by the feature learning module. It consists of two convolutional layers with a convolution kernel of 1×1 and a stride of 1, a batch normalization layer with a dimension of 64, and a ReLU layer. The input and output channels of the two convolutional layers are (64, 64) and (64, 1), respectively. The heat map output by the node detection head and the weight map output by the weight learning module are then fused through a weighted summation operation. The resulting 1×H×W heat map is used to obtain the predicted node through the non-maximum suppression algorithm. The feature sampling module uses the grid sampling method to obtain the features of the corresponding area from the feature map output by the feature learning module according to the coordinates of the obtained prediction nodes, that is, the visual descriptor; Step 3: Design a multi-task constrained loss function to train the TopDiG model in step 2. The multi-task loss function includes node detection loss function and adjacency graph prediction loss function; The node detection loss function is as follows: Where M(·) represents the operation of finding the mean value, represents the heat map of the true value node, and h represents the heat map of the predicted node; The directed adjacency graph prediction loss function is as follows: Where p represents the predicted adjacency graph, represents the adjacency graph label; Therefore, the final expression of the loss function integrating multiple tasks is as follows: L total =L node +L graph (4) Step 4: After the training is completed, the trained TopDiG model is used to extract remote sensing objects from the image prediction library to obtain the vector topology structure of the target objects in the remote sensing image.

2. The method for extracting the vector topological structure of remote sensing objects according to claim 1, wherein: The specific implementation of the dynamic tag supervision strategy module is as follows; The dynamic label supervision strategy module dynamically generates adjacency graph labels based on the predicted nodes to supervise the learning of the directed adjacency graph. The module receives the unordered predicted nodes v = {v i |i=1,2,…,k},v i ={(x i ,y i )} and ordered truth nodes x i ,y i Respectively represent the horizontal and vertical coordinates of node i; first traverse each true value node To find its nearest prediction node v i , With v i The matching relationship is one-to-one, that is, each or v i The corresponding v i or The number does not exceed 1. After the matching is completed, the directed adjacency graph label is constructed. The first step is to create a k×k zero-value matrix. The row and column numbers of the matrix are the numbers of the k predicted nodes. The second step is to traverse all the predicted nodes. If the predicted node v i There is a corresponding truth node Find all the adjacent nodes of the true value node. If these adjacent nodes have corresponding predicted nodes, then these predicted nodes are the v being traversed. i Adjacent nodes; According to the first and second steps, the adjacency relationship of the predicted nodes needs to be recorded in the zero value matrix constructed in the first step. The recording method is: Assume that the predicted node v being traversed i The sequence number in the prediction node list v is m, and the found v i The serial number of an adjacent prediction node is n, then the value of the element with coordinate (m,n) in the zero value matrix is ​​modified to 1; such as v i If there is no adjacent node, the value of the element at coordinate (m,m) in the zero-value matrix is ​​modified to 1; after traversing all predicted nodes according to the above process, the directed adjacency graph label is constructed.

3. The method for extracting the vector topological structure of remote sensing objects according to claim 1, wherein: The directed adjacency graph generator is composed of a position encoding module, a multi-layer perceptron module, a connectivity network, and an adjacency graph prediction module in sequence; The position encoding module converts the normalized node image coordinates (x, y) = {(x i ,y i )|i=1,2,…,k} and visual descriptor d={d i |i=1,2,…k},d∈R k×64 , d i ∈R 64 , stacked along the feature dimension as an integrated descriptor d emb ∈R k×(64+2) , R represents all real numbers, k represents the number of predicted nodes; the multilayer perceptron module contains two submodules, each of which consists of a fully connected layer with an input and output channel number of (66, 768), a LN layer with a dimension of 64, and a ReLU layer. The output of the multilayer perceptron module is the feature descriptor d ′ ∈R 320×768 The connectivity network consists of 6 Transformer encoder modules, each of which contains a multi-head attention layer, a feedforward layer, an MLP layer, and a LN layer. The adjacency graph prediction module includes two parallel adjacency graph prediction heads and an optimization module. Each graph prediction head contains four convolution blocks. The first three convolution blocks are composed of a convolution layer with a convolution kernel size of 1×1 and a stride of 1, a BN layer, and a ReLU layer. The last convolution block does not contain a ReLU layer. The input and output channel numbers of the four convolution blocks are (768, 256), (256, 128), (128, 64), and (64, 1), respectively. Finally, the prediction result of the first prediction head and the transpose of the prediction result of the second prediction head are added to obtain a predicted directed adjacency graph, which represents the connection relationship between the nodes. When the extracted feature is a planar feature, the predicted adjacency graph is optimized using an optimization module based on the Sinkhorn algorithm. The Sinkhorn algorithm iteratively normalizes each row and column of the adjacency matrix, and the number of iterations is 100.

4. The method for extracting the vector topological structure of remote sensing objects according to claim 3, wherein: The multi-head attention layer follows the formula: In the formula, Q, K, and V represent query, index, and content vectors respectively. Q, K, and V are the d vectors processed by three independent fully connected layers with input and output feature dimensions of 768. ′ indicates; d k is the feature dimension of the feature descriptor d′, 768, softmax represents the activation function, and T represents the transpose of the matrix; the multi-head attention layer sets a 10% random neuron inactivation ratio for the result of the softmax operation; the feedforward layer sequentially contains 1 fully connected layer with an input and output channel number of (768, 3072), Gaussian error linear units (GELU) and 1 fully connected layer with an input and output channel number of (3072, 768); the MLP layer sequentially consists of 1 fully connected layer with an input and output channel number of (1536, 1536), 1 LN layer with a dimension of 1536, 1 ReLU layer and an input and output channel number of (1536, 768).

5. The method for extracting the vector topological structure of remote sensing objects according to claim 3, wherein: The final vector topology structure of the target object can be generated by the predicted nodes and the directed adjacency graph. The specific method is as follows: the non-zero value on the off-diagonal line in the directed adjacency graph represents a connected pair of predicted nodes. The corresponding predicted nodes are taken out according to the row and column numbers and connected. The zero value and the non-zero value on the diagonal in the adjacency graph are ignored.

6. The method for extracting the vector topological structure of remote sensing objects according to claim 1, wherein: In step 3, the Adam optimizer is used to train the TopDiG model.

Images