A multi-modal disaster spatiotemporal graph construction method, system, device and storage medium

Abstract

The application discloses a multi-modal disaster space-time graph construction method, system, device and storage medium, and relates to the technical field of intelligent water conservancy. The method comprises the following steps: a physical evolution disaster chain field ontology is constructed by relying on a basin water system topology; unstructured disaster text is subjected to zero-sample element extraction, absolute geographic coordinates and high-frequency time stamps are bound as space-time attribute, and micro-dynamic disaster point elements are generated; water body range vectors are extracted based on radar remote sensing images and are vectorized as macro-physical envelope surface elements; under a unified geographic information system, a space-time tolerance alignment mechanism of macro-surface constraint and micro-point interpolation is adopted, a spatial physical buffer tolerance is set, a cross-modal physical trust edge is established when the space-time constraint is met, and finally, a graph is constructed and cross-scale deduction is performed. The application solves the problems that the existing disaster graph is separated from the physical space mapping, the traditional extraction model is prone to chain breakage, and the multi-modal data space-time matching is difficult, and improves the knowledge extraction accuracy.

Description

Technical Field

[0001] This invention relates to the field of smart water conservancy technology, and in particular to a method, system, device and storage medium for constructing a multimodal disaster spatiotemporal map. Background Technology

[0002] Rainstorms and floods, and the resulting basin-wide major floods, are characterized by rapid onset, complex causative factors, and rapid hydrodynamic evolution. With the deepening of "smart water conservancy" and the construction of digital twin river basins, the core challenge in improving intelligent decision-making capabilities for flood prevention and disaster reduction lies in accurately extracting disaster evolution chains and combining them with spatiotemporal reasoning based on the underlying physical topology of the river basin, in the face of massive amounts of hydrological monitoring data, remote sensing images, and unstructured disaster information.

[0003] In recent years, knowledge graphs, as a cognitive intelligence technology that reveals the underlying semantic relationships of massive heterogeneous data, have been widely explored in the field of natural disasters. However, when faced with the complex dynamic evolution of flood disasters at the watershed scale (such as "upstream mountain runoff generation and confluence—midstream plain confluence and backwater—downstream depression and flood storage"), existing knowledge graph construction paradigms and multimodal fusion strategies still have two significant limitations and technical bottlenecks: First, the entities in disaster maps are detached from absolute physical space, leading to performance bottlenecks in traditional extraction algorithms. Existing automated extraction of disaster information largely relies on traditional supervised sequence labeling models (such as BiLSTM-CRF). These models not only face the constraint of massive corpus labeling costs but are also prone to entity boundary truncation when processing complex causal long texts with highly colloquial language and nested elements. More critically, traditional methods often extract "suspended semantic networks" detached from the real geographical topology of water systems, lacking mapping between underlying absolute physical coordinates and high-frequency timestamps. This makes it difficult for disaster maps to support meaningful hydrological inferences and source tracing within a real Geographic Information System (GIS) framework. Existing disaster maps often use subjective time slices for modeling, making it difficult to accurately depict the real hydrodynamic cascade transmission process.

[0004] Second, multimodal data suffers from a spatiotemporal scale matching paradox, making in-depth alignment of images and text challenging. The evolution of flood disasters exhibits high-frequency, non-steady-state characteristics of "sharp rises and falls." Existing multimodal alignment research (such as using contrastive learning methods for pure semantic alignment) often attempts to simply "hard-stitch" together the semantic similarity of optical or radar satellite remote sensing images and social media texts on an absolute time scale. However, limited by long satellite revisit cycles and cloud cover during extreme weather, remote sensing images often present as low-frequency, lagging, macroscopic discrete areal data; while disaster texts with timestamps represent high-frequency, dynamic microscopic areal feedback. Forcing alignment while ignoring this objective constraint of physical spatiotemporal scale easily leads to spurious associations, making it difficult to achieve credible verification of heterogeneous data in the dynamic evolution of real disasters. Summary of the Invention

[0005] The purpose of this invention is to provide a method, system, device and storage medium for constructing a multimodal disaster spatiotemporal map, aiming to solve or improve at least one of the above-mentioned technical problems.

