Methods for Establishing Spatiotemporal Topology of Geographic Elements in Large-Scale Models for Marine Forecasting and Disaster Mitigation

Abstract

This invention provides a method for establishing the spatiotemporal topology of geographic elements for a large-scale marine forecasting and disaster reduction model, belonging to the field of marine observation and forecasting technology. This invention achieves real-time fusion and cleaning of multi-source heterogeneous marine data, utilizes a temperature-salinity-current collaborative interpolation model and graph neural networks to complete sparse data with physical constraints, establishes a spatial index based on quadtree Z-curve encoding, and achieves millisecond-level query response through parallel execution of master and auxiliary filters using a GPU dual-layer sub-thread grid. It employs spatial partitioning and an event-driven mechanism to perform local incremental updates of topological relationships, and introduces Bayesian Skrigin interpolation and Monte Carlo simulation to add confidence labels to topological nodes to quantify uncertainty. Finally, it outputs standardized spatiotemporal topological relationship data, solving the technical problem of the large-scale marine forecasting and disaster reduction model's difficulty in understanding and efficiently constructing the spatiotemporal topological relationships of geographic elements under observation and forecasting data conditions.

Description

Technical Field

[0001] This invention belongs to the field of marine observation and forecasting technology. Specifically, it relates to a method for establishing the spatiotemporal topology of geographic elements for a large-scale marine forecasting and disaster reduction model. Background Technology

[0002] Marine forecasting and disaster mitigation rely on multi-source heterogeneous data, including satellite remote sensing, buoy observations, and numerical weather prediction, to construct spatiotemporal topological relationships of geographic elements. Traditional methods employ unified formatting and spatial indexing techniques for data integration and querying, establishing topological relationships between geographic elements through R-trees or grid indexes, using Kriging interpolation to complete sparse observation area data, and storing the topological network structure based on relational databases or file systems. However, traditional methods suffer from insufficient spatiotemporal registration accuracy when dealing with multi-scale nested features of the marine environment. Batch processing noise reduction based on statistical thresholds cannot respond to dynamic changes in real time. Independent parameter interpolation ignores the physical coupling relationship between temperature, salinity, and current velocity. Global topology reconstruction suffers from excessive computational load and response delays when updating massive amounts of real-time data. The lack of quantitative assessment of interpolation uncertainty affects the reliability of decision-making. In existing technologies, due to the lack of efficient spatiotemporal topology construction and dynamic maintenance mechanisms for sparse observation data, traditional methods struggle to simultaneously guarantee the real-time performance of topology construction and the reliability of interpolation results. In other words, existing technologies face the technical challenge of understanding and efficiently constructing spatiotemporal topological relationships of geographic elements under observational and forecast data conditions in large-scale marine forecasting and disaster mitigation models. Summary of the Invention

[0003] In view of this, the present invention provides a method for establishing the spatiotemporal topology of geographic elements for a large-scale marine forecasting and disaster reduction model, which can solve the technical problem in the prior art that it is difficult to efficiently construct the spatiotemporal topological relationship of geographic elements under the condition of sparse observation data in marine forecasting and disaster reduction models.

[0004] This invention is implemented as follows: It provides a method for establishing the spatiotemporal topology of geographic elements for a large-scale marine forecasting and disaster reduction model. This method includes multi-source heterogeneous data fusion and integration, geographic information transcoding and standardization, spatiotemporal indexing and topological relationship construction, topological relationship reasoning and dynamic process object-oriented modeling, dynamic maintenance and incremental updating of topological relationships, quantification of uncertainty in sparse observation data interpolation, and spatial topological relationship output steps. It employs a vortex sensing alignment algorithm to achieve multi-scale nested dynamic coupling, utilizes LSTM-Autoencoder anomaly detection technology combined with edge computing to achieve real-time streaming cleaning, introduces a temperature-salinity-current collaborative interpolation model and graph neural network to learn nonlinear relationships, activates the maximum number of CUDA streams on the GPU, and within each CUDA stream, activates two layers of sub-threaded meshes to perform the minimum bounding rectangle fast exclusion operation of the main filter and the precise geometric intersection judgment operation of the auxiliary filter, respectively. It uses spatial partitioning technology to divide the topological relationship network into independent sub-regions and combines an event-driven mechanism to perform local reconstruction only on the affected areas. It uses the Bayesian Skrigin interpolation method combined with Monte Carlo simulation to evaluate the interpolation uncertainty range and attach confidence labels.

[0005] In the multi-source heterogeneous data fusion and integration step, spatiotemporal registration adopts adaptive grid reconstruction based on flow field characteristics and vortex sensing alignment algorithm to achieve dynamic coupling of multi-scale nesting.

[0006] In the multi-source heterogeneous data fusion and integration step, the noise reduction process adopts a spatiotemporal continuity test model and LSTM-Autoencoder anomaly detection technology, combined with edge computing to achieve real-time streaming cleaning and feedback learning.

[0007] In the multi-source heterogeneous data fusion and integration step, the interpolation method introduces a four-dimensional variational interpolation with physical constraints and a temperature-salt-flow collaborative interpolation model, and uses a graph neural network to learn nonlinear relationships.

[0008] In the geographic information transcoding and standardization process, the Earth ellipsoid mathematical model adopts the WGS84 coordinate system, and the triangulation method determines the spatial coordinates of the target point by solving a system of nonlinear equations based on at least three known reference points.

[0009] In the geographic information transcoding and standardization process, the spatial database standard follows the OGC SimpleFeatures specification, supporting the storage indexing of point, line, and polygon geometric objects and spatial relationship query operations.

[0010] In the spatiotemporal index and topology relationship construction steps, a quadtree combined with Z-curve encoding is used for spatial dimensionality reduction, and a B-tree index is used to index the time dimension.

[0011] In the spatiotemporal index and topology relationship construction step, the first-layer sub-thread grid consists of multiple thread blocks, and each thread block contains multiple threads for parallel execution of the minimum bounding rectangle fast exclusion operation of the main filter.

[0012] In the spatiotemporal index and topology relationship construction step, the second-layer sub-thread grid starts after the first-layer sub-thread grid outputs the candidate set and performs the precise geometric intersection judgment operation of the auxiliary filter.

[0013] In the topological relationship reasoning and dynamic process objectification modeling steps, based on the object-oriented data model, storm surge overflow is abstracted into dynamic line objects and a topological adjacency relationship is established with the surrounding coastline.

[0014] In the dynamic maintenance and incremental update steps of topology relationships, the event-driven mechanism monitors changes in observed data and automatically starts the processing flow when the triggering conditions are met.

[0015] In the dynamic maintenance and incremental update steps of topology relationships, the message queue uses Kafka or RabbitMQ to implement the producer-consumer pattern, and the distributed cache uses Redis or Memcached to store frequently accessed data in memory.

[0016] In the uncertainty quantification step of sparse observation data interpolation, the variogram describes the variation of the difference between observations between two points in space as the distance increases. Common forms include spherical models, exponential models, and Gaussian models.

[0017] In the step of quantifying the uncertainty of sparse observation data interpolation, Monte Carlo simulation generates a large number of random samples that conform to the prior distribution and repeatedly performs interpolation calculations, and statistically analyzes the mean, variance and confidence interval of the interpolation results.

[0018] In the spatial topology output step, the row and column indices of the adjacency matrix correspond to the node numbers, the matrix element values ​​indicate whether there is a connection relationship between the nodes and the relationship weight, and the edge list records each edge as a triplet of the start point, end point and weight.

[0019] In the spatial topology output step, the attenuation rate dynamic rule adjusts the strength of the topology relationship according to the distance-time difference or the characteristics of the propagation medium. Common forms are exponential attenuation functions or power-law attenuation functions.

