A multi-source spatio-temporal data driven flood impact intelligent analysis method and system

Abstract

This invention discloses a multi-source spatiotemporal data-driven intelligent flood impact analysis method, relating to the fields of smart water conservancy and geographic information technology. It includes the following steps: S1, constructing a multi-scale spatiotemporal index benchmark; S2, constructing an associated dual knowledge base system; S3, extracting the affected fine-grained grid set; S4, parallel querying of the dual knowledge bases; S5, calculating the flood impact index through dual-model collaborative calibration; S6, generating fine-grained analysis results; S7, iterative optimization of the model and knowledge base. This invention, through a two-level unified spatiotemporal index framework of "coarse grid management - fine grid calculation" using BeiDou grid codes, transforms complex spatial relationship calculations into rapid matching of standardized grid codes, unifying cross-source data to the BeiDou spatiotemporal benchmark, achieving a balance between large-area coverage and high-precision requirements, significantly improving the efficiency and accuracy of flood impact analysis, and providing reliable high-performance support for real-time flood impact analysis.

Description

Technical Field

[0001] This invention relates to the fields of smart water conservancy and geographic information technology, specifically to a method and system for intelligent analysis of flood impact driven by multi-source spatiotemporal data. Background Technology

[0002] Floods are one of the major natural disasters facing my country, characterized by their suddenness and wide-ranging impact, posing a serious threat to people's lives and property and social stability. Timely and accurate flood impact analysis is the core scientific basis for flood warning issuance, emergency command and dispatch, disaster loss assessment, and post-disaster recovery and reconstruction, directly affecting the efficiency of flood emergency response and the effectiveness of disaster reduction.

[0003] With the rapid development of technologies such as Geographic Information Systems (GIS), remote sensing, the Internet of Things (IoT), big data, and artificial intelligence (AI), flood impact analysis using multi-source spatiotemporal data has become a mainstream approach. However, existing technologies still face key technical bottlenecks when dealing with massive amounts of heterogeneous data and handling complex, dynamic disaster scenarios, severely restricting the timeliness, accuracy, and effectiveness of decision support in the analysis:

[0004] 1. Low efficiency in multi-source spatiotemporal data fusion retrieval and computational analysis: Spatiotemporal data from different sources have problems such as inconsistent spatiotemporal benchmarks and heterogeneous encoding formats, and lack standardized association indexing mechanisms, resulting in a cumbersome and time-consuming data fusion process; traditional methods such as latitude and longitude grids and custom rule grids do not perform dynamic mapping association between grids and geographic entities, making it impossible to achieve real-time and rapid flood impact analysis. Especially in scenarios with large-area and high-precision requirements, data retrieval and computational overhead become performance bottlenecks, making it difficult to support real-time analysis needs over large areas.

[0005] 2. The disconnect between real-time data analysis and historical experience leads to insufficient model generalization ability and accuracy: Existing methods mostly rely on physical or statistical models based on current real-time data for analysis, or simply query historical cases. The two are not deeply integrated: On the one hand, purely data-driven models experience a sharp drop in reliability when encountering unfamiliar scenarios not covered by training data; on the other hand, unstructured historical experience (such as text reports and images) is not transformed into calculable and reasonable digital knowledge, making it impossible for intelligent systems to effectively reuse it. This results in analysis results lacking historical depth and experience reference, making it difficult to guarantee accuracy when facing complex scenarios. Furthermore, existing analysis methods cannot continuously learn from historical flood events.

[0006] 3. The system lacks dynamic evolution capabilities, and model performance degrades over time: Most existing analysis models have fixed parameters and static knowledge bases, making it impossible to optimize them using real data from post-disaster feedback. Model performance degrades over time and with changes in the disaster environment, failing to achieve intelligent evolution that improves accuracy with use, and thus failing to meet the long-term, reliable disaster analysis and decision support needs.

[0007] 4. The spatial granularity of the analysis results is too coarse, and the operability of risk positioning and emergency measures is insufficient: Traditional analysis results are mostly described in terms of large units such as administrative divisions (such as townships and administrative villages), which cannot accurately identify the specific locations of high risks. Personnel evacuation and emergency resource allocation lack precise targeting, and protection measures in key areas are difficult to implement accurately, affecting the effectiveness of emergency response.

[0008] 5. Fixed multi-model fusion logic, lacking scene adaptation capability and extreme scenario anomaly protection mechanism: Existing solutions have implemented data and knowledge fusion schemes, which generally adopt static fusion logic of fixed weighting and simple voting. They cannot dynamically adjust the fusion strategy according to the matching degree between the current flood scenario and historical patterns and the complexity of geographical entities in the affected area. When extreme or new flood scenarios not covered by training data occur, they are easily misled by low-correlation historical experience, resulting in serious deviations in the analysis results, which cannot meet the high reliability decision-making requirements in flood emergency scenarios.

[0009] 6. Model and knowledge base optimization are disconnected and lack a full-link collaborative evolution mechanism: Existing solutions can only achieve independent parameter optimization of a single model, and cannot achieve end-to-end joint optimization of data-driven models, empirical reasoning models and fusion calibration modules, let alone drive the dynamic iterative update of the knowledge base through model optimization. The model and knowledge base are in a state of "independent optimization and no collaboration" for a long time. Even if the parameters are updated, the performance degradation problem cannot be solved from the root, and a complete intelligent closed loop of "analysis-verification-optimization-update" cannot be formed.

[0010] In summary, the key technical challenges to improving the accuracy and reliability of flood impact analysis are how to achieve efficient fusion and retrieval of massive multi-source data, achieve deep collaboration between real-time data and historical knowledge, design analytical models with adaptive calibration and continuous evolution capabilities, and improve the precision of high-risk area positioning.

[0011] As a standardized spatial coding system independently developed in my country, the BeiDou grid code possesses the characteristics of global uniqueness, clear hierarchy, and cross-scale compatibility, providing an innovative technical path for the unified management and efficient fusion computation and analysis of multi-source spatiotemporal data. In existing technologies, the BeiDou grid code is mainly applied to spatial positioning and data organization; there is currently no known solution that uses it as a unified index key to drive collaborative queries of heterogeneous knowledge bases and achieve deep integration of micro-entity computation and macro-experience reasoning in flood impact analysis. Based on this, this invention proposes an intelligent analysis scheme that integrates a unified spatiotemporal index based on the BeiDou grid code, dual-knowledge-base collaboration, dual-model calibration, and closed-loop optimization. This scheme aims to improve the accuracy, efficiency, and practicality of flood impact analysis across the entire chain from data management and knowledge computation to result generation. Summary of the Invention

[0012] To address the problems in existing flood impact analysis technologies, such as difficulties in multi-source data fusion, insufficient data and knowledge synergy, limited model accuracy, lack of closed-loop optimization mechanisms, and insufficient positioning precision, this invention provides a multi-source spatiotemporal data-driven intelligent flood impact analysis method and system. The core of this invention lies in constructing a complete technical system based on a unified framework using BeiDou grid codes as a unified index, dual-database collaboration, dual-model calibration, and closed-loop optimization, to achieve intelligent analysis of the entire flood impact analysis process, from data fusion and intelligent computation to result generation and system evolution. This invention is not a simple superposition of existing technical modules, but rather produces unexpected technical effects through the following collaborative mechanisms:

[0013] (1) The Beidou grid code, as a unified spatial key, enables parallel association retrieval between the geographic entity map database and the flood scenario vector knowledge base. While significantly improving the data retrieval and computation efficiency, it also enables the precise calculation of micro-entities and macro-experience reasoning to be deeply integrated at the same spatial granularity.

[0014] (2) Using the fine Beidou grid as the anchor point, the geographic entity map database provides micro-geographic entity information, and the flood scenario vector knowledge base provides historical experience. The two are aligned at the same spatial granularity, which solves the pain point of "spatiotemporal mismatch between real-time data and historical experience" in existing technologies, and provides the analysis model with source-based and comparable input data.

[0015] (3) An adaptive fusion mechanism based on consistency measurement and average semantic similarity, combined with a circuit breaker strategy, enables the system to automatically avoid low-quality experience interference in extreme scenarios and ensure the reliability of analysis.

[0016] (4) Using the Beidou grid code as the anchor point, the joint loss function is used to simultaneously drive the optimization of the analysis model parameters and the updating of the content of the geographic entity map database and the flood scenario vector knowledge base, forming an intelligent closed loop of "analysis-verification-optimization-update", which solves the problem of long-term degradation of model performance from the root.

[0017] The technical effects produced by the above-mentioned collaborative mechanism will go beyond the simple summation of the independent functions of each module, and will have unexpected technical effects.

[0018] To achieve the above objectives, the present invention adopts the following technical solution:

[0019] This invention provides a multi-source spatiotemporal data-driven intelligent analysis method for flood impact, comprising the following steps: S1, constructing a multi-scale spatiotemporal index benchmark; determining the geographical boundary of the target analysis area, and discretizing the area into layers based on the hierarchical coding characteristics of the BeiDou grid code, constructing a coarse grid layer and its nested fine grid layers to form a unified spatiotemporal index framework; S2, constructing an associated dual knowledge base system; constructing and maintaining dual knowledge bases spatially associated with the fine grids, including a geographic entity map database and a flood scenario vector knowledge base; S3, extracting the affected fine grid set; receiving current flood inundation feature data, and extracting the affected fine grid set within the inundation range through spatial overlay analysis based on the spatiotemporal index framework; S4, parallel querying of the dual knowledge bases; based on the BeiDou grid code corresponding to the affected fine grid set, querying the geographic entity map database and the flood scenario vector knowledge base in parallel to obtain the corresponding data for each fine grid. The system consists of: S5, calculating the flood impact index using dual-model collaborative calibration; S6, generating fine-grained analysis results; S7, generating flood impact analysis results with fine grids as the smallest analysis unit; S8, iterating and optimizing the model and knowledge base; S9, receiving actual post-disaster impact data (I_a), constructing a joint loss function based on the difference between (I_a) and (I_c), jointly optimizing the parameters of the data-driven model, historical experience model, and dual-model collaborative calibration module, and synchronously updating the associated data and vector features in the dual knowledge bases.

