Land subsidence risk assessment method based on time-series insar and knowledge graph evidence weighting

Abstract

The application discloses a ground subsidence risk assessment method based on time series InSAR and knowledge graph evidence weighting, relates to the technical field of urban geological disaster monitoring, disaster prevention and reduction planning and infrastructure safety assessment, and utilizes PS-InSAR technology to process Sentinel-1 images, extracts surface vertical deformation and spatial deformation gradient; constructs a ground subsidence risk knowledge graph based on a risk-vulnerability-exposure framework, establishes a six-factor edge weight model to quantize correlation weights; integrates various graph algorithms and combines AHP prior constraints to determine index weights; through a macro grid and a micro building double-scale framework, hierarchical weighted calculation is performed to complete comprehensive ground subsidence risk assessment and grading. The application realizes explicit description of index correlation and evidence traceability of weight, reduces subjective bias, considers regional risk identification and single building fine assessment, is good in robustness through multidimensional verification, is suitable for urban ground subsidence disaster prevention and reduction and risk management and control, and has strong universality and expandability.

Description

Technical Field

[0001] This invention relates to the fields of urban geological disaster monitoring, disaster prevention and mitigation planning, and infrastructure safety assessment, specifically a ground subsidence risk assessment method based on time-series InSAR and knowledge graph evidence weighting. Background Technology

[0002] Land subsidence is a geological hazard caused by the combined effects of natural and human factors, resulting in a continuous decline in surface elevation. It can damage infrastructure, exacerbate urban flooding, and pose a significant threat to urban geological safety. Synthetic Aperture Radar Interferometry (InSAR) offers advantages such as all-weather, wide-area, and high-precision monitoring. Temporal InSAR technology can extract millimeter-level surface deformation information and has become a core method for land subsidence monitoring.

[0003] Existing methods for assessing land subsidence risk have significant drawbacks. First, the determination of weights is highly subjective and the evidence is not traceable. Traditional analytic hierarchy process (AHP) relies on expert subjective experience to construct the judgment matrix, and the source of the weights cannot be verified. Second, the characterization of indicator correlations is lacking. The entropy weight method assigns weights only based on the degree of data dispersion, treating indicators as independent individuals and completely ignoring the multi-layered causal transmission relationship of land subsidence: "groundwater level change - compressible layer compression - surface deformation - building damage". Third, robustness verification is lacking. Existing methods do not quantify the uncertainty of the weight results and cannot determine the tolerance of the weights to disturbances in the input data. Fourth, the assessment scale is singular, and analysis can only be carried out at a single scale such as administrative region or grid, failing to take into account both macro-regional identification and micro-level refined building assessment.

[0004] The core reason for the above-mentioned defects is that existing technologies have not unified and structured the causal knowledge, literature evidence and indicator system in the field of land subsidence into a computable model. They have not used graph structures to explicitly express the semantic relationships and cross-layer propagation logic between indicators, nor have they established an objective weighting and traceability mechanism based on literature evidence. At the same time, they have not designed a technical framework that is suitable for multi-scale assessment, resulting in risk assessment results that are highly subjective, have weak physical meaning, and whose accuracy and applicability are difficult to meet the actual needs of urban disaster prevention and mitigation.

[0005] Based on this, the present invention proposes a dual-scale risk assessment method for land subsidence that integrates temporal InSAR and knowledge graph evidence weighting to address the core pain points of existing technologies. Summary of the Invention

[0006] 1. The technical problem to be solved by the present invention

[0007] The purpose of this invention is to propose a land subsidence risk assessment method based on temporal InSAR and knowledge graph evidence weighting to solve the following problems existing in the prior art: (1) How to encode the multi-layered causal propagation relationship of "groundwater level-compressible layer-deformation-disaster-exposure" in the field of land subsidence into a computable graph structure, and assign traceable weights to each relation edge in the graph based on literature evidence; (2) How to automatically derive the weights of each evaluation index from the constructed knowledge graph, while integrating domain prior knowledge constraints to ensure the physical interpretability of the derivation results; (3) How to achieve dual-scale assessment of ground subsidence risk at both the macro-grid scale and the micro-building scale, so as to meet the dual needs of regional risk zoning and refined risk identification at the building level.