[0006] To achieve the above objectives, the present invention provides the following solution: A method for constructing a multimodal disaster spatiotemporal map includes: Acquire basic geographic water system topology data of the target watershed, unstructured disaster information text data, and radar remote sensing image data for the time period matching the unstructured disaster information text data; the unstructured disaster information text data is data with original geographic location and timestamp attached; Based on the aforementioned basic geographic water system topology data, a disaster chain domain ontology model is constructed with hydrodynamic physical evolution as the logical main line. Using the disaster chain domain ontology model as a semantic constraint framework, zero-sample element extraction is performed on the unstructured disaster text data, converting the original geographic location into absolute geographic coordinates. The absolute geographic coordinates and the corresponding timestamp are then bound as spatiotemporal meta-attributes to the edge relations of the disaster triple, generating micro-dynamic disaster point elements with absolute spatiotemporal coordinates. The maximum inundation water body range is extracted based on the radar remote sensing image data, and the maximum inundation water body range is vectorized into macroscopic physical envelope surface elements under a unified geographic information system reference system. Under the unified geographic information system reference system, the macroscopic physical envelope surface features are used as spatial constraint benchmarks, and the microscopic dynamic disaster point features are used as verification objects. A spatiotemporal tolerance alignment mechanism of macroscopic surface constraints and microscopic point interpolation is adopted to set a preset distance of spatial physical buffer tolerance for the macroscopic physical envelope surface features. The spatial topological relationship between the microscopic dynamic disaster point features and the macroscopic physical envelope surface features is calculated. When the preset spatiotemporal constraint conditions are met, cross-modal physical trust edges are established to realize the spatiotemporal alignment and fusion of multi-source heterogeneous data. The micro-dynamic disaster point elements, macro-physical envelope elements, and cross-modal physical trust edges that have completed spatiotemporal alignment are jointly stored in a graph database to construct a multimodal spatiotemporal knowledge graph. Based on spatial topology and temporal path queries, cross-scale disaster chain evolution inference is performed.

[0007] Optionally, the disaster chain domain ontology model includes: disaster-inducing environment, disaster-causing factor, disaster-bearing body, and disaster damage and emergency response. Entities in each cluster are constrained in the physical causal chain through preset directed relation edges and are anchored to the underlying absolute geographic space.

[0008] Optionally, the generation process of the micro-dynamic disaster point elements specifically includes: Inject system-level prompt words with disaster physical evolution logic into a large language model, constrain the large language model to parse the unstructured disaster text data according to the physical causal chain of the disaster chain domain ontology model, and output the disaster triple in a standard format; The original geographical location is subjected to coordinate system correction processing and converted into latitude and longitude coordinates of the standard geographic coordinate system to obtain the absolute geographic coordinates; The absolute geographic coordinates, the corresponding timestamps, and the topological branches of the main stream and tributaries of the watershed are used as spatiotemporal element attributes and injected into the edge relationships of the disaster triplet to generate micro-dynamic disaster point elements.

[0009] Optionally, establishing cross-modal physical trust edges when preset spatiotemporal constraints are met specifically includes: Determine whether the micro-dynamic disaster point element meets the first set condition in spatial coordinates, and whether the timestamp of the micro-dynamic disaster point element meets the second set condition; the first set condition is that it falls inside the macro-physical envelope surface element or within the range of the spatial physical buffer tolerance; the second set condition is that it is located within the set disaster receding period or evolution period time window; If the first and second conditions are met, the cross-modal physical trust edge connecting the micro-dynamic disaster point elements and the macro-physical envelope surface elements will be automatically generated at the bottom layer of the graph database. If any of the set conditions are not met, the micro-dynamic disaster point element is determined to be outlier data and is rejected from being injected into the multimodal spatiotemporal knowledge graph.

[0010] Optionally, the disaster chain evolution simulation includes macro-basin disaster-causing simulation and micro-local emergency simulation; The macro-basin disaster induction is as follows: performing a time-series path query in the multimodal spatiotemporal knowledge graph, reconstructing and outputting the basin-wide cascade disaster trajectory that starts from meteorological rainfall, crosses space to cause runoff in mountainous areas, converges and backs up in plains, and finally leads to flood storage and detention in downstream depressions; The micro-level local emergency simulation is as follows: based on the multimodal spatiotemporal knowledge graph, the spatial reference is located to the set core defense area or flood storage area, the flood control emergency response actions and disaster-bearing body response events under spatiotemporal constraints are extracted and connected, and the cross-regional emergency resource allocation trajectory with embedded high-frequency timestamps and absolute geographical coordinates is dynamically associated.