[0020] Preferably, in step S1, the coarse grid is a grid cell selected from the first to third levels of the BeiDou grid code, used for distributed storage indexing of spatiotemporal data to reduce I / O overhead during data retrieval; the fine grid is a grid cell selected from the fourth to seventh levels of the BeiDou grid code, serving as the smallest spatial calculation unit for flood impact analysis; the coarse grid and the fine grid are associated and located through the BeiDou grid code.

[0021] Preferably, in step S2, the geographic entity map database is used to store the topological structure and attribute characteristics of various geographic entities within the target area, as well as the spatial association between the geographic entities and the fine grid, and constructing the geographic entity map database includes:

[0022] S21. Acquire multi-source spatiotemporal data related to flood impact analysis and unify them to a spatiotemporal reference based on the BeiDou grid code. Extract geographic entity object data after cutting and segmenting according to the coarse grid. The multi-source spatiotemporal data includes basic geographic information data, land feature vector data, and social operation vector data. Through format conversion, spatial matching, and attribute mapping operations, convert the land feature vector data and social operation vector data into geographic entity object data. The geographic entity object data contains the spatial geometric information, attribute feature information, and grid association information of the entity.

[0023] S22. Geographic entities are mapped to fine grids through GIS spatial analysis. For areal or linear geographic entities spanning multiple fine grids, spatial segmentation is performed, and "entity-grid" association weight factors are dynamically generated based on the entity type and its geometric proportion within the current grid. A graph data model is constructed, in which the fine grid serves as the master node and the associated geographic entity objects serve as the subordinate nodes. The association weight factors are used for the weighted aggregation of geographic entity features in the subsequent flood impact index model, as well as the weighted calculation of the geographic entity-level preliminary impact index and the geographic entity-level historical benchmark index.

[0024] S23. Update geographic entity data based on the interpretation of remote sensing image data at a preset period, and update population distribution data in social operation vector data based on the analysis of mobile communication signaling data at a preset period.

[0025] Preferably, in step S2, the flood scenario vector knowledge base is used to store the semantic feature vectors and hydrological parameters of historical flood disaster cases, as well as the mapping relationship between the historical flood cases and the affected grids, forming a historical experience vector database. Constructing the flood scenario vector knowledge base includes:

[0026] S24. Multimodal fusion encoding is performed on text reports and image data of historical flood cases to generate high-dimensional semantic vectors representing the overall scenario of the cases. The multimodal fusion encoding adopts a Transformer-based two-stream network structure, extracting semantic feature vectors from text reports and visual feature vectors from image data, and then performing feature-level fusion and splicing through an attention mechanism. The pre-training weights of the two-stream network are fine-tuned based on flood-specific corpus and remote sensing image data.

[0027] S25. Based on the inundation range recorded in historical flood case data or through spatial overlay analysis, determine the fine grid set affected by the case.

[0028] S26. Establish a reverse index mapping relationship between the semantic vector of each case and the set of fine-grid BeiDou grid codes it affects.

[0029] Preferably, in step S4, the geographic entity dataset is the geographic entity dataset within each affected fine grid.

[0030] The set of similar historical scenarios is a set of historical flood scenarios that are similar to the current flood scenario in terms of hydrological semantics.

[0031] Preferably, the data-driven model is a flood impact index model, and the historical experience model is a scenario experience analysis model. The dual-model collaborative analysis in step S5 specifically includes:

[0032] S51. Input the geographic entity dataset and the current flood inundation feature data into the flood impact index model to calculate the preliminary flood impact index (I_p).

[0033] S52. Input the set of similar historical scenario information into the scenario experience analysis model to calculate the experience benchmark index (I_b).

[0034] S53. Input (I_p) and (I_b) into the dual-model collaborative calibration module, calculate the consistency measure (C) between the preliminary flood impact index (I_p) and the empirical benchmark index (I_b) and the average semantic similarity (S) between the current scenario and the historical scenario, determine the fusion weight based on C and S, perform dynamic weighted fusion of (I_p) and (I_b), and finally output the calibrated flood impact index (I_c).

[0035] Preferably, the flood impact index model is a supervised machine learning regression model based on the LightGBM gradient boosting decision tree framework, using the actual impact data I_a verified after the disaster for each fine grid in historical flood events as training labels, where the actual impact data I_a is the post-disaster actual impact data mentioned in step S7; the scenario experience analysis model is a case-based reasoning model, which is configured to obtain the historical experience benchmark index (I_b) in the following way: using the semantic similarity between each historical scenario and the current scenario in the similar historical scenario information set as weights, and weighting and aggregating the corresponding actual impact data.

[0036] Preferably, when the scenario experience analysis model performs weighted aggregation, it only includes historical scenarios whose semantic similarity to the current flood scenario is greater than a preset minimum threshold, and the weight is positively correlated with the semantic similarity of the corresponding historical scenario; the preset minimum threshold is determined by statistical analysis of historical flood case datasets, and the value range is [0.4, 0.7].

[0037] Preferably, the dual-model collaborative calibration module performs adaptive dynamic weighted fusion in the following manner:

[0038] Based on the set of similar historical scenarios obtained in step S4, which have a semantic similarity greater than a preset minimum threshold with the current flood scenario, the number K of qualified similar historical scenarios is determined. If K is less than the preset threshold, I_c=I_p is directly output, and the subsequent fusion calibration process in this step is terminated. The preset threshold is determined by statistical analysis of the historical flood case dataset, and the value range is [3,10]. It can be dynamically adjusted according to the richness of historical cases in the region.

[0039] Otherwise, calculate the average semantic similarity S between the current flood scenario and the aforementioned qualified similar historical scenarios. The calculation formula is: S = Where K is the number of qualified similar historical scenarios mentioned above. Let be the semantic similarity between the i-th qualified similar historical scenario and the current scenario;

[0040] The consistency measure (C) between the preliminary flood impact index (I_p) and the historical experience benchmark index (I_b) is calculated using the following formula: When max(I_p,I_b)=0, C=1 is defined. At this time, it is determined that the current fine mesh is not affected by flood, and I_c=0 is directly output.

[0041] Calculate the current confidence factor α, α=w1×C+w2×S, where w1 and w2 are preset weight coefficients, with default values ​​of w1=0.5 and w2=0.5, which can be dynamically adjusted according to regional flood characteristics and the richness of historical cases. The values ​​range are [0.2,0.8] and w1+w2=1.

[0042] A dynamic confidence threshold T is set, which is adaptively adjusted based on the complexity of geographic entities within the current fine grid. The quantification formula for the complexity F is as follows: ,in The number of geographic entity types within a fine-grained grid. F represents the total number of geographic entities within the fine grid, and is normalized to the range [0,1]. The adjustment rule is T=T0×(1+μ×F), where T0 is the basic threshold with a value range of [0.5,0.8], which can be dynamically adjusted according to the flood warning level and emergency response level. μ is the adjustment coefficient, used to control the impact of geographic entity complexity on the threshold, with a value range of [-0.3,0.2]. When the importance of the protected objects in the area is high, a positive value is taken to increase the threshold and improve the fusion credibility requirements. When the richness of historical cases is low, a negative value is taken to decrease the threshold to adapt to scenarios with insufficient historical experience.

[0043] Based on C and S, as well as the confidence factor α and the dynamic confidence threshold T, the calibrated flood impact index (I_c) is calculated:

[0044] If α is greater than or equal to T, and C+S≠0, then according to the formula... Calculate; if α is less than T, or C+S=0, then it is determined that the current flood scenario is significantly different from the historical pattern, and the current fusion result has low reliability. In this case, I_c is set to be equal to I_p.

[0045] Preferably, in step S7, the joint loss function (L_t) is defined as: ;

[0046] Where I_a represents grid-level post-disaster actual flood impact data, To calibrate the output error, n is the total number of geographic entities in the current fine grid, and g is the g-th geographic entity in this fine grid. This is the association weight factor for the entity in the current fine grid (the association weight factor is generated based on the entity type and its geometric proportion in the current fine grid). and These are the entity-level preliminary impact index and the entity-level historical baseline index, calculated for the entity using the flood impact index model and the scenario experience analysis model, respectively. The actual impact data of the corresponding geographic entities are obtained through post-disaster field verification and remote sensing image interpretation; β is a balance coefficient used to adjust the weight between the overall grid-level error and the consistency control of entity granularity deviation, and its value is determined by cross-validation on historical flood case datasets within the interval [0.1, 1.0]; the joint optimization process aims to minimize L_t.

[0047] Preferably, in step S6, the process of generating the flood impact analysis results includes: firstly, screening high-risk areas based on the calibrated flood impact index I_c (fine grids with I_c ≥ a preset risk threshold are high-risk areas, the preset risk threshold is determined through statistical analysis of historical flood cases, and this threshold can be dynamically adjusted according to the regional flood control level and the importance of the protected objects), and then generating analysis results with the set of fine grid BeiDou grid codes constituting the high-risk areas as the core spatial identifier; the analysis results include at least one of the following:

[0048] (a) The scope of the high-risk area identified by the Beidou grid code set;

[0049] (b) Key geographic entity information associated with the high-risk area;

[0050] (c) Visual representation based on the flood impact index I_c;

[0051] (d) Statistical analysis data and risk assessment conclusions based on the flood impact index I_c;

[0052] (e) Emergency response recommendations based on information on high-risk areas and key geographic entities.