[0008] 2. Technical Solution To achieve the above objectives, the present invention provides the following technical solution: A land subsidence risk assessment method based on temporal InSAR and knowledge graph evidence weighting includes the following steps: S1, Temporal InSAR Processing and Deformation Gradient Extraction: Using temporal InSAR technology to process remote sensing images and extract ground vertical deformation and deformation gradient information. S2, Construction of Ground Subsidence Risk Knowledge Graph: Construct a ground subsidence risk knowledge graph, perform evidence weighting calculation on the directed edges in the knowledge graph, and obtain the edge weights; S3, Construct a multi-factor edge weight calculation model: Based on core supporting literature, calculate the basic weight of each directed edge in the knowledge graph; S4: Employ a multi-algorithm integration approach combined with domain knowledge constraints to derive the final weights of each node in the knowledge graph; S5: Based on the dual-scale assessment framework, and combining the above deformation information, edge weights, and node weights, a ground subsidence risk assessment is completed.

[0009] Preferably, the temporal InSAR processing in S1 employs PS-InSAR technology, and the remote sensing image is a multi-temporal Sentinel-1 SAR image. PS-InSAR processing includes precise orbit correction, flat-earth effect removal, differential interferometry, spatiotemporal filtering, and phase unwrapping. This process obtains the line-of-sight (LOS) temporal deformation field and projects it as vertical cumulative deformation. The deformation gradient information is the vertical spatial deformation gradient between adjacent ground features, and its calculation formula is... , where Δd is the cumulative vertical deformation difference between adjacent features, and l is the distance between adjacent features, used to characterize the local differential settlement intensity.

[0010] Preferably, the construction of the ground subsidence risk knowledge graph in S2 is based on the "hazard (H)-vulnerability (V)-exposure (E)" framework, constructing a risk knowledge graph containing several semantic nodes and directed edges:

[0011] In the formula, V is the semantic node set, including H1-H6 hazard layer nodes, V1-V3 vulnerability layer nodes, and E1-E3 exposure layer nodes, totaling 12 nodes; E is the directed edge set, totaling 29 edges, used to describe the cross-layer propagation relationship between nodes; R is the relation type set, including four types: causal driving, trigger activation, amplification enhancement, and spatial coupling, with relation type coefficients of 1.00, 0.90, 0.75, and 0.55, respectively; L is the core literature evidence set supporting each directed edge, which includes 45 core documents in this embodiment. The semantic nodes are divided into three layers: hazard layer nodes, vulnerability layer nodes, and exposure layer nodes. The hazard layer nodes include groundwater level, faults, compressible layer thickness, ground fissures, vertical cumulative deformation, and vertical spatial deformation gradient. The vulnerability layer nodes include building height, road network density, and nighttime light remote sensing intensity. The exposure layer nodes include population density, GDP, and subway density. The directed edges are used to describe the cross-layer propagation relationship between nodes, and the relationship types include four categories: causal driving, trigger activation, amplification enhancement, and spatial coupling.

[0012] Preferably, the construction of the multi-factor edge weight model in S3 includes constructing a six-factor edge weight calculation model. The six-factor edge weight calculation model calculates the basic weight bW(e) of each directed edge based on the core supporting literature, using the following formula:

[0013] In the formula, , , , , , The coefficient is , and + + + + + =1.00, This is expressed as literature support. This is expressed as the average quality of the literature. Represented as relation type coefficient; This is represented as temporal novelty; This is represented as citation influence; This is represented as the consistency of multi-source evidence based on the DS evidence theory; where, The evaluation is conducted from four dimensions: relevance, novelty, reliability, and completeness.

[0014] Preferably, the multi-algorithm integration method in S4 employs five graph algorithms: PageRank, HITS, betweenness centrality, eigenvector centrality, and GraphSAGE. After calculating the importance of each node, the algorithms are weighted and integrated according to a preset ratio to obtain the initial weight of the nodes. The domain knowledge constraint is the literature-guided AHP prior weight. The final weight w_final of each node is obtained through a linear fusion formula; the linear fusion formula is as follows:

[0015] In the formula, λ is the fusion coefficient.

[0016] Preferably, the dual-scale assessment framework in S5 includes a macro-scale and a micro-scale; the macro-scale adopts a complete "hazard (H)-vulnerability (V)-exposure (E)" framework, and the micro-scale adopts a "hazard (H)-vulnerability (V)" adapted framework.

