Multi-source remote sensing data-based and nlp specific text matching positioning method and system

Abstract

The application provides a multi-source remote sensing data and NLP specific text matching positioning method and system, which comprises the following steps: S1, extracting the fuzzy time and place of a case text, constructing a three-level geographical name knowledge base and a time-space hybrid index, and realizing standardization organization of multi-source remote sensing data; S2, adopting a time window and a space buffer joint retrieval model, combining a Γ3 sensitive target library for semantic-image double-layer disambiguation; S3, extracting multi-scale features based on U-Net++, introducing a case severity score ψ to adjust the attention weight, and adaptively fusing satellite and unmanned aerial vehicle image details; S4, constructing a seven-tuple judicial knowledge graph, associating the remote sensing interpretation result with the case information semantics, and quantitatively evaluating the evidence chain reliability by using an analytic hierarchy process.

Description

Technical Field

[0001] This invention relates to the field of data analysis, and in particular to a method and system for locating specific texts in NLP based on multi-source remote sensing data. Background Technology

[0002] In handling environmental public interest litigation cases, although NLP (Natural Language Processing) technology has enabled the rapid extraction of core case information, the following technical bottlenecks still exist:

[0003] Low efficiency in matching spatiotemporal data: Manually comparing case text descriptions with remote sensing images is time-consuming and labor-intensive, making it difficult to meet the rapid evidence collection needs of cross-regional and long-term environmental damage cases. For example, in the case of "pollution of a section of the Yangtze River," manually searching for remote sensing images of the corresponding time period and locating the polluted area would take several hours to several days;

[0004] Insufficient matching accuracy: Traditional keyword-based retrieval matching methods cannot handle ambiguous spatiotemporal descriptions (such as "a certain area upstream in the past three years"), and it is difficult to resolve ambiguities of places with the same name (such as "Jialing River" may correspond to multiple river sections), resulting in a matching error rate of over 30%.

[0005] Multi-source data fusion is difficult: Multi-source spatiotemporal data such as satellite remote sensing, UAVs, and ground monitoring have problems such as inconsistent coordinate systems and inconsistent time resolutions, lack efficient fusion mechanisms, and are difficult to form a complete spatiotemporal evidence chain.

[0006] Lack of dynamic change analysis: Existing technologies cannot automatically complete the comparative analysis of remote sensing images from different years, making it difficult to intuitively present the spatiotemporal evolution of environmental damage and weakening the persuasiveness of the evidence. Summary of the Invention

[0007] This invention aims to at least solve the technical problems existing in the prior art, and in particular, it innovatively proposes a method and system for localization based on multi-source remote sensing data and NLP specific text matching.

[0008] To achieve the above-mentioned objectives of this invention, this invention provides a method for localization based on multi-source remote sensing data and NLP-specific text matching, comprising:

[0009] S1, extract the fuzzy time and location from the case text, construct a three-level place name knowledge base and a spatiotemporal hybrid index, and realize the standardized organization of multi-source remote sensing data;

[0010] S2 employs a joint retrieval model combining time window and spatial buffer, and combines the Γ3 sensitive target database for semantic-image dual-layer disambiguation.

[0011] S3 extracts multi-scale features based on U-Net++, introduces a case severity score ψ to adjust attention weights, and adaptively fuses satellite and UAV imagery details.

[0012] S4. Construct a seven-tuple judicial knowledge graph, semantically link remote sensing interpretation results with case information, and use the analytic hierarchy process to quantitatively assess the credibility of the evidence chain.

[0013] In a preferred embodiment of the above technical solution, S1 includes:

[0014] S1-1, Complete the time information standardization conversion model, construct a time parsing function for environmental cases, and convert the time text described in natural language into a computable spatiotemporal interval;

[0015] Define the time transformation function model:

[0016] τ = Φ(θ, κ, δ);

[0017] Where τ represents the set of standardized time intervals of the output, τ={τ1,τ2,...,τ} i ,…,τ n}, each time interval It is a closed interval; Let be the start time of the i-th standardized time interval. Let θ represent the end time of the i-th standardized time interval, θ represent the input fuzzy time text string containing the terms "last year," "flood season," and "dry season," and κ represent the case baseline time parameter. This serves as a reference date for the filing of a case or the discovery of environmental damage; The year as the base time, The month as the base time, The reference time is a specific day; δ represents the regional hydrological calendar database, δ={δ1,δ2,...,δ j ,…,δ m}, where each δ j Includes the start and end dates of the flood season and the rules for dividing the wet and dry seasons in region j.