[0053] A multi-source spatiotemporal data-driven intelligent flood impact analysis system is applied to the aforementioned multi-source spatiotemporal data-driven intelligent flood impact analysis method. The multi-source spatiotemporal data-driven intelligent flood impact analysis system includes:

[0054] The BeiDou grid code spatial service engine provides two-level spatial grid division, encoding calculation, and hierarchical mapping services based on BeiDou grid codes, supporting rapid encoding conversion, spatial relationship analysis, and grid mapping retrieval for different grid levels.

[0055] The geographic entity map database management module is used to perform the access of multi-source geospatial data, organize it into blocks according to coarse grids, segment geographic entities and dynamically weight the attributes based on fine grids, and build and maintain the map data model.

[0056] The flood scenario vector knowledge management module is used to perform multimodal fusion encoding of historical flood cases, high-dimensional semantic vector storage, and retrieval based on vector similarity and grid code reverse indexing, so as to realize the construction and maintenance of the flood scenario vector knowledge base; the pre-trained weights of the dual-stream network for multimodal fusion encoding are fine-tuned based on flood-specific corpus and remote sensing image data;

[0057] A dual-model collaborative analysis engine is used to concurrently drive parallel queries on the geographic entity map database and the flood scenario vector knowledge base, using the BeiDou grid code of the affected fine grid as a unified search key. It calls the data-driven model and the historical experience model to calculate I_p and I_b respectively, and uses its internal dual-model collaborative calibration module to adaptively and dynamically weight and fuse I_p and I_b, outputting the calibrated flood impact index I_c. The data-driven model is the flood impact index model, and the historical experience model is the scenario experience analysis model.

[0058] The analysis results generation module is used to generate analysis results with a fine-grid Beidou grid code set as the core identifier based on the calibrated flood impact index I_c;

[0059] The model and knowledge base joint optimization module is used to drive the joint optimization of the data-driven model, historical experience model and dual-model collaborative calibration module based on the joint loss function L_t described in step S7 of this specification, and simultaneously realize the dynamic update of geographic entity data in the geographic entity map database and the update of semantic vector and reverse index mapping in the flood scenario vector knowledge base; the joint loss function is configured to introduce entity-level error terms and their associated weight factors for collaborative optimization on the basis of grid-level error.

[0060] The beneficial effects of this invention are:

[0061] 1. This invention addresses the challenge of inefficient multi-source data fusion retrieval calculations, achieving a breakthrough in efficiency for high-precision analysis over large areas: Responding to the pain points of existing technologies such as inconsistent spatiotemporal benchmarks, heterogeneous coding, and high retrieval computational overhead for multi-source data, this invention utilizes a two-level unified spatiotemporal indexing framework of BeiDou grid codes—"coarse grid management - fine grid calculation"—to transform complex spatial relationship calculations into rapid matching using standardized grid codes. Cross-source data is unified to the BeiDou spatiotemporal benchmark. Compared to traditional latitude and longitude grid schemes, this significantly reduces data retrieval I / O overhead, effectively improves the efficiency of cross-source data fusion calculations, and achieves a balance between large-area coverage and high-precision requirements, providing performance support for real-time flood impact analysis.

[0062] 2. This invention breaks down the barriers between data and historical experience, improving the accuracy and generalization ability of full-scenario analysis: Addressing the pain points of existing technologies where real-time data and historical experience are not deeply integrated and unstructured experience cannot be reused, this invention constructs a dual-database collaborative mechanism of a geographic entity map database and a flood scenario vector knowledge base. It achieves deep integration of micro-level precise calculation and macro-level experience-based reasoning through a unified spatial key using BeiDou grid codes. Simultaneously, it transforms unstructured historical cases into high-dimensional semantic vectors, combining semantic similarity retrieval to achieve experience reuse, solving the problem of weak generalization ability in purely data-driven models. This allows the analysis results to have both real-time data support and historical experience reference, significantly improving accuracy.

[0063] 3. This invention breaks through the bottleneck of static and fixed models, achieving continuous self-evolution of system analysis performance: Addressing the pain points of fixed model parameters, static knowledge bases, and performance degradation over time in existing technologies, this invention designs an end-to-end joint optimization mechanism driven by a joint loss function. It utilizes actual post-disaster impact data to simultaneously optimize the flood impact index model, scenario experience analysis model, and dual-model collaborative calibration module, while dynamically updating the dual knowledge bases. This breaks the limitations of "independent optimization of a single model and separation of model and knowledge base," forming an intelligent closed loop of "analysis-verification-optimization-update," truly realizing the intelligent evolution of the system to become "more accurate with use," and meeting the long-term reliable decision support needs.

[0064] 4. This invention achieves refined risk positioning and enhances the operability of emergency decision-making: Addressing the pain points of coarse spatial granularity and vague risk positioning in existing technical analysis results, this invention uses the BeiDou fine grid as the smallest calculation and expression unit, and the analysis results use the BeiDou grid code set as the core spatial identifier, penetrating the definition of high-risk areas to the scale of blocks, key facilities and even building clusters, accurately identifying the location of risk objects, so that personnel evacuation routes and emergency resource allocation plans have clear spatial orientation, significantly improving the efficiency of emergency response measures and solving the problem of insufficient operability of traditional administrative division-level results.

[0065] 5. In this invention, an innovative adaptive fusion logic is constructed to build an extreme scenario anomaly protection mechanism: Addressing the pain points of existing technologies where multi-model fusion logic is fixed and lacks scene adaptation capabilities, this invention is the first to create an adaptive dynamic weighted fusion algorithm based on consistency metric (C) and average semantic similarity (S), which adaptively adjusts the dynamic threshold T based on geographic entity complexity and introduces a "circuit breaker mechanism"; when the current scenario is detected to be significantly inconsistent with historical patterns, it automatically avoids low-correlation experience interference, solves the problem of performance degradation of traditional fixed-weight fusion under extreme and new flood scenarios, and ensures high-reliability decision-making requirements.

[0066] 6. In this invention, a full-link collaborative evolution system is established to solve the performance degradation problem at its root: Addressing the pain points of existing technology model and knowledge base optimization being fragmented and lacking a collaborative evolution mechanism, this invention achieves deep collaboration between model optimization and knowledge base updates. The joint loss function not only optimizes the overall error at the grid level, but also forces the two models to align and balance at the entity granularity through entity-level regularization constraints. At the same time, model optimization drives the dynamic updating of geographic entity data and the iteration of historical case semantic vectors, solving the performance degradation problem caused by "individual optimization and lack of collaboration", forming a full-link collaborative evolution system to ensure the reliability of analysis in the long term. Attached Figure Description

[0067] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0068] Figure 1 This is a flowchart illustrating a multi-source spatiotemporal data-driven intelligent analysis method for flood impacts according to the present invention.

[0069] Figure 2 This is a schematic diagram of the module architecture of a multi-source spatiotemporal data-driven intelligent analysis system for flood impacts according to the present invention. Detailed Implementation

[0070] 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0071] like Figure 1 As shown, this embodiment provides a multi-source spatiotemporal data-driven intelligent analysis method for flood impact, including the following steps: S1, constructing a multi-scale spatiotemporal index benchmark; determining the geographical boundary of the target analysis area, and discretizing the area into layers based on the hierarchical coding characteristics of the BeiDou grid code, constructing a coarse grid layer and its nested fine grid layers to form a unified spatiotemporal index framework; step S1 also includes the following steps:

[0072] S11. In this embodiment, the target area is a river basin at the provincial level. Based on its spatial range and management requirements, the third-level (L3) grid unit of the BeiDou grid code is selected as the coarse grid. This level of grid has a moderate coverage range and is suitable as a distributed block index for massive basic data, which can significantly reduce the I / O overhead during data retrieval. At the same time, the sixth-level (L6) grid unit of the BeiDou grid code is selected as the fine grid. This level of grid scale can reach the meter level, which can accurately match the positioning accuracy of micro-geographic entities such as houses, road sections, and key facilities, thus serving as the smallest spatial calculation unit for flood impact analysis.

[0073] S12. Through the inherent hierarchy and coding system of the BeiDou grid code, a rapid mapping relationship is established from macroscopic coarse grid to microscopic fine grid, forming a unified spatiotemporal index framework. This framework serves as the unified spatial benchmark and correlation basis for data storage, querying, analysis, and result expression in all subsequent steps.

[0074] S2. Construct a dual knowledge base system; construct and maintain dual knowledge bases that are respectively associated with the fine grid space, including a geographic entity map database and a flood scenario vector knowledge base;

[0075] S3. Extract the set of affected fine grids; Receive the current flood inundation feature data, and extract the set of affected fine grids within the inundation range based on the spatiotemporal index framework through spatial overlay analysis;

[0076] In this invention, step S3 further includes the following sub-steps:

[0077] S31. Receive current flood inundation characteristic data from hydrodynamic model simulation, remote sensing monitoring and other channels. The flood inundation characteristic data includes inundation range polygons and core characteristic parameters associated with each polygon, such as inundation depth, flow velocity and duration.