[0020] This invention proposes a spatiotemporal topology construction method that combines multi-source data fusion, GPU-accelerated indexing, local incremental updates, and Bayesian uncertainty quantification. The invention addresses spatiotemporal registration accuracy issues through multi-scale dynamic coupling using a vortex sensing alignment algorithm; overcomes batch processing latency with real-time streaming noise reduction using LSTM-Autoencoder anomaly detection and edge computing; compensates for limitations in independent parameter processing by introducing a temperature-salinity-flow collaborative interpolation model and graph neural network learning of nonlinear relationships; reduces query response time from seconds to milliseconds by utilizing CUDA dual-layer sub-thread grid parallel execution of master and auxiliary filters; reduces global computational burden by employing spatial partitioning and event-driven mechanisms to perform local reconstruction only on affected regions; and quantifies interpolation uncertainty by adding confidence labels to each topology node through Bayesian Skrigin interpolation and Monte Carlo simulation. In summary, this invention solves the technical problem mentioned in the background art of the difficulty in understanding and efficiently constructing the spatiotemporal topological relationships of geographic elements under the conditions of observation and forecast data by establishing a multi-scale spatiotemporal registration system for vortex sensing, a real-time streaming noise reduction mechanism, a physical constraint collaborative interpolation model, a GPU-accelerated millisecond-level index response, event-driven local incremental updates, and a Bayesian uncertainty quantification framework. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention.

[0022] Figure 2 This is a network structure diagram of the topological relationships between objects in three-dimensional space.

[0023] Figure 3 The comparison diagrams before and after the grid topology update include two sub-diagrams: (A) is a schematic diagram of the grid topology before the update, and (B) is a schematic diagram of the grid topology after the update.

[0024] Figure 4 This is a confidence label distribution map for the entire sea area. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0026] like Figure 1 The diagram shown is a flowchart of a method for establishing the spatiotemporal topology of geographic elements for a large-scale marine forecasting and disaster reduction model provided by this invention. This method includes the following steps:

[0027] S1. Multi-source heterogeneous data fusion and integration: Satellite remote sensing data, buoy observation data and forecast data are formatted using a unified data standard. Data comparability is ensured through data cleaning, format conversion and spatiotemporal registration. Noise reduction, interpolation and feature extraction methods are used to preprocess the raw data. Spatiotemporal alignment and semantic association techniques are used to map multi-source data to a unified coordinate system to establish logical relationships between data and output a standardized marine dataset.

[0028] S2. Geographic Information Transcoding and Standardization Processing: Using a word segmenter and address component parsing, the semantic information of the textual description in the standardized marine dataset is converted into a structured standard format. Based on the Earth ellipsoid mathematical model and the Global Positioning System, the structured standard format is converted into precise latitude and longitude coordinates using triangulation. The latitude and longitude coordinates are then standardized, stored, and managed using spatial database standards to form spatial location data.

[0029] S3. Spatiotemporal Indexing and Topological Relationship Construction: Establish an efficient indexing mechanism for the spatial location data. Use quadtrees combined with Z-curve encoding for spatial dimensionality reduction, and use B+ trees to index the time dimension. Quickly filter the candidate set through the main filter and then perform precise geometric calculations through the auxiliary filter. Start the maximum number of CUDA streams of the GPU. Each CUDA stream starts two layers of sub-thread grids, including a first layer and a second layer. The first layer of sub-thread grids performs the minimum bounding rectangle fast exclusion operation of the main filter, and the second layer of sub-thread grids performs the precise geometric intersection judgment operation of the auxiliary filter. Reduce the query response time from seconds to milliseconds and output the spatiotemporal indexing results.

[0030] S4. Topological Relationship Reasoning and Dynamic Process Object-Oriented Modeling: Based on the spatiotemporal index results, construct a geometric and semantic association network of three-dimensional spatial objects to support the calculation of adjacency, intersection, and inclusion relationships. Based on the object-oriented data model, abstract storm surge overtopping into dynamic line objects and establish topological adjacency relationships with the surrounding coastline. Encapsulate time series data into spatiotemporal objects with life cycle attributes and track the evolution characteristics of the spatiotemporal objects through a version management mechanism to output the initial topological relationship network.

[0031] S5. Dynamic maintenance and incremental update of topology: The initial topology network is divided into independent sub-regions using spatial partitioning technology. Only the affected regions are locally reconstructed. Updates are triggered by an event-driven mechanism. Asynchronous parallel processing is achieved using message queues and distributed caching technology. When the change in the observed data of a certain independent sub-region exceeds a preset threshold, the local topology reconstruction process is triggered to reduce the computational load and response latency. The updated topology network is then output.

[0032] S6. Quantification of interpolation uncertainty of sparse observation data: A prior probability distribution is established by combining the Bayesian Skrigin interpolation method with historical statistical data. Spatial correlation is quantified by variogram analysis. The range of interpolation uncertainty is evaluated by Monte Carlo simulation. Confidence labels are added to the updated topological relationship network to provide risk assessment basis for downstream decision-making. The topological relationship network with confidence is output.

[0033] S7. Spatial Topology Output: Through standardized abstraction, the geographic elements in the confidence-based topology network are transformed into node and edge structures. Parameter attributes are assigned to the nodes, and the spatial topology is encoded into an adjacency matrix or edge list using graph theory methods. The network weights are adjusted by incorporating dynamic rules based on decay rate. The input is adapted to JSON or CSV format to ensure the integrity of node ID, connection relationship, and quantified feature fields. Standardized topology data is output for use by the marine forecasting and disaster reduction large model.

[0034] Among these, unified data standards refer to standardized processing guidelines established to address the differences in format, accuracy, and spatiotemporal resolution of multi-source heterogeneous data, ensuring comparability of data from different sources within the same framework. Spatiotemporal registration refers to the process of aligning data from different time bases and spatial reference systems to a unified reference frame, employing adaptive mesh reconstruction based on flow field characteristics and vortex sensing alignment algorithms to achieve dynamic coupling across multiple scales. Noise reduction processing utilizes a spatiotemporal continuity verification model and LSTM-Autoencoder anomaly detection technology, combined with edge computing to achieve real-time streaming cleaning and feedback learning, surpassing batch processing modes based on statistical thresholds.

[0035] The interpolation methods incorporate physically constrained four-dimensional variational interpolation and a temperature-salinity-current co-interpolation model, utilizing graph neural networks to learn nonlinear relationships and overcome the limitations of traditional Kriging interpolation and independent parameter processing. Spatiotemporal alignment refers to the technique of mapping data from different spatiotemporal references to a unified spatiotemporal framework through mathematical transformations. Semantic association techniques establish logical connections between data through ontology and knowledge graphs. Standardized ocean datasets include satellite remote sensing data, buoy observation data, and forecast data that have undergone format standardization, quality control, and spatiotemporal registration.

[0036] The Earth ellipsoid mathematical model uses the WGS84 coordinate system to describe the Earth's shape. Triangulation, based on distance or angle observations from at least three known reference points, determines the spatial coordinates of a target point by solving a system of nonlinear equations. The spatial database standard follows the OGC Simple Features specification, supporting the storage, indexing, and spatial relationship querying of point, line, and surface geometric objects. Latitude and longitude coordinates are in degrees, with longitude ranging from -180° to 180° and latitude ranging from -90° to 90°. Spatial location data includes latitude and longitude coordinates, elevation information, and a timestamp.

[0037] A quadtree is a tree-like data structure that recursively divides a two-dimensional space into four quadrants, with each node representing a spatial region. It is suitable for fast retrieval of spatial objects. Z-curve encoding maps multi-dimensional spatial coordinates to a one-dimensional sequence along a Z-shaped path, preserving spatial proximity. The dimensionality-reduced sequence is then efficiently indexed using a B+ tree. The main filter uses the minimum bounding rectangle to quickly eliminate obviously non-intersecting spatial objects, while the auxiliary filter performs precise point-to-face inclusion judgment, line segment intersection calculation, and polygon overlay analysis on the candidate set.

[0038] CUDA streams are execution queues in the NVIDIA GPU architecture used to manage asynchronous operations. Each CUDA stream independently executes a series of GPU operations without interference. The first-layer sub-threaded grid consists of multiple thread blocks, each containing multiple threads, used to perform the minimum bounding rectangle fast elimination operation of the main filter in parallel. The second-layer sub-threaded grid starts after the candidate set is output by the first-layer sub-threaded grid, performing the precise geometric intersection judgment operation of the auxiliary filter. The spatiotemporal index results contain the identifiers of spatial objects that meet the query conditions after fast filtering and precise calculation, along with their spatiotemporal attributes.

[0039] Object-oriented data models abstract geographic entities into objects with attributes and methods. Dynamic line objects represent linear geographic features that change over time, and topological adjacency describes the contact or proximity states between spatial objects. Spatiotemporal objects are data entities that simultaneously possess spatial location and temporal attributes. Lifecycle attributes record the complete process of an object from its creation to its destruction, and version management mechanisms track the historical evolution of object states through timestamps or version numbers. Adjacency refers to two spatial objects sharing a boundary but not overlapping; intersection refers to two spatial objects having a common part; and containment refers to one spatial object being completely located inside another. The initial topological relationship network refers to the network structure initially constructed that includes all geographic features and their spatial relationships.