[0011] This invention also provides a multimodal disaster spatiotemporal mapping system, comprising: The data acquisition module is used to acquire basic geographic water system topology data of the target watershed, unstructured disaster text data, and radar remote sensing image data of the time period matching the unstructured disaster text data; the unstructured disaster text data is data with original geographic location and timestamp attached. The joint extraction module is used to construct a disaster chain domain ontology model based on the basic geographic water system topology data, with the hydrodynamic physical evolution process as the logical main line, and to extract zero-sample elements from the unstructured disaster text data using the disaster chain domain ontology model as a semantic constraint framework. The original geographic location is converted into absolute geographic coordinates, and the absolute geographic coordinates and the corresponding timestamp are bound to the edge relationship of the disaster triple as spatiotemporal meta-attributes to generate micro-dynamic disaster point elements with absolute spatiotemporal coordinates. The macroscopic surface extraction module is used to extract the maximum inundated water body range based on the radar remote sensing image data, and vectorize the maximum inundated water body range into macroscopic physical envelope surface elements under a unified geographic information system reference system. The spatiotemporal alignment and mutual verification module is used to, under the unified geographic information system reference system, take the macroscopic physical envelope surface elements as the spatial constraint benchmark and the microscopic dynamic disaster point elements as the verification objects, and adopt a spatiotemporal tolerance alignment mechanism of macroscopic surface constraints and microscopic point interpolation to set a preset distance of spatial physical buffer tolerance for the macroscopic physical envelope surface elements, and calculate the spatial topological relationship between the microscopic dynamic disaster point elements and the macroscopic physical envelope surface elements. When the preset spatiotemporal constraint conditions are met, a cross-modal physical trust edge is established to realize the spatiotemporal alignment and fusion of multi-source heterogeneous data. The graph inference module is used to store the spatiotemporally aligned micro-dynamic disaster point elements, macro-physical envelope elements, and cross-modal physical trust edges into the graph database, construct a multimodal spatiotemporal knowledge graph, and perform cross-scale disaster chain evolution inference based on spatial topology and temporal path queries.

[0012] The present invention also provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the multimodal disaster spatiotemporal mapping construction method described above.

[0013] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the multimodal disaster spatiotemporal mapping construction method as described above.

[0014] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: This invention discloses a method, system, device, and storage medium for constructing a multimodal disaster spatiotemporal map. The method includes: constructing a physical evolution disaster chain ontology based on the topology of a watershed; extracting zero-sample elements from unstructured disaster text, binding absolute geographic coordinates and high-frequency timestamps as spatiotemporal meta-attributes to generate micro-dynamic disaster point elements; extracting water body range vectorization from radar remote sensing images into macro-physical envelope surface elements; and, under a unified geographic information system, employing a spatiotemporal tolerance alignment mechanism of macro-surface constraints and micro-point interpolation, setting spatial physical buffer tolerance, and establishing cross-modal physical trust edges while satisfying spatiotemporal constraints, ultimately constructing the map and performing cross-scale extrapolation. This invention solves the problems of existing disaster maps being detached from physical spatial mapping, traditional extraction models being prone to chain breakage, and difficulties in spatiotemporal matching of multimodal data, and improves the accuracy of knowledge extraction, providing underlying knowledge support with physical constraints for flood prevention and disaster reduction. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating the technical route of the multimodal disaster spatiotemporal map construction method in this embodiment; Figure 2 This is a schematic diagram of the "macroscopic surface constraint-micro viewpoint interpolation" multimodal spatiotemporal tolerance alignment mechanism and spatial geometry determination in this embodiment; Figure 3 This is a structural block diagram of the multimodal disaster spatiotemporal mapping system in this embodiment; Figure 4 This is a hardware structure block diagram of an electronic device provided in this embodiment. Detailed Implementation

[0017] 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.

[0018] The purpose of this invention is to provide a method, system, device and storage medium for constructing a multimodal disaster spatiotemporal map, aiming to solve or improve at least one of the above-mentioned technical problems.

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] In a first aspect, the present invention provides a method for constructing a multimodal disaster spatiotemporal map, comprising: Step S1: Acquire basic geographic water system topology data of the target watershed, unstructured disaster information text data, and radar remote sensing image data for the time period matching the unstructured disaster information text data; the unstructured disaster information text data is data with original geographic location and timestamp. The radar remote sensing image data includes synthetic aperture radar (SAR) image data.

[0021] Step S2: Based on the basic geographic water system topology data, construct a disaster chain domain ontology model with hydrodynamic physical evolution as the logical main line. Using the disaster chain domain ontology model as a semantic constraint framework, perform zero-sample feature extraction on the unstructured disaster text data, convert the original geographic location into absolute geographic coordinates, and bind the absolute geographic coordinates and corresponding timestamps as spatiotemporal meta-attributes to the edge relations of the disaster triple, generating micro-dynamic disaster point elements with absolute spatiotemporal coordinates. The disaster chain domain ontology model includes: disaster-inducing environment, disaster-causing factors, disaster-bearing bodies, and disaster damage and emergency response. Entities in each cluster are constrained in the physical causal chain through preset directed relation edges and anchored to the underlying absolute geographic space.

[0022] The generation process of the micro-dynamic disaster point elements specifically includes: System-level prompts with disaster physical evolution logic are injected into a large language model, constraining the model to parse the unstructured disaster text data according to the physical causal chain of the disaster chain domain ontology model, and outputting the disaster triple in a standard format. The original geographical location is subjected to coordinate system correction processing and converted to latitude and longitude coordinates in a standard geographic coordinate system to obtain the absolute geographic coordinates. The absolute geographic coordinates, the corresponding timestamp, and the topological branches of the main stream and tributaries of the watershed are used as spatiotemporal meta-attributes and injected into the edge relations of the disaster triple to generate micro-dynamic disaster point elements.

