A forest fire risk intelligent prediction method and system based on a heterogeneous graph state space model

Abstract

The application provides a forest fire risk intelligent prediction method and system based on a heterogeneous graph state space model, comprising: constructing a unified grid multi-dimensional feature set; the unified grid multi-dimensional feature set comprises feature information of a plurality of grid points distributed in a target area; obtaining a human activity risk index; based on the unified grid multi-dimensional feature set and the human activity risk index, constructing a meteorological-grid heterogeneous graph state space model; based on the meteorological-grid heterogeneous graph state space model, determining the forest fire occurrence probability of each grid point. The method provided in the application constructs a folk custom and agricultural activity dynamic space-time calendar, uses an asymmetric Gaussian space-time kernel function to represent the human fire risk evolution, and reflects the fire risk law of "accumulating before the festival and remaining after the festival"; a physically constrained meteorological-grid heterogeneous graph state space model is constructed, interpolation distortion and error accumulation are avoided, and the forest fire prediction accuracy is improved by coupling double risks through an enhanced step length.

Description

Technical Field

[0001] This application relates to the field of fire monitoring and early warning technology, and in particular to a method and system for intelligent prediction of wildfire risk based on a heterogeneous graph state space model. Background Technology

[0002] Wildfire forecasting, as a crucial preliminary measure in forest fire prevention, enables early assessment of fire risks and spread trends. This provides a scientific basis for fire source control, patrol deployment, emergency response force dispatch, and personnel evacuation, achieving early detection, early suppression, and early extinguishment, effectively reducing the probability of fires and the scale of disasters. Simultaneously, it minimizes casualties, protects forest resources, residents' property, and critical infrastructure, and prevents problems such as vegetation destruction, biodiversity loss, soil erosion, and air pollution. It also mitigates post-disaster ecological restoration and economic losses, playing a vital role in maintaining regional ecological security, public safety, and social stability.

[0003] Existing wildfire risk prediction technologies suffer from two main shortcomings. First, they do not adequately consider peak daily fire usage periods, lacking dynamic representation of the temporal continuity of fire risk changes during modeling, making it difficult to accurately reflect the objective evolutionary pattern of fire risk "accumulating before holidays and lingering after holidays." Second, existing methods generally employ a two-stage approach of "interpolation first, prediction later," typically using algorithms such as inverse distance weighting and Kriging interpolation to interpolate discrete meteorological station data into a regular prediction grid before inputting it into the model for prediction. These interpolation methods are all based on the assumption of spatial stationarity, which easily leads to significant spatial smoothing distortion in complex terrain areas with significant microclimate abrupt changes, such as ridges, windward slopes, and leeward slopes. Furthermore, they cannot effectively constrain information transmission by incorporating prior physical information such as terrain shielding and wind direction, resulting in the continuous accumulation of prediction errors across multiple stages. Summary of the Invention

[0004] This application provides a method and system for intelligent prediction of wildfire risk based on a heterogeneous graph state-space model. To solve the above-mentioned technical problems, this application adopts the following technical methods: In a first aspect, this application provides a method for intelligent prediction of wildfire risk based on a heterogeneous graph state-space model, characterized by comprising: Construct a unified grid multidimensional feature set; the unified grid multidimensional feature set includes feature information of multiple grid points distributed in the target area; Obtain a risk index for human activities; Based on the unified grid multidimensional feature set and the human activity risk index, a meteorological-grid heterogeneous graph state space model is constructed. Based on the meteorological-grid heterogeneous graph state space model, the probability of wildfire occurrence at each grid point is determined.

[0005] Optionally, the step of constructing a unified grid multidimensional feature set includes the following: Construct a unified spatial prediction grid covering the target area; Multi-source heterogeneous data are spatially aligned and fused, and mapped to the unified spatial prediction grid to form a unified grid multi-dimensional feature set; the multi-source heterogeneous data includes meteorological data, combustible load data, topography data, and historical labels.