[0017] As a preferred approach, a regular grid is used as the evaluation unit at the macro scale. After standardizing the indicator observations corresponding to each node, the comprehensive risk index RI is calculated by weighting the nodes according to their final weights. The calculation formula is as follows:

[0018] In the formula, The total weights for the hazard layer, vulnerability layer, and exposure layer are respectively obtained by linearly fusing the weights from multiple algorithms with the AHP prior; in this embodiment, The percentages were 52.54%, 25.97%, and 21.49%, respectively. These are the risk index, vulnerability index, and exposure index, which are obtained by weighting the indicators at each level according to their final weights after being normalized, robustly standardized, and 0-1 normalized.

[0019] As a preferred approach, the micro-scale uses individual buildings as the assessment unit, and the building's hazard level inherits the macro-hazard level of the grid in which it is located; the building's vulnerability is calculated by weighting the building height and building volume according to preset weights.

[0020] A land subsidence risk assessment system based on time-series InSAR and knowledge graph evidence weighting, the system sequentially connects to: InSAR Deformation Extraction Module: Input multi-temporal SAR images, output vertical cumulative deformation field and spatial deformation gradient field; Knowledge graph construction module: Based on the HVE framework and literature evidence, a ground subsidence risk knowledge graph containing multi-layer nodes and cross-layer directed edges is constructed; Six-factor edge weight calculation module: Calculates the six-factor weighted basic weights of directed edges in a knowledge graph based on documentary evidence; Multi-algorithm integration weighting module: Performs integrated calculations of five graph algorithms on knowledge graph nodes and integrates AHP priors to obtain the final index weights; Dual-scale risk assessment module: Based on the final weights, risk index calculation and level classification are performed at both the macro grid scale and the micro building scale.

[0021] As a preferred option, the graph algorithms used in the multi-algorithm integration weighting module include PageRank, HITS, betweenness centrality, eigenvector centrality, and GraphSAGE, and are integrated with weights according to a preset ratio.

[0022] Compared with existing technologies, the land subsidence risk assessment method based on temporal InSAR and knowledge graph evidence weighting provided by this invention has the following beneficial effects: (1) Evidence traceability: The weight of each edge in the knowledge graph can be traced back to the specific literature source, quality score and relation type, which overcomes the defect of the lack of transparency of subjective weighting evidence in traditional AHP and supports the review and auditing of weight results.

[0023] Furthermore, this scheme does not isolate and stack temporal InSAR, knowledge graph weighting, and dual-scale risk assessment, but forms a continuous technical chain of "deformation observation input - knowledge graph evidence propagation - node weight derivation - dual-scale risk output". Specifically, the vertical cumulative deformation and vertical spatial deformation gradient output by temporal InSAR are entered into the knowledge graph as observation nodes H5 and H6, respectively; the knowledge graph explicitly encodes the cross-layer propagation relationships between groundwater level, compressible layer, ground fissures, deformation response, vulnerability of the disaster-bearing body, and exposure through directed edges; the six-factor edge weight model transforms document support, document quality, relationship type, temporal novelty, citation influence, and evidence consistency into traceable edge weights; the multi-algorithm integrated weighting module then transforms the edge weight propagation results into node weights; finally, the same set of node weights drives the calculation of macro-grid and micro-building risks. Therefore, the technical effects produced by this solution cannot be obtained by applying knowledge graphs alone, monitoring InSAR alone, or simply replacing traditional AHP / EWM. Instead, it achieves a synergistic effect in the context of ground subsidence risk assessment, including verifiable evidence chains, calculable causal propagation, quantifiable weight perturbations, and consistent output of results at both regional and building scales.

[0024] (2) Cross-layer semantic propagation: By explicitly depicting the multi-layer causal propagation chain of "groundwater level - compressible layer - deformation - disaster-bearing body - exposure" through graph structure, it overcomes the limitation of EWM treating indicators as independent and can more naturally reflect the causal transmission and scale nesting characteristics of ground subsidence.

[0025] (3) Quantification of uncertainty: Through multi-dimensional tests such as Bootstrap literature perturbation, λ sensitivity analysis and α-ζ six-factor Monte Carlo perturbation, the uncertainty of the weighting results was quantitatively assessed, which made up for the lack of robustness verification in existing methods.

[0026] (4) Dual-scale assessment capability: Risk assessment can be carried out at both the macro grid scale and the micro building scale, which can not only realize regional risk zoning identification, but also support the refined risk quantification at the building level, thus meeting the multi-level management needs of urban disaster prevention and mitigation.