[0040] Spatial partitioning technology divides geographic space into several grid units according to latitude and longitude or administrative divisions. Each unit maintains a local topological relationship index, reducing the global computational burden. Independent sub-regions refer to non-overlapping geographic units divided by spatial partitioning technology. An event-driven mechanism monitors changes in observed data and automatically initiates the processing flow when trigger conditions are met. The message queue uses Kafka or RabbitMQ to implement a producer-consumer pattern, supporting asynchronous message delivery and traffic shaping. Distributed caching utilizes Redis or Memcached to store frequently accessed data in memory, improving read speed and reducing database pressure. Preset thresholds are determined based on the characteristics of marine environmental changes and business needs, including quantitative indicators such as temperature change amplitude, current velocity change rate, and water level fluctuation range. An update operation is triggered when the deviation between real-time observation values ​​and historical benchmark values ​​exceeds the preset threshold. The updated topological relationship network refers to the topological relationship network reflecting the latest observation data after local reconstruction.

[0041] Bayesian skrygian interpolation treats the interpolation problem as a stochastic process, fusing prior information and observed data through a Bayesian statistical framework. The prior probability distribution is established based on the statistical characteristics of historical observation data from the same period, reflecting the initial understanding of observations at unknown locations. The variogram describes the variation of the difference between observed values ​​between two points in space as distance increases, quantifying the strength of spatial autocorrelation; common forms include spherical models, exponential models, and Gaussian models. Monte Carlo simulation generates a large number of random samples conforming to the prior distribution, repeatedly performs interpolation calculations, and statistically analyzes the mean, variance, and confidence interval of the interpolation results, quantifying the range of uncertainty. Confidence labels are numerical indicators representing the reliability of interpolation results, ranging from 0 to 1, and are affected by the density of observation sites, the strength of spatial autocorrelation, and the accuracy of the interpolation method, providing a quantitative basis for decision-makers to assess data quality. A topological relationship network with confidence labels refers to the network structure after attaching confidence labels to each node and edge of the updated topological relationship network.

[0042] Nodes represent the spatial location of geographic features in the topological network, edges represent the types of topological relationships between features, and parameter attributes include the physical characteristic values, timestamps, and metadata information of the features. The adjacency matrix is ​​a square matrix where row and column indices correspond to node numbers, and matrix element values ​​indicate whether a connection exists between nodes and the weight of that connection; it is suitable for storing and quickly querying dense graphs. The edge list records each edge as a triplet of start, end, and weight, suitable for efficient storage of sparse graphs. The decay rate dynamic rule adjusts the strength of topological relationships based on distance, time differences, or propagation medium characteristics, simulating the decay characteristics of marine environmental influences propagating spatially; common forms are exponential decay functions or power-law decay functions. Node IDs are unique identifiers for nodes in the topological network, connection relationships describe the edges between nodes and their attributes, and quantified features include the physical parameters and confidence values ​​of the nodes. Standardized topological relationship data refers to structured data that conforms to JSON or CSV format specifications and contains complete node information and connection relationships.

[0043] Optionally, the present invention also provides a computer-based method for establishing a spatiotemporal topology system for geographic elements in a large-scale marine forecasting and disaster reduction model. The computer is equipped with a readable storage medium that stores program instructions. When the program instructions are run on the computer, they execute the aforementioned method for establishing the spatiotemporal topology of geographic elements for a large-scale marine forecasting and disaster reduction model.

[0044] The specific implementation methods of the above steps are described in detail below.

[0045] The specific implementation of step S1 is as follows: First, the satellite remote sensing data, buoy observation data, and forecast data are formatted and classified to determine whether the data is in NetCDF, CSV, or binary format. Based on the identification results, the corresponding parser is called to uniformly convert the heterogeneous formats into an internal standard format. The standard format includes timestamps, spatial coordinates, observation elements, and quality identifier fields. During the conversion process, XML schema is used to define document constraints to ensure consistency. Next, data cleaning is performed. A spatiotemporal continuity test model is used to detect jump points and missing values ​​in the observation sequence. The spatiotemporal continuity test model establishes constraints based on the rate of change of observation values ​​at adjacent times and adjacent spatial locations. When the rate of change exceeds the physically achievable range, it is marked as an outlier. LSTM-Autoencoder anomaly detection technology is used to learn the spatiotemporal evolution pattern of normal observation data. By reconstructing errors, abnormal data points that deviate from the normal pattern are identified. The cleaned data then enters the spatiotemporal registration stage. Spatiotemporal registration employs an adaptive grid reconstruction algorithm based on flow field characteristics. This algorithm dynamically adjusts the grid density according to the vortex structure and frontal position of the ocean current field, increasing the grid resolution in regions with large flow field gradients and decreasing the grid density in uniform flow field regions, thus optimizing the allocation of computational resources. The vortex sensing alignment algorithm detects the center position, rotation direction, and intensity characteristics of vortices from different data sources, aligning the corresponding vortex structures from multiple sources to a unified reference frame, eliminating spatial position offsets caused by differences in observation time. Interpolation processing uses a physically constrained four-dimensional variational interpolation method. This method introduces ocean dynamics equations as constraints during interpolation, ensuring that the interpolation results satisfy the laws of mass conservation, momentum conservation, and energy conservation. The temperature-salinity-current co-interpolation model utilizes the physical coupling relationship between temperature, salinity, and flow velocity, learning the nonlinear correlation between the three through a graph neural network. It simultaneously considers the constraints of other factors when interpolating one element, improving interpolation accuracy. The feature extraction stage uses wavelet transform to separate signal components at different time scales and extracts features such as tides, storm surges and long-term trends. Semantic association technology establishes logical relationships between physical quantities such as temperature, salinity and current velocity by constructing an ontology of marine elements. Finally, it outputs a standardized marine dataset containing a unified spatiotemporal benchmark, complete quality labels and semantic associations.

[0046] The specific implementation of step S2 is as follows: First, lexical analysis is performed on the semantic information of the text descriptions in the standardized marine dataset. The word segmenter uses the maximum matching algorithm to segment the continuous text into word sequences, identifying address components such as provinces, cities, districts, and streets. The address component parser maps the word sequences into a hierarchical address structure based on the administrative division tree structure. The hierarchical address structure includes national, provincial, municipal, county, and township-level geographical units in order from coarse to fine, forming a structured standard format. In the spatial positioning stage, the spatial coordinates corresponding to the address are calculated using the Earth ellipsoid mathematical model. The Earth ellipsoid mathematical model uses the WGS84 coordinate system to define a rotating ellipsoid with a major semi-axis of 6,378,137 meters and a minor semi-axis of 6,356,752 meters to describe the shape of the Earth. The triangulation method is based on the known latitude and longitude coordinates of the reference point and the azimuth and distance observations of the target address relative to the reference point. The latitude and longitude coordinates of the target point are calculated by solving the geodetic forward calculation problem. The solution process uses the Gauss-Kruger projection to convert the ellipsoidal coordinates into planar coordinates for calculation, achieving sub-meter accuracy. During the normalization storage phase, the geometric types of spatial objects are defined according to the OGC Simple Features specification. Point objects store single latitude and longitude coordinate pairs, line objects store ordered coordinate point sequences, and polygon objects store closed coordinate loops. Each spatial object is appended with a spatial reference system identifier and a timestamp. The spatial location data is organized and stored using an R-tree spatial index structure. The R-tree organizes the minimum bounding rectangle of the spatial objects in a hierarchical structure, supports rectangle range queries and nearest neighbor queries, and the query efficiency is logarithmically related to the amount of data, forming spatial location data consistent with the spatiotemporal benchmark of the standardized ocean dataset.