[0023] Step S3: Extract the maximum inundation water body range based on the radar remote sensing image data, and vectorize the maximum inundation water body range into macroscopic physical envelope surface elements under a unified geographic information system reference system.

[0024] Step S4: Under the unified geographic information system reference system, using the macroscopic physical envelope surface features as spatial constraint benchmarks and the microscopic dynamic disaster point features as verification objects, a spatiotemporal tolerance alignment mechanism of macroscopic surface constraints and microscopic point interpolation is adopted to set a preset distance of spatial physical buffer tolerance for the macroscopic physical envelope surface features, and to calculate the spatial topological relationship between the microscopic dynamic disaster point features and the macroscopic physical envelope surface features. When the preset spatiotemporal constraint conditions are met, cross-modal physical trust edges are established to realize the spatiotemporal alignment and fusion of multi-source heterogeneous data.

[0025] The spatial physical buffer tolerance of the preset distance is set to 500 meters. The spatial physical buffer tolerance is used to accommodate the mixed pixel effect error of the radar remote sensing image data, the mobile terminal positioning drift error when acquiring text data, and the actual physical behavior and movement trajectory of the disaster-bearing body when it is covered to avoid harm and seek benefits during a disaster.

[0026] Furthermore, the process of establishing the aforementioned cross-modal physical trust edge specifically includes: Determine whether the micro-dynamic disaster point element meets the first set condition in spatial coordinates, and whether the timestamp of the micro-dynamic disaster point element meets the second set condition; the first set condition is that it falls inside the macro-physical envelope surface element or within the range of the spatial physical buffer tolerance; the second set condition is that it is located within the set disaster receding period or evolution period time window.

[0027] If the first and second conditions are met, a cross-modal physical trust edge connecting the micro-dynamic disaster point elements and the macro-physical envelope surface elements is automatically generated at the bottom layer of the graph database; if neither condition is met, the micro-dynamic disaster point elements are determined to be outlier data and are rejected from being injected into the multimodal spatiotemporal knowledge graph.

[0028] Step S5: Store the spatiotemporally aligned micro-dynamic disaster point elements, macro-physical envelope elements, and cross-modal physical trust edges into a graph database to construct a multimodal spatiotemporal knowledge graph, and perform cross-scale disaster chain evolution inference based on spatial topology and temporal path queries.

[0029] The disaster chain evolution simulation includes macro-basin disaster simulation and micro-local emergency simulation.

[0030] The macro-basin disaster simulation involves performing a temporal path query in the multimodal spatiotemporal knowledge graph to reconstruct and output the cascading disaster trajectory of the basin, starting from meteorological rainfall, traversing space to cause runoff generation in mountainous areas, backwater effect at the confluence of plains, and flood storage and detention in downstream depressions. The micro-local emergency simulation involves locating the spatial benchmark to the designated core defense area or flood storage area based on the multimodal spatiotemporal knowledge graph, extracting and connecting flood control emergency response actions and disaster-bearing body response events under spatiotemporal constraints, and dynamically associating cross-regional emergency resource allocation trajectories with embedded high-frequency timestamps and absolute geographic coordinates.

[0031] Based on the above technical solution, the following is provided: Figures 1-2 The specific embodiments shown are as follows.

[0032] like Figure 1 As shown, this embodiment provides a method for constructing a multimodal disaster spatiotemporal map that integrates physical evolution constraints and a large-scale model. Taking a severe rainstorm and flood disaster in a certain watershed as an empirical example, the method specifically includes the following core steps: S1, Constructing a disaster chain domain ontology based on the laws of physical evolution: The evolution of flood disasters is not an isolated textual semantic event, but rather a cascading response strictly controlled by the underlying hydrogeographic environment of the watershed. This embodiment relies on real geographic hydrological data of the target watershed (such as DEM elevation, water system vectors of main and tributary flood channels, and distribution of key flood storage and detention areas) provided by the National Geographic Information Public Service Platform to construct a domain ontology model with "disaster chain" as the core logic.

[0033] This ontology model comprises four core class clusters, which are anchored to the underlying absolute geospatial space: (1) Disaster-prone environment: the basic hydrogeographical background of the disaster-stricken area (such as windward slope topography, specific main and tributary water systems, etc.).

[0034] (2) Disaster-causing factors: Dynamic elements that drive the evolution of disasters (such as typhoon factors, extreme heavy rainfall, rapid rise in water level, overflow and breach, flood backwater, etc.).

