Forestry fire risk assessment method and system based on multi-source data fusion

Abstract

The present application relates to the technical field of forestry disaster early warning, in particular to a forestry fire risk assessment method and system based on multi-source data fusion, comprising: integrating meteorological, thermal infrared and soil moisture multi-source data of the target forest area, and forming a unified standardized feature field through assimilation processing. Through a structure self-adaptive analysis engine, the correlation dimensions in the feature field are dynamically decoupled and reorganized to construct a multi-dimensional fire risk situation profile reflecting a specific risk driving mechanism. Based on this profile, combined with a historical fire case library, scenario matching and deduction operation are carried out to generate a dynamic prediction result of fire threat indicating a specific spatial position and a predictive evolution path. The method realizes dynamic diagnosis of the internal mechanism of forest fire risk and accurate prediction of the spatio-temporal evolution process of the threat, providing a direct basis for targeted prevention and control resource allocation.

Description

Technical Field

[0001] This invention relates to the field of forestry disaster early warning technology, and in particular to a method and system for forestry fire risk assessment based on multi-source data fusion. Background Technology

[0002] Existing forestry fire risk assessment methods are typically based on multi-source data integration and analysis. The mainstream technical approach involves pre-assigning fixed weights to factors such as meteorology, remote sensing, and the environment, or establishing static statistical and empirical models for calculation. While these methods can output a comprehensive risk level, their model structure is rigid and cannot effectively depict the dynamic coupling relationships and nonlinear dominant patterns among various risk-driving factors under different spatiotemporal contexts. Because they employ a static factor superposition logic, the assessment results essentially reflect the average contribution of each element, making it difficult to deeply diagnose the unique intrinsic driving mechanisms of forest fire occurrence and development in specific periods and regions, resulting in insufficient depth and timeliness in risk interpretation.

[0003] In the risk prediction stage, existing technologies mainly generate static risk probability maps or risk zoning maps for specific moments or short future periods. While these results can identify high-risk areas, they cannot further reveal where the threat will specifically originate or its subsequent spatiotemporal evolution path. This leads to the allocation of prevention and control resources being largely based on macro-level probabilistic judgments, lacking precise and forward-looking insights into the specific locations and spread directions of potential fire sources, thus limiting the targeted and proactive nature of measures.

[0004] Current technological systems need to overcome the limitations of static models and develop an assessment method capable of analyzing the dynamic structural relationships between risk factors in real time and further simulating and deducing the specific spatiotemporal evolution paths of threats. This invention aims to address this need. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a forestry fire risk assessment method and system based on multi-source data fusion.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a forestry fire risk assessment method based on multi-source data fusion, comprising: Receive and integrate various types of raw data about the target forest area, wherein the various types of raw data at least cover periodic meteorological observation records, surface thermal infrared image sequences, and soil moisture sampling values ​​at different depth levels; Data assimilation processing is performed on the integrated raw data of various types to generate a set of standardized feature fields with a unified spatiotemporal reference. The standardized feature fields characterize the state of abnormal heat accumulation, the rate of continuous water loss of surface combustibles, and the potential activity of near-surface turbulent motion. Using a structural adaptive analysis engine, the interrelated dimensions in the standardized feature field are dynamically decoupled and recombined to construct a multidimensional fire risk profile. The multidimensional fire risk profile is used to indicate the combination of internal driving factors and external expansion conditions of forest fires. Based on event patterns in the historical fire case database, scenario matching and extrapolation calculations are performed on the multidimensional fire risk profile to output fire threat prediction results with explicit spatial location and future evolution path. Based on the spatial distribution pattern of the fire threat prediction results, a targeted prevention and control measure configuration scheme with hierarchical differences is derived in reverse.

[0007] Preferably, the data assimilation processing performed on the integrated original data of various types generates a set of standardized feature fields with a unified spatiotemporal reference. The specific implementation method is as follows: The temporal variation curves of wind speed vector, relative humidity, and precipitation were separated from the periodic meteorological observation records. The pixel-level brightness temperature in the surface thermal infrared image sequence is converted into the actual surface temperature, and the variability distribution map of the surface temperature field between consecutive images is calculated. By associating soil moisture sampling values ​​at different depth levels with their corresponding geographical locations, three-dimensional raster data of soil moisture content is generated through spatial interpolation. An energy-moisture exchange balance equation is established that synchronously constrains the wind speed vector, air relative humidity, and the variability distribution map of the surface temperature field. Solving the energy-moisture exchange balance equation yields a quantitative characterization of the abnormal heat accumulation state. The time-series variation curve of the precipitation is used as input to drive a water transport model. The water transport model combines the three-dimensional grid data of soil moisture content to simulate the spatial distribution of the continuous water loss rate of surface combustibles. By utilizing the pulsating characteristics of wind speed vectors and the spatial heterogeneity of the surface temperature field, a microscale turbulence parameterization scheme is constructed, which outputs a classification of the potential activity level of near-surface turbulent motion.