[0047] The specific implementation of step S3 is as follows: First, the spatial location data is decomposed into a quadtree. The quadtree decomposition starts from the global range represented by the root node and recursively divides the rectangular area represented by each node into four sub-quadrants. The decomposition stops when the number of spatial objects contained in a node is less than a set threshold or the node depth reaches the maximum number of layers. The set threshold reference value is 64 objects, and the maximum number of layers reference value is 16 layers. The leaf nodes of the quadtree record the list of spatial object identifiers they contain. Z-curve encoding converts the two-dimensional coordinates of the quadtree nodes into one-dimensional encoding. The encoding process alternates the binary bits of the coordinates according to a Z-shaped path to ensure that spatially adjacent nodes are also adjacent in the one-dimensional encoding sequence. The encoded one-dimensional sequence is used as the index key of the B+ tree. The time dimension of the B+ tree index uses the timestamp as the second key of the composite index. The composite index supports spatiotemporal range joint queries. During the main filter execution phase, all available CUDA streams on the GPU are launched. The number of CUDA streams is determined by the GPU hardware, typically ranging from 16 to 32. Each CUDA stream is assigned a query task. The first-layer sub-thread grid performs an intersection test between the minimum bounding rectangle of the query range and the spatial range of the quadtree node. The intersection test determines whether the two rectangles overlap by comparing the coordinates of the lower left and upper right corners. Nodes that do not overlap and their contained objects are directly excluded, while objects contained in overlapping nodes are added to the candidate set. Each thread in the first-layer sub-thread grid processes one quadtree node, and the number of threads reaches thousands to tens of thousands to achieve high parallelism. During the auxiliary filter execution phase, the second-layer sub-thread mesh performs precise geometric calculations on each spatial object in the candidate set. The point and polygon inclusion judgment uses the ray method to count the number of intersections between rays emitted from the query point in any direction and the polygon boundary. If the number of intersections is odd, the point is inside the polygon. The line segment intersection calculation determines the intersection relationship by judging whether two line segments cross each other's extensions. Each thread of the second-layer sub-thread mesh processes one object in the candidate set. The two layers of sub-thread mesh exchange data in GPU shared memory to avoid frequent access to global memory. Finally, the set of spatial object identifiers that meet the query conditions is output as the spatiotemporal index result.

[0048] The specific implementation of step S4 is as follows: A geometric and semantic association network of three-dimensional spatial objects is constructed based on the spatiotemporal index results. Geometric association is established by calculating the Euclidean distance, azimuth, and relative elevation between spatial objects. Semantic association is established by matching the attribute labels and functional classifications of objects. Adjacency relationships determine whether the minimum bounding rectangles of two objects share a boundary and do not overlap internally. Intersection relationships determine whether the geometric representations of two objects have a common set of points. Inclusion relationships determine whether all vertices of one object are located inside another object. The object-oriented data model abstracts the storm surge overtopping process into dynamic line objects. These dynamic line objects contain a coordinate sequence and a timestamp sequence representing the evolution of the water level over time. Each timestamp corresponds to a water level position. Topological adjacency relationships are established by detecting the intersections and overlapping segments between dynamic line objects and coastline objects. The intersection coordinates record the location of the overtopping, and the length of the overlapping segment reflects the range of the overtopping's impact. Time series data is encapsulated as spatiotemporal objects containing attributes such as spatial location, observation sequence, lifecycle start and end time, and evolution rate. A version management mechanism assigns globally unique identifiers and incremental version numbers to spatiotemporal objects. A new version record is created each time the object state is updated, and historical versions are retained to support time-backtracking queries. Evolution features are extracted by calculating the rate of change of observations between adjacent versions and the speed of spatial location movement. The output initial topological relationship network is stored in a graph structure, where nodes represent spatiotemporal objects, edges represent topological relationships, and the weight of the edges is quantified by the relationship strength.

[0049] The specific implementation of step S5 is as follows: Spatial segmentation technology divides the geographic space covered by the initial topological relationship network into rectangular grids according to longitude and latitude. The grid size is determined based on data density and update frequency. Data-dense areas use smaller grid sizes (reference value: 0.1 degrees x 0.1 degrees), while sparse data areas use larger grid sizes (reference value: 1 degree x 1 degree). Each grid serves as an independent sub-region, maintaining a local topological relationship index. An event-driven mechanism monitors the updates of observed data within each independent sub-region, calculating the deviation between the new observation and the historical baseline. The deviation calculation uses a standardized difference method, subtracting the historical mean from the observed value and dividing by the historical standard deviation. When the absolute value of the standardized difference exceeds a preset threshold, local topological reconstruction is triggered. The preset threshold reference value is 3 times the standard deviation, corresponding to a 99.7% confidence interval. Local topological reconstruction only processes the independent sub-region that triggered the event and its adjacent regions. The reconstruction process recalculates the topological relationships between spatiotemporal objects within the region, updates edge weights and node attributes, and ensures boundary consistency during the processing of adjacent regions. The message queue utilizes the Kafka distributed stream processing platform. The producer module encapsulates triggering events into messages and sends them to the topic queue. The consumer module retrieves messages from the queue and initiates topology reconstruction tasks. The message queue supports multiple consumers processing reconstruction tasks for different regions in parallel. Distributed caching leverages the Redis key-value storage system to cache frequently accessed topology relationship data in memory. The cache key uses an independent sub-region identifier, and the cache value is a serialized topology relationship subgraph. The cache expiration time is set to twice the observation data update cycle, with a reference value of 30 minutes. Asynchronous parallel processing manages the concurrent execution of multiple reconstruction tasks through a thread pool, outputting an updated topology relationship network that reflects the spatial relationships of the latest observation data.

[0050] The specific implementation of step S6 is as follows: The input parameters of the Bayesian Skrigin interpolation method include the spatial coordinates of the observation station, observed values, and historical statistical data. The output parameters are the estimated value and confidence label of the location to be interpolated. The prior probability distribution uses a Gaussian distribution to describe the probability characteristics of the observed values ​​at the unknown location. The mean and variance of the distribution are determined by the statistics of historical data from the same period. The variogram analysis calculates the variance of observed pairs at different distance intervals and fits the parameters of a spherical model, an exponential model, or a Gaussian model. The spherical model is suitable for cases where spatial correlation disappears after a certain distance, the exponential model is suitable for cases where spatial correlation decays exponentially with distance, and the Gaussian model is suitable for cases where spatial correlation changes smoothly. The model selection is determined by comparing prediction errors through cross-validation. Monte Carlo simulation generates a random sample set following a prior probability distribution, with a reference sample size of 10,000. Kriging interpolation is performed on each sample, and the interpolation results of all samples are statistically analyzed. The mean is calculated as the final estimate, and the standard deviation is calculated as an uncertainty measure. The confidence label is defined as the reciprocal of the ratio of the standard deviation of the estimated value to the standard deviation of the observed value. The smaller the ratio, the greater the interpolation uncertainty and the lower the confidence. The confidence label ranges from 0 to 1. The corresponding confidence label is attached to each node and edge of the updated topological relationship network, and the topological relationship network with confidence is output to provide a quantitative risk assessment for marine forecasting and disaster reduction decision-making.

[0051] The specific implementation of step S7 is as follows: Standardized abstraction maps spatiotemporal objects in the topological relationship network with confidence to nodes. Node attributes include the object's spatial coordinates, physical feature values, timestamp, and confidence label. Physical feature values ​​include observations such as temperature, salinity, and current velocity. The timestamp records the valid moment of the object's state. Topological relationships are mapped to edges. Edge attributes include relationship type, strength weight, and confidence label. The relationship type uses enumerated values ​​to represent adjacent, intersecting, or contained relationships. The strength weight is calculated according to the dynamic decay rate rule. The dynamic decay rate rule uses an exponential decay function. The function input is the spatial distance and time difference between nodes, and the output is the decayed relationship strength. The spatial distance is calculated using geodetic distance, and the time difference is calculated using timestamp differences. The decay coefficient is determined based on the characteristics of the marine environment, with a reference value of 0.1 per kilometer for spatial decay and 0.05 per hour for time decay. The adjacency matrix construction process assigns a unique integer identifier to each node in the network as the matrix row and column index. The matrix element values ​​are the weights of the edges between the corresponding nodes. Nodes without connections have an element value of 0. The adjacency matrix is ​​suitable for networks with a small number of nodes and dense connections. The edge list construction process records each edge as a triple containing a start identifier, an end identifier, and a weight. Edge lists are suitable for networks with a large number of nodes and sparse connections; the storage space complexity is proportional to the number of edges. The JSON format output organizes the node and edge lists into nested object structures. Each element in the node list contains a node ID, coordinate field, attribute field, and confidence field. Each element in the edge list contains a start ID, end ID, relationship type, weight, and confidence field. The CSV format output stores the node and edge data as two separate table files. The node table contains columns for node ID, longitude, latitude, attribute, and confidence. The edge table contains columns for start ID, end ID, relationship type, weight, and confidence. The output standardized topological relationship data is directly used as input to the large-scale marine forecasting and disaster reduction model, supporting spatial queries, path planning, and disaster impact assessment.