[0035] (3) Disaster-bearing entities: specific physical entities threatened by floods (subdivided into infrastructure such as dams and dikes, settlements, transportation networks, etc.).

[0036] (4) Disaster damage and emergency response: cascading damage caused by disasters and human intervention actions (such as flooding and power outages, house destruction, relocation and resettlement of people, opening of floodgates for flood discharge, adjustment of flood control emergency response level, etc.).

[0037] By defining directed relation edges such as "trigger", "cause", "flood", "threat", and "start", the above entities are strictly constrained in the physical causal chain of "disaster-inducing environment → disaster-causing factor → disaster-bearing body → disaster damage and emergency response", providing a rigorous prior expert knowledge framework for the targeted extraction of large language models.

[0038] S2, Spatiotemporal attribute joint extraction based on a large language model: Traditional supervised named entity recognition and relation extraction models are prone to boundary truncation when dealing with complex, colloquial, and nested long texts on social media. This embodiment introduces the zero-shot deep reasoning capability of Large Language Model (LLM).

[0039] First, unstructured disaster information text data with original point of interest (POI) and timestamps within a specific time period is crawled using web scraping and other technologies. System-level expert hints are injected into the LLM (Local Management Module) to constrain it to discard simple listings of literal meanings and parse the text according to the disaster physical causal chain established in step S1, outputting standard JSON format triples.

[0040] Furthermore, for the parsed abstract entity relationships, coordinate system transformation and correction are performed (e.g., converting the GCJ-02 coordinates obtained from the mobile device to WGS84 standard latitude and longitude coordinates) to obtain absolute geographic coordinates. These absolute geographic coordinates, high-frequency dynamic time, and the corresponding watershed topology branch are then injected as "spatiotemporal meta-attributes" into the edge relationships of this triple, generating micro-dynamic disaster point elements with absolute spatiotemporal coordinates.

[0041] Through stratified sampling cross-ablation experiments, the LLM-based extraction method in this embodiment significantly improved the F1 score of complex disaster long text extraction from 81.2% to 92.6% compared to the traditional BiLSTM-CRF model. Thousands of multi-hop disaster chain triples with spatiotemporal attributes were successfully extracted and stored in the database across the entire basin.

[0042] S3, Macroscopic physical envelope surface extraction: Watershed-level floods are characterized by "sharp rises and falls," while satellite remote sensing imagery often presents low-frequency, lagging, macroscopic discrete data. This embodiment utilizes the spatial demarcation advantages of visual imagery for dimensionality reduction. During critical periods of flood evolution (such as the passage of the flood peak and the receding period), based on synthetic aperture radar imagery (such as Sentinel-1 SAR) capable of penetrating extreme weather cloud layers, the true maximum inundation area is extracted. This area is vectorized into polygon features in the GIS system, i.e., macroscopic physical envelope polygon features, serving as a macroscopic objective verification base for map multimodal fusion.

[0043] S4, Multimodal Spatiotemporal Tolerance Alignment and Mutual Verification: like Figure 2As shown, in response to the spatiotemporal matching paradox between low-frequency lagging remote sensing surface features and high-frequency micro-disaster text point features, this embodiment innovatively proposes a spatiotemporal tolerance alignment method of "macro-surface constraint + micro-point interpolation" under a unified GIS reference system.

[0044] This embodiment sets a spatial physical buffer tolerance of 500 meters outward from the macroscopic physical envelope element. This physical tolerance setting scientifically accommodates the following multi-dimensional objective realities: (1) Boundary blurring error caused by the mixed pixel effect of radar remote sensing images themselves.

[0045] (2) GPS positioning drift error when mobile communication devices acquire text disaster data.

[0046] (3) Most importantly, the actual physical behavior trajectory of the disaster-bearing body (disaster victims and rescuers) when the disaster occurs (the location of the posting for help or checking in has mostly retreated to the high ground or dike at the edge of the water that has not been submerged, rather than the center of the submerged deep water area).

[0047] The specific judgment rules for performing spatiotemporal tolerance alignment and mutual verification are as follows: Determine whether the absolute geographic coordinates of the micro-dynamic disaster point elements fall precisely within the remote sensing macro-physical envelope elements, or within their 500-meter tolerance buffer; and determine whether the high-frequency timestamp of the text node is within the set flood evolution or receding period time window.

[0048] When the above spatial and temporal constraints are simultaneously met, the underlying layer of the knowledge graph automatically generates a cross-modal physical trust edge of: [textual micro-disaster node] — <spatial verification> — [remote sensing physical envelope element]. If the conditions are not met, it is judged as long-distance outlier data or invalid location noise, and is rejected from being injected into the knowledge graph. Quantitative experiments in a typical flood storage area verification section show that the micro-viewpoint topological mutual verification matching rate reaches 88.5%, successfully eliminating spurious association noise.