[0027] (5) The method is highly versatile: the framework of this method does not depend on data from a specific region. It can be migrated to other cities or geological disaster risk assessment scenarios by changing knowledge graph nodes and literature support, and has strong versatility and scalability. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the technical process of the ground subsidence risk assessment method based on temporal InSAR and knowledge graph evidence weighting of the present invention. Figure 2 This is a schematic diagram of the knowledge graph structure for ground subsidence risk in this invention; Figure 3 This is a sensitivity analysis diagram of the AHP constraint coefficient λ of the present invention; Figure 4 This is a bar chart showing the integrated node weights of the present invention. Figure 5 This is a comparison chart of the high-risk area ratio and node weight ranking correlation between the scheme in Embodiment 2 of the present invention and Comparative Example 1 (AHP prior weight); Figure 6 This is a comparison chart of the high-risk area ratio and node weight ranking correlation between the scheme in Embodiment 2 of the present invention and Comparative Example 2 (EWM Entropy Weight Method). Detailed Implementation

[0029] 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, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0030] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0031] Example 1, please refer to Figures 1 to 6 As shown: Image and text association explanation: Figure 1This corresponds to the overall technical process of the present invention: "data acquisition - graph construction - edge weight calculation - node weighting - risk assessment"; Figure 2 Corresponding HVE risk knowledge graph structure and cross-layer directed edge relationships; Figure 3 Sensitivity analysis corresponding to the AHP constraint coefficient λ is used to illustrate the stable range of the domain prior fusion ratio; Figure 4 The corresponding integrated node weight results are used to explain the contribution of key indicators such as H5 and H6. In Example 1, the vertical cumulative deformation and vertical spatial deformation gradient output in step one are respectively entered into... Figure 2 Nodes H5 and H6 are shown; steps two and three form... Figure 2 The directed edges and their weights are shown; Step 4 output. Figure 4 The node weights are shown, and through Figure 3 Provide a description of λ sensitivity; Step 5 follows Figure 1 The process completes the macro-grid and micro-building risk classification.

[0032] To address the problems mentioned in the technical solutions, this application provides a ground subsidence risk assessment method based on temporal InSAR and knowledge graph evidence weighting. Step 1: Temporal InSAR Processing and Deformation Gradient Extraction. Multi-temporal Sentinel-1 SAR images are processed using PS-InSAR technology. Through precise orbit correction, flat-earth effect removal, differential interferometry, spatiotemporal filtering, and phase unwrapping, the line-of-sight (LOS) temporal deformation field is obtained and projected as the vertical cumulative deformation. Based on this, the vertical spatial deformation gradient β=Δd / l between adjacent ground features is calculated to characterize the intensity of local differential subsidence.

[0033] Step Two: Construction of a Ground Subsidence Risk Knowledge Graph. Within the "Hazard (H) - Vulnerability (V) - Exposure (E)" framework, a risk knowledge graph G=(V,E,R,L) is constructed, containing several semantic nodes and directed edges. Nodes are categorized into hazard layer nodes (e.g., groundwater level, faults, compressible layer thickness, ground fissures, vertical cumulative deformation, vertical spatial deformation gradient), vulnerability layer nodes (e.g., building height, road network density, nighttime light remote sensing intensity), and exposure layer nodes (e.g., population density, GDP, subway density). Directed edges describe the cross-layer propagation relationships between nodes, with relationship types including causal driving, trigger activation, amplification enhancement, and spatial coupling.

[0034] Step 3: Six-Factor Edge Weight Calculation Model. Based on core supporting literature, calculate the basic weight for each directed edge in the knowledge graph.

[0035] The six factors are: literature support, etc. Average quality of literature Relationship type coefficient Temporal novelty Influence of references and based on Consistency of Multiple Sources of Evidence in Evidence Theory The sum of all coefficients is 1.00. Document quality (Q) is assessed from four dimensions: relevance, novelty, reliability, and completeness.

[0036] Step 4: Multi-algorithm integration, weight derivation, and domain knowledge constraints. Five graph algorithms—PageRank, HITS, betweenness centrality, eigenvector centrality, and GraphSAGE—are used to calculate the importance of each node and then integrated with weights according to a preset ratio. Literature-guided AHP priors are introduced as domain knowledge constraints, and linear fusion is employed. Obtain the final node weight.