[0052] It should be noted that the key technical concepts of this invention include a multi-layered CUDA parallel computing architecture and an incremental topology update strategy. The multi-layered CUDA parallel computing architecture achieves pipelined parallel processing of computational tasks by performing coarse screening of the main filter through a first-layer sub-thread grid and precise calculation of the auxiliary filter through a second-layer sub-thread grid. Compared to traditional CPU serial processing, utilizing the parallel capabilities of thousands of GPU computing cores reduces the spatiotemporal index query response time from seconds to milliseconds, significantly improving the real-time performance of marine disaster early warning. The multi-layered architecture avoids submitting all candidate objects for precise calculation, reducing computational resource waste and improving overall processing efficiency. The incremental topology update strategy decomposes the global topology network into independent sub-regions using spatial partitioning technology. Combined with an event-driven mechanism, it triggers reconstruction only for local regions where data changes are significant. Compared to the traditional global reconstruction method, incremental updates avoid processing most regions that have not changed. This reduces the computational complexity of topology maintenance from being proportional to the size of the global network to being proportional to the size of local changes. In the dynamic scenario of marine environment changes, local changes usually only account for 5% to 10% of the global changes. The incremental update strategy can reduce redundant computation by more than 90%. At the same time, message queues and distributed caching support asynchronous parallel updates of multiple local regions, further reducing response latency and ensuring that the topology network continuously reflects the latest observation data. The synergistic effect of the two technical approaches lies in the fact that multi-layer CUDA parallel computing provides efficient geometric computing capabilities for incremental topology updates, enabling local reconstruction processes to be completed in milliseconds. Incremental topology updates reduce the scale of data to be processed and lower the demand for GPU computing resources. The combination of the two enables real-time indexing and querying of massive spatiotemporal data and dynamic topology maintenance. Compared with the offline batch processing mode of existing technologies, this invention supports the marine forecasting and disaster reduction big model to respond instantly to sudden disaster events, improving the disaster warning timeliness from hours to minutes, and buying valuable time for emergency decision-making.

[0053] It should be noted that this invention also addresses the following technical problem: traditional marine data processing systems suffer from data fusion error accumulation due to inconsistent spatiotemporal references when dealing with multi-source heterogeneous data. This invention establishes an adaptive grid reconstruction algorithm based on flow field characteristics, identifying dynamic features such as eddies and fronts in the marine environment as registration references, replacing traditional geometric coordinate alignment methods. This enables accurate mapping of data from different sources at the physical level. Simultaneously, ontology and knowledge graph technologies are used to establish semantic relationships between data, ensuring the comparability of measurement parameters from different observation platforms within a unified physical framework. This significantly reduces the systematic bias introduced by spatiotemporal reference differences during multi-source data fusion, providing a high-quality, standardized data foundation for subsequent topology construction. Furthermore, this invention also solves the problem of excessive computational resource consumption in traditional global topology update mechanisms under conditions of massive real-time observation data. This invention divides the geographic space into several independent sub-regions using spatial partitioning technology and maintains a local topology index for each sub-region. Combined with an event-driven mechanism to monitor changes in observed data, when the data change in a certain region exceeds a preset threshold, only the local topology reconstruction process of that region is triggered. Asynchronous parallel processing is achieved using message queues and distributed caching, avoiding the computational redundancy of rebuilding the entire global topology network for any data update in traditional methods. While ensuring the accuracy of topological relationships, the computational load is reduced to a local proportion of the original method, significantly improving the system's real-time response capability to dynamic marine environments.

[0054] Specifically, the principle of this invention is as follows: The invention solves the aforementioned technical problems by establishing a complete technical chain from data fusion to topology output. The eddy-sensing alignment algorithm adaptively reconstructs the mesh by identifying ocean eddy characteristics, enabling precise registration of data at different resolutions at the physical field feature level, overcoming the limitation of traditional geometric alignment that ignores dynamic features. The LSTM-Autoencoder identifies outliers in real time by learning spatiotemporal continuity patterns, and the edge computing architecture cleans the data source at the data source end, avoiding batch transmission delays. Temperature-salinity-current co-interpolation uses oceanographic physical equations as constraints, and graph neural networks capture complex nonlinear relationships between parameters, significantly improving interpolation accuracy in sparse regions. The GPU dual-layer sub-thread mesh pipelines the coarse screening and precise computation of spatial queries through hardware parallelization, achieving millisecond-level response. Spatial partitioning technology decomposes the global topology into independent sub-regions, and the event-driven mechanism monitors data changes, triggering only local reconstruction, avoiding waste of full computational resources. The Bayesian framework integrates historical statistical priors with current observations, and Monte Carlo simulation evaluates the probability distribution of interpolation results through random sampling, providing quantitative reliability indicators for each node. This multi-layered technological collaboration enables the system to output highly reliable spatiotemporal topological relationships while ensuring real-time performance.

[0055] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0056] The specific implementation of step S1 involves formatting satellite remote sensing data, buoy observation data, and forecast data using a unified data standard. Data comparability is ensured through data cleaning, format conversion, and spatiotemporal registration. Noise reduction, interpolation, and feature extraction methods are used for preprocessing the raw data. Spatiotemporal alignment and semantic association techniques are used to map multi-source data to a unified coordinate system, establishing logical relationships between the data and outputting a standardized ocean dataset. Specifically, spatiotemporal registration employs adaptive grid reconstruction based on flow field characteristics and vortex sensing alignment algorithms to achieve multi-scale nested dynamic coupling. Noise reduction utilizes a spatiotemporal continuity verification model and LSTM autoencoder anomaly detection technology, combined with edge computing to achieve real-time streaming cleaning and feedback learning. The interpolation method incorporates physically constrained four-dimensional variational interpolation and a temperature-salinity-current co-interpolation model, utilizing graph neural networks to learn nonlinear relationships.

[0057] The specific implementation of step S2 involves using a word segmenter and address component parsing to convert the semantic information of the textual descriptions in the standardized marine dataset into a structured standard format. Based on the Earth ellipsoid mathematical model and the Global Positioning System (GPS), the structured standard format is converted into precise latitude and longitude coordinates using triangulation. Spatial database standards are then used to standardize, store, and manage these latitude and longitude coordinates to form spatial location data. The Earth ellipsoid mathematical model uses the WGS84 coordinate system to describe the Earth's shape. The triangulation method determines the spatial coordinates of the target point based on distance or angle observations from at least three known reference points. The latitude coordinates are specifically represented as follows:

[0058] ;

[0059] In the formula, These are latitude coordinates, in degrees. This is the distance from the target point to the first reference point, in meters. This is the distance from the target point to the second reference point, in meters. This is the distance between the first and second reference points, in meters. Pi, with a value of 3.14159. The azimuth angle is calculated as follows:

[0060] ;

[0061] In the formula, This is the azimuth angle, in radians; and These are the projected coordinates of the target point, in meters; and These are the projected coordinates of the first reference point, in meters. Longitude coordinates are expressed as:

[0062] ;

[0063] In the formula, These are longitude coordinates, in degrees. The longitude of the first reference point, in degrees; The latitude of the first reference point, in degrees; This is the equatorial radius of the Earth's ellipsoid, with a default value of 6,378,137 meters. Longitude coordinates range from -180 degrees to 180 degrees, and latitude coordinates range from -90 degrees to 90 degrees.

