An InSAR building PS point extraction method for geographic prior map learning

Abstract

The application discloses an InSAR building PS point extraction method based on geographical priori graph learning, and belongs to the technical field of remote sensing data processing; by encoding the building geographical priori as continuous spatial membership priori and building attribute vectors, multi-level coupling constraints are formed in three levels of pseudo label generation, graph structure construction and graph attention message passing, so that building PS points are extracted from all PS point sets with high precision under the condition of a small amount of artificial labels. The application adopts the above method, under the condition that there is an error between the building contour and the PS point position, the PS point attribute dimension is high, and the artificial label is limited, a calculable geographical priori is constructed, consistent coupling constraints are formed in pseudo label construction, graph message passing and training target, and then building PS points are extracted from all PS point sets with high precision, especially the extraction accuracy of PS points near the building boundary is improved, and reviewable confidence information is provided for the extraction result.

Description

Technical Field

[0001] This invention relates to the field of remote sensing data processing technology, and in particular to an InSAR building PS point extraction method based on geographic prior map learning. Background Technology

[0002] Multi-Temporal Interferometric Synthetic Aperture Radar (MT-InSAR) obtains the deformation rate and time series of surface targets by registering, interferometrically interferometrically, unwrapping phases, correcting errors, and retrieving time series from multiple SAR images. Urban areas typically contain a large number of Persistent Scatterer (PS) points, scattered across various terrain features such as buildings, roads, ground, and bare land. In building deformation monitoring scenarios, it is first necessary to extract a subset of PS points belonging to buildings from the entire PS point set—that is, to distinguish between building PS points and non-building PS points—and establish the attribution relationship between PS points and building objects. However, the main drawbacks of existing technologies are as follows: (1) Unreliable spatial matching leads to low extraction accuracy: Binary overlay / buffered matching is sensitive to SAR geometric distortion and contour error, resulting in serious mis-extraction and omission of boundary PS points.

[0003] (2) Insufficient use of geographical priors: Geographical information such as building outlines, heights, and structural types are only used as spatial filters in the matching stage and fail to form a consistent multi-level constraint in label construction, feature learning, and extraction decisions.

[0004] (3) The topological relationships of PS points are missing: the topological relationships between discrete PS points, such as belonging to the same building and distinguishing between buildings, are not explicitly modeled and utilized, and PS points of adjacent features are easily confused.

[0005] (4) Lack of verifiability of extraction results: Hard binary output lacks indication of extraction confidence and systematic verification mechanism, making it difficult to systematically discover and correct erroneous PS points. Summary of the Invention

[0006] The purpose of this invention is to provide an InSAR building PS point extraction method based on geographic prior graph learning. The method encodes the geographic priors of buildings (building vector outline, height level, structure type, etc.) into continuous spatial membership priors and building attribute vectors, forming multi-level coupling constraints at three levels: pseudo-label generation, graph structure construction, and graph attention message passing. This enables high-precision extraction of building PS points from the entire PS point set with a small number of manual labels.