[0037] Step 5: HVE Two-Scale Risk Assessment. The macro-scale assessment uses the complete HVE framework, with a regular grid as the assessment unit. Standardized and weighted values ​​of each indicator are then used to calculate the comprehensive risk index. At the microscale, the HV adaptation framework is used, with individual buildings as the assessment unit. The building's hazard level is inherited from the macro-hazard level of the grid in which it is located, and vulnerability is calculated by weighting the building's height and volume.

[0038] A further proposed approach is a land subsidence risk assessment method based on temporal InSAR and knowledge graph evidence weighting, whose modules are sequentially connected to include: (1) InSAR Deformation Extraction Module: The input is multi-temporal SAR images (such as Sentinel-1), and the output is the vertical cumulative deformation field and the spatial deformation gradient field, which serve as the data source for nodes H5 and H6 in the knowledge graph.

[0039] (2) Knowledge graph construction module: Based on the "ontology constraint + literature evidence driven" approach, each indicator in the HVE three-layer indicator system is used as a semantic node, and directed edges are established according to the literature support relationship to form a computable knowledge graph structure.

[0040] (3) Six-factor edge weight calculation module: Taking the core supporting literature as input, the edge weight of each directed edge is calculated from six dimensions: literature support, literature quality, relationship type, temporal novelty, citation influence and evidence consistency.

[0041] (4) Multi-algorithm integration weighting module: Taking the edge weight knowledge graph as input, it runs five graph algorithms: PageRank, HITS, betweenness centrality, eigenvector centrality and GraphSAGE respectively, and integrates them to obtain the node weights. The final weights are then linearly fused with the prior knowledge of AHP domain.

[0042] (5) Dual-scale risk assessment module: Based on the final weight, risk index is calculated and level is classified at both the macro grid scale and the micro building scale.

[0043] The above five modules are connected in sequence to form a complete technical chain of "data acquisition → graph construction → edge weight calculation → node weighting → risk assessment".

[0044] Example 2: Based on Example 1, but with some differences, the following describes a ground subsidence risk assessment method based on temporal InSAR and knowledge graph evidence weighting proposed in this invention, with reference to specific examples and accompanying drawings. The specific content is as follows.

[0045] This embodiment uses both a traditional weighted risk model and a simplified risk model as comparative examples to illustrate that this solution is not a simple combination of existing technologies. The comparative examples use the same 500m×500m grid, the same batch of standardized indicators, and the same study area as this embodiment, only replacing the weight derivation method or the risk synthesis method.

[0046] I. Data Preparation Eighty-four Sentinel-1A up-orbit SLC images (orbit P142) from May 2017 to May 2023 were collected, with a spatial resolution of approximately 5m in the range direction and approximately 20m in the azimuth direction. Simultaneously, multi-source auxiliary data were collected, including groundwater level, fault vectors, compressible layer thickness, ground fissure vectors, building height / volume, road network / metro vectors, nighttime light remote sensing data, population density, and GDP raster data.

[0047] II. InSAR Timing Processing PS-InSAR processing was performed using STAMPS software. The image dated June 15, 2020, was selected as the master image. The temporal baseline threshold was -1000 days, the spatial baseline threshold was 200m, and the coherence threshold was 0.3, resulting in approximately 350,000 PS points. After spatiotemporal filtering and phase unwrapping to remove atmospheric delay, the time-series deformation from LOS was obtained, and the result was expressed as d... vertical =d LOS / cosθ, where θ≈34° is the projection of the vertical deformation. Calculate the spatial deformation gradient β on a 500m×500m grid.

[0048] III. Knowledge Graph Construction A knowledge graph containing 12 semantic nodes was constructed: a hazard layer with 6 nodes (H1 groundwater level, H2 fault, H3 compressible layer thickness, H4 ground fissure, H5 vertical cumulative deformation, H6 vertical spatial deformation gradient); a vulnerability layer with 3 nodes (V1 building height, V2 road network density, V3 nighttime light remote sensing intensity); and an exposure layer with 3 nodes (E1 population density, E2 GDP, E3 subway density). 29 directed edges were established, with the following relationship type coefficients: causal driving 1.00, triggered activation 0.90, amplification enhancement 0.75, and spatial coupling 0.55.