[0064] The specific implementation of step S3 involves establishing an efficient indexing mechanism for spatial location data. This is achieved by using a quadtree combined with Z-curve encoding for spatial dimensionality reduction, and leveraging a B-tree index for the time dimension. A main filter quickly filters the candidate set, followed by an auxiliary filter for precise geometric calculations. The maximum number of CUDA streams on the GPU is activated, with each CUDA stream initiating two layers of sub-threaded grids: a first-layer sub-threaded grid and a second-layer sub-threaded grid. The first-layer sub-threaded grid performs the main filter's minimum bounding rectangle elimination operation, while the second-layer sub-threaded grid performs the auxiliary filter's precise geometric intersection judgment operation. This reduces query response time from seconds to milliseconds, outputting the spatiotemporal indexing results. Specifically, Z-curve encoding maps multidimensional spatial coordinates to a one-dimensional sequence along a Z-shaped path, as shown below:

[0065] ;

[0066] In the formula, The Z-curve encoding value is dimensionless. This represents the number of coordinate bits, typically 16. The x-axis is the first A binary value, which can be either 0 or 1; The vertical axis is the first A binary value, taking the value 0 or 1. The time dimension index is represented as:

[0067] ;

[0068] In the formula, This is a time index value, a dimensionless integer; This is the current timestamp, in seconds. For reference, the start timestamp is in seconds; The time resolution is in seconds, with a default value of 3600 seconds. This represents the floor function. The minimum bounding rectangle criterion for the main filter is expressed as:

[0069] ;

[0070] In the formula, The result of the rectangle overlap judgment is either true or false. and These are the minimum and maximum x-coordinates of the query rectangle, in meters; and These are the minimum and maximum x-coordinates of the candidate rectangles, respectively, in meters; and These are the minimum and maximum ordinates of the query rectangle, in meters; and These are the minimum and maximum ordinates of the candidate rectangle, respectively, in meters; As a normalized length benchmark, the empirical value is rice; Represents the logical AND operator.

[0071] The specific implementation of step S4 is to construct a geometric and semantic association network of three-dimensional spatial objects based on the spatiotemporal index results, supporting the calculation of adjacency, intersection, and containment relationships. Based on an object-oriented data model, storm surge breaches are abstracted as dynamic line objects and topological adjacency relationships are established with surrounding coastlines. Time-series data is encapsulated into spatiotemporal objects with lifecycle attributes, and the evolution characteristics of these objects are tracked through a version management mechanism, outputting an initial topological relationship network. The adjacency relationship judgment is represented as follows:

[0072] ;

[0073] In the formula, The result of the adjacent relationship judgment is either true or false. and They are space objects and The representative point coordinate vector, in meters; For reference distance, the empirical value is 1000 meters; This is the distance threshold, which defaults to 10 meters. and They are space objects and The geometric range; This represents the Euclidean distance between two points, in meters. This represents the set intersection operator; This represents the empty set. The inclusion relation is expressed as:

[0074] ;

[0075] In the formula, The result of the inclusion relationship judgment is either true or false; and They are space objects and The area, in square meters; For reference area, the empirical value is square meters; This is the area ratio coefficient, which is usually taken as 0.95; Symbols representing subset relations.

[0076] The specific implementation of step S5 involves using spatial partitioning technology to divide the initial topology network into independent sub-regions, performing local reconstruction only on affected regions, triggering updates using an event-driven mechanism, and utilizing message queues and distributed caching technology to achieve asynchronous parallel processing. When the change in observed data in a certain independent sub-region exceeds a preset threshold, a local topology reconstruction process is triggered to reduce computational load and response latency, outputting the updated topology network. The normalized deviation value is calculated as follows:

[0077] ;

[0078] In the formula, The normalized deviation value is dimensionless. The number of observation parameters; For the first New observations of each parameter; For the first Historical baseline values ​​for each parameter; For the first The standard deviation of each parameter. The preset threshold trigger judgment is expressed as:

[0079] ;

[0080] In the formula, The value is either true or false to trigger the judgment result; This is the threshold coefficient, with an empirical value of 3. The calculation method is as follows:

[0081] ;

[0082] In the formula, For the first The standard deviation of each parameter; This represents the number of historical observation samples. For the first The parameter of the first Previous historical observation value; For the first The historical mean of each parameter.

[0083] The specific implementation of step S6 involves establishing a prior probability distribution using Bayesian skryz interpolation combined with historical statistical data, quantifying spatial correlation through variogram analysis, evaluating the range of interpolation uncertainty using Monte Carlo simulation, and adding confidence labels to the updated topological relationship network to provide risk assessment basis for downstream decision-making, outputting a topological relationship network with confidence scores. The variogram is expressed as:

[0084] ;

[0085] In the formula, The values ​​are dimensionless and represent the variogram. This refers to spatial distance, expressed in kilometers. The distance is The number of observation point pairs; For position Observations at; Distance for Observations at; For reference variance, the empirical value is the global variance of the observed data; The reference observation value is used for normalization; the root mean square value of historical observation data is taken. Let be a two-dimensional spatial coordinate vector containing longitude and latitude components, in degrees. The spherical variogram model is represented as:

[0086] ;

[0087] In the formula, The variogram value for the spherical model is dimensionless. The nugget value represents the microscale variation at zero distance and is dimensionless. The sill value represents the spatial correlation strength and is dimensionless. For the range, represents the effective action distance of spatial autocorrelation, in kilometers. The exponential variation function model is expressed as:

[0088] ;

[0089] The parameters in the formula have the same meaning as in the spherical model. The Gaussian variation function model is expressed as:

[0090] ;

[0091] The parameters in the formula have the same meaning as in the spherical model. The confidence label calculation is expressed as:

[0092] ;

[0093] In the formula, This is the confidence level label value, ranging from 0 to 1; The variance of the interpolation results obtained from the Monte Carlo simulation; This is the mean of the interpolation results obtained from the Monte Carlo simulation, and its units are consistent with the units of the observed parameters to be interpolated. This is the stabilization constant, with a default value of 0.01, and its unit is... The units are consistent. Among them... and The results were obtained by generating a large number of random samples that conform to the prior distribution and repeatedly performing interpolation calculations.

[0094] The specific implementation of step S7 involves transforming geographical elements in a confidence-based topological relationship network into a node and edge structure through standardized abstraction. Nodes are assigned parameter attributes, and spatial topological relationships are encoded into adjacency matrices or edge lists using graph theory methods. Network weights are adjusted by incorporating dynamic rules based on decay rates. The system is adapted to JSON or CSV input formats to ensure the integrity of node IDs, connectivity relationships, and quantified feature fields. The output is standardized topological relationship data for use in the marine forecasting and disaster reduction large-scale model. The adjacency matrix for the exponential decay mode is represented as follows:

[0095] ;

[0096] In the formula, The first in the adjacency matrix Line number The elements of a column represent nodes. With nodes The connection weights between them are dimensionless. The base weight is usually set to 1; For nodes With nodes The Euclidean distance between them, in kilometers; The attenuation length scale is empirically set to 50 km. This formula characterizes the exponential decay of spatial topological relationship strength with increasing distance. The basic weight term determines the initial connectivity strength, the exponential decay term simulates the weakening effect of marine environmental influences propagating through space, and the attenuation length scale controls the rate of weight decrease. The adjacency matrix of the power-law decay mode is expressed as:

[0097] ;

[0098] In the formula, The connection weights in the power-law decay mode are dimensionless. The normalized distance is empirically defined as 1 kilometer. The power law exponent typically ranges from 1.5 to 3; The function takes the smaller of the two values ​​to ensure that the weight does not exceed the base weight.

[0099] To better understand and implement this invention, a specific application scenario is provided below as Example 2: A forecasting center receives a typhoon warning and needs to conduct a real-time assessment of the storm surge risk of breaching the seawall in a certain sea area of ​​the North Sea to provide decision support for disaster prevention and mitigation in coastal cities. The sea area covers a region of approximately 500 km by 300 km, involving 5 buoy observation stations, 3 satellite remote sensing data sources, and a forecasting database. The technical team uses the geographic element spatiotemporal topology establishment method of this invention to fuse multi-source heterogeneous data and construct topological relationships.