[0007] To achieve the above objectives, this invention provides a method for extracting InSAR building PS points based on geographic prior map learning, comprising the following steps: Step S1: Acquire all permanent scatterer (PS) point data. This involves processing the multi-temporal SAR imagery using Multi-temporal Synthetic Aperture Radar Interferometry (MT-InSAR) to obtain all PS point data within the observed scene. ,include Points ; Step S2: From point data Each point in Three types of features are extracted: radiation features, deformation features, and spatial context features. After processing, these three types of features are concatenated to obtain the node feature vector. ; Step S3: Obtain the building vector outline, and use the perpendicular line method to obtain the distance of each PS point relative to the nearest building outline. and will Mapping to spatial membership prior Spatial membership prior characterizes the probability that the PS point is located inside a building; the two-dimensional spatial location, height, and material properties of the nearest building are encoded into attribute vectors. Based on the distance between PS points and the building affiliation relationships, a sparse spatial constraint graph is constructed. ,in, Let E represent a node, and E represent an edge. For each PS point, define an edge E including the relation type. and border rights ; Step S4: Generate pseudo-labels, reduce the dimensionality of multi-dimensional node features and perform density clustering to obtain the clusters and the probability of cluster membership for each PS point; for each cluster, determine whether the PS point corresponding to the cluster is a building or a non-building PS point by combining the statistics of the prior spatial membership within the cluster and the spatial purity; generate pseudo-labels and sample confidence scores for the PS points within the cluster that meet the high confidence condition. PS points that do not meet the high confidence condition are marked as uncertain samples. Step S5: Construct an extraction network containing a geographic knowledge encoder and a geographic bias coupling graph attention layer, specifically as follows: For intra-cluster PS points that satisfy high-confidence downsampling, use the geographic knowledge encoder to... , , and missing mask fusion into enhanced node representation For uncertain PS points, perform self-training expansion; when calculating the attention score of node i to neighbor j in the graph attention layer, introduce a method... , , and The resulting geographical bias function :when Relationship between the same building and and When both are high, increase the attention weight, among which This indicates that the spatial membership of neighbor j is prior. ;when When distinguishing relationships between adjacent buildings, the attention weight is reduced, and the extraction network is trained using a combination loss. Step S6: Output the probability and extraction uncertainty of each PS point being extracted as a building PS point. and the a priori spatial affiliation Located in the preset fuzzy range The boundary PS points within the area are refined and determined.

[0008] Preferably, in step S1, the specific execution process is as follows: Data of all PS points within the observation scene obtained by processing multi-temporal SAR images using multi-temporal synthetic aperture radar interferometry (MT-InSAR). as follows: ; in, This represents the data for the i-th point, where each point includes at least spatial coordinates, elevation, deformation time series, and amplitude time series.

[0009] Preferably, in step S2, the specific process of constructing multi-dimensional node features is as follows: from each point The three types of features extracted are as follows: (1) Radiation features, used to characterize scattering intensity and stability, including average amplitude and amplitude deviation index, extracted based on point data. (2) Deformation features, used to characterize trends, fluctuations and periodicity, including linear deformation rate and deformation time series standard deviation, extracted based on point data. The temporal InSAR analysis results for each point in the data, corresponding to the extracted linear deformation rate and deformation time series standard deviation; (3) spatial context features, used to characterize the statistical distribution of neighborhood PS points, including the statistics of radiation and deformation features of local neighborhood PS points, for points The surrounding radiation and deformation features, after missing value processing and scale normalization, are concatenated into a node feature vector. .

[0010] Preferably, step S3 is specifically implemented as follows: Obtain building vector outlines and their available attributes, perform geographic prior generation, and use the building vector outlines to obtain the distance of each PS point relative to the nearest building outline using the perpendicular line method. When point PS is inside a building, the signed distance is set to a positive value; when it is outside a building, the signed distance is set to a negative value. Mapping to spatial membership priors using monotonically non-decreasing functions , The character corresponds to the first The probability that a PS point is located inside a building is given by the following formula: , It is a Sigmoid type function. For scale parameters, The symbolic distance mapping bias parameter encodes at least one of the following attributes—the two-dimensional spatial location, height, and material properties—of the nearest building into an attribute vector. For PS points without corresponding buildings, a missing mask is set, so that subsequent encoders can replace attribute vectors with learnable default vectors under the indication of the missing mask. And concatenate all attribute vectors; Based on the PS point representation as a point state distribution, a corresponding spatial constraint graph is constructed within a preset search radius. Within each PS point, select no more than One neighbor; Relationship type At least including relationships within the same building, relationships distinguishing adjacent buildings, and relationships outside the building's vicinity; boundary rights The normalized reciprocal of the distance between PS points, the prior consistency of spatial membership, and the relation type By jointly determining the points PS that are in the same building and have the same prior knowledge, higher edge weights are obtained. .

[0011] Preferably, in step S4, the multidimensional node features are dimensionality reduced and density clustering is performed, wherein the dimensionality reduction method is principal component analysis and the density clustering method is K-means.