[0078] S32. Perform spatial overlay analysis on the flooding range data and the L6 level fine grid layer, filter out all flooded or partially flooded grids, and form a set of affected fine grids, denoted as G={g1,g2,...,gn}, where each g is a unique BeiDou grid code.

[0079] S4. Parallel query of dual knowledge bases: Based on the BeiDou grid code corresponding to the affected fine grid set, the geographic entity map database and the flood scenario vector knowledge base are queried in parallel to obtain the geographic entity dataset and similar historical scenario information set corresponding to each fine grid.

[0080] In this invention, step S4 further includes the following sub-steps:

[0081] S41. Using each grid code in set G as the key, query the geographic entity map database in parallel to obtain the type, quantity, weighted attribute values, etc. of all geographic entities in the grid, and construct the geographic entity dataset of the grid.

[0082] S42. Transform the overall situational characteristics of the current flood (data such as inundation depth, flow velocity, duration, and impact range obtained from step S3) into a semantic query vector comparable to historical cases. Specifically, the transformation is as follows: First, according to a preset standardized text template, the above hydrological parameters are structured into standardized descriptive text consistent with the format of historical flood case reports. Then, this standardized descriptive text is input into the text encoding stream of the dual-stream network described in S222 (based on a BERT model fine-tuned for the flood domain), generating a 768-dimensional text semantic vector. If real-time remote sensing imagery is available, the imagery can also be input into the image encoding stream of the dual-stream network (based on a ViT model fine-tuned for flood remote sensing imagery) to extract visual feature vectors. These vectors are then fused with the text semantic vectors through a cross-modal attention layer to generate the final real-time situational semantic query vector. If no imagery is available, the text semantic vector is used directly as the query vector. After generating the query vector, a similarity search (using cosine similarity calculation) is performed in the flood scenario vector knowledge base, combined with the BeiDou grid codes in the affected fine grid set G, to find the K most similar historical cases (in this embodiment, the initial value of K is set to 5, and it serves as the threshold value for determining whether to start the calibration process in step S531), thus forming a similar historical scenario information set. In this set, each case information includes its semantic similarity to the current scenario, as well as the actual post-disaster verified impact data of the case at the current grid location in its historical context.

[0083] S5. Calculation of flood impact index by dual-model collaborative calibration: For each affected fine grid, based on the geographic entity dataset, similar historical scenario information set and flood inundation characteristic data, the data-driven model and historical experience model are collaboratively analyzed, and the dual-model collaborative calibration module dynamically weights and fuses the analysis results of the two models to output the calibrated flood impact index (I_c).

[0084] In this invention, step S5 further includes the following sub-steps:

[0085] S51. Calculate the preliminary flood impact index I_p, specifically by constructing a flood impact index system encompassing flood characteristic indicators, land feature indicators, and social operation indicators. Flood characteristic indicators include water depth, flow velocity, and duration, with values ​​directly extracted from flood inundation characteristic data. Land feature indicators include land use, buildings, water conservancy projects, transportation facilities, power facilities, and communication facilities. Social operation indicators include settlements, population, hospitals, schools, factories, and cultural relics. The values ​​of land feature indicators and social operation indicators are obtained by spatial statistics and attribute quantification of the corresponding entities in the geographic entity dataset obtained in S4. After normalizing all indicators to the [0,1] interval using the min-max normalization method, a standardized feature vector for this grid is formed and input into a pre-trained flood impact index model. This model is a regression model based on gradient boosting decision tree (LightGBM), capable of learning a nonlinear mapping from complex features to the degree of impact, and outputs a preliminary impact index I_p between [0,1].

[0086] The model training dataset is constructed according to the following rules: inundation feature data of historical flood events in the target area and corresponding grid geographic entity data are collected as input features, and the grid-level actual flood impact index verified after the flood is used as the training label; the dataset is divided into training set and validation set in a 7:3 ratio, and the hyperparameters are optimized by five-fold cross-validation. All input features are normalized to the [0,1] interval by the min-max normalization method to eliminate the influence of units.

[0087] S52. Calculate the historical experience benchmark index (I_b): This is done using the scenario experience analysis model (Case-Based Reasoning, CBR). Retrieve K similar historical cases whose semantic similarity to the current flood scenario is greater than a preset minimum threshold. The semantic similarity of the i-th similar historical case is used as the benchmark index. As the weight, the actual impact data on the i-th similar historical case. The weighted aggregation is calculated using the following formula: The preset minimum threshold is determined through statistical analysis of historical flood case datasets, with a value range of [0.4, 0.7]. Cases below this threshold are excluded due to excessively large scenario differences, thus avoiding the introduction of noise. In a specific embodiment of the present invention, this threshold is optimized to a value of 0.5.

[0088] For the g-th geographic entity within each affected fine-grained grid, the entity-level impact index is calculated simultaneously: Entity-level Preliminary Impact Index By combining the entity's attribute characteristics with the current flood inundation characteristic data, and inputting this data into the flood impact index model, the inherent impact index of the entity is generated. This index is then multiplied by the entity's association weight factor in the current grid. Obtain; Entity-level historical benchmark index By retrieving historical cases matching the entity type and location from a set of similar historical context information, and weighting the historical actual impact data with semantic similarity as the weight, an inherent historical benchmark index is obtained, which is then multiplied by an association weight factor. get;

[0089] S53. Input the preliminary flood impact index and the historical experience benchmark index into the dual-model collaborative calibration module for fusion calibration, and output the calibrated flood impact index. This module contains a rigorous logical judgment process:

[0090] S531. First, based on the similar historical scenario information set obtained in step S4, determine the number of similar historical scenarios K. If K < 5, directly output I_c = I_p and terminate the calibration process. This threshold value of 5 can be dynamically adjusted according to the richness of historical cases in the region. When the number of historical cases in the region is greater than 10, the threshold is increased to 8-10; when the number of historical cases is 5-10, the threshold is set to 5-8; when the number of historical cases is less than 5, the threshold is decreased to 3-5. This step ensures that the fusion is statistically significant and avoids fusion bias caused by small sample data.

[0091] S532. Otherwise, calculate the average semantic similarity S: S = The S value ranges from [0,1] and is used to quantify the overall essential similarity between the current scenario and the historical scenario set. The higher the S value, the more similar the current flood scenario is to the retrieved historical scenarios in terms of causes, evolution patterns, and impact characteristics, and the higher the reference value of historical experience.

[0092] S533. Calculate the consistency measure C between the preliminary flood impact index I_p and the historical experience benchmark index I_b. The calculation formula is as follows: When max(I_p,I_b)=0, C=1 is defined, and I_c=0 is directly output; this scenario represents that the current grid has no risk of flooding and no subsequent calibration calculation is required.

[0093] The value range of C is [0, 1]. The higher the C value, the higher the degree of coincidence between the calculation result of the real-time model and the value judged by historical experience, and the higher the credibility of the fusion result;

[0094] S534. Calculate the confidence factor α = w1×C + w2×S. In this embodiment, w1 and w2 are defaulted to 0.5, and can be dynamically adjusted within the range of [0.2, 0.8] according to the regional flood characteristics and the richness of historical cases, and satisfy w1 + w2 = 1;

[0095] S535. Set the dynamic confidence threshold T. First, calculate the geographical entity complexity of the current fine grid , and normalize it to [0, 1]. Then, calculate T according to the formula T = T0×(1 + μ×F), where μ is the adjustment coefficient, and the value range is [-0.3, 0.2], which is used to control the influence amplitude of the geographical entity complexity on the threshold. In this embodiment, let T0 = 0.7 and μ = 0.2. The specific values of T0 and μ can be optimized and determined on the validation set by the grid search method according to the richness of regional historical flood cases and the complexity of the geographical environment. If there are many types and large quantities of geographical entities in the grid (F is high), then T will be correspondingly increased, which means a higher threshold for fusion and more cautious decision-making;

[0096] S536. Perform adaptive dynamic weighted fusion and output the calibrated flood impact index I_c. This step includes two layers of judgment logic:

[0097] If α≥T and C + S≠0, then calculate according to the formula The preset threshold can be determined by cross-validation on the flood historical case dataset with the goal of maximizing the comprehensive analysis accuracy in the historical scenario;

[0098] If α < T or C + S = 0, then it is determined that there is a significant statistical difference between the current flood scenario and historical laws, and it is very likely to belong to extreme or new disaster events not covered by the training data. At this time, to prevent low-quality or irrelevant historical experience from misleading the calibration result, the system triggers a "fuse mechanism" and forcibly sets I_c to be equal to I_p;

[0099] S6. Generate fine-grained analysis results; based on the calibrated flood impact index (I_c), generate flood impact analysis results with the fine grid as the smallest analysis unit;

[0100] In this invention, step S6 further includes the following sub-steps:

[0101] S61. The preset flood impact index risk threshold is 0.85. This threshold can be determined through statistical analysis of historical flood disaster data. For example, the lower quartile of the grid I_a value corresponding to events historically assessed as "significant losses" or higher can be selected as the benchmark. Based on the calibrated I_c values ​​of all affected grids, fine grids with I_c ≥ 0.85 are selected and identified as high-risk areas, and their BeiDou grid code sets are recorded. The risk threshold can be dynamically adjusted within the range of [0.8, 1.0] according to the regional flood control level and the importance of the protected objects.

[0102] S62. Extract key risk objects within the grid of high-risk areas, including schools, hospitals, residential areas, main transportation routes, water conservancy facilities, power and communication facilities, etc.

[0103] S63. Based on the I_c values ​​of all grids, render and generate a flood risk classification thematic map according to three levels: 0-0.3 (low risk), 0.3-0.85 (medium risk), and 0.85-1.0 (high risk).