[0006] Optionally, the step of spatially aligning and fusing multi-source heterogeneous data and mapping it to the unified spatial prediction grid to form a unified grid multidimensional feature set includes: Acquire time-series observation data of temperature, humidity, wind speed, and precipitation from meteorological stations within the target area, while preserving the discrete station observation characteristics; The combustible load data is acquired, and the combustible load pixel values ​​are aligned to the corresponding grid points of the unified spatial prediction grid through spatial matching and resampling. Based on DEM data, elevation, slope, and aspect topographic factors are extracted for each grid point, and historical label data is overlaid. Through the physical constraint weighted aggregation mechanism of heterogeneous graph neural network, the discrete observation features of meteorological stations are directly mapped to the prediction grid nodes to form a unified grid multidimensional feature set.

[0007] Optionally, the process of determining the human activity risk index includes the following steps: Based on information on folk festivals and agricultural activities in the target region, a dynamic spatiotemporal calendar of folk customs and agricultural activities is constructed. Based on the aforementioned dynamic spatiotemporal calendar of folk customs and agricultural activities, the risk radiation value of all risk events in the target area to the surrounding spatiotemporal space is determined; the risk radiation value is calculated by an asymmetric Gaussian spatiotemporal kernel function, which uses differentiated precursor and dissipation time scale parameters to characterize the asymmetric evolution law of human fire activities over time. Based on the aforementioned risk radiation value, a human activity risk index is determined.

[0008] Optionally, the risk radiation value includes folk custom risk radiation value and agricultural risk radiation value; the step of determining the risk radiation value of all risk events in the target area to the surrounding spacetime based on the dynamic spatiotemporal calendar of folk customs and agriculture includes: Based on the aforementioned dynamic spatiotemporal calendar of folk customs and agricultural activities, determine the working dates and folk festivals corresponding to agricultural activities; Obtain meteorological data for the working days corresponding to the agricultural activities, and determine whether the corresponding grid points meet the conditions for agricultural burning based on the meteorological data; If the conditions are met, the agricultural fire risk event at the corresponding grid point within the target area is activated, and the agricultural risk radiation value of that grid point is calculated. Based on the aforementioned folk festivals, calculate the folk risk radiation value of the corresponding grid points within the target area.

[0009] Optionally, determining the human activity risk index based on the risk radiation value includes: The agricultural risk radiation value and the folk custom risk radiation value are superimposed, and then... Function normalization is used to determine the risk index of human activities.

[0010] Optionally, the step of constructing a meteorological-grid heterogeneous graph state-space model based on the unified grid multidimensional feature set and the human activity risk index includes: Based on the unified grid multidimensional feature set and the human activity risk index, a source node set, a target node set, and an edge set are constructed. The source node set consists of multiple discrete meteorological station nodes, each of which is bound to real-time temperature, humidity, wind direction, and wind speed features. The target node set consists of multiple grid nodes, each of which is bound to combustible material load, terrain, and human activity risk index features. The edge set contains only directed edges from meteorological station nodes to predicted grid nodes. There are no connecting edges between meteorological station nodes or between predicted grid nodes, and each predicted grid node only establishes directed edges with a preset number of meteorological station nodes with the closest horizontal projection distance. The information transmission weight of the directed edges is calculated by fusing distance attenuation, altitude correction, and physical similarity scores of wind field direction. The information transmission intensity conforms to the physical diffusion law of meteorological elements. Based on the source node set, the target node set, and the edge set, a meteorological-grid heterogeneous graph state space model is constructed.

[0011] Optionally, determining the probability of wildfire occurrence at each grid point based on the meteorological-grid heterogeneous graph state-space model includes: Based on the meteorological-grid heterogeneous graph state space model, the information transmission weights of directed edges are calculated; Based on the information transmission weights, calculate the complete spatial representation vector of the grid node; Based on the complete spatial representation vector, the probability of wildfire occurrence at each grid point is determined.

[0012] Optionally, determining the probability of wildfire occurrence at each grid point based on the complete spatial representation vector includes: The complete spatial representation vector is input into the Mamba state space model to output high-dimensional spatiotemporal features; The high-dimensional spatiotemporal features are input into the output prediction head, which outputs the probability of wildfire occurrence at each grid point.