[0018] In a preferred embodiment of the above technical solution, S1 includes:

[0019] S1-2, the process of geocoding and disambiguation preprocessing of location information, is to establish a three-level place name knowledge base Γ, from administrative divisions to ecological functional zones, and then to sensitive targets, containing 2000+ place name entities;

[0020] The geocoding function is defined as:

[0021] Λ = Ψ(L, Γ, ε);

[0022] Where: Λ represents the set of output geographical areas. Each The latitude and longitude bounding box corresponding to the WGS84 coordinate system;

[0023] L represents the input location text string, such as: A diesel leak occurred 500 meters upstream of the water intake of a water plant in a certain section of the Jialing River in Liangjiang New Area;

[0024] Γ represents a three-level place name knowledge base, Γ = Γ1∪Γ2∪Γ3, where Γ1 is the administrative division database, Γ2 is the ecological function zone database, and Γ3 is the environmentally sensitive target database.

[0025] In a preferred embodiment of the above technical solution, S1 includes:

[0026] S1-3, Establish a unified spatiotemporal index structure for multi-source remote sensing data, construct a three-dimensional hybrid index tree H, and integrate the spatial R-tree index and the temporal B-tree index;

[0027] The index structure is defined as follows:

[0028] ;

[0029] in: The spatial extent (MBR) for storing remote sensing images, i.e., the minimum bounding rectangle;

[0030] Storing images and obtaining timestamps Arranged in ascending order using a B-tree;

[0031] Storage image metadata attribute set .

[0032] In a preferred embodiment of the above technical solution, step S2 includes:

[0033] S2-1, Establish a spatiotemporal coupled retrieval model with joint constraints of time window and spatial buffer to replace the traditional separate retrieval;

[0034] Define the spatiotemporal joint matching degree function:

[0035] ;

[0036] in: This represents the overall matching score, with a value range of [0,1]. The time-matching sub-score is calculated as follows: ; For image acquisition time, The core time of the case is represented by Δτ, which is the half-width of the time window. When the text confidence score is >0.9, it can be increased to 0.7 to expand the retrieval recall rate.

[0037] The spatial matching sub-score is calculated as follows: , For image coverage area, For query buffer;

[0038] The score represents the spatiotemporal coupling enhancement score. It is set to 1 when the time difference is less than 3 days and the spatial overlap is greater than 0.8, and 0 otherwise. α, β, and γ are weighting coefficients.

[0039] In a preferred embodiment of the above technical solution, step S2 includes:

[0040] S2-2, a semantic-feature dual-layer place name disambiguation algorithm, constructs a dual-layer disambiguation mechanism to address the problem of multiple names or duplicate names for the same place.

[0041] First layer: Semantic context disambiguation defines the semantic similarity function:

[0042] ;

[0043] Where: C represents the case text context vector. c i Keywords include: sewage outlet, ecological red line; This represents the attribute vector of candidate entities in the place name knowledge base. , where gⱼ contains entity type, associated functional area, and historical event tags;

[0044] Similarity is calculated using a weighted Jaccard coefficient:

[0045] ;

[0046] Where I is the indicator function, and wᵢ is the keyword weight;

[0047] Second layer: When performing remote sensing image feature disambiguation, when > 0.7 When multiple candidates exist, initiate image feature matching;

[0048] Define a feature template library Φ, containing spectral-texture features of typical environmental violation scenarios:

[0049] ;

[0050] in Let be the spectral feature vector in the i-th land feature template. Let be the texture co-occurrence matrix statistics in the i-th land feature template. The landscape index is the feature template for the i-th land cover.

[0051] Feature matching degree calculation:

[0052] .

[0053] In a preferred embodiment of the above technical solution, step S3 includes:

[0054] S3-1, a dual-branch encoder structure is constructed based on the multi-scale feature extraction network of U-Net++ to process satellite imagery and UAV imagery respectively;

[0055] According to the final spatial range of S2-2 By cropping the original satellite imagery and drone imagery, a set of candidate satellite images is obtained. and drone candidate image set

[0056] Network structure is defined as:

[0057] ;

[0058] in: For the satellite image coding branch, the input is... C s Number of satellite image bands;

[0059] For the UAV image coding branch, the input is The spatial resolution H×W is K times that of the satellite imagery, preferably K≥4; Cu is the number of bands in the UAV imagery.

[0060] Concatenate represents a channel cascading operation. Indicates dense skip connections;

[0061] For the decoder, output fused feature map L represents the number of environmental damage categories;

[0062] S3-2, in the case severity-driven attention fusion mechanism, establish a case importance parameter ψ and dynamically adjust the feature fusion weights.