[0049] IV. Six-Factor Boundary Weight Calculation Based on 45 core articles, a six-factor weighted baseline weight was calculated for each edge: literature support (α=0.25), average literature quality (β=0.20), relation type coefficient (γ=0.18), temporal novelty (δ=0.13), citation influence (ε=0.09), and DS evidence consistency (ζ=0.15). Literature quality was assessed using four dimensions: relevance (0.35), novelty (0.22), reliability (0.22), and completeness (0.21), with independent scoring by two researchers. All scores had an ICC higher than 0.80.

[0050] V. Weighting of Multi-Algorithm Integration Five graph algorithms were integrated: PageRank (25%, damping coefficient 0.85, source prior H1:H2=7:3), HITS (20%, authority-based), betweenness centrality (20%), eigenvector centrality (20%), and GraphSAGE (15%, two-layer message passing). An AHP prior constraint was introduced, with a fusion coefficient λ=0.15. The final weighting results were: 52.54% for the hazard layer, 25.97% for the vulnerability layer, and 21.49% for the exposure layer; the top five highest-weighted nodes were H5 (14.28%), H6 (13.76%), V1 (10.99%), H1 (9.63%), and V2 (7.93%).

[0051] VI. Two-Scale Risk Assessment Macro scale: Taking a 500m×500m grid as a unit, calculate the comprehensive risk index RI, and divide it into five levels from R0 (low) to R4 (high) by using the Jenks natural breakpoint method or pre-set grading thresholds. Suppose the four level breakpoints of RI after 0-1 normalization are t1, t2, t3, t4 in turn, and 0≤t1<t2<t3<t4≤1. Then the R0 level is 0≤RI≤t1, the R1 level is t1<RI≤t2, the R2 level is t2<RI≤t3, the R3 level is t3<RI≤t4, and the R4 level is t4<RI≤1. After the calculation is completed for the same study area and the same batch of grids, write t1-t4 into the risk grading threshold table and lock it for preservation to ensure consistent grading results when implemented repeatedly; when implemented in other cities or other periods, t1-t4 can be recalculated and solidified based on the continuous values of RI in the corresponding area. In this embodiment, the area proportions of R0-R4 are 8.20%, 31.90%, 31.13%, 20.78% and 7.98% respectively, and the total proportion of the high-risk areas of R3-R4 accounts for 28.77% of the total area of the plain area, with an area of about 1766.50 km². Micro scale: Taking 351415 individual buildings as evaluation units, using the H-V adaptation framework, the numbers of R0, R1, R2 and R3 are 38975, 166831, 114355 and 31254 respectively. The total of R2-R3 is 145609, accounting for 41.44%, which is mainly used for relative risk screening at the building scale and priority ranking of on-site verification.

[0052] VII. Robustness verification Bootstrap literature perturbation (n = 200) shows that the CV of all 12 nodes is lower than 2%, and the average CV is 0.68%. The λ sensitivity analysis shows that the Top-5 nodes remain unchanged within λ∈[0, 0.20]. The mean of Spearman ρ of the six-factor Monte Carlo perturbation (500 groups of random parameters) is 0.995, and the minimum value is 0.965.

[0053] VIII. Comparative example verification Set Comparative Example 1 as only using the literature-guided AHP prior weights, Comparative Example 2 as only using the entropy weight method EWM to assign weights according to the index dispersion, and Comparative Example 3 as a simplified model without constructing a knowledge graph propagation chain. Under the same standardized grid index conditions, the proportion of the high-risk area of R3-R4 identified by this scheme is 28.77%, that of Comparative Example 1 is 20.10%, that of Comparative Example 2 is 10.57%, and that of Comparative Example 3 is 8.60%; the Spearman correlation coefficients of Comparative Example 1, Comparative Example 2 and Comparative Example 3 with respect to the node weight ranking of this scheme are approximately 0.67, -0.05 and 0.52 respectively. Among them, the verification diagrams of Comparative Example 1 and Comparative Example 2 are respectively as Figure 5 and Figure 6As shown in the figures, Comparative Example 1, while reflecting some domain experience, lacks edge-by-edge literature evidence tracing and cross-layer propagation; Comparative Example 2 is mainly dominated by the spatial dispersion of indicators, making it difficult to express the settlement disaster chain of "groundwater level - compressible layer - settlement response - disaster-bearing body"; Comparative Example 3 cannot simultaneously output the literature evidence chain, edge weights, and node propagation paths. The above comparative examples demonstrate that the high-risk identification results and weight structure of this scheme originate from the organic coupling of InSAR observation nodes, knowledge graph evidence edges, multi-algorithm node weighting, and dual-scale risk assessment, rather than the mechanical superposition of any single technology or multiple existing technologies.