[0013] This application has the following beneficial effects: The method proposed in this application constructs a dynamic spatiotemporal calendar of folk agricultural activities, uses an asymmetric Gaussian spatiotemporal kernel function to characterize the evolution of human fire risk, and reflects the pattern of fire risk "accumulation before the festival and residual fire after the festival". It also constructs a physically constrained meteorological-grid heterogeneous graph state space model to avoid interpolation distortion and error accumulation, and improves the accuracy of wildfire prediction by enhancing the step size coupling of dual risks. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating a method for intelligent wildfire risk prediction based on a heterogeneous graph state-space model, provided in an embodiment of this application. Detailed Implementation

[0015] To facilitate understanding by those skilled in the art, the present application will be further described below in conjunction with embodiments and accompanying drawings. The content mentioned in the embodiments is not intended to limit the present application.

[0016] To solve the above technical problems, such as Figure 1 As shown, this application proposes a wildfire risk intelligent prediction method based on a heterogeneous graph state-space model, including: Step S101: Construct a unified grid multidimensional feature set; the unified grid multidimensional feature set includes feature information of multiple grid points distributed in the target area; A unified spatial prediction network covering the target area is constructed. This target area is typically a forest area, mountainous region, or key forest fire prevention and control zone, containing multiple grid points (M grid points). Multi-source heterogeneous data, including meteorological data, combustible load data, topographic data, and historical labels, are spatially aligned and fused, mapped to the unified spatial prediction grid to form a unified grid multi-dimensional feature set. The alignment and fusion process is as follows: Obtain the distribution within the target area Time-series observation data from several meteorological stations are collected, including temperature, humidity, wind speed, and precipitation. Temperature data typically includes daily average temperature and daily maximum / minimum temperature; humidity data typically includes daily relative humidity; and wind speed data includes daily average / maximum wind speed. The discrete station observation characteristics are preserved. Combustible material load data is acquired, and through spatial matching and resampling, the combustible material load pixel values ​​are aligned to the corresponding grid points in the unified spatial prediction grid. Based on DEM (Digital Elevation Model) data, elevation, slope, and aspect topographic factors are extracted for each grid point, and historical label data is overlaid. Through a physical constraint weighted aggregation mechanism of a heterogeneous graph neural network, the discrete observation features of the meteorological stations are directly mapped to the prediction grid nodes, forming a unified grid multidimensional feature set.

[0017] At this point, each grid point contains different feature information, which facilitates subsequent data processing.

[0018] Step S102: Obtain the human activity risk index; During folk festivals, people often engage in activities involving fire, such as sacrificial rites and incense burning; while during busy agricultural seasons, productive fires such as burning stubble and straw are also concentrated. These human-caused fire activities significantly increase the fire hazard in surrounding areas. Therefore, when conducting wildfire probability assessments, the fire risk factors caused by human activities should be fully considered. This study primarily employs a dual-engine triggering mechanism of "knowledge-driven (timetable) + data-driven (meteorological)" to transform folk and agricultural activities into a Human Activity Risk Index (HARI) field with clearly defined "latency period" and "dissipation period." The specific process is as follows: Based on information on folk festivals and agricultural activities in the target region, a dynamic spatiotemporal calendar for folk customs and agriculture is constructed. Using this calendar, the risk radiation value of all risk events within the target region on the surrounding spatiotemporal area is determined. This calendar incorporates and dynamically maintains a spatiotemporal knowledge graph based on a traditional calendar. It consists of three parts: (1) Folk festival mapping: Based on the lunar calendar, determine the peak period for fire use in sacrificial rites, such as the Spring Festival (the first to fifteenth day of the first lunar month), Qingming Festival, Zhongyuan Festival (the fifteenth day of the seventh lunar month), and Hanyi Festival (the first day of the tenth lunar month).