[0049] S5, Spatiotemporal Mapping and Cross-Scale Extrapolation: The extracted and spatially verified micro-viewpoint elements, macro-level elements, and cross-modal trust edges are stored in graph databases such as Neo4j to construct a physically constrained multimodal spatiotemporal knowledge graph. Cypher spatial topology and temporal path queries are then performed in the graph database, supporting dual-scale, multi-hop tracing and deduction. (1) Macro-basin disaster-causing dimension extrapolation: The map clearly reconstructs the cascading disaster-causing trajectory of hydrodynamics. For example, starting with [typhoon (disaster-causing factor)], it triggers [extremely heavy rain], which is forced by specific [windward slope topography] to cause [sharp rise in river level], ultimately leading to [large-scale flooding and damage] downstream. At the logical level, the underlying dynamic transmission path of "meteorological rainfall - runoff generation in mountainous areas - confluence and backwater in plains - flood discharge in depressions" is reproduced.

[0050] (2) Micro-level local emergency simulation: The map shifts the spatial benchmark down to the core defense area downstream of the flood (such as specific flood storage and detention areas). It accurately captures and connects dense flood control emergency and disaster-bearing body response actions: [the city flood control headquarters raising the emergency response level] leads to [regional emergency relocation and resettlement], and dynamically links multiple parallel events such as [cross-provincial and municipal allocation of flood control materials for assistance] and [multi-department participation in dike patrol and hazard elimination]. The high-frequency timestamps and absolute coordinates embedded in each evolutionary edge in the map provide fine-grained underlying knowledge support for flood control departments to intuitively analyze emergency dispatch trajectories and optimize flood control plans.

[0051] In summary, this method achieves the following beneficial effects: (1) A basin disaster chain ontology and spatiotemporal map base with real physical constraints were reconstructed. This invention abandons the subjective stage division mode of traditional maps that lack hydrodynamic constraints. Based on the real river network topology of the target basin, a disaster evolution ontology with "disaster-inducing environment - disaster-causing factors - disaster-bearing body - disaster damage and emergency response" as the underlying logic is constructed. Not only are the causal relationships of disasters clarified, but the underlying dynamic transmission of "meteorological rainfall - mountain runoff - plain confluence and backwater - depression storage and flood discharge" is also reproduced at the logical level. This makes the constructed map no longer a "suspended semantic network", but a spatiotemporal dynamic deduction sand table that conforms to the constraints of the physical laws of the basin.

[0052] (2) The large language model breaks through the performance bottleneck of complex colloquial text extraction, completely eliminating the cost of massive corpus annotation. For complex disaster-related long texts with severe colloquialisms and nested elements, the zero-shot instruction reasoning capability of the large model (LLM) replaces the traditional supervised sequence labeling model (such as BiLSTM-CRF). Comparative experiments show that compared with the F1 score of approximately 81.2% of the traditional BiLSTM-CRF model, the knowledge extraction F1 score of this method jumps to 92.6%. At the same time, it innovatively injects absolute latitude and longitude and high-frequency timestamps as meta-attributes into the triplet association edges, giving the abstract semantic association an absolute physical space mapping.

[0053] (3) The invention pioneered a multimodal spatiotemporal tolerance alignment mechanism of "macroscopic surface constraint + microscopic viewpoint interpolation", which fundamentally resolves the paradox of spatiotemporal scale matching. Addressing the pain point that low-frequency satellite remote sensing area images and high-frequency microscopic disaster text point data are difficult to synchronize on an absolute time scale, the invention clarifies the superior and inferior division of labor among heterogeneous data. Using the maximum submerged water body in the reduced-dimensional radar image as the "macroscopic static physical envelope", GIS spatial topology interpolation is performed on the "high-frequency dynamic microscopic text nodes" extracted by LLM. In particular, this invention innovatively introduces spatial physical buffer tolerance (e.g., preferably 500 meters), replacing the "hard splicing" of traditional pure text semantic vectors with objective topological overlap at the geographic spatial geometric scale, perfectly accommodating radar mixed pixel effects, terminal GPS positioning drift, and the objective movement trajectory of humans seeking advantage and avoiding harm during disasters (e.g., retreating to high ground at the edge of the water to check in). Quantitative verification confirms that after setting the tolerance, the microscopic viewpoint surface topology mutual verification matching rate reaches as high as 88.5%, achieving high-reliability fusion of multimodal heterogeneous data on a unified platform.

[0054] As a second aspect, the present invention also provides, for example Figure 3 The system shown is a multimodal disaster spatiotemporal mapping system, which is constructed based on the method described in Example 1, and includes: The data acquisition module is used to acquire basic geographic water system topology data of the target watershed, unstructured disaster text data, and radar remote sensing image data of the time period matching the unstructured disaster text data; the unstructured disaster text data is data with original geographic location and timestamp.