[0104] S64: Automatically generate structured analysis reports. The core content of the report includes: (a) defining the scope of the high-risk area using a fine-grained BeiDou grid code set; (b) listing key geographic entities associated with the high-risk area (such as schools, hospitals, and substations); (c) risk statistics and visualization charts generated based on I_c; (d) providing statistical analysis data and risk assessment conclusions based on I_c; and (e) retrieving and generating preliminary emergency response suggestions from the rule base based on information about the high-risk area and key geographic entities.

[0105] S7. Iterative optimization of model and knowledge base: Receive actual post-disaster impact data (I_a), construct a joint loss function based on the difference between (I_a) and (I_c), jointly optimize the parameters of the flood impact index model, scenario experience analysis model and dual-model collaborative calibration module, and synchronously update the associated data and vector features in the dual knowledge base;

[0106] In this invention, step S7 further includes the following sub-steps:

[0107] S71. Collect actual flood impact data through on-site investigation, remote sensing assessment and other means, including actual inundation range and area, actual affected risk objects, number of affected people, and the area of ​​various types of land affected, number of different facilities, etc. Using the same index system and method as steps S51 and S52, calculate the grid-level actual flood impact index I_a (normalized to the range of [0,1]).

[0108] S72. Construct and calculate the joint loss function:

[0109] ,

[0110] Where β is the balance coefficient, which is determined by cross-validation on a historical flood case dataset within the interval [0.1, 1.0]. In this embodiment, β is set to 0.5. The optimization process of β is as follows: traverse [0.1, 1.0] on the validation set with a step size of 0.1, and select the β value that minimizes the validation set L_t.

[0111] S73. Joint Optimization of Model Parameters: With the objective of minimizing the joint loss function L_t, the Adam gradient descent algorithm is used to perform end-to-end joint updates on the tree structure and leaf node weights of the flood impact index model, the similarity weighting rules of the scenario experience analysis model, and the w1, w2 weight coefficients and μ adjustment coefficients of the dual-model collaborative calibration module. The batch size is optimized to 64, the number of iterations to 50, and the learning rate is set to 0.001. If the validation set loss does not decrease for five consecutive iterations, training is terminated early to prevent overfitting. This process aims to improve the overall analytical performance of the model without altering any existing records in the knowledge base.

[0112] S74. Based on the complete data and model performance of this flood event, initiate a collaborative update process for the two knowledge bases:

[0113] S741. New Case Entry: This flood event will be stored as a new case in the flood scenario vector knowledge base. The multimodal fusion coding, impact grid matching, and reverse index establishment will be completed according to the complete process of S24-S26 to expand the case coverage of the knowledge base.

[0114] S742. Calculate the deviation between the historical experience benchmark I_b and the actual value I_a. If the deviation exceeds the preset threshold of 20%, add a "low reference value" mark to the key historical cases on which this retrieval is based, and associate it with the event ID of this retrieval. This mark is only used for subsequent retrieval weight adjustment and does not change the original data of the original cases.

[0115] S743. Cases marked with "low reference value" are centrally reviewed according to a preset cycle (e.g., every six months). The knowledge base is only modified when it is confirmed by experts in the field of water conservancy or meets the preset rules. The modification methods include adding correction notes and adjusting its weight coefficient in similarity retrieval.

[0116] S744. Based on the damage and reconstruction of ground features identified in the post-disaster remote sensing images, update the status and attribute information of the corresponding geographic entities in the geographic entity map database to ensure the timeliness and accuracy of geographic entity data.

[0117] The collaborative update triggering conditions for the dual knowledge bases are as follows: after each flood event ends and complete post-disaster actual impact data is obtained, a full update is performed; at the same time, incremental updates of geographic entity data are performed according to a preset quarterly cycle, and case review and weight adjustment of the flood scenario vector knowledge base are performed according to a semi-annual cycle.

[0118] At this point, the system has completed a full intelligent closed loop of 'perception-analysis-decision-verification-evolution', and its analysis model and knowledge base have been enhanced, realizing the system's self-evolution.

[0119] In this invention, in step S1, the coarse grid is a grid cell selected from the first to third levels of the BeiDou grid code, used for distributed storage indexing of spatiotemporal data to reduce I / O overhead during data retrieval; the fine grid is a grid cell selected from the fourth to seventh levels of the BeiDou grid code, serving as the smallest spatial calculation unit for flood impact analysis; the coarse grid and the fine grid are associated and located through the BeiDou grid code. In step S2, the geographic entity map database is used to store the topological structure, attribute characteristics, and spatial association between the geographic entities and the fine grid within the target area. Constructing the geographic entity map database includes: S21, acquiring multi-source spatiotemporal data related to flood impact analysis and unifying it to a spatiotemporal reference based on the BeiDou grid code; extracting geographic entity object data after trimming and dividing it into blocks according to the coarse grid; the multi-source spatiotemporal data includes basic geographic information data, land feature vector data, and social operation vector data; through format conversion, spatial matching, and attribute mapping operations, the land feature vector data and social operation vector data are converted into geographic entity object data, which contains the spatial geometric information, attribute characteristic information, and grid association information of the entities.

[0120] It should be noted that the target area can be one of the following: national, river basin, province, city, district (county), or natural geographic unit; the basic geographic information data includes remote sensing imagery, administrative divisions, and POI data; the geographic feature vector data includes vector data of land use, buildings, water conservancy projects, transportation facilities, power facilities, and communication facilities; among them, the geometric type of land use data is polygon, and the attributes are land use type and area of ​​the map patch; the geometric type of building data is polygon, and the attribute is building area; the geometric type of water conservancy projects, power facilities, and communication facilities data is point, and the attribute is type; the geometric type of transportation facilities is line, and the attribute is length; the social operation vector data includes vector data of settlements, population, hospitals, schools, factories, and cultural relics, and the data is verified and fused through administrative divisions and POI data; among them, the geometric type of settlements, hospitals, schools, factories, and cultural relics data is point, and the attributes are name and level; the geometric type of population data is point, and the attribute is number of people.

[0121] In this invention, step S21, constructing the geographic entity map database, is as follows:

[0122] S211. Obtain basic geographic information data and vector data of land features and social operations within the target area from industry management departments such as water resources, natural resources, housing and construction, public security, ecological environment, agriculture and rural affairs, statistics, and tourism. Perform standardization processing such as data format conversion, coordinate conversion, and data fusion on the data, and unify all data to a spatiotemporal reference based on Beidou grid code.

[0123] S212. The multi-source data after unified benchmarking is clipped and segmented according to the range of L3 level coarse grid. Distributed storage technology (such as HBase) is used to fragment the multi-source spatial data processed in step S211 according to the coarse grid unit, with each coarse grid corresponding to one data fragment, improving data storage and access efficiency. Through format conversion, spatial matching, and attribute mapping operations, the standardized ground feature vector data and social operation vector data are converted into geographic entity object data. The geographic entity object data contains three core dimensions: spatial geometric information, attribute feature information, and grid association information of the entity.

[0124] S213. Through GIS spatial analysis, map all geographic entity objects (area, line, and point) to L6 fine-grained grids. For area or line entities spanning multiple grids (such as a cultivated land patch or a main road), perform spatial segmentation and dynamically generate "entity-grid" association weight factors. For area entities, the weight factor is the ratio of the entity's area within the grid to its total area; for line entities, the weight factor is the ratio of the entity's length within the grid to its total length. The attributes of geographic entities (such as area and road length) will be weighted according to this weight factor. For example, a road with a total length of 1000 meters has a length of 300 meters in grid A (weight factor 0.3) and 700 meters in grid B (weight factor 0.7). When constructing the graph model, the node in grid A is associated with a "road" entity with a "length" feature value of 300; the node in grid B is associated with an entity feature value of 700. Point entities are directly associated with a weight factor of 1.

[0125] S214. Using each L6 fine grid as the master node, the geographic entities contained within the grid (or after segmentation) as the subordinate nodes, and the weighted attribute values ​​as the node attribute features, construct a "grid-entity" graph model, and import it into graph databases such as Neo4j to complete storage and index construction.

[0126] S215. According to the preset cycle of each quarter, use semantic segmentation models such as U-Net to interpret remote sensing images and update the change patches in the vector data of ground features, such as land use, buildings, water conservancy projects, transportation facilities, power facilities, communication facilities, etc.

[0127] S216. Analyze mobile communication signaling data according to a preset monthly cycle, and calculate the population distribution data of communication base station locations through methods such as kernel density estimation. The population distribution data includes the latitude and longitude coordinates of the base station and the number of residents, employed people, and migrants. Dynamically aggregate the population distribution data to a fine grid, calculate the population distribution data at the fine grid scale, and dynamically update the population distribution data in the social operation vector data accordingly.

[0128] S22. Geographic entities are mapped to fine grids through GIS spatial analysis. For areal or linear geographic entities spanning multiple fine grids, spatial segmentation is performed, and "entity-grid" association weight factors are dynamically generated based on the entity type and its geometric proportion within the current grid. A graph data model is constructed, in which the fine grid serves as the master node and the associated geographic entity objects serve as the subordinate nodes. The association weight factors are used for the weighted aggregation of geographic entity features in the subsequent flood impact index model, as well as the weighted calculation of the geographic entity-level preliminary impact index and the geographic entity-level historical benchmark index.

[0129] In this invention, step S22 further includes the following sub-steps:

[0130] S221. Collect case data on 30 historical flood events in the target area over the past 15 years, including disaster assessment reports, remote sensing inundation images, on-site photographs, and data on inundation depth, duration, impact range, and emergency response measures for each case. The 30 cases selected in this embodiment are merely examples; in practical applications, the number of cases, time span, and K-value can be configured based on data completeness and computing resources.