[0008] Preferably, the step of using a structural adaptive analysis engine to dynamically decouple and recombine the interrelated dimensions in the standardized feature field to construct a multidimensional fire risk profile includes the following steps: Identify local peak regions in the quantitative characterization of the abnormal heat accumulation state, and extract the geometric center, temperature gradient direction, and energy accumulation rate of each peak region; The spatial distribution of the continuous water loss rate of the surface combustibles was analyzed in relation to the topographic relief and vegetation type layers, and sub-regions with different combustible dryness sensitivity levels were divided. The classification results of the potential activity level of the near-surface turbulent motion are mapped onto a three-dimensional spatial grid to form a three-dimensional structure of the turbulent energy field; Design a coupling feedback mechanism that enables bidirectional information flow among the energy accumulation rate, the combustible material drying sensitivity level, and the three-dimensional structure of the turbulent energy field. Through the aforementioned coupling feedback mechanism, the contribution weights of each element are iteratively adjusted until the system state converges, thereby generating a stable multidimensional fire risk profile that reflects the interaction of multiple factors.

[0009] Preferably, the process of performing scenario matching and extrapolation calculations on the multidimensional fire risk profile based on event patterns in the historical fire case database, and outputting fire threat prediction results with explicit spatial location and future evolution path, is executed sequentially: Retrieve historical cases from the historical fire case database that are similar to the current multidimensional fire risk profile in terms of the combination of driving factors, and extract the coordinates of the ignition point, the initial spread rate and the final fire shape of the historical cases; The local peak regions and sub-regions with different combustible dryness sensitivity levels in the current situation profile are compared with the retrieved historical cases to calculate a similarity matrix; Based on the similarity matrix, one or more of the most similar historical cases are assigned as reference templates for each currently identified potential risk area; Using the initial spread rate and final burned shape of the reference template as a baseline, and combining the currently acquired real-time wind speed vector with the three-dimensional structure of the turbulent energy field, dynamic simulation of fire behavior is performed to predict multiple development paths for each potential risk area. By integrating the predicted paths of all potential risk areas and overlaying the terrain barrier effect, the fire threat prediction results are generated, which are labeled with the probability of fire hazard range under different time slices.

[0010] Preferably, the step of deriving a targeted prevention and control measure configuration scheme with hierarchical differences based on the spatial distribution pattern of the fire threat prediction results further includes: The evolution of the fire hazard range probability under different time slices in the fire threat prediction results was analyzed to identify the core transmission channel with the fastest increase in fire hazard range probability and the areas with a high probability of stable maintenance. For the core transmission channel, its spatial orientation, width variation and spread resistance are calculated, and a dynamically variable width physical isolation zone is planned accordingly. The width of the planned route is positively correlated with the rate of increase of the probability of fire hazard range. For areas with a high probability of stability, assess their internal combustible material load, water accessibility, and transportation accessibility to generate a tiered water storage point layout plan and site selection recommendations for forward fire station points. The spatial overlay analysis of the planned route of the dynamically variable width physical isolation zone and the layout scheme of the graded water storage points is performed to identify the overlapping areas and blank areas of the measures, and the intensity of the measures is redistributed. The final document is a configuration plan for the targeted prevention and control measures, which includes the type of measure, the location of implementation, the intensity level, and the priority of execution.

[0011] Preferably, the design of the dynamically variable width physical isolation strip route is implemented in the following ways: Obtain the continuous centerline of the core transmission channel, as well as the surface cover type data within a preset range on both sides of the channel; Along the continuous centerline, sampling points are set at fixed intervals. At each sampling point, the probability value of the fire risk range and its gradient of the corresponding location in the fire threat prediction result are read. Based on the fire hazard range probability value and its gradient, the reference width of the physical isolation zone at the sampling point is calculated using a preset width mapping function. Based on the land cover type data at the sampling points, the baseline width is corrected: if it is flammable vegetation, the width is increased; if it is a natural barrier zone, the width is decreased. Connect the corrected width values ​​of all sampling points to form a dynamically variable width physical isolation strip planning line with a smooth width change.

[0012] Preferably, the method further includes a model self-evolution step, which is automatically triggered after each risk assessment task is completed, and includes: Record all the various types of raw data used in this risk assessment task, the intermediately generated standardized feature fields, the multidimensional fire risk situation profile, and the final output fire threat prediction results; Continuously monitor the actual fire situation in the target forest area, and once a real fire point is detected, immediately capture its occurrence time, accurate coordinates, and initial spread trajectory; The occurrence time and coordinates of the actual fire points are compared with all the fire threat prediction results output in the historical assessment tasks to find the prediction records where there is fire risk at the corresponding time and location. If a matching prediction record is found, all the data and intermediate states that the prediction record relied on when it was generated are extracted, and it is bundled with the actual development data of the real fire point and stored as a new positive sample case in the historical fire case library. Periodically perform cluster analysis on all cases in the historical fire case database and update the feature weights and similarity calculation rules used for scenario matching.