[0055] The joint extraction module is used to construct a disaster chain domain ontology model based on the basic geographic water system topology data, with the hydrodynamic physical evolution process as the logical main line. Using the disaster chain domain ontology model as a semantic constraint framework, it performs zero-sample feature extraction on the unstructured disaster text data, converts the original geographic location into absolute geographic coordinates, and binds the absolute geographic coordinates and the corresponding timestamp as spatiotemporal meta-attributes to the edge relations of the disaster triple, generating micro-dynamic disaster point elements with absolute spatiotemporal coordinates.

[0056] The macroscopic surface extraction module is used to extract the maximum inundation water body range based on the radar remote sensing image data, and vectorize the maximum inundation water body range into macroscopic physical envelope surface features under a unified geographic information system reference system.

[0057] The spatiotemporal alignment and mutual verification module is used under the unified geographic information system reference system, with the macroscopic physical envelope surface elements as the spatial constraint benchmark and the microscopic dynamic disaster point elements as the verification objects. It adopts a spatiotemporal tolerance alignment mechanism of macroscopic surface constraints and microscopic point interpolation to set a preset distance of spatial physical buffer tolerance for the macroscopic physical envelope surface elements, and calculates the spatial topological relationship between the microscopic dynamic disaster point elements and the macroscopic physical envelope surface elements. When the preset spatiotemporal constraint conditions are met, a cross-modal physical trust edge is established to realize the spatiotemporal alignment and fusion of multi-source heterogeneous data.

[0058] The graph inference module is used to store the spatiotemporally aligned micro-dynamic disaster point elements, macro-physical envelope elements, and cross-modal physical trust edges into the graph database, construct a multimodal spatiotemporal knowledge graph, and perform cross-scale disaster chain evolution inference based on spatial topology and temporal path queries.

[0059] As a third aspect, the present invention also provides, for example Figure 4 An electronic device is shown, comprising a memory and a processor. The memory stores a computer program, and the processor runs the computer program to enable the electronic device to execute the multimodal disaster spatiotemporal mapping construction method described above. This electronic device can be a server node in a smart water conservancy flood control command center, an edge computing industrial control computer, or a cloud computing cluster.

[0060] As a fourth aspect, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the multimodal disaster spatiotemporal mapping construction method described above. The readable storage medium may include various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0061] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0062] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for constructing a multimodal disaster spatiotemporal map, characterized in that, include: Acquire basic geographic water system topology data of the target watershed, unstructured disaster information text data, and radar remote sensing image data for the time period matching the unstructured disaster information text data; the unstructured disaster information text data is data with original geographic location and timestamp attached; Based on the aforementioned basic geographic water system topology data, a disaster chain domain ontology model is constructed with hydrodynamic physical evolution as the logical main line. Using the disaster chain domain ontology model as a semantic constraint framework, zero-sample element extraction is performed on the unstructured disaster text data, converting the original geographic location into absolute geographic coordinates. The absolute geographic coordinates and the corresponding timestamp are then bound as spatiotemporal meta-attributes to the edge relations of the disaster triple, generating micro-dynamic disaster point elements with absolute spatiotemporal coordinates. The maximum inundation water body range is extracted based on the radar remote sensing image data, and the maximum inundation water body range is vectorized into macroscopic physical envelope surface elements under a unified geographic information system reference system. Under the unified geographic information system reference system, the macroscopic physical envelope surface features are used as spatial constraint benchmarks, and the microscopic dynamic disaster point features are used as verification objects. A spatiotemporal tolerance alignment mechanism of macroscopic surface constraints and microscopic point interpolation is adopted to set a preset distance of spatial physical buffer tolerance for the macroscopic physical envelope surface features. The spatial topological relationship between the microscopic dynamic disaster point features and the macroscopic physical envelope surface features is calculated. When the preset spatiotemporal constraint conditions are met, cross-modal physical trust edges are established to realize the spatiotemporal alignment and fusion of multi-source heterogeneous data. The micro-dynamic disaster point elements, macro-physical envelope elements, and cross-modal physical trust edges that have completed spatiotemporal alignment are jointly stored in a graph database to construct a multimodal spatiotemporal knowledge graph. Based on spatial topology and temporal path queries, cross-scale disaster chain evolution inference is performed.

2. The method for constructing a multimodal disaster spatiotemporal map according to claim 1, characterized in that, The disaster chain domain ontology model includes: disaster-inducing environment, disaster-causing factors, disaster-bearing bodies, and disaster damage and emergency response. Entities in each cluster are constrained in the physical causal chain through pre-defined directed relation edges and are anchored to the underlying absolute geographic space.