[0131] S222. Employing a two-stream network based on the Transformer architecture for multimodal fusion encoding is key to transforming unstructured historical experience into high-quality computable knowledge.

[0132] (a) Text Encoding Stream: Using a BERT model fine-tuned for the flood domain, deep semantic understanding is performed on disaster report texts to extract text feature vectors T containing information such as disaster causes, evolution process, impact degree, and handling experience. The core parameters of the model are: hidden layer dimension 768, number of attention heads 12, number of network layers 12, and maximum length of input text 512.

[0133] (b) Image Coding Stream: The ViT model, finely tuned from flood remote sensing images, is used to perform visual analysis on flooded images and on-site photos to extract visual feature vectors V that reflect the flooding range, flooding degree, and damage status of land features. The core parameters of the model are: image input size 224×224, patch size 16, hidden layer dimension 768, number of attention heads 12, and number of network layers 12.

[0134] (c) Cross-modal fusion and fine-tuning: A cross-modal attention layer is introduced, enabling textual features T and visual features V to perform interactive attention calculations, generating a unified, high-dimensional contextual semantic vector F. Specifically, the pre-trained weights of the dual-stream network are fine-tuned based on flood-specific corpus and remote sensing image data, enabling it to extract semantic features of flood scenarios more accurately.

[0135] S223. Based on the inundation range recorded in historical flood case data or through spatial overlay analysis, determine the L6 level fine grid set and corresponding BeiDou grid code affected by each case.

[0136] S224. A Milvus vector database is used to store the semantic vectors of 30 cases. An IVF_FLAT index is used to create a reverse index from each semantic vector to the corresponding BeiDou grid code set of the influencing grid, thus completing the construction of a flood scenario vector knowledge base. This allows for the rapid retrieval of a set of historical cases similar to the current scenario by calculating vector similarity.

[0137] S23. Update geographic entity data based on the interpretation of remote sensing image data at a preset period, and update population distribution data in social operation vector data based on the analysis of mobile communication signaling data at a preset period.

[0138] In step S2, the flood scenario vector knowledge base is used to store the semantic feature vectors and hydrological parameters of historical flood disaster cases, as well as the mapping relationship between the historical flood cases and the affected grids, forming a historical experience vector database. Constructing the flood scenario vector knowledge base includes:

[0139] S24. Multimodal fusion encoding is performed on text reports and image data of historical flood cases to generate high-dimensional semantic vectors representing the overall scenario of the cases. The multimodal fusion encoding adopts a Transformer-based two-stream network structure, extracting semantic feature vectors from text reports and visual feature vectors from image data, and then performing feature-level fusion and splicing through an attention mechanism. The pre-training weights of the two-stream network are fine-tuned based on flood-specific corpus and remote sensing image data.

[0140] S25. Based on the inundation range recorded in historical flood case data or through spatial overlay analysis, determine the fine grid set affected by the case.

[0141] S26. Establish a reverse index mapping relationship between the semantic vector of each case and the set of fine-grid BeiDou grid codes it affects.

[0142] In step S4, the geographic entity dataset is the geographic entity dataset within each affected fine grid.

[0143] The set of similar historical scenarios is a set of historical flood scenarios that are similar to the current flood scenario in terms of hydrological semantics.

[0144] In step S5, the dual-model collaborative analysis specifically includes:

[0145] S51. Input the geographic entity dataset and the current flood inundation feature data into the flood impact index model to calculate the preliminary flood impact index (I_p).

[0146] S52. Input the historical similar scenario information set into the scenario experience analysis model to calculate the experience benchmark index (I_b).

[0147] S53. Input (I_p) and (I_b) into the dual-model collaborative calibration module, calculate the consistency measure (C) between the preliminary flood impact index (I_p) and the empirical benchmark index (I_b) and the average semantic similarity (S) between the current scenario and the historical scenario, determine the fusion weight based on C and S, perform dynamic weighted fusion of (I_p) and (I_b), and finally output the calibrated flood impact index (I_c).

[0148] The flood impact index model is a supervised machine learning regression model based on the LightGBM gradient boosting decision tree framework. It uses the actual impact data I_a verified after the disaster for each fine grid in historical flood events as training labels. The actual impact data I_a is the actual impact data after the disaster mentioned in step S7. The scenario experience analysis model is a case-based reasoning model, which is configured to obtain the historical experience benchmark index (I_b) in the following way: the actual impact data corresponding to each historical scenario in the similar historical scenario information set is weighted and aggregated.

[0149] When the scenario experience analysis model performs weighted aggregation, it only includes historical scenarios whose semantic similarity to the current flood scenario is greater than a preset minimum threshold, and the weight is positively correlated with the semantic similarity of the corresponding historical scenario. The preset minimum threshold is determined through statistical analysis of the historical flood case dataset, and its value ranges from [0.4, 0.7]. This threshold is used to filter out historical cases with too low similarity to the current scenario to avoid introducing irrelevant noise, and its specific value is determined by analyzing the average similarity distribution among cases in the historical case set.

[0150] The dual-model collaborative calibration module performs adaptive dynamic weighted fusion in the following manner:

[0151] Based on the similar historical scenario information set obtained in step S4, the number of qualified similar historical scenarios K is determined. If K < 5, I_c = I_p is directly output, and the subsequent fusion calibration process in this step is terminated. The critical value for the number of similar historical scenarios is 5. This critical value is determined by statistical analysis of the historical flood case dataset and can be dynamically adjusted according to the richness of historical cases in the region.

[0152] Otherwise, calculate the average semantic similarity S between the current flood scenario and the aforementioned qualified similar historical scenarios. The calculation formula is: S = Where K is the number of qualified similar historical scenarios mentioned above. Let be the semantic similarity between the i-th qualified similar historical scenario and the current scenario;

[0153] The consistency measure (C) between the preliminary flood impact index (I_p) and the historical experience benchmark index (I_b) is calculated using the following formula: When max(I_p,I_b)=0, C=1 is defined. At this time, it is determined that the current fine mesh is not affected by flood, and I_c=0 is directly output.

[0154] Calculate the current confidence factor α, α=w1×C+w2×S, where w1 and w2 are preset weight coefficients, with default values ​​of w1=0.5 and w2=0.5, which can be dynamically adjusted according to regional flood characteristics and the richness of historical cases. The values ​​range are [0.2,0.8] and w1+w2=1.

[0155] A dynamic confidence threshold T is set, which is adaptively adjusted based on the complexity of geographic entities within the current fine grid. The quantification formula for the complexity F is as follows: ,in The number of geographic entity types within a fine-grained grid. F represents the total number of geographic entities within the fine grid, and is normalized to the range [0,1]. The adjustment rule is T=T0×(1+μ×F), where T0 is the basic threshold with a value range of [0.5,0.8], which can be dynamically adjusted according to the flood warning level and emergency response level. μ is the adjustment coefficient, used to control the impact of geographic entity complexity on the threshold, with a value range of [-0.3,0.2]. When the importance of the protected objects in the area is high, a positive value is taken to increase the threshold and improve the fusion credibility requirements. When the richness of historical cases is low, a negative value is taken to decrease the threshold to adapt to scenarios with insufficient historical experience.

[0156] Based on C and S, as well as the confidence factor α and the dynamic confidence threshold T, the calibrated flood impact index (I_c) is calculated:

[0157] If α is greater than or equal to T, and C+S≠0, then according to the formula... Calculate; if α is less than T, or C+S=0, then it is determined that there is a significant difference between the current flood scenario and historical patterns, and the current fusion result has low reliability. In this case, I_c is set to be equal to I_p.

[0158] It should be noted that the core innovative value of the dual-model collaborative calibration module of this invention lies in constructing a hybrid intelligent decision-maker based on confidence assessment and possessing anomaly protection capabilities. It achieves, for the first time, fully adaptive dynamic adjustment of the fusion weights—the fusion weights change in real time with the values ​​of the consistency metric C and the average semantic similarity S, completely abandoning the fixed-weight fusion logic of existing technologies. First, a fusion threshold is set by judging K<5 to ensure statistical significance. Second, the calibration weights are defined as functions of two technical indicators (C and S) with clear physical meaning. C measures the numerical consistency between real-time data and historical experience knowledge, and S measures the essential similarity between the current scenario and historical patterns. Furthermore, by introducing a confidence factor α and a dynamic threshold T adaptively based on geographic entity complexity F, the calibration behavior becomes dynamically adaptive and robust: when α is sufficiently high (i.e., the combined confidence of C and S is high) and C+S is non-zero, the dynamic ratio of C and S is used for weighted fusion; when α is lower than the adaptive threshold T, the system triggers a "circuit breaker mechanism," forcing reliance on the real-time model to prevent interference from erroneous knowledge in complex geographical environments. This solves the problem of performance degradation of traditional weighted average or voting methods when the scenario is mismatched or the geographical environment is complex.

[0159] In step S7, the joint loss function (L_t) is defined as:

[0160] ;

[0161] Where I_a represents grid-level post-disaster actual flood impact data, To calibrate the output error, n is the total number of geographic entities in the current fine grid, and g is the g-th geographic entity in this fine grid. This is the association weight factor for the entity in the current fine grid (the association weight factor is generated based on the entity type and its geometric proportion in the current fine grid). and These are the entity-level preliminary impact index and the entity-level historical baseline index, calculated for the entity using the flood impact index model and the scenario experience analysis model, respectively. The actual impact data of the corresponding geographic entities are obtained through post-disaster field verification and remote sensing image interpretation; β is a balance coefficient used to adjust the weight between the overall grid-level error and the consistency control of entity granularity deviation, and its value is determined by cross-validation on historical flood case datasets within the interval [0.1, 1.0]; the joint optimization process aims to minimize L_t.