[0063] In a preferred embodiment of the above technical solution, step S4 includes:

[0064] S4-1, Construction of a Seven-Tuple Evidence Knowledge Graph Model: A knowledge graph structure suitable for judicial evidence presentation, defined as a seven-tuple:

[0065] KG = {Q | Q = (E, R, T, S, P, V, Rulebase)}

[0066] in:

[0067] E represents the entity set. It covers four core entities: cases, responsible parties, remote sensing images, and damaged areas;

[0068] R represents a set of relations. , respectively representing three types of judicial causal relationships: occurring in, related to, and leading to;

[0069] T represents the time attribute function, T: E → The time attribute function maps entities to positive timestamps;

[0070] S represents the spatial attribute function, S: E → (lat, lon), which assigns WGS84 coordinates to each entity;

[0071] P represents the set of entity attributes. Record key parameters such as image resolution, pollutant type, and damaged area;

[0072] V represents the evidence link vector. , where each v i Points to the paths of the raw remote sensing image data, interpretation reports, and algorithm confidence log files stored in the evidence repository;

[0073] Represents the environmental regulations and rules library. Each ρ j A logical rule of the form IF(S∈EcologicalRedLine) AND (E =IndustrialFacility) THEN (Legality=0) is used to automatically verify the illegality of harmful acts.

[0074] In a preferred embodiment of the above technical solution, step S4 includes:

[0075] S4-2, A multi-dimensional quantitative evaluation model for evidence credibility is developed, employing the Analytic Hierarchy Process (AHP) to construct the evaluation system and defining the credibility scoring function:

[0076] ;

[0077] in: The weights of the primary indicators, i∈{1,2,3,4}, correspond to the four dimensions of data quality, interpretation accuracy, spatiotemporal matching, and domain consistency, respectively, satisfying the following conditions: The values ​​were determined to be (0.25, 0.35, 0.25, 0.15) using the expert weighting method; ωᵢⱼ represents the weights of the secondary indicators. = 1; x ij The scores for the secondary indicators are standardized to the [0,1] interval, n i Let n1 be the number of secondary indicators in the i-th dimension, where n1=n2=n3=3 and n4=2.

[0078] The present invention also discloses a computer system, comprising:

[0079] processor;

[0080] Memory used to store processor-executable instructions;

[0081] The processor is configured to implement the method for matching and locating specific text based on multi-source remote sensing data and NLP when executing the executable instructions.

[0082] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0083] Construct a dedicated spatiotemporal semantic system for environmental justice, coupling the flood season and dry season of ambiguous time texts with the regional hydrological calendar database to achieve precise time quantification; establish a three-level place name knowledge base of administrative divisions, ecological functional zones, and environmentally sensitive targets to solve the shortcomings of traditional geographic coding in identifying key judicial elements such as water intakes and ecological red lines.

[0084] The original spatiotemporal coupling retrieval model proposes a three-weight joint matching degree function of time window + spatial buffer + spatiotemporal coupling enhancement, which breaks through the efficiency bottleneck of traditional separate retrieval and reduces the retrieval response time from hours to seconds.

[0085] By introducing a case severity score ψ as an attention weight adjustment parameter, the model can automatically focus on key areas, achieving a deep integration of judicial priority and algorithmic decision-making.

[0086] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0087] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0088] Figure 1 This is a schematic diagram of the overall invention. Detailed Implementation

[0089] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0090] like Figure 1 As shown, this invention discloses a method and system for localization based on multi-source remote sensing data and NLP-specific text matching, including the following:

[0091] S1, Spatiotemporal standardization of core information in environmental cases and unified indexing of multi-source remote sensing data;

[0092] S1-1, Complete the time information standardization conversion model, construct a time parsing function for environmental cases, and convert the time text described in natural language into a computable spatiotemporal interval;

[0093] Define the time transformation function model:

[0094] τ = Φ(θ, κ, δ);

[0095] Where τ represents the set of standardized time intervals of the output, τ={τ1,τ2,...,τ} i ,…,τ n}, each time interval It is a closed interval; Let be the start time of the i-th standardized time interval. Let θ represent the end time of the i-th standardized time interval, θ represent the input fuzzy time text string containing the terms "last year," "flood season," and "dry season," and κ represent the case baseline time parameter. This serves as a reference date for the filing of a case or the discovery of environmental damage; The year as the base time, The month as the base time, The reference time is a specific day; δ represents the regional hydrological calendar database, δ={δ1,δ2,...,δ j ,…,δ m}, where each δ j Includes the start and end dates of the flood season and the rules for dividing the wet and dry seasons in region j.

[0096] The corresponding conversion rules are as follows:

[0097] When θ includes last year: ;

[0098] When θ includes the flood season: Query the flood season records for the corresponding region in δ to obtain...