[0012] Preferably, in step S5, the total loss function of the combined loss is as follows: Total loss function: ; In the above formula, Extract loss weights for confidence level weighting. For space Figure 1 Weight of consistency loss The weight for geographical prior consistency loss. The loss is weighted by physical consistency, and the loss is extracted using confidence weighting. :pass The error weighting of each pseudo-labeled sample is used to obtain the spatial data. Figure 1 Sexual damage : Through the For pairs of nodes connected by edges that share the same building relationship, the probability of building extraction is constrained to be similar, and the constraint strength is related to the edge weight. Positive correlation; geographical prior consistency loss Only for Above the first threshold The PS point constraint ensures that the probability of extracting the corresponding building is not lower than a preset high probability value. Below the second threshold The PS point is constrained to have an extraction probability no higher than a preset low probability value, corresponding to... ;right In and The loss is not applied to the PS points between them; physical consistency loss The attention formula, which includes at least one of the following constraints: deformation pattern similarity constraint within the same building, scattering stability ranking constraint, and temperature response correlation constraint, is as follows when calculating the attention score of node i for its neighbor j in the graph attention layer: ; ; In the above formula, For attention parameter vectors, This represents the attention score of node i to its neighbor j. This indicates that the input geographic knowledge encoder produces the i-th augmented node. This indicates that the input geographic knowledge encoder produces the j-th augmented node. It is a linear transformation matrix. This represents vector concatenation. Let i be the set of neighbors of node i. This represents the geographic bias encoding function for nodes i and j.

[0013] Preferably, in step S5, self-training expansion is performed on the PS points for which no pseudo-labels were generated in step S4: using the current network to predict, only those prediction confidence levels reach a preset threshold are included. And with PS points that are consistent with the neighborhood predictions are added to the training set along with their prediction labels.

[0014] Preferably, step S6 is implemented as follows: Prior knowledge of spatial affiliation Located in the preset fuzzy range For boundary PS points within the building, a refinement process is performed: after the graph attention network outputs the initial extraction probability, the deformation time series correlation coefficient and spatial distance between the PS point and the PS points of spatially neighboring high-confidence buildings are calculated; when the correlation coefficient is higher than a threshold and the distance is lower than a threshold, the point is refined into a building point and extracted as a set of PS points of the same building, and the consistency of deformation rate within the building is calculated. and extraction probability dispersion Buildings that do not meet the consistency or dispersion conditions are marked as pending review.

[0015] Therefore, the InSAR building PS point extraction method based on geographic prior map learning described above has the following advantages: (1) Significantly improved extraction accuracy: In an experiment in a built-up area of ​​a city, the building PS point extraction accuracy of the method of the present invention reached more than 95%, which is about 11% higher than the traditional buffer matching method and about 7% higher than the graph neural network method without coupled geographic prior.

[0016] (2) Strong verifiability: About 5% of PS points were marked for verification due to extraction uncertainty > 0.5, and 12% of buildings were marked due to non-compliance with consistency or dispersion conditions. Among them, about 80% were confirmed by manual verification to be contour errors or incorrect extraction. This systematic quality feedback mechanism is not available in traditional buffer matching methods.

[0017] (3) High labeling efficiency: Only about 8% of the PS points in the clustered uncertain samples need to be manually verified to obtain high-quality pseudo-labels sufficient for training the extraction network, which significantly reduces the cost of manual labeling.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] Figure 1 This is a general framework diagram of the InSAR building PS point extraction method based on geographic prior map learning according to the present invention; Figure 2 This is a flowchart illustrating the specific process of an InSAR building PS point extraction method based on geographic prior map learning according to the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Specific model specifications need to be selected and determined according to the actual specifications of the device, etc. The specific selection calculation method adopts existing technology in the art, and therefore will not be described in detail.