[0100] During the multi-source heterogeneous data fusion and integration phase, the technical team acquired a total of 1200 sea surface temperature images from satellite remote sensing data, buoy observation data including minute-by-minute temperature, salinity, and current velocity observation records from 5 stations, and forecast data including storm surge numerical simulation results for the next 72 hours. The original format of the satellite remote sensing data was HDF5, the buoy observation data was in txt format, and the forecast data was in NetCDF format. Through format identification and unified conversion, a standardized data structure containing timestamps, latitude and longitude coordinates, observation elements, and quality identifiers was formed. During the data cleaning process, a temperature jump was detected at buoy station 07 at a certain time period, with the observed value abruptly changing from 18.5℃ to 32.7℃. The spatiotemporal continuity test model calculated the rate of change to be 14.2℃ per hour, exceeding the physically achievable threshold of 3℃ per hour. This data was marked as an outlier and removed. The LSTM-Autoencoder anomaly detection technology identified a systematic shift in the salinity observation sequence of station 12, with a reconstruction error reaching 1.8 psu, significantly higher than the normal level of 0.3 psu. This was determined to be sensor drift and calibration was performed. Spatiotemporal registration employs an adaptive grid reconstruction algorithm based on flow field characteristics. In the eddy region of the Kuroshio Extension, the grid resolution is refined from 10 km to 2 km, while maintaining a 10 km resolution in the uniform flow field region of the open ocean. An eddy-sensing alignment algorithm identifies corresponding mesoscale eddies in satellite and buoy data, and spatial alignment is achieved by matching the eddy center position offset of 7.3 km. A physically constrained four-dimensional variational interpolation method interpolates sea surface temperature in sparse data regions, with the interpolation results satisfying the heat conservation constraint. A temperature-salinity-current co-interpolation model utilizes the temperature-salinity relationship curve, simultaneously considering temperature gradient information when interpolating salinity, improving interpolation accuracy by 42% compared to independent interpolation. The processed output standardized ocean dataset contains 1200 spatiotemporal grid nodes, each storing eight physical elements including temperature, salinity, current velocity, and water level.

[0101] During the geographic information transcoding and standardization phase, the technical team processed address information for coastal cities, including textual descriptions such as a province, city, district, and street. A word segmenter identified provincial, municipal, district, and street-level geographic units, and an address component parser constructed a four-layer address hierarchy. Triangulation, based on the coordinates of three known reference points and the observed azimuth and distance of the target address relative to these reference points, calculated the spatial coordinates of the target point with an accuracy of 0.8 meters. The spatial database stores 278 coastal geographic features according to the OGC Simple Features specification, including 112 port point objects, 85 coastline objects, and 81 breakwater objects. Each object is appended with a spatial reference system identifier (EPSG: 4326) and a timestamp. An R-tree spatial index organizes the objects into a six-layer tree structure based on the minimum bounding rectangle, with each leaf node containing an average of 46 objects.

[0102] During the spatiotemporal indexing and topology construction phase, the technical team decomposed 1200 spatiotemporal grid nodes into quadtrees, recursively dividing them from the global scope represented by the root node to the 14th layer, resulting in 3856 leaf nodes. Each leaf node contained between 12 and 68 spatial objects, with an average of 31. Z-curve encoding converted the two-dimensional coordinates of the quadtree nodes into a one-dimensional encoded sequence with a length of 28 bits. The B+ tree index used hourly timestamps as the second key of the composite index, supporting queries on the spatial distribution at any time within the next 72 hours. The GPU was equipped with 24 CUDA streams, each allocated to process 50 spatiotemporal grid nodes. The first-layer sub-thread grid contained 512 thread blocks, each containing 256 threads, for a total of 131072 threads executing the main filter operation in parallel. The main filter performed an intersection test between the minimum bounding rectangle of the query range and the 3856 quadtree nodes, with a test time of 3.7 milliseconds, selecting 428 overlapping nodes to add to the candidate set. The second-layer sub-threaded mesh contains 256 thread blocks, each containing 128 threads, for a total of 32,768 threads executing auxiliary filter operations in parallel. It performs precise geometric calculations on 6,784 spatial objects in the candidate set. The point and face inclusion judgment takes 8.2 milliseconds, the line segment intersection calculation takes 11.5 milliseconds, and the total query response time is 23.4 milliseconds. It outputs the identifiers of 1,893 spatial objects that meet the query conditions as the spatiotemporal index results.

[0103] In the topological relationship reasoning and dynamic process object-oriented modeling stage, a three-dimensional spatial object geometric and semantic association network containing 1893 nodes and 7421 edges is constructed based on the spatiotemporal index results, such as... Figure 2As shown, there are 2847 adjacent edges, 3156 intersecting edges, and 1418 containing edges. The storm surge overtopping process is abstracted as a dynamic line object, containing a sequence of hourly water level coordinates for the next 72 hours, with 72 timestamps and corresponding 72 sets of coordinate points. The dynamic line object establishes topological adjacency relationships with 85 coastline objects, detecting 23 overtopping intersections. The intersection coordinates record the location of the overtopping, and the total length of the overlapping segment is 37.8 kilometers, reflecting the overtopping impact range. The tidal change time series is encapsulated as a spatiotemporal object, with the lifecycle start and end times being the start and end times of the forecast period. The evolution rate is obtained by calculating the water level difference between adjacent hours, with a maximum high tide rate of 0.85 meters per hour and a maximum low tide rate of 0.72 meters per hour. The version management mechanism assigns each spatiotemporal object a globally unique identifier in UUID format and an incremental version number. Historical version storage supports backtracking and querying the ocean status of the past 24 hours.

[0104] During the dynamic maintenance and incremental update phase of topological relationships, the spatial segmentation technology divides the sea area into 50 rectangular grids, each with a size of 50 km by 30 km. Table 1 shows the changes in observation data for some of the grids.

[0105] Table 1. Changes in Grid Observation Data

[0106]

[0107] The preset threshold is set to 3 times the standard deviation. When the standardized deviation of grids G17, G31, and G44 exceeds the threshold, local topology reconstruction is triggered. Reconstruction processes these 3 triggering grids and their 9 adjacent grids, totaling 12 grids, representing 24% of all grids. An event-driven mechanism encapsulates the triggering event into a message and sends it to a Kafka topic queue. The message queue supports parallel processing by 4 consumers, each handling the topology reconstruction task for 3 grids. The reconstruction process recalculates the topological relationships between 286 spatiotemporal objects and updates the weights of 453 edges, such as... Figure 3 The diagram shows a comparison of the topology of a grid before and after an update, including two subgraphs: (A) is a schematic diagram of the grid topology before the update, and (B) is a schematic diagram of the grid topology after the update. In Figure (B), green represents the grid topology coordinates after the update. Redis distributed cache stores the topology subgraphs of 50 grids, achieving a cache hit rate of 87%. The cache expiration time is set to 30 minutes. Asynchronous parallel processing enables the reconstruction of 4 grids to be completed within 18.6 seconds. The output updated topology network contains 1893 nodes and 7421 edges, reflecting the spatial relationships of the latest observation data.

[0108] In the uncertainty quantification stage of sparse observation data interpolation, the Bayesian Skrigin interpolation method was used to process a 500 km x 300 km sea area covered by 18 buoy stations. The prior probability distribution adopted a Gaussian distribution with a mean of 19.3℃ and a standard deviation of 2.1℃. A spherical model was fitted using variogram analysis, with model parameters including a nugget value of 0.8, a sill value of 3.6, and a range of 58 km. Monte Carlo simulation generated 10,000 random samples and interpolated at a sparse data point in the center of the sea area. The interpolation results of all samples were statistically analyzed, with a mean of 18.7℃ used as the final estimate and a standard deviation of 1.3℃ used as the uncertainty measure. The confidence label was calculated as 2.1 divided by the reciprocal of 1.3, which equals 0.62. Figure 4 The figure shows the distribution of confidence labels across the entire sea area. Confidence labels near buoy sites are above 0.85, while those in sparsely populated areas are below 0.55. Confidence labels are added to the 1893 nodes of the updated topological network to provide decision-makers with a basis for identifying high-risk and low-confidence areas.

[0109] In the spatial topology output stage, standardized abstraction maps 1893 spatiotemporal objects to nodes. Node attributes include spatial coordinates, temperature, salinity, flow velocity, water level, timestamp, and confidence label. 7421 topological relationships are mapped to edges, with edge attributes including relationship type, strength weight, and confidence label. The dynamic decay rate rule adopts an exponential decay function, with a spatial decay coefficient set to 0.1 per kilometer and a temporal decay coefficient set to 0.05 per hour. The relationship strength is calculated when the distance between nodes is 15 kilometers and the time difference is 3 hours. The decayed weight is 0.223 times the original weight. The adjacency matrix is ​​constructed as a 1893-row, 1893-column square matrix with 7421 non-zero elements and a sparsity of 99.79%. The edge list records the start identifier, end identifier, and weight of each of the 7421 edges. The JSON output file is 8.7MB in size and contains two top-level objects: a node list and an edge list. Each element in the node list contains the fields id, longitude, latitude, temperature, salinity, velocity, water level, timestamp, and confidence. Each element in the edge list contains the fields source_id, target_id, relation_type, weight, and confidence. Standardized topological relationship data is input into a large-scale marine forecasting and disaster reduction model. The model performs spatial queries to identify 112 ports and 81 breakwaters within the storm surge impact area. Path planning calculates the shortest path length of the storm surge from the open sea to the coast to be 78 kilometers. Disaster impact assessment predicts the inundation depth at 23 breakwater inundation points to be between 0.5 meters and 2.3 meters, providing a decision-making basis for the emergency command center to formulate evacuation plans and material allocation plans.