[0099] , This is the month when the flood season begins. It was a certain day when the flood season began;

[0100] , The month that marks the end of the flood season. It refers to a day when the flood season ends;

[0101] When θ includes the dry season: take the non-flood season time period in δ as τ, and subtract the flood season time period from the annual time domain of the region in δ;

[0102] Final output timestamp sequence: ;

[0103] Where Δω is the preset sampling interval (unit: days), and k is the time series density coefficient, which takes integer values ​​of 1, 2, ... to ensure continuous time coverage.

[0104] S1-2, the process of geocoding and disambiguation preprocessing of location information, is to establish a three-level place name knowledge base Γ, from administrative divisions to ecological functional zones, and then to sensitive targets, containing 2000+ place name entities;

[0105] The geocoding function is defined as:

[0106] ;

[0107] Where: Λ represents the set of output geographical areas. Each The latitude and longitude bounding box corresponding to the WGS84 coordinate system;

[0108] L represents the input location text string, such as: A diesel leak occurred 500 meters upstream of the water intake of a water plant in a certain section of the Jialing River in Liangjiang New Area;

[0109] Γ represents a three-level place name knowledge base, where Γ = Γ1∪Γ2∪Γ3, where Γ1 is the administrative division database, Γ2 is the ecological function zone database, and Γ3 is the environmentally sensitive target database. When the text L contains sensitive words such as "water source protection area," "ecological red line," and "core area of ​​nature reserve," the Γ3 database is queried first. If the results returned by Γ3 spatially overlap with the geographic entities of Γ2, the boundary of Γ3 is used as the core constraint, and the range of Γ2 is used as the extended constraint. The weight of the overlapping area is increased by 50% to ensure that sensitive targets are identified first.

[0110] Γ3 = {Target Name, Protection Level, Spatial Boundaries, Legal Terms};

[0111] ε represents the space buffer parameter, which is set to 500 meters by default;

[0112] First, analyze the administrative entity A in L, for example: Liangjiang New Area, and then query the basic coordinate range of Γ1. ;

[0113] Then, analyze the geographic entity E, for example, a section of the Jialing River, and query Γ2 to obtain the linear feature coordinate sequence. ; r is the number of nodes of geographic entity E, that is, the number of linear feature coordinate pairs;

[0114] Querying the Γ3 environmentally sensitive target database: Querying the water intake of the water plant yields:

[0115] Target Name: Primary Protection Zone of Drinking Water Source of a Certain Water Plant; Protection Level: Level 1 (Highest Level); Spatial Boundary: Polygon enclosed by the four boundary coordinates [(106.48,29.68), (106.49,29.69), (106.50,29.68), (106.49,29.67)]; Legal Provisions: Article 64 of the Water Pollution Prevention and Control Law, 2017 Fourth Revision;

[0116] ;

[0117] The parameters are defined as follows:

[0118] The basic scope of the administrative division database output The weighting coefficient is μ1 = 0.2;

[0119] : The linear buffer buffer(Ξ, ε) output by the ecological functional zone database, with a weight μ2 = 0.3;

[0120] Precise boundaries output by the sensitive target library Point features generate a buffer zone with a radius of ε / 2, and a weight of μ3 = 0.5.

[0121] ε: Spatial buffer parameter, default 500 meters; automatically adjusted to εsensitive = 200 meters for Γ3 point features;

[0122] μ is the dynamic weight vector (μ1, μ2, μ3) of the three-level knowledge base. When Γ3 matches successfully, μ3 is automatically increased to 0.6 and μ2 is decreased to 0.2.

[0123] Fusion computing rules:

[0124] When Γ3 is not empty, it means the text matches a sensitive target:

[0125] buffer(Ξ,εsensitive) represents the intersection of the sensitive area and the area adjacent to the river.

[0126] That is

[0127] ;

[0128] in For spatial union operation, For spatial intersection operations,

[0129] When Γ3 is empty, it degenerates into the formula for calculating the buffer range:

[0130] ;

[0131] Buffer is a buffer algorithm that extends ε meters along the normal direction of geographic entity E;

[0132] S1-3, Establish a unified spatiotemporal index structure for multi-source remote sensing data, construct a three-dimensional hybrid index tree H, and integrate the spatial R-tree index and the temporal B-tree index;

[0133] The index structure is defined as follows:

[0134] ;

[0135] in: The spatial extent (MBR) for storing remote sensing images, i.e., the minimum bounding rectangle;

[0136] Storing images and obtaining timestamps Arranged in ascending order using a B-tree;

[0137] Storage image metadata attribute set .