[0021] Example like Figures 1-2 As shown, this invention provides a method for extracting InSAR building PS points based on geographic prior map learning. To demonstrate the effectiveness of the method, a specific experimental procedure is provided below, taking a location approximately 5km away... 2 This will be illustrated using 1m resolution TerraSAR-X data as an example: Step S1: Approximately 50,000 PS points are obtained through MT-InSAR processing. Each PS point contains UTM coordinates, elevation, and approximately 11 periods of line-of-sight (LOS) deformation time series and amplitude time series between 2007 and 2008. Several PS points are distributed on different land features such as buildings, roads, and the ground.

[0022] Step S2: Extract three types of features and construct an approximately 12-dimensional node feature vector. Radiation features: mean amplitude and amplitude deviation index; deformation features: linear deformation rate and deformation time series standard deviation; spatial context features: PS point density, mean elevation difference, and deformation rate variation coefficient within a local neighborhood (within an 80 m radius). All features are imputed (using the feature median) and normalized using Z-score.

[0023] Step S3: Using building vector outline data (including height and type attributes) from a 1:2000 topographic map as input, obtain the symbolic distance for each PS point using the perpendicular line method. And mapped to spatial membership priors Scale parameters are taken Symbolic distance mapping bias parameters Building attribute coding The height is encoded using one-hot encoding, and the function type is encoded using word embedding. The search radius is then concatenated. Maximum number of neighbors Construct a sparse space constraint graph.

[0024] Relationship type The three classifications are specifically defined in the embodiments of this application as follows: if both PS points are located within the same building outline and the distance is less than 1000 meters... If the nearest neighbor distance is twice the distance between the two nearest neighbors, they are considered to be within the same building; if they belong to different buildings and the distance between their outlines is less than 20m, they are considered to be adjacent buildings; otherwise, they are non-building neighborhoods. The formula for calculating the corresponding edge weights is as follows: In the above formula, For Euclidean distance, The value is 1.0 when the two buildings are in the same relationship, 0.4 when they are adjacent, and 0.2 when they are not in a building neighborhood.

[0025] Step S4: PCA reduces dimensionality to 4 dimensions and performs DBSCAN clustering, corresponding to , Then calculate each cluster Mean, standard deviation, and probability of cluster membership for each PS point. Spatial purity = upper bound estimate of the proportion of PS points of the same building within a cluster. Set a high confidence condition, where the high confidence condition is... When the conditions are met, pseudo-labels are assigned to clusters with high confidence. The formula for calculating the confidence level is: ; The confidence level formula is truncated to [0,1].

[0026] Step S5: Construct a 2-layer geographic bias map attention extraction network, with 128 hidden dimensions in each layer and 4 multi-head attention heads. The geographic knowledge encoder is a 2-layer MLP (128→256→128), and learnable default vectors are used to replace PS points with missing mask indicators. Training was performed using the Adam optimizer with a learning rate of 0.001. Training for 200 epochs. Self-training expansion is performed every 20 epochs, with a confidence threshold. , , .

[0027] Step S6: Output the extraction results. Approximately 85% of the PS points have an extraction probability >0.92 (high confidence level), approximately 10% have an extraction probability between 0.5 and 0.92 (moderate confidence level), and approximately 5% have an extraction uncertainty >0.5 (marked for further verification). The spatial membership prior is in the fuzzy range. For the boundary PS points (approximately 8% of the total PS points), calculate their deformation Pearson correlation coefficient r and normalized spatial distance with the nearest high-confidence building PS point. : satisfying r>0.7 and Points less than 15 m are extracted as building points; otherwise, they are classified as non-building points. Building-level verification: Aggregate by building outline and calculate deformation rate consistency. (Normalized reciprocal of the coefficient of variation of deformation rate) and the discreteness of extraction probability (Standard deviation normalized value) or The building markings are pending review.

[0028] Technical performance data: In the above embodiments, the building PS point extraction accuracy is 95%; 12% of the buildings were marked with building-level verification, of which about 80% were confirmed by manual verification to be contour errors or PS point misextraction.