[0019] (2) Mapping of agricultural activities: Based on the twenty-four solar terms, the basic time window for agricultural fire use (such as burning stubble and straw burning) is defined. For example, the period from "Awakening of Insects" to "Qingming" is set as the peak period for burning stubble during spring plowing and preparation; the period from "Autumnal Equinox" to "Frost's Descent" is set as the peak period for concentrated burning of straw during autumn harvest.

[0020] (3) Dynamic maintenance table: The weight coefficients of the above timetable are dynamically modified according to the policies issued by the local agricultural or emergency departments (such as "comprehensive ban on burning" or "orderly restriction of burning window").

[0021] Since agricultural fires are highly dependent on weather conditions, a meteorological coupling activation logic is introduced. This involves determining the corresponding work dates and folk festivals for agricultural activities based on the dynamic spatiotemporal calendar of folk customs and agricultural activities. Then, meteorological data for these work dates is determined, and based on this data, it is judged whether the corresponding grid points meet the conditions for agricultural fire burning. (1) Number of consecutive rainless days (If there is no rain for more than 3 consecutive days and the combustibles are dry); (2) Daily average wind speed (The wind is moderate, suitable for farmers to carry out controlled burning). If the above conditions for agricultural burning are met, the agricultural fire hazard event at the corresponding grid point within the target area will be officially activated. And calculate the agricultural risk radiation value of the grid point.

[0022] Based on folk festivals, calculate the folk risk radiation value of the corresponding grid points in the target area.

[0023] For any activated folk event (fixed at) or agricultural events The risk radiation values ​​of folk customs and agricultural activities are calculated using an asymmetric Gaussian spatiotemporal kernel function. This function employs differentiated precursor and dissipation timescale parameters to characterize the asymmetric temporal evolution of human fire use activities, as shown in the following formula: (1) In the formula, This can represent the agricultural risk radiation value and the folk custom risk radiation value of this grid point. The definitions and sources of each parameter are explained below: (Risk Amplitude): Represents the basic strength or historical statistical weight of this type of event (e.g., a higher value of 2.5 can be set for the Qingming Festival when activities are concentrated; a lower value can be set for regular weekends).

[0024] (Spatial distance): Evaluation grid points The Euclidean distance or road network distance from the center of the risk (such as a cemetery, farmland, or road boundary).

[0025] (Spatial diffusion scale): The range (bandwidth) of the risk of control in geographic space. The larger the value, the wider the reach of human activities at the fire source.

[0026] (Current Time): The specific date and time of the current model evaluation.

[0027] (Peak event time): Legal holidays or the best weather days for farming (such as the day of Qingming Festival) locked by the knowledge calendar.

[0028] (Time-varying decay standard deviation): The core control parameter of this patent adopts an asymmetric design based on piecewise functions to reflect the time tail effect of fire hazards. (2) The parameter table for symmetrical folk events is shown in Table 1: ; Parameter tables for other types of folk events can be obtained from the corresponding websites, and will not be elaborated here.

[0029] The agricultural event parameter table is shown in Table 2: ; After superimposing the above-mentioned agricultural risk radiation values ​​and folk custom risk radiation values, through... After function normalization, the human activity risk index is obtained, as shown in the following formula: (3) in, for The total number of risk events that are currently in an active phase. This is determined by the amplitude of each risk kernel function. Furthermore, all exponential terms are always positive, and the sum of values ​​is... , The purpose of normalization is to softly compress the non-negative superposition values ​​to... The interval is used to avoid numerical explosion when multiple high-risk events overlap, while maintaining a non-linear response (i.e., the total risk accelerates after multiple medium-risk events overlap). The output HARI tensor has a dimension of [missing information]. .

[0030] Step S103: Based on the unified grid multidimensional feature set and the human activity risk index, construct a meteorological-grid heterogeneous graph state space model; Based on the aforementioned unified grid multidimensional feature set and the human activity risk index, a source node set, a target node set, and an edge set are constructed, specifically including: Source node set It consists of multiple discrete weather station nodes, each of which is bound to real-time temperature, humidity, wind direction and wind speed characteristics.