[0138] The retrieval efficiency constraint is:

[0139] , Execution time satisfy: , where Q() is the query request structure;

[0140] in =10 seconds, To query the time range, To query a spatial range, The attribute constraints are used to determine the search results of index tree H, which is a set of candidate images. , as the input of S2-1;

[0141] S2, spatiotemporal joint intelligent matching and two-layer place name disambiguation;

[0142] S2-1, Establish a spatiotemporal coupled retrieval model with joint constraints of time window and spatial buffer to replace the traditional separate retrieval;

[0143] Define the spatiotemporal joint matching degree function:

[0144] ;

[0145] in: This represents the overall matching score, with a value range of [0,1]. The time-matching sub-score is calculated as follows: ; For image acquisition time, The core timeframe of the case is represented by Δτ, which is the half-width of the time window, for example, 30 days by default. When the text confidence score is >0.9, it can be increased to 0.7 to expand the retrieval recall rate.

[0146] The spatial matching sub-score is calculated as follows: , For image coverage area, For the query buffer, the aforementioned To query the spatial range;

[0147] The score represents the enhancement of spatiotemporal coupling. It is set to 1 when the time difference is less than 3 days and the spatial overlap is greater than 0.8, and 0 otherwise.

[0148] α, β, γ are weight coefficients, satisfying α + β + γ = 1 and α = β = 0.35, γ = 0.3.

[0149] The search optimization goal is:

[0150] ;

[0151] ;

[0152] in The matching threshold is set to 0.6. Set the maximum number of results returned in a single query to 100 to prevent information overload.

[0153] S2-2, a semantic-feature dual-layer place name disambiguation algorithm, constructs a dual-layer disambiguation mechanism to address the problem of multiple names or duplicate names for the same place.

[0154] First layer: Semantic context disambiguation defines the semantic similarity function:

[0155] ;

[0156] Where: C represents the case text context vector. , cᵢ is a keyword, such as: sewage outlet, ecological red line; This represents the attribute vector of candidate entities in the place name knowledge base. ,in It includes entity types, associated functional areas, and historical event tags.

[0157] Similarity is calculated using a weighted Jaccard coefficient:

[0158] ;

[0159] Where I is the indicator function, w iTo assign keyword weights, environmentally sensitive words are weighted at 2.0, while ordinary geographical words are weighted at 1.0.

[0160] Second layer: When performing remote sensing image feature disambiguation, when > 0.7 When multiple candidates exist, initiate image feature matching.

[0161] Define a feature template library Φ, containing spectral-texture features of typical environmental violation scenarios:

[0162] ;

[0163] in Let be the spectral feature vector in the i-th land feature template. Let be the texture co-occurrence matrix statistics in the i-th land feature template. is the landscape index in the i-th land feature template.

[0164] Feature matching degree calculation:

[0165] ;

[0166] in: The feature vector of the image to be matched is extracted from remote sensing slices of the candidate locations; The target feature templates associated with case types are obtained by mapping the tags for industrial pollution and ecological damage from the knowledge base. The characteristic tolerance parameter is set to 0.15.

[0167] The final spatial extent disambiguation result is: ;

[0168] Where η=0.6 is the semantic-feature disambiguation tradeoff coefficient, which can be increased to 0.7 when the text confidence is >0.9.

[0169] S3, Dynamic Fusion of Multi-Source Data Driven by Case Importance

[0170] S3-1 is a dual-branch encoder structure built on the multi-scale feature extraction network of U-Net++ to process satellite imagery and UAV imagery respectively.

[0171] According to the final spatial range of S2-2 By cropping the original satellite imagery and drone imagery, a set of candidate satellite images is obtained. and drone candidate image set

[0172] Network structure is defined as:

[0173] ;

[0174] in: For the satellite image coding branch, the input is... C s Number of satellite image bands;

[0175] For the UAV image coding branch, the input is The spatial resolution H×W is K times that of the satellite imagery, preferably K≥4; C u Number of image bands for UAVs;

[0176] Concatenate represents a channel cascading operation. Indicates dense skip connections;

[0177] For the decoder, output fused feature map L represents the number of environmental damage categories.

[0178] S3-2, In the attention fusion mechanism driven by case severity, a case importance parameter ψ is established, and the feature fusion weights are dynamically adjusted;

[0179] Channel attention weight calculation:

[0180] ;

[0181] Where: ψ is the severity score of the case, ψ∈[1,5], which is calculated by comprehensively considering the type of pollution event, the area affected, and the distance to sensitive targets; , , Here, σ is the learnable weight matrix; σ is the Sigmoid activation function. , These are the feature maps output by the two encoders, respectively. To calculate the average over the spatial dimension H×W;

[0182] Spatial attention weight calculation: ;

[0183] Where [;] indicates channel cascading, It has a 7×7 convolution kernel.

[0184] Fusion feature output:

[0185] ;

[0186] Here, ⊙ represents element-wise multiplication, enabling an adaptive trade-off between macroscopic features of satellite imagery and detailed features of UAV imagery based on the importance of the case.