[0029] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for extracting InSAR building PS points based on geographic prior map learning, characterized in that: Includes the following steps: Step S1: Acquire all permanent scatterer (PS) point data. This involves processing the multi-temporal SAR imagery using Multi-temporal Synthetic Aperture Radar Interferometry (MT-InSAR) to obtain all PS point data within the observed scene. ,include Points ; Step S2: From point data Each point in Three types of features are extracted: radiation features, deformation features, and spatial context features. After processing, these three types of features are concatenated to obtain the node feature vector. ; Step S3: Obtain the building vector outline, and use the perpendicular line method to obtain the distance of each PS point relative to the nearest building outline. and will Mapping to spatial membership prior Spatial membership prior characterizes the probability that the PS point is located inside a building; the two-dimensional spatial location, height, and material properties of the nearest building are encoded into attribute vectors. ; Based on the distance between PS points and the building affiliation relationships, a sparse spatial constraint graph is constructed. ,in, Let E represent a node, and E represent an edge. For each PS point, define an edge E including the relation type. and border rights ; Step S4: Generate pseudo-labels, reduce the dimensionality of multi-dimensional node features and perform density clustering to obtain the clusters and the probability of cluster membership for each PS point; for each cluster, determine whether the PS point corresponding to the cluster is a building or a non-building PS point by combining the statistics of the prior spatial membership within the cluster and the spatial purity; generate pseudo-labels and sample confidence scores for the PS points within the cluster that meet the high confidence condition. PS points that do not meet the high confidence condition are marked as uncertain samples. Step S5: Construct an extraction network containing a geographic knowledge encoder and a geographic bias coupling graph attention layer, specifically as follows: For intra-cluster PS points that satisfy high-confidence downsampling, use the geographic knowledge encoder to... , , and missing mask fusion into enhanced node representation For uncertain PS points, perform self-training expansion; when calculating the attention score of node i to neighbor j in the graph attention layer, introduce a method... , , and The resulting geographical bias function :when Relationship between the same building and and When both are high, increase the attention weight, among which This indicates that the spatial membership of neighbor j is prior. ;when When distinguishing relationships between adjacent buildings, the attention weight is reduced, and the extraction network is trained using a combination loss. Step S6: Output the probability and extraction uncertainty of each PS point being extracted as a building PS point. and the a priori spatial affiliation Located in the preset fuzzy range The boundary PS points within the area are refined and determined.

2. The InSAR building PS point extraction method based on geographic prior map learning according to claim 1, characterized in that: In step S1, the specific execution process is as follows: Data of all PS points within the observation scene obtained by processing multi-temporal SAR images using multi-temporal synthetic aperture radar interferometry (MT-InSAR). as follows: ； in, This represents the data for the i-th point, where each point includes at least spatial coordinates, elevation, deformation time series, and amplitude time series.

3. The InSAR building PS point extraction method based on geographic prior map learning according to claim 1, characterized in that: In step S2, the specific process of constructing multi-dimensional node features is as follows: from each point The three types of features extracted are as follows: (1) Radiation features, used to characterize scattering intensity and stability, including average amplitude and amplitude deviation index, extracted based on point data. All points in the image are used to extract radiometric features from multi-temporal SAR amplitude images. (2) Deformation features, used to characterize trends, fluctuations and periodicity, including linear deformation rate and deformation time series standard deviation, extracted based on point data. The temporal InSAR analysis results for each point in the data, corresponding to the extracted linear deformation rate and deformation time series standard deviation; (3) spatial context features, used to characterize the statistical distribution of neighborhood PS points, including the statistics of radiation and deformation features of local neighborhood PS points, for points The surrounding radiation and deformation features, after missing value processing and scale normalization, are concatenated into a node feature vector. .