[0031] target node set It consists of multiple grid nodes; each grid node is bound to characteristics such as combustible material load, terrain, and human activity risk index.

[0032] Edge set E contains only directed edges from meteorological station nodes to prediction grid nodes. There are no connecting edges between meteorological station nodes and between prediction grid nodes. Each prediction grid node establishes directed edges only with a preset number of meteorological station nodes with the closest horizontal projection distance. The information transmission weight of the directed edges is calculated by fusing distance attenuation, altitude correction and physical similarity scores of wind field direction. The information transmission intensity conforms to the physical diffusion law of meteorological elements.

[0033] By combining the source node set, target node set, and edge set, a meteorological-grid heterogeneous graph state-space model can be constructed. .

[0034] For each prediction grid node The K meteorological station nodes with the closest horizontal projection distance are selected to form its neighborhood set. ,but: .

[0035] The meteorological-grid heterogeneous graph state-space model transforms spatial interpolation into a physically constrained and end-to-end learnable graph convolution process. The edge weights (information transmission strength) of the graph are determined by physical similarity scores (incorporating physical priors such as elevation difference blocking and wind field along-wind direction gain). Furthermore, the weight coefficients (e.g.) , The algorithm will be updated through backpropagation based on the final fire prediction loss.

[0036] Step S104: Based on the meteorological-grid heterogeneous graph state space model, determine the probability of wildfire occurrence at each grid point.

[0037] The core of graph convolution lies in calculating weather stations. For prediction grid Information transmission weight This application uses physical similarity scores normalized to the same value range. At each grid point Local Softmax normalization is performed within the nearest neighbor range.

[0038] First, for each prediction grid Select the nearest distance based on the two-dimensional horizontal projection distance (planar distance excluding elevation difference). Each weather station forms a local neighborhood set. Calculate attention weights within the set: (4) Similarity score Taking into account distance, altitude, and the physical effects of wind field, the formula is: (5) in: For learnable weight coefficients, satisfying During training, this is achieved by performing Softmax normalization on the three free parameters.

[0039] For weather station With predicted grid points The absolute horizontal spatial distance between them can be determined by extracting information from weather stations during implementation. With grid points The geographic coordinates (latitude and longitude) of the center are obtained using the spherical distance calculation method, or the horizontal Euclidean distance between the two is calculated under a unified projected coordinate system. The normalized parameter for the maximum distance in the region. The difference in elevation between two points This is an altitude-scale parameter (e.g., 500m). The wind field along-wind gain function: (6) The parameters are defined as follows: Weather station Measured wind speed at the location (unit: m / s); Normalized reference wind speed, which is the maximum wind speed recorded in historical meteorological observations within the study area (unit: m / s). The measured wind direction angle (unit: °) at meteorological station i follows meteorological standard conventions; : The azimuth angle (in °) from meteorological station i to prediction grid j, with true north as the reference and clockwise as positive.

[0040] Based on the information transmission weights obtained above We perform weighted aggregation of meteorological characteristics from local neighboring nodes. For any prediction grid... , its in Spatial fusion meteorological features received in real time The calculation is as follows: (7) in, For weather station exist The observed feature vector at time step.

[0041] Subsequently, the aggregated meteorological features were combined with the grid. By combining its own static terrain features and combustible load, a grid of nodes is formed. Complete spatial representation vector Each grid point obtains a high-dimensional spatial feature vector that integrates local micro-meteorological, topographical, and static attributes. Over time, the heterogeneous graph network outputs these spatial features daily, generating a continuous feature sequence of length T for each grid point. This continuous feature sequence will serve as the direct input for the next step of time series modeling.

[0042] To capture the long-term cumulative effects of meteorological elements, the complete spatial representation vector sequence of the above grid nodes is converted into the input of the time series model Mamba. Let the grid... The input time series is .

[0043] To deeply couple human activity risks with meteorological risks, this step designs a HARI modulation gating in the core state update equation of Mamba, adjusting the original discretized time step of Mamba. Replace with enhanced step size : (8) in For a two-layer perceptron: .