[0162] In this embodiment, the design of the joint loss function L_t essentially achieves multi-task collaborative optimization under a single objective: the first term The second term directly optimizes the final output accuracy. It is an entity-level regularization constraint, the effect of which is that during training, it utilizes entity weight factors... At a single fine grid granularity, this forces the prediction errors of the flood impact index model and the scenario empirical analysis model to increase. and They move closer together. This means that the optimization process not only pursues the minimum total error but also drives the two heterogeneous models to align and balance at a finer entity granularity, thereby promoting implicit synergy and co-evolution of the two models in sharing knowledge representations. The parameter β is optimized through cross-validation, which essentially adjusts the weights of the two optimization objectives: 'final accuracy' and 'balance between entity-level models'.

[0163] Furthermore, to support refined analysis and joint optimization at the geographic entity level, step S5 also includes the calculation of the geographic entity-level impact index:

[0164] For the g-th geographic entity within each affected fine grid, its entity-level preliminary impact index The index is obtained by combining the entity's attribute features with the current flood inundation feature data, inputting it into the flood impact index model to generate the entity's inherent impact index, and then multiplying it by the entity's correlation weight factor in the current grid. ,get .

[0165] For the g-th geographic entity within each affected fine grid, its entity-level historical benchmark index The data is obtained by: retrieving historical cases related to the entity type and location from the set of similar historical scenarios; weighting the historical actual impact data with semantic similarity as the weight to obtain the entity's inherent historical benchmark index; and then multiplying it by the association weight factor. ,get .

[0166] In step S6, the process of generating the flood impact analysis results includes: firstly, screening high-risk areas based on the calibrated flood impact index I_c (fine grids with I_c ≥ a preset risk threshold are high-risk areas, the preset risk threshold is determined through statistical analysis of historical flood cases, and the value range is [0.8, 1.0], which can be dynamically adjusted according to the regional flood control level and the importance of the protected objects), and then generating analysis results with the set of fine grid Beidou grid codes constituting the high-risk areas as the core spatial identifier; the analysis results include at least one of the following:

[0167] (a) The scope of the high-risk area identified by the Beidou grid code set;

[0168] (b) Key geographic entity information associated with the high-risk area;

[0169] (c) Visual representation based on the flood impact index I_c;

[0170] (d) Statistical analysis data and risk assessment conclusions based on the flood impact index I_c;

[0171] (e) Emergency response recommendations based on information on high-risk areas and key geographic entities.

[0172] like Figure 1 As shown, a multi-source spatiotemporal data-driven intelligent flood impact analysis system is used to execute the multi-source spatiotemporal data-driven intelligent flood impact analysis method. The multi-source spatiotemporal data-driven intelligent flood impact analysis system includes:

[0173] The BeiDou grid code spatial service engine provides two-level spatial grid division, encoding calculation, and hierarchical mapping services based on BeiDou grid codes, supporting rapid encoding conversion, spatial relationship analysis, and grid mapping retrieval for different grid levels.

[0174] The geographic entity map database management module is used to perform the access of multi-source geospatial data, organize it into blocks according to coarse grids, segment geographic entities and dynamically weight the attributes based on fine grids, and build and maintain the map data model.

[0175] The flood scenario vector knowledge management module is used to perform multimodal fusion encoding of historical flood cases, high-dimensional semantic vector storage, and retrieval based on vector similarity and grid code reverse indexing, so as to realize the construction and maintenance of the flood scenario vector knowledge base; the pre-trained weights of the dual-stream network for multimodal fusion encoding are fine-tuned based on flood-specific corpus and remote sensing image data;

[0176] The dual-model collaborative analysis engine uses the BeiDou grid code of the affected fine grid as a unified retrieval key to concurrently drive parallel queries on the geographic entity map database and the flood scenario vector knowledge base. It calls the flood impact index model and the scenario experience analysis model to calculate I_p and I_b respectively, and uses the dual-model collaborative calibration module to adaptively and dynamically weight and fuse I_p and I_b to output the calibrated flood impact index I_c.

[0177] The analysis results generation module is used to generate analysis results with a fine-grid Beidou grid code set as the core identifier based on the calibrated flood impact index I_c;

[0178] The model and knowledge base joint optimization module is used to drive the joint optimization of the flood impact index model, the scenario experience analysis model, and the dual-model collaborative calibration module based on the joint loss function L_t described in step S7 of this specification. It simultaneously realizes the dynamic update of geographic entity data in the geographic entity map database and the update of semantic vector and reverse index mapping in the flood scenario vector knowledge base. The joint loss function is configured to introduce entity-level error terms and their associated weight factors for collaborative optimization on the basis of grid-level error.

[0179] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0180] The GIS spatial analysis, machine learning model training, vector similarity retrieval, and other related techniques not elaborated in this specification are all conventional techniques well-known to those skilled in the art. Those skilled in the art can fully reproduce the technical solutions of this invention based on the content disclosed in this specification.

Claims

1. A multi-source spatiotemporal data-driven intelligent analysis method for flood impact, characterized in that, Includes the following steps: S1. Constructing a multi-scale spatiotemporal index benchmark The geographical boundaries of the target analysis area are determined, and the area is discretized into layers based on the hierarchical coding characteristics of the Beidou grid code. A coarse grid layer and its nested fine grid layers are constructed to form a unified spatiotemporal indexing framework. S2. Constructing a related dual knowledge base system Construct and maintain a dual knowledge base that is associated with the fine grid space, including a geographic entity map database and a flood scenario vector knowledge base; S3, Extract the affected fine mesh set Receive current flood inundation characteristic data, and extract the set of affected fine grids within the inundation range through spatial overlay analysis based on the spatiotemporal index framework; S4, Parallel query with dual knowledge bases Based on the BeiDou grid code corresponding to the affected fine grid set, the geographic entity map database and flood scenario vector knowledge base are queried in parallel to obtain the geographic entity dataset and similar historical scenario information set corresponding to each fine grid. S5. Calculation of Flood Impact Index under Dual-Model Co-calibration For each affected fine grid, based on the geographic entity dataset, similar historical scenario information set and flood inundation characteristic data, the data-driven model and historical experience model are analyzed in collaboration, and the analysis results of the two models are dynamically weighted and fused by the dual-model collaborative calibration module to output the calibrated flood impact index (I_c). S6. Generate fine-grained analysis results Based on the calibrated flood impact index (I_c), flood impact analysis results are generated with fine grid as the smallest analysis unit; S7, Iterative Optimization of Model and Knowledge Base Receive the actual post-disaster impact data (I_a), construct a joint loss function based on the difference between (I_a) and (I_c), jointly optimize the parameters of the data-driven model, historical experience model and dual-model collaborative calibration module, and synchronously update the associated data and vector features in the dual knowledge bases.

2. The method for intelligent analysis of flood impact driven by multi-source spatiotemporal data according to claim 1, characterized in that, In step S1, the coarse grid is a grid cell selected from the first to third levels of the BeiDou grid code, used for distributed storage indexing of spatiotemporal data to reduce I / O overhead during data retrieval; the fine grid is a grid cell selected from the fourth to seventh levels of the BeiDou grid code, serving as the smallest spatial calculation unit for flood impact analysis; the coarse grid and the fine grid are associated and located through the BeiDou grid code.

3. The method for intelligent analysis of flood impact driven by multi-source spatiotemporal data according to claim 1, characterized in that, In step S2, the geographic entity map database is used to store the topological structure and attribute characteristics of various geographic entities within the target area, as well as the spatial association between the geographic entities and the fine grid. Constructing the geographic entity map database includes: S21. Acquire multi-source spatiotemporal data related to flood impact analysis and unify them to a spatiotemporal reference based on the BeiDou grid code. Extract geographic entity object data after cutting and segmenting according to the coarse grid. The multi-source spatiotemporal data includes basic geographic information data, land feature vector data, and social operation vector data. Through format conversion, spatial matching, and attribute mapping operations, convert the land feature vector data and social operation vector data into geographic entity object data. The geographic entity object data contains the spatial geometric information, attribute feature information, and grid association information of the entity. S22. Geographic entities are mapped to fine grids through GIS spatial analysis. For areal or linear geographic entities spanning multiple fine grids, spatial segmentation is performed, and "entity-grid" association weight factors are dynamically generated based on the entity type and its geometric proportion within the current grid. A graph data model is constructed, in which the fine grid serves as the master node and the associated geographic entity objects serve as the subordinate nodes. The association weight factors are used for the weighted aggregation of geographic entity features in the subsequent flood impact index model, as well as the weighted calculation of the geographic entity-level preliminary impact index and the geographic entity-level historical benchmark index. S23. Update geographic entity data based on the interpretation of remote sensing image data at a preset period, and update population distribution data in social operation vector data based on the analysis of mobile communication signaling data at a preset period.