4. The InSAR building PS point extraction method based on geographic prior map learning according to claim 1, characterized in that: The specific implementation process in step S3 is as follows: Obtain building vector outlines and their available attributes, perform geographic prior generation, and use the building vector outlines to obtain the distance of each PS point relative to the nearest building outline using the perpendicular line method. When point PS is inside a building, the signed distance is set to a positive value; when it is outside a building, the signed distance is set to a negative value. Mapping to spatial membership priors using monotonically non-decreasing functions , The character corresponds to the first The probability that a PS point is located inside a building is given by the following formula: ； It is a Sigmoid type function. For scale parameters, The symbolic distance mapping bias parameter encodes at least one of the following attributes—the two-dimensional spatial location, height, and material properties—of the nearest building into an attribute vector. For PS points without corresponding buildings, a missing mask is set, so that subsequent encoders can replace attribute vectors with learnable default vectors under the indication of the missing mask. And concatenate all attribute vectors; Based on the PS point representation as a point state distribution, a corresponding spatial constraint graph is constructed within a preset search radius. Within each PS point, select no more than One neighbor; Relationship type At least including relationships within the same building, relationships distinguishing adjacent buildings, and relationships outside the building's vicinity; boundary rights The normalized reciprocal of the distance between PS points, the prior consistency of spatial membership, and the relation type By jointly determining the points PS that are in the same building and have the same prior knowledge, higher edge weights are obtained. .

5. The InSAR building PS point extraction method based on geographic prior map learning according to claim 1, characterized in that: In step S4, the multidimensional node features are dimensionality reduced and density clustering is performed. Principal component analysis is selected as the dimensionality reduction method, and K-means is selected as the density clustering method.

6. The InSAR building PS point extraction method based on geographic prior map learning according to claim 1, characterized in that: In step S5, the total loss function of the combined loss is as follows: Total loss function: ; In the above formula, Extract loss weights for confidence level weighting. Weights for spatial graph consistency loss. The weight for geographical prior consistency loss. The loss is weighted by physical consistency, and the loss is extracted using confidence weighting. :pass The error weighting of each pseudo-labeled sample is used to obtain the spatial graph consistency loss. : Through the For pairs of nodes connected by edges that share the same building relationship, the probability of building extraction is constrained to be similar, and the constraint strength is related to the edge weight. Positive correlation; geographical prior consistency loss Only for Above the first threshold The PS point constraint ensures that the probability of extracting the corresponding building is not lower than a preset high probability value. Below the second threshold The PS point is constrained to have an extraction probability no higher than a preset low probability value, corresponding to... ;right In and The loss is not applied to the PS points between them; physical consistency loss The attention formula, which includes at least one of the following constraints: deformation pattern similarity constraint within the same building, scattering stability ranking constraint, and temperature response correlation constraint, is as follows when calculating the attention score of node i for its neighbor j in the graph attention layer: ； ； In the above formula, For attention parameter vectors, This represents the attention score of node i to its neighbor j. This indicates that the input geographic knowledge encoder produces the i-th augmented node. This indicates that the input geographic knowledge encoder produces the j-th augmented node. It is a linear transformation matrix. This represents vector concatenation. Let i be the set of neighbors of node i. This represents the geographic bias encoding function for nodes i and j.

7. The InSAR building PS point extraction method based on geographic prior map learning according to claim 6, characterized in that: In step S5, self-training expansion is performed on the PS points for which no pseudo-labels were generated in step S4: using the current network to predict, only those prediction confidence levels reach a preset threshold are included. And with PS points that are consistent with the neighborhood predictions are added to the training set along with their predicted labels.

8. The InSAR building PS point extraction method based on geographic prior map learning according to claim 7, characterized in that: In step S6, the specific implementation is as follows: Prior knowledge of spatial affiliation Located in the preset fuzzy range For boundary PS points within the building, a refinement process is performed: after the graph attention network outputs the initial extraction probability, the deformation time series correlation coefficient and spatial distance between the PS point and the PS points of spatially neighboring high-confidence buildings are calculated; when the correlation coefficient is higher than a threshold and the distance is lower than a threshold, the point is refined into a building point and extracted as a set of PS points of the same building, and the consistency of deformation rate within the building is calculated. and extraction probability dispersion Buildings that do not meet the consistency or dispersion conditions are marked as pending review.

Images