[0044] Modulation intensity coefficient (set) Physical meaning: When (In a period of extreme folk custom risk) It can approach the upper bound. The model's state step size increases by approximately 50%, forcing the network to rapidly update its internal state and making it extremely sensitive to small inputs (such as a sudden gust of wind); when At that time, the model tends to revert to a normal, stable long memory state.

[0045] The enhancement step size and input time series are as follows Substituting into the following formula, the high-dimensional spatiotemporal features of the Mamba model output are calculated as follows: (To clearly identify the corresponding grid node, it is denoted as) ): (9) (10) (11) (12) In the formula, A is the continuous-time parameter matrix (state transition matrix), B is the input matrix, and A and B are both learnable parameters of the model. It is the identity matrix. Output parameter matrix for continuous time.

[0046] The high-dimensional spatiotemporal features are input into the output prediction head (which includes a mapping layer and a Sigmoid activation function), and the output is the probability of wildfire occurrence for M grid points in the next T days: (13) in, , and These are the learnable parameters of the prediction head.

[0047] In summary, the method proposed in this application, on the one hand, constructs a dynamic spatiotemporal calendar of folk customs and agricultural activities and uses an asymmetric Gaussian spatiotemporal kernel function to determine the human activity risk index, thereby achieving a dynamic representation of the asymmetric evolution of human fire activity over time and truly reflecting the objective law of fire risk "accumulating before the festival and lingering after the festival." On the other hand, it constructs a meteorological-grid heterogeneous graph state-space model constrained by physical laws, directly mapping the discrete observation characteristics of meteorological stations to prediction grid nodes, effectively avoiding spatial smoothing distortion and cross-stage error accumulation caused by traditional spatial interpolation under complex terrain. At the same time, by introducing an enhanced step size into the state-space model, human activity risk and meteorological risk are deeply coupled, significantly improving the accuracy of wildfire probability prediction.

[0048] In some embodiments, this application also provides a computer system including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0049] This application also provides a computer-readable storage medium for storing a computer program. This computer-readable storage medium can be applied to a computer device, and the computer program causes the computer device to execute the corresponding processes in the methods described above in the embodiments of this application; for brevity, further details are omitted here.

[0050] The above embodiments are preferred implementations of this application. In addition, this application can be implemented in other ways. Any obvious substitutions without departing from the concept of this technical solution are within the protection scope of this application.

[0051] To facilitate understanding by those skilled in the art of the improvements made by this application compared to the prior art, some of the accompanying drawings and descriptions have been simplified, and for clarity, some other elements have been omitted from this application. Those skilled in the art should realize that these omitted elements may also constitute the content of this application.

Claims

1. A method for intelligent prediction of wildfire risk based on a heterogeneous graph state-space model, characterized in that, include: Construct a unified grid multidimensional feature set; The unified grid multidimensional feature set includes feature information of multiple grid points distributed in the target area; Obtain a risk index for human activities; Based on the unified grid multidimensional feature set and the human activity risk index, a meteorological-grid heterogeneous graph state space model is constructed. Based on the meteorological-grid heterogeneous graph state space model, the probability of wildfire occurrence at each grid point is determined.

2. The method according to claim 1, characterized in that, The steps for constructing a unified grid multidimensional feature set include: the following: Construct a unified spatial prediction grid covering the target area; Multi-source heterogeneous data are spatially aligned and fused, and mapped to the unified spatial prediction grid to form a unified grid multi-dimensional feature set; the multi-source heterogeneous data includes meteorological data, combustible load data, topography data, and historical labels.

3. The method according to claim 2, characterized in that, The step of spatially aligning and fusing multi-source heterogeneous data and mapping it to the unified spatial prediction grid to form a unified grid multidimensional feature set includes: Acquire time-series observation data of temperature, humidity, wind speed, and precipitation from meteorological stations within the target area, while preserving the discrete station observation characteristics; The combustible load data is acquired, and the combustible load pixel values ​​are aligned to the corresponding grid points of the unified spatial prediction grid through spatial matching and resampling. Based on DEM data, elevation, slope, and aspect topographic factors are extracted for each grid point, and historical label data is overlaid. Through the physical constraint weighted aggregation mechanism of heterogeneous graph neural network, the discrete observation features of meteorological stations are directly mapped to the prediction grid nodes to form a unified grid multidimensional feature set.