[0187] S4, Seven-tuple judicial evidence chain and credibility quantification assessment;

[0188] The fused feature map output by S3-2 After decoding, the target detection result and semantic segmentation mask are generated, which are used as the target detection result and semantic segmentation mask in entity E. The inputs for spatial attributes S and P;

[0189] S4-1, Construction of a Seven-Tuple Evidence Knowledge Graph Model: A knowledge graph structure suitable for judicial evidence presentation, defined as a seven-tuple:

[0190] KG = {Q | Q = (E, R, T, S, P, V, Rule base )}

[0191] in:

[0192] E represents the entity set. It covers four core entities: cases, responsible parties, remote sensing images, and damaged areas;

[0193] R represents a set of relations. , respectively representing three types of judicial causal relationships: occurring in, related to, and leading to;

[0194] T represents the time attribute function, T: E → The time attribute function maps entities to positive timestamps;

[0195] S represents the spatial attribute function, S: E → (lat, lon), which assigns WGS84 coordinates to each entity;

[0196] P represents the set of entity attributes. Record key parameters such as image resolution, pollutant type, and damaged area;

[0197] V represents the evidence link vector. , where each v i Points to the paths of the raw remote sensing image data, interpretation reports, and algorithm confidence log files stored in the evidence repository;

[0198] Represents the environmental regulations and rules library. Each ρⱼ is a logical rule of the form IF(S∈Ecological Red Line) AND (E =Industrial Facilities) THEN (Legality = 0), used to automatically verify the illegality of harmful behavior.

[0199] S4-2, A multi-dimensional quantitative evaluation model for evidence credibility is developed, employing the Analytic Hierarchy Process (AHP) to construct the evaluation system and defining the credibility scoring function:

[0200] ;

[0201] in: The weights of the primary indicators, i∈{1,2,3,4}, correspond to the four dimensions of data quality, interpretation accuracy, spatiotemporal matching, and domain consistency, respectively, satisfying the following conditions: The values ​​were determined to be (0.25, 0.35, 0.25, 0.15) using the expert weighting method; ω ij The weights of the secondary indicators = 1; x ij The scores for the secondary indicators are standardized to the [0,1] interval, n i Let n1 be the number of secondary indicators in the i-th dimension, where n1=n2=n3=3 and n4=2.

[0202] The specific definitions of the secondary indicators are as follows:

[0203] Dimension 1: Data Quality (x 1j )

[0204] x 11 Spatial resolution score, x 11 = max(0,1-r / 2), where r is the spatial resolution of the image in meters. Full marks are awarded when r ≤ 0.5 meters and 0 marks are awarded when r > 2.0 meters.

[0205] x 12 Time coverage score , To the actual number of images acquired, The minimum number of images required in theory is calculated based on 1 scene per month;

[0206] x 13 Radiation correction accuracy score, x 13 = 1 - RMSE / 0.1, where RMSE is the root mean square error of the radiometric correction.

[0207] Dimension 2: Interpretation Precision (x2ⱼ)

[0208] x 21 : Object detection score, x 21 =mAP@0.5, where mAP is the average accuracy of the YOLOv8 model on the validation set;

[0209] x 22 Semantic segmentation score, x 22 = IoUdamaged, where IoU is the crossover ratio of the damaged area;

[0210] x 23 : Change detection false alarm rate score, x 23 =1-FPrate, where FPrate is the false alarm rate for change detection.

[0211] Dimension 3: Spatiotemporal matching (x3ⱼ)

[0212] x 31 Time matching score, x 31 =1-Δt / 30, where Δt is the time interval (days) between the image acquisition time and the occurrence of the case.

[0213] x 32 Spatial matching score, x 32 =Overlaprate, which is the overlap ratio between the image coverage area and the query buffer;

[0214] x 33 Multi-source data consistency score , where represents the spectral correlation coefficient between images from different sensors.

[0215] Dimension 4: Domain Consistency (x 4j )

[0216] x 41 Compliance score is determined by the environmental regulations and rules database. Automatically determine whether the condition is met; a score of 1 is given if the condition is not met, and a score of 0 is given if the condition is not met.

[0217] x 42 Ecological rationality score is determined by ecological evolution rationality thresholds preset by domain experts.

[0218] The final criteria for determining the validity of the chain of evidence are:

[0219] ≥ 85 and x 41 =1;

[0220] Only when both the credibility threshold and regulatory compliance are met can a chain of evidence be accepted by the judiciary.

[0221] Example: A case of illegal sewage discharge by a company in the Chongqing section of the Yangtze River Basin.

[0222] Step 1: Extract the core information from the case file: "During last year's flood season, a large amount of milky white wastewater appeared in the Yuelai section of the Jialing River." Call the time conversion function T and combine it with the Chongqing hydrological database δ to get τ = [2022-06-5, 2022-09-5]; call the geocoding function Ψ to get λ = (106.52°E, 106.58°E, 29.58°N, 29.62°N).