[0110] The technological advancements of this invention compared to traditional methods are reflected in both the real-time processing of spatiotemporal data and the high efficiency of topological relationship maintenance. Traditional methods employ CPU serial processing for spatial index queries, with computational complexity increasing linearly with the amount of data. For query tasks involving thousands of spatiotemporal grid nodes, response times can reach the order of seconds or even tens of seconds, failing to meet the real-time requirements of marine disaster early warning. This invention utilizes a GPU multi-layer CUDA parallel computing architecture. The first-layer sub-thread grid performs coarse screening of the main filter, while the second-layer sub-thread grid performs precise calculations of the auxiliary filter. By leveraging the parallel capabilities of tens of thousands of threads, the query response time is reduced to the order of milliseconds, enabling instant retrieval of spatiotemporal data and saving valuable time for emergency decision-making. Traditional methods require rebuilding the global topological relationship network when the marine environment changes. The computational load is proportional to the square of the network size, resulting in update delays of several minutes to tens of minutes, and the topological relationships cannot reflect the latest observational data in a timely manner. This invention employs an incremental topology update strategy, which decomposes the global network into independent sub-regions using spatial partitioning technology. Combined with an event-driven mechanism, reconstruction is triggered only for local regions where data changes are significant. The computational cost is proportional to the scale of local changes. In scenarios of dynamic changes in the marine environment, local changes typically account for only 5% to 10% of the global changes. Incremental updates avoid processing the majority of regions that have not changed, reducing update latency to the second level. This ensures that the topology network continuously reflects the latest marine conditions, supports the real-time response of the marine forecasting and disaster reduction model to sudden disaster events, and significantly improves the timeliness and accuracy of disaster prevention and mitigation.

[0111] It should be noted that the variables involved in this invention are explained in detail in Tables 2 and 3.

[0112] Table 2. Variable Explanation Table (Part 1)

[0113]

[0114] Table 3. Variable Explanation Table (Part Two)

[0115]

[0116] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for establishing the spatiotemporal topology of geographic elements for a large-scale marine forecasting and disaster mitigation model, characterized in that, Includes the following steps: S1. Multi-source heterogeneous data fusion and integration: Satellite remote sensing data, buoy observation data and forecast data are formatted using a unified data standard. Data comparability is ensured through data cleaning, format conversion and spatiotemporal registration. Noise reduction, interpolation and feature extraction methods are used to preprocess the raw data. Spatiotemporal alignment and semantic association techniques are used to map multi-source data to a unified coordinate system to establish logical relationships between data and output a standardized marine dataset. S2. Geographic Information Transcoding and Standardization Processing: Using a word segmenter and address component parsing, the semantic information of the textual description in the standardized marine dataset is converted into a structured standard format. Based on the Earth ellipsoid mathematical model and the Global Positioning System, the structured standard format is converted into precise latitude and longitude coordinates using triangulation. The latitude and longitude coordinates are then standardized, stored, and managed using spatial database standards to form spatial location data. S3. Spatiotemporal Indexing and Topological Relationship Construction: Establish an efficient indexing mechanism for the spatial location data. Use quadtrees combined with Z-curve encoding for spatial dimensionality reduction, and use B+ trees to index the time dimension. Quickly filter the candidate set through the main filter and then perform precise geometric calculations through the auxiliary filter. Start the maximum number of CUDA streams of the GPU. Each CUDA stream starts two layers of sub-thread grids, including a first layer and a second layer. The first layer of sub-thread grids performs the minimum bounding rectangle fast exclusion operation of the main filter, and the second layer of sub-thread grids performs the precise geometric intersection judgment operation of the auxiliary filter. Reduce the query response time from seconds to milliseconds and output the spatiotemporal indexing results. S4. Topological Relationship Reasoning and Dynamic Process Object-Oriented Modeling: Based on the spatiotemporal index results, construct a geometric and semantic association network of three-dimensional spatial objects to support the calculation of adjacency, intersection, and inclusion relationships. Based on the object-oriented data model, abstract storm surge overtopping into dynamic line objects and establish topological adjacency relationships with the surrounding coastline. Encapsulate time series data into spatiotemporal objects with life cycle attributes and track the evolution characteristics of the spatiotemporal objects through a version management mechanism to output the initial topological relationship network. S5. Dynamic maintenance and incremental update of topology: The initial topology network is divided into independent sub-regions using spatial partitioning technology. Only the affected regions are locally reconstructed. Updates are triggered by an event-driven mechanism. Asynchronous parallel processing is achieved using message queues and distributed caching technology. When the change in the observed data of a certain independent sub-region exceeds a preset threshold, the local topology reconstruction process is triggered to reduce the computational load and response latency. The updated topology network is then output. S6. Quantification of interpolation uncertainty of sparse observation data: A prior probability distribution is established by combining the Bayesian Skrigin interpolation method with historical statistical data. Spatial correlation is quantified by variogram analysis. The range of interpolation uncertainty is evaluated by Monte Carlo simulation. Confidence labels are added to the updated topological relationship network to provide risk assessment basis for downstream decision-making. The topological relationship network with confidence is output. S7. Spatial Topology Output: Through standardized abstraction, the geographic elements in the confidence-based topology network are transformed into node and edge structures. Parameter attributes are assigned to the nodes, and the spatial topology is encoded into an adjacency matrix or edge list using graph theory methods. The network weights are adjusted by incorporating dynamic rules based on decay rate. The input is adapted to JSON or CSV format to ensure the integrity of node ID, connection relationship, and quantified feature fields. Standardized topology data is output for use by the marine forecasting and disaster reduction large model.

2. The method for establishing the spatiotemporal topology of geographic elements for a large-scale marine forecasting and disaster reduction model according to claim 1, characterized in that, In the multi-source heterogeneous data fusion and integration step, spatiotemporal registration adopts adaptive grid reconstruction based on flow field characteristics and vortex sensing alignment algorithm to achieve dynamic coupling of multi-scale nesting.

3. The method for establishing the spatiotemporal topology of geographic elements for a large-scale marine forecasting and disaster reduction model according to claim 2, characterized in that, In the multi-source heterogeneous data fusion and integration process, the noise reduction process adopts a spatiotemporal continuity test model and LSTM-Autoencoder anomaly detection technology, combined with edge computing to achieve real-time streaming cleaning and feedback learning.

4. The method for establishing the spatiotemporal topology of geographic elements for a large-scale marine forecasting and disaster reduction model according to claim 3, characterized in that, In the multi-source heterogeneous data fusion and integration step, the interpolation method introduces a four-dimensional variational interpolation with physical constraints and a temperature-salinity-flow collaborative interpolation model, and uses a graph neural network to learn nonlinear relationships.

5. The method for establishing the spatiotemporal topology of geographic elements for a large-scale marine forecasting and disaster reduction model according to claim 4, characterized in that, In the geographic information transcoding and standardization process, the Earth ellipsoid mathematical model adopts the WGS84 coordinate system, and the triangulation method determines the spatial coordinates of the target point by solving a system of nonlinear equations based on at least three known reference points.

6. The method for establishing the spatiotemporal topology of geographic elements for a large-scale marine forecasting and disaster reduction model according to claim 5, characterized in that, In the geographic information transcoding and standardization process, the spatial database standard follows the OGC Simple Features specification, supporting the storage indexing of point, line, and polygon geometric objects and spatial relationship query operations.

7. The method for establishing the spatiotemporal topology of geographic elements for a large-scale marine forecasting and disaster reduction model according to claim 6, characterized in that, The first layer of sub-threaded grid consists of multiple thread blocks, each containing multiple threads for parallel execution of the minimum bounding rectangle fast exclusion operation of the main filter.

8. The method for establishing the spatiotemporal topology of geographic elements for a large-scale marine forecasting and disaster reduction model according to claim 7, characterized in that, The second-level sub-thread grid starts after the first-level sub-thread grid outputs its candidate set.

Images