3. The method for constructing a multimodal disaster spatiotemporal map according to claim 1, characterized in that, The generation process of the micro-dynamic disaster point elements specifically includes: Inject system-level prompt words with disaster physical evolution logic into a large language model, constrain the large language model to parse the unstructured disaster text data according to the physical causal chain of the disaster chain domain ontology model, and output the disaster triple in a standard format; The original geographical location is subjected to coordinate system correction processing and converted into latitude and longitude coordinates of the standard geographic coordinate system to obtain the absolute geographic coordinates; The absolute geographic coordinates, the corresponding timestamps, and the topological branches of the main stream and tributaries of the watershed are used as spatiotemporal element attributes and injected into the edge relationships of the disaster triplet to generate micro-dynamic disaster point elements.

4. The method for constructing a multimodal disaster spatiotemporal map according to claim 1, characterized in that, The establishment of cross-modal physical trust edges when preset spatiotemporal constraints are met specifically includes: Determine whether the micro-dynamic disaster point element meets the first set condition in spatial coordinates, and whether the timestamp of the micro-dynamic disaster point element meets the second set condition; the first set condition is that it falls inside the macro-physical envelope surface element or within the range of the spatial physical buffer tolerance; the second set condition is that it is located within the set disaster receding period or evolution period time window; If the first and second conditions are met, the cross-modal physical trust edge connecting the micro-dynamic disaster point elements and the macro-physical envelope surface elements will be automatically generated at the bottom layer of the graph database. If any of the set conditions are not met, the micro-dynamic disaster point element is determined to be outlier data and is rejected from being injected into the multimodal spatiotemporal knowledge graph.

5. The method for constructing a multimodal disaster spatiotemporal map according to claim 1, characterized in that, The disaster chain evolution simulation includes macro-basin disaster simulation and micro-local emergency simulation; The macro-basin disaster induction is as follows: performing a time-series path query in the multimodal spatiotemporal knowledge graph, reconstructing and outputting the basin-wide cascade disaster trajectory that starts from meteorological rainfall, crosses space to cause runoff in mountainous areas, converges and backs up in plains, and finally leads to flood storage and detention in downstream depressions; The micro-level local emergency simulation is as follows: based on the multimodal spatiotemporal knowledge graph, the spatial reference is located to the set core defense area or flood storage area, the flood control emergency response actions and disaster-bearing body response events under spatiotemporal constraints are extracted and connected, and the cross-regional emergency resource allocation trajectory with embedded high-frequency timestamps and absolute geographical coordinates is dynamically associated.

6. A multimodal disaster spatiotemporal mapping system, characterized in that, include: The data acquisition module is used to acquire basic geographic water system topology data of the target watershed, unstructured disaster text data, and radar remote sensing image data of the time period matching the unstructured disaster text data; the unstructured disaster text data is data with original geographic location and timestamp attached. The joint extraction module is used to construct a disaster chain domain ontology model based on the basic geographic water system topology data, with the hydrodynamic physical evolution process as the logical main line, and to extract zero-sample elements from the unstructured disaster text data using the disaster chain domain ontology model as a semantic constraint framework. The original geographic location is converted into absolute geographic coordinates, and the absolute geographic coordinates and the corresponding timestamp are bound to the edge relationship of the disaster triple as spatiotemporal meta-attributes to generate micro-dynamic disaster point elements with absolute spatiotemporal coordinates. The macroscopic surface extraction module is used to extract the maximum inundated water body range based on the radar remote sensing image data, and vectorize the maximum inundated water body range into macroscopic physical envelope surface elements under a unified geographic information system reference system. The spatiotemporal alignment and mutual verification module is used to, under the unified geographic information system reference system, take the macroscopic physical envelope surface elements as the spatial constraint benchmark and the microscopic dynamic disaster point elements as the verification objects, and adopt a spatiotemporal tolerance alignment mechanism of macroscopic surface constraints and microscopic point interpolation to set a preset distance of spatial physical buffer tolerance for the macroscopic physical envelope surface elements, and calculate the spatial topological relationship between the microscopic dynamic disaster point elements and the macroscopic physical envelope surface elements. When the preset spatiotemporal constraint conditions are met, a cross-modal physical trust edge is established to realize the spatiotemporal alignment and fusion of multi-source heterogeneous data. The graph inference module is used to store the spatiotemporally aligned micro-dynamic disaster point elements, macro-physical envelope elements, and cross-modal physical trust edges into the graph database, construct a multimodal spatiotemporal knowledge graph, and perform cross-scale disaster chain evolution inference based on spatial topology and temporal path queries.

7. An electronic device, characterized in that, The device includes a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to cause the electronic device to perform the multimodal disaster spatiotemporal mapping construction method according to any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, It stores a computer program, which, when executed by a processor, implements the multimodal disaster spatiotemporal mapping construction method as described in any one of claims 1-5.

Images