[0054] As shown in Example 2, this scheme achieves at least the following synergistic technical effects: First, by using PS-InSAR to extract the H5 vertical cumulative deformation and H6 vertical spatial deformation gradient, the risk assessment has observable settlement response input; Second, by using the HVE knowledge graph and 29 directed edges to structure the propagation relationship between groundwater level, stratum compression, deformation response, vulnerability of the disaster-bearing body, and exposure, the problem that traditional independent weighting of indicators cannot express cross-layer causal chains is solved; Third, by integrating six-factor edge weights and five types of graph algorithms, the evidence from 45 core literature articles is transformed into recalculated node weights, and the stability of the weight ranking is proved through Bootstrap, λ sensitivity, and Monte Carlo perturbation; Fourth, the macro-grid results are used for regional risk zoning, and the micro-building results are used for hotspot screening. Both share the same evidence weighting system, thereby avoiding the separation of macro and micro evaluation standards. Therefore, this embodiment demonstrates that the technical solution claimed in this application forms a continuous coupling relationship between data, knowledge, weights and scales in the scenario of ground subsidence risk assessment. Its technical effect cannot be expected to be obtained by simply combining a single existing technology such as time-series InSAR monitoring, knowledge graph modeling, AHP constraints or building risk classification.

[0055] Please refer to the above work process and comparative verification. Figures 1 to 6 .

[0056] It should be noted that the term "comprising" or any other variation thereof is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0057] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for assessing land subsidence risk based on temporal InSAR and knowledge graph evidence weighting, characterized in that, Includes the following steps: S1, Temporal InSAR Processing and Deformation Gradient Extraction: Using temporal InSAR technology to process remote sensing images and extract ground vertical deformation and deformation gradient information. S2, Construction of Ground Subsidence Risk Knowledge Graph: Construct a ground subsidence risk knowledge graph, perform evidence weighting calculation on the directed edges in the knowledge graph, and obtain the edge weights; S3, Construct a multi-factor edge weight calculation model: Based on core supporting literature, calculate the basic weight of each directed edge in the knowledge graph; S4: Employ a multi-algorithm integration approach combined with domain knowledge constraints to derive the final weights of each node in the knowledge graph; S5: Based on a dual-scale framework of macro-grid and micro-building, and combining vertical deformation, deformation gradient, edge weight and node weight, a comprehensive risk assessment and classification with hierarchical weighting is achieved.

2. The method for land subsidence risk assessment based on temporal InSAR and knowledge graph evidence weighting as described in claim 1, characterized in that, The temporal InSAR processing in S1 employs PS-InSAR technology, and the remote sensing image is a multi-temporal Sentinel-1 SAR image. PS-InSAR processing includes precise orbit correction, flat-earth effect removal, differential interferometry, spatiotemporal filtering, and phase unwrapping. This process acquires the line-of-sight (LOS) temporal deformation field and projects it as vertical cumulative deformation. The deformation gradient information is the vertical spatial deformation gradient between adjacent ground features, and its calculation formula is... , where Δd is the cumulative vertical deformation difference between adjacent features, and l is the distance between adjacent features, used to characterize the local differential settlement intensity.

3. The method for land subsidence risk assessment based on temporal InSAR and knowledge graph evidence weighting as described in claim 1, characterized in that, The construction of the ground subsidence risk knowledge graph in S2 is based on the "hazard (H)-vulnerability (V)-exposure (E)" framework, and constructs a risk knowledge graph containing several semantic nodes and directed edges: In the formula, V is the semantic node set, including H1-H6 hazard layer nodes, V1-V3 vulnerability layer nodes and E1-E3 exposure layer nodes, totaling 12; E is the directed edge set, totaling 29 edges, used to describe the cross-layer propagation relationship between nodes; R represents the set of relation types, including four categories: causal driving, trigger activation, amplification enhancement, and spatial coupling, with relation type coefficients of 1.00, 0.90, 0.75, and 0.55, respectively. L represents the set of core literature evidence supporting each directed edge, including 45 core documents in this embodiment. The semantic nodes are divided into three layers: hazard layer nodes, vulnerability layer nodes, and exposure layer nodes. The hazard layer nodes include groundwater level, faults, compressible layer thickness, ground fissures, vertical cumulative deformation, and vertical spatial deformation gradient. The vulnerability layer nodes include building height, road network density, and nighttime light remote sensing intensity. The exposure layer nodes include population density, GDP, and subway density. The directed edges are used to describe the cross-layer propagation relationships between nodes, and the relation types include causal driving, trigger activation, amplification enhancement, and spatial coupling.