4. The method according to claim 3, characterized in that, The process of determining the human activity risk index includes the following steps: Based on information on folk festivals and agricultural activities in the target region, a dynamic spatiotemporal calendar of folk customs and agricultural activities is constructed. Based on the aforementioned dynamic spatiotemporal calendar of folk customs and agricultural activities, the risk radiation value of all risk events in the target area to the surrounding spatiotemporal space is determined; the risk radiation value is calculated by an asymmetric Gaussian spatiotemporal kernel function, which uses differentiated precursor and dissipation time scale parameters to characterize the asymmetric evolution law of human fire activities over time. Based on the aforementioned risk radiation value, a human activity risk index is determined.

5. The method according to claim 4, characterized in that, The risk radiation value includes the folk custom risk radiation value and the agricultural risk radiation value; The determination of the risk radiation value of all risk events within the target area to the surrounding spacetime based on the aforementioned dynamic spatiotemporal calendar of folk customs and agricultural activities includes: Based on the aforementioned dynamic spatiotemporal calendar of folk customs and agricultural activities, determine the working dates and folk festivals corresponding to agricultural activities; Obtain meteorological data for the working days corresponding to the agricultural activities, and determine whether the corresponding grid points meet the conditions for agricultural burning based on the meteorological data; If the conditions are met, the agricultural fire risk event at the corresponding grid point within the target area is activated, and the agricultural risk radiation value of that grid point is calculated. Based on the aforementioned folk festivals, calculate the folk risk radiation value of the corresponding grid points within the target area.

6. The method according to claim 5, characterized in that, The determination of the human activity risk index based on the risk radiation value includes: The agricultural risk radiation value and the folk custom risk radiation value are superimposed, and then... Function normalization is used to determine the risk index of human activities.

7. The method according to claim 6, characterized in that, The construction of a meteorological-grid heterogeneous graph state-space model based on the unified grid multidimensional feature set and the human activity risk index includes: Based on the unified grid multidimensional feature set and the human activity risk index, a source node set, a target node set, and an edge set are constructed. The source node set consists of multiple discrete meteorological station nodes, each of which is bound to real-time temperature, humidity, wind direction, and wind speed features. The target node set consists of multiple grid nodes, each of which is bound to combustible material load, terrain, and human activity risk index features. The edge set contains only directed edges from meteorological station nodes to predicted grid nodes. There are no connecting edges between meteorological station nodes or between predicted grid nodes, and each predicted grid node only establishes directed edges with a preset number of meteorological station nodes with the closest horizontal projection distance. The information transmission weight of the directed edges is calculated by fusing distance attenuation, altitude correction, and physical similarity scores of wind field direction. The information transmission intensity conforms to the physical diffusion law of meteorological elements. Based on the source node set, the target node set, and the edge set, a meteorological-grid heterogeneous graph state space model is constructed.

8. The method according to claim 7, characterized in that, The determination of the wildfire occurrence probability at each grid point based on the meteorological-grid heterogeneous graph state space model includes: Based on the meteorological-grid heterogeneous graph state space model, the information transmission weights of directed edges are calculated; Based on the information transmission weights, calculate the complete spatial representation vector of the grid node; Based on the complete spatial representation vector, the probability of wildfire occurrence at each grid point is determined.

9. The method according to claim 8, characterized in that, The determination of the wildfire occurrence probability at each grid point based on the complete spatial representation vector includes: The complete spatial representation vector is input into the Mamba state space model to output high-dimensional spatiotemporal features; The high-dimensional spatiotemporal features are input into the output prediction head, which outputs the probability of wildfire occurrence at each grid point.

10. A computer system, characterized in that, include: Memory is used to store instructions that can be executed by the processor; A processor for executing the instructions to implement the method as described in any one of claims 1 to 9.

Images