4. The method for intelligent analysis of flood impact driven by multi-source spatiotemporal data according to claim 1, characterized in that, In step S2, the flood scenario vector knowledge base is used to store the semantic feature vectors and hydrological parameters of historical flood disaster cases, as well as the mapping relationship between the historical flood cases and the affected grids, forming a historical experience vector database. Constructing the flood scenario vector knowledge base includes: S24. Multimodal fusion encoding is performed on text reports and image data of historical flood cases to generate high-dimensional semantic vectors representing the overall scenario of the cases. The multimodal fusion encoding adopts a Transformer-based two-stream network structure, extracting semantic feature vectors from text reports and visual feature vectors from image data, and then performing feature-level fusion and splicing through an attention mechanism. The pre-training weights of the two-stream network are fine-tuned based on flood-specific corpus and remote sensing image data. S25. Based on the inundation range recorded in historical flood case data or through spatial overlay analysis, determine the fine grid set affected by the case. S26. Establish a reverse index mapping relationship between the semantic vector of each case and the set of fine-grid BeiDou grid codes it affects.

5. The method for intelligent analysis of flood impact driven by multi-source spatiotemporal data according to claim 1, characterized in that, In step S4, the geographic entity dataset is the geographic entity dataset within each affected fine grid. The set of similar historical scenarios is a set of historical flood scenarios that are similar to the current flood scenario in terms of hydrological semantics.

6. The method for intelligent analysis of flood impact driven by multi-source spatiotemporal data according to claim 1, characterized in that, The data-driven model is a flood impact index model, and the historical experience model is a scenario-based experience analysis model. The dual-model collaborative analysis in step S5 specifically includes: S51. Input the geographic entity dataset and the current flood inundation feature data into the flood impact index model to calculate the preliminary flood impact index (I_p). S52. Input the set of similar historical scenario information into the scenario experience analysis model to calculate the experience benchmark index (I_b). S53. Input (I_p) and (I_b) into the dual-model collaborative calibration module, calculate the consistency measure (C) between the preliminary flood impact index (I_p) and the empirical benchmark index (I_b) and the average semantic similarity (S) between the current scenario and the historical scenario, determine the fusion weight based on C and S, perform dynamic weighted fusion of (I_p) and (I_b), and finally output the calibrated flood impact index (I_c).

7. The method for intelligent analysis of flood impact driven by multi-source spatiotemporal data according to claim 6, characterized in that, The flood impact index model is a supervised machine learning regression model based on the LightGBM gradient boosting decision tree framework. It uses the actual impact data I_a verified after the disaster for each fine grid in historical flood events as training labels. The actual impact data I_a is the actual impact data after the disaster mentioned in step S7. The scenario experience analysis model is a case-based reasoning model, which is configured to obtain the historical experience benchmark index (I_b) in the following way: the actual impact data corresponding to each historical scenario in the similar historical scenario information set is weighted and aggregated.

8. The method for intelligent analysis of flood impact driven by multi-source spatiotemporal data according to claim 7, characterized in that, When the scenario experience analysis model performs weighted aggregation, it only includes historical scenarios whose semantic similarity to the current flood scenario is greater than a preset minimum threshold, and the weight is positively correlated with the semantic similarity of the corresponding historical scenario; the preset minimum threshold is determined by statistical analysis of historical flood case datasets, and the value range is [0.4, 0.7].

9. The method for intelligent analysis of flood impact driven by multi-source spatiotemporal data according to claim 8, characterized in that, The dual-model collaborative calibration module performs adaptive dynamic weighted fusion in the following manner: Based on the set of similar historical scenarios obtained in step S4, which have a semantic similarity greater than a preset minimum threshold with the current flood scenario, the number K of qualified similar historical scenarios is determined. If K is less than the preset threshold, I_c=I_p is directly output, and the subsequent fusion calibration process in this step is terminated. The preset threshold is determined by statistical analysis of the historical flood case dataset, and the value range is [3,10]. It can be dynamically adjusted according to the richness of historical cases in the region. Otherwise, calculate the average semantic similarity S between the current flood scenario and the aforementioned qualified similar historical scenarios. The calculation formula is: S = Where K is the number of qualified similar historical scenarios mentioned above. Let be the semantic similarity between the i-th qualified similar historical scenario and the current scenario; The consistency measure (C) between the preliminary flood impact index (I_p) and the historical experience benchmark index (I_b) is calculated using the following formula: When max(I_p,I_b)=0, C=1 is defined. At this time, it is determined that the current fine mesh is not affected by flood, and I_c=0 is directly output. Calculate the current confidence factor α, α=w1×C+w2×S, where w1 and w2 are preset weight coefficients, with default values ​​of w1=0.5 and w2=0.5, which can be dynamically adjusted according to regional flood characteristics and the richness of historical cases. The values ​​range are [0.2,0.8] and w1+w2=1. A dynamic confidence threshold T is set, which is adaptively adjusted based on the complexity of geographic entities within the current fine grid. The quantification formula for the complexity F is as follows: ,in The number of geographic entity types within a fine-grained grid. F represents the total number of geographic entities within the fine grid, and is normalized to the range [0,1]. The adjustment rule is T=T0×(1+μ×F), where T0 is the basic threshold with a value range of [0.5,0.8], which can be dynamically adjusted according to the flood warning level and emergency response level. μ is the adjustment coefficient, used to control the impact of geographic entity complexity on the threshold, with a value range of [-0.3,0.2]. When the importance of the protected objects in the area is high, a positive value is taken to increase the threshold and improve the fusion credibility requirements. When the richness of historical cases is low, a negative value is taken to decrease the threshold to adapt to scenarios with insufficient historical experience. Based on C and S, as well as the confidence factor α and the dynamic confidence threshold T, the calibrated flood impact index (I_c) is calculated: If α is greater than or equal to T, and C+S≠0, then according to the formula... Calculate; if α is less than T, or C+S=0, then it is determined that the current flood scenario is significantly different from the historical pattern, and the current fusion result has low reliability. In this case, I_c is set to be equal to I_p.

10. The method for intelligent analysis of flood impact driven by multi-source spatiotemporal data according to claim 1, characterized in that, In step S7, the joint loss function (L_t) is defined as: ; Where I_a represents grid-level post-disaster actual flood impact data, To calibrate the output error, n is the total number of geographic entities in the current fine grid, and g is the g-th geographic entity in this fine grid. This is the association weight factor for the entity in the current fine mesh, which is generated based on the entity type and its geometric proportion within the current fine mesh. and These are the entity-level preliminary impact index and the entity-level historical baseline index, calculated for the entity using the flood impact index model and the scenario experience analysis model, respectively. The actual impact data of the corresponding geographic entities are obtained through post-disaster field verification and remote sensing image interpretation; β is a balance coefficient used to adjust the weight between the overall grid-level error and the consistency control of entity granularity deviation, and its value is determined by cross-validation on historical flood case datasets within the interval [0.1, 1.0]; the joint optimization process aims to minimize L_t.

11. The method for intelligent analysis of flood impact driven by multi-source spatiotemporal data according to claim 1, characterized in that, In step S6, the process of generating the flood impact analysis results includes: firstly, screening high-risk areas based on the calibrated flood impact index I_c (fine grids with I_c ≥ a preset risk threshold are high-risk areas, the preset risk threshold is determined through statistical analysis of historical flood cases, and this threshold can be dynamically adjusted according to the regional flood control level and the importance of the protected objects); then generating analysis results with the set of fine grid BeiDou grid codes constituting the high-risk areas as the core spatial identifier; the analysis results include at least one of the following: (a) The scope of the high-risk area identified by the Beidou grid code set; (b) Key geographic entity information associated with the high-risk area; (c) Visual representation based on the flood impact index I_c; (d) Statistical analysis data and risk assessment conclusions based on the flood impact index I_c; (e) Emergency response recommendations based on information on high-risk areas and key geographic entities.

12. A multi-source spatiotemporal data-driven intelligent flood impact analysis system, applied to the multi-source spatiotemporal data-driven intelligent flood impact analysis method according to any one of claims 1-11, characterized in that, A multi-source spatiotemporal data-driven intelligent flood impact analysis system includes: The BeiDou grid code spatial service engine provides two-level spatial grid division, encoding calculation, and hierarchical mapping services based on BeiDou grid codes, supporting rapid encoding conversion, spatial relationship analysis, and grid mapping retrieval for different grid levels. The geographic entity map database management module is used to perform the access of multi-source geospatial data, organize it into blocks according to coarse grids, segment geographic entities and dynamically weight the attributes based on fine grids, and build and maintain the map data model. The flood scenario vector knowledge management module is used to perform multimodal fusion encoding of historical flood cases, high-dimensional semantic vector storage, and retrieval based on vector similarity and grid code reverse indexing, so as to realize the construction and maintenance of the flood scenario vector knowledge base; the pre-trained weights of the dual-stream network for multimodal fusion encoding are fine-tuned based on flood-specific corpus and remote sensing image data; A dual-model collaborative analysis engine is used to concurrently drive parallel queries on the geographic entity map database and the flood scenario vector knowledge base, using the BeiDou grid code of the affected fine grid as a unified search key. It calls the data-driven model and the historical experience model to calculate I_p and I_b respectively, and uses its internal dual-model collaborative calibration module to adaptively and dynamically weight and fuse I_p and I_b, outputting the calibrated flood impact index I_c. The data-driven model is the flood impact index model, and the historical experience model is the scenario experience analysis model. The analysis results generation module is used to generate analysis results with a fine-grid Beidou grid code set as the core identifier based on the calibrated flood impact index I_c; The model and knowledge base joint optimization module is used to drive the joint optimization of the data-driven model, historical experience model and dual-model collaborative calibration module through the joint loss function L_t. It simultaneously realizes the dynamic update of geographic entity data in the geographic entity map database and the update of semantic vector and inverse index mapping in the flood scenario vector knowledge base. The joint loss function is configured to introduce entity-level error terms and their associated weight factors for collaborative optimization on the basis of grid-level error.

Images