[0223] Step 2: Perform a spatiotemporal joint search in the PostGIS database, setting the time window Δτ=30 days and the spatial buffer ε=500 meters, and search for and match 3 Pleiades-1A images and 5 UAV images, with an average response time of 8.3 seconds.

[0224] Step 3: Perform radiometric calibration and atmospheric correction on the retrieved image using the FLAASH model. After correction, RMSE = 0.05, obtaining x. 13 =0.5.

[0225] Step 4: Input the case severity score ψ=4.2, set it as a major pollution event, the U-Net++ fusion network automatically increases the weight of the UAV image to 0.68, YOLOv8 detects 3 sewage outlets, mAP@0.5=0.94, DeepLabv3+ segments the polluted area of ​​12.3 hectares, IoU=0.91.

[0226] Step 5: Construct a seven-tuple knowledge graph. Entities include: Case 2023-001, sewage outlets A / B / C, and the Yuelai section of the Jialing River. The relationship is that Case 2023-001 caused water pollution. The time attribute is T=2022-07-18, the spatial attribute is S=(106.55°E,29.60°N), the evidence link V points to the original image and the interpretation report, and the rule Ω determines that the sewage outlet is located in the secondary protection zone of the drinking water source, with illegality = 1.

[0227] Step 6: Calculate the credibility score: Data quality score 0.85, interpretation accuracy score 0.93, spatiotemporal matching score 0.88, domain consistency score 1.0, and the overall score is... =89.2, which meets the judicial acceptance standard.

[0228] The final result was a three-dimensional visualization evidence package that showed the dynamic process of pollution diffusion from June to September 2022. The evidence chain was accepted by the court, and the case was successfully concluded.

[0229] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for localization based on multi-source remote sensing data and NLP-specific text matching, characterized in that, include: S1, extract the fuzzy time and location from the case text, construct a three-level place name knowledge base and a spatiotemporal hybrid index, and realize the standardized organization of multi-source remote sensing data; S1 includes: S1-1, Complete the time information standardization conversion model, construct a time parsing function for environmental cases, and convert the time text described in natural language into a computable spatiotemporal interval; Define the time transformation function model: τ = Φ(θ, κ, δ); Where τ represents the set of standardized time intervals of the output, τ={τ1,τ2,...,τ} i ,…,τ n }, each time interval For a closed interval, Let be the start time of the i-th standardized time interval. Let θ represent the end time of the i-th standardized time interval, θ represent the input fuzzy time text string containing the terms "last year," "flood season," and "dry season," and κ represent the case baseline time parameter. This serves as a reference date for the filing of a case or the discovery of environmental damage; The year as the base time, The month as the base time, The reference time is a specific day; δ represents the regional hydrological calendar database, δ={δ1,δ2,...,δ j ,…,δ m }, where each δ j Includes the start and end dates of the flood season and the rules for dividing wet and dry seasons in region j; S1-2, the process of geocoding and disambiguation preprocessing of location information, is to establish a three-level place name knowledge base Γ, from administrative divisions to ecological functional zones, and then to sensitive targets, containing 2000+ place name entities; The geocoding function is defined as: Λ = Ψ(L, Γ, ε); Where: Λ represents the set of output geographical areas. Each The corresponding latitude and longitude bounding box in the WGS84 coordinate system; L represents the input location text string; Γ represents the three-level place name knowledge base, Γ= Γ1∪Γ2∪Γ3, where Γ1 is the administrative division database, Γ2 is the ecological function zone database, and Γ3 is the environmentally sensitive target database; S1-3, Establish a unified spatiotemporal index structure for multi-source remote sensing data, construct a three-dimensional hybrid index tree H, and integrate the spatial R-tree index and the temporal B-tree index; The index structure is defined as follows: ； in: The spatial extent (MBR) for storing remote sensing images, i.e., the minimum bounding rectangle; Storing images and obtaining timestamps Arranged in ascending order using a B-tree; Storage image metadata attribute set ; S2 employs a joint retrieval model combining time window and spatial buffer, and combines the Γ3 sensitive target database for semantic-image dual-layer disambiguation. S2 includes: S2-1, Establish a spatiotemporal coupled retrieval model with joint constraints of time window and spatial buffer to replace the traditional separate retrieval; Define the spatiotemporal joint matching degree function: ； in: This represents the overall matching score, with a value range of [0,1]. The time-matching sub-score is calculated as follows: ; For image acquisition time, The core time of the case is represented by Δτ, which is the half-width of the time window. When the text confidence score is >0.9, it can be increased to 0.7 to expand the retrieval recall rate. The spatial matching sub-score is calculated as follows: , For image coverage area, For query buffer; The score represents the enhancement of spatiotemporal coupling. It is set to 1 when the time difference is less than 3 days and the spatial overlap is greater than 0.8, and 0 otherwise. α, β, and γ are weighting coefficients. S2-2, a semantic-feature dual-layer place name disambiguation algorithm, constructs a dual-layer disambiguation mechanism to address the problem of multiple names or duplicate names for the same place. First layer: Semantic context disambiguation defines the semantic similarity function: ； Where: C represents the case text context vector. cᵢ is the keyword; This represents the attribute vector of candidate entities in the place name knowledge base. , where g j Includes entity type, associated functional areas, and historical event tags; Similarity is calculated using a weighted Jaccard coefficient: ； Where I is the indicator function, w i Keyword weight; Second layer: When performing remote sensing image feature disambiguation, when > 0.7 When multiple candidates exist, initiate image feature matching; Define a feature template library Φ, containing spectral-texture features of typical environmental violation scenarios: ； in Let be the spectral feature vector in the i-th land feature template. Let be the texture co-occurrence matrix statistics in the i-th land feature template. The landscape index is the feature template of the i-th land cover. Feature matching degree calculation: ； S3 extracts multi-scale features based on U-Net++, introduces a case severity score ψ to adjust attention weights, and adaptively fuses satellite and UAV imagery details. S4. Construct a seven-tuple judicial knowledge graph, semantically link remote sensing interpretation results with case information, and use the analytic hierarchy process to quantitatively assess the credibility of the evidence chain.