4. The method for land subsidence risk assessment based on temporal InSAR and knowledge graph evidence weighting as described in claim 1, characterized in that, The construction of the multi-factor edge weight model in S3 includes the construction of a six-factor edge weight calculation model. The six-factor edge weight calculation model calculates the basic weight of each directed edge based on the core supporting literature. The calculation formula is: In the formula, , , , , , The coefficient is , and + + + + + =1.00, This is expressed as literature support. This is expressed as the average quality of the literature. Represented as relation type coefficient; This is represented as temporal novelty; This is represented as citation influence; This is represented as the consistency of multi-source evidence based on the DS evidence theory; where, The evaluation is conducted from four dimensions: relevance, novelty, reliability, and completeness.

5. The method for land subsidence risk assessment based on temporal InSAR and knowledge graph evidence weighting as described in claim 1, characterized in that, The multi-algorithm integration method in S4 employs five graph algorithms: PageRank, HITS, betweenness centrality, eigenvector centrality, and GraphSAGE. After calculating the importance of each node, they are weighted and integrated according to a preset ratio to obtain the initial weight of the nodes. The domain knowledge constraint is the literature-guided AHP prior weight. The final weight w_final of each node is obtained through a linear fusion formula; the linear fusion formula is as follows: In the formula, λ is the fusion coefficient.

6. The method for land subsidence risk assessment based on temporal InSAR and knowledge graph evidence weighting as described in claim 1, characterized in that, The dual-scale assessment framework in S5 includes a macro-scale and a micro-scale; the macro-scale adopts the complete "hazard (H)-vulnerability (V)-exposure (E)" framework, and the micro-scale adopts the "hazard (H)-vulnerability (V)" adapted framework.

7. A ground subsidence risk assessment method based on temporal InSAR and knowledge graph evidence weighting as described in claim 11, characterized in that, At the macro level, a regular grid is used as the evaluation unit. After standardizing the indicator observations corresponding to each node, the comprehensive risk index is calculated by weighting the nodes according to their final weights. The calculation formula is: In the formula, The total weights for the hazard layer, vulnerability layer, and exposure layer are respectively obtained by linearly fusing the weights from multiple algorithms with the AHP prior; in this embodiment, The percentages were 52.54%, 25.97%, and 21.49%, respectively. These are the risk index, vulnerability index, and exposure index, which are obtained by weighting the indicators at each level according to their final weights after being normalized, robustly standardized, and 0-1 normalized.

8. A ground subsidence risk assessment method based on temporal InSAR and knowledge graph evidence weighting as described in claim 11, characterized in that, At the micro scale, individual buildings are used as the assessment unit, and the building’s risk level inherits the macro risk level of the grid in which it is located; the vulnerability of a building is calculated by weighting the building height and building volume according to preset weights.

9. A land subsidence risk assessment system based on temporal InSAR and knowledge graph evidence weighting, applicable to the land subsidence risk assessment method based on temporal InSAR and knowledge graph evidence weighting as described in any one of claims 1-8, characterized in that, The system is connected sequentially to include: InSAR Deformation Extraction Module: Input multi-temporal SAR images, output vertical cumulative deformation field and spatial deformation gradient field; Knowledge graph construction module: Based on the HVE framework and literature evidence, a ground subsidence risk knowledge graph containing multi-layer nodes and cross-layer directed edges is constructed; Six-factor edge weight calculation module: Calculates the six-factor weighted basic weights of directed edges in a knowledge graph based on documentary evidence; Multi-algorithm integration weighting module: Performs integrated calculations of five graph algorithms on knowledge graph nodes and integrates AHP priors to obtain the final index weights; Dual-scale risk assessment module: Based on the final weights, risk index calculation and level classification are performed at both the macro grid scale and the micro building scale.

10. The ground subsidence risk assessment system based on temporal InSAR and knowledge graph evidence weighting according to claim 9, characterized in that, The graph algorithms used in the multi-algorithm integration and weighting module include PageRank, HITS, betweenness centrality, eigenvector centrality, and GraphSAGE, and are integrated with weights according to a preset ratio.

Images