2. The method for localization based on multi-source remote sensing data and NLP specific text matching according to claim 1, characterized in that, S3 includes: S3-1, a dual-branch encoder structure is constructed based on the multi-scale feature extraction network of U-Net++ to process satellite imagery and UAV imagery respectively; According to the final spatial range of S2-2 By cropping the original satellite imagery and drone imagery, a set of candidate satellite images is obtained. and drone candidate image set ; Network structure is defined as: ; in: For the satellite image coding branch, the input is... C s Number of satellite image bands; For the UAV image coding branch, the input is The spatial resolution H×W is K times that of the satellite imagery, where K≥4; C u Number of image bands for UAVs; Concatenate represents a channel cascading operation. Indicates dense skip connections; For the decoder, output fused feature map L represents the number of environmental damage categories; S3-2, in the case severity-driven attention fusion mechanism, establish a case importance parameter ψ and dynamically adjust the feature fusion weights.

3. The method for localization based on multi-source remote sensing data and NLP specific text matching according to claim 1, characterized in that, S4 includes: S4-1, Construction of a Seven-Tuple Evidence Knowledge Graph Model: A knowledge graph structure suitable for judicial evidence presentation, defined as a seven-tuple: KG = {Q | Q = (E, R, T, S, P, V, Rule base )} in: E represents the entity set. It covers four core entities: cases, responsible parties, remote sensing images, and damaged areas; R represents a set of relations. , respectively representing three types of judicial causal relationships: occurring in, related to, and leading to; T represents the time attribute function, T: E → The time attribute function maps entities to positive timestamps; S represents the spatial attribute function, S: E → (lat, lon), which assigns WGS84 coordinates to each entity; P represents the set of entity attributes. Record key parameters such as image resolution, pollutant type, and damaged area; V represents the evidence link vector. , where each v i Points to the paths of the raw remote sensing image data, interpretation reports, and algorithm confidence log files stored in the evidence repository; Represents the environmental regulations and rules library. Each ρⱼ is a logical rule of IF(S∈Ecological Red Line) AND (E =Industrial Facilities) THEN (Legality = 0), used to automatically verify the illegality of the damaging behavior.

4. The method for localization based on multi-source remote sensing data and NLP specific text matching according to claim 1, characterized in that, S4 includes: S4-2, A multi-dimensional quantitative evaluation model for evidence credibility is developed, employing the Analytic Hierarchy Process (AHP) to construct the evaluation system and defining the credibility scoring function: ； in: The weights of the primary indicators, i∈{1,2,3,4}, correspond to the four dimensions of data quality, interpretation accuracy, spatiotemporal matching, and domain consistency, respectively, satisfying the following conditions: The values ​​were determined to be (0.25, 0.35, 0.25, 0.15) using the expert weighting method; ω ij The weights of the secondary indicators = 1; x ij The scores for the secondary indicators are standardized to the [0,1] interval, n i Let n1 = n2 = n3 = 3, n4 = 2.

5. A computer system, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to implement the method for matching and locating specific text based on multi-source remote sensing data and NLP as described in any one of claims 1 to 4 when executing the executable instructions.

Images