[0013] Preferably, the step of retrospectively comparing the occurrence time and coordinates of the actual fire point with all the fire threat prediction results output in the historical assessment task, using a spatiotemporal convolutional matching algorithm, includes the following steps: A four-dimensional search space is established with the coordinates of the actual fire point as the center. The four dimensions include longitude, latitude, time window, and prediction probability threshold. Within the four-dimensional search space, a spatiotemporal convolution kernel of a preset size is slid, which is given high weight in the near time dimension and Gaussian decay in the spatial dimension. For each probability grid in the fire threat prediction result, calculate its matching response value with the current spatiotemporal convolution kernel. The matching response value takes into account spatial distance, time difference and probability value itself. When a grid cell with a matching response value exceeding a set threshold is found in the search space, it is determined to be a successful prediction backtracking. Record all successful backtracking matching events and associate them with the corresponding evaluation task metadata for subsequent model evolution analysis.

[0014] Preferably, before receiving and integrating various types of raw data about the target forest area, a preprocessing stage for data quality enhancement and completion is further included, wherein the preprocessing stage performs: Outlier detection is performed on the input periodic meteorological observation records, and the missing or abnormal data is interpolated using the spatiotemporal correlation of neighboring stations to form a continuous and complete meteorological data sequence. The surface thermal infrared image sequence is subjected to cloud mask removal and atmospheric correction, and a terrain correction model is used to eliminate radiation distortion caused by slope and aspect, generating an accurate surface temperature product. For the soil moisture sampling values ​​at different depth levels, their spatial distribution uniformity is checked, and a surrogate model based on the topographic moisture index is introduced to simulate sparsely sampled areas in order to increase data density. The processed meteorological data sequences, surface temperature products, and enhanced soil moisture data are uniformly resampled to the same spatial resolution and map projection coordinate system to complete the data preparation work.

[0015] Preferably, the present invention also includes a forestry fire risk assessment system based on multi-source data fusion, the system including a processor and a memory that are communicatively connected to each other, the memory storing a computer program, and the processor being configured to execute the computer program to implement the forestry fire risk assessment method based on multi-source data fusion as described above.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: By introducing a structure-adaptive analysis engine, dynamic decoupling and recombination of multi-dimensional information in a standardized feature field are achieved. This engine can automatically identify real-time correlation structures and dominant coupling patterns between dimensions such as heat anomalies, combustible water loss rates, and turbulence activity based on real-time data streams. It separates and reconstructs complex, intertwined signals into a situation profile reflecting the current specific risk mechanism. This breaks free from the constraints of fixed model structures, enabling risk assessment to respond sensitively to unique risk combinations under different weather systems and underlying surface conditions, improving the depth and accuracy of risk diagnosis from superficial observations to underlying driving mechanisms.

[0017] Based on a historical fire case database, scenario matching and extrapolation calculations are performed to compare and simulate the current dynamically diagnosed risk profile with the "event patterns" of historical fire occurrences and developments. The output of this process is no longer an abstract risk level, but a fire threat prediction result containing a clear geographical location and predictable time trajectory. This technical approach achieves a leap from outputting "risk status" to predicting "threat processes," directly providing spatial location information of potential fire sources and their possible spread directions and paths. This enables fire command to plan barrier deployment, forward deployment of teams, and monitoring priorities based on a clear and forward-looking dynamic threat map.

[0018] The combination of structural adaptive analysis and case scenario simulation forms a closed loop of "dynamic mechanism diagnosis - process-based threat prediction." The former ensures the scientific rigor and sensitivity of risk identification, while the latter transforms scientific identification into spatiotemporal intelligence that can be directly used for action decisions. This technical framework transforms the results of forest fire risk assessment from guiding risk background judgments to actionable, process-based, and spatially explicit threat early warnings, directly supporting the formulation of more precise, proactive, and timely hierarchical and categorized prevention and control strategies. Attached Figure Description

[0019] Figure 1 This is a flowchart of the forestry fire risk assessment method based on multi-source data fusion as described in this invention; Figure 2 A three-dimensional visualization of the multi-dimensional fire risk profile; Figure 3 A flowchart for generating standardized feature fields for data assimilation processing; Figure 4 A flowchart for constructing a multidimensional fire risk profile for structural adaptive analysis; Figure 5 A comparative bar chart showing the weighting adjustments for fire risk characteristics; Figure 6A heatmap showing the standardization level of features from multiple data sources. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0021] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0022] See Figure 1 The system receives and integrates various types of raw data from the target forest area, including periodic meteorological observation records, surface thermal infrared image sequences, and soil moisture sampling values ​​at different depths. Data assimilation processing is performed on these integrated raw data to generate a set of standardized feature fields unified on a spatiotemporal basis. These feature fields quantitatively describe the state of abnormal heat accumulation, the rate of continuous water loss from surface combustibles, and the potential activity level of near-surface turbulent motion. Using a structural adaptive analysis engine, the interrelated dimensions in the standardized feature fields are dynamically decoupled and recombined to construct a multidimensional fire risk profile that indicates the combination of internal driving factors and external expansion conditions of forest fires. Based on event patterns stored in a historical fire case database, scenario matching and extrapolation calculations are performed on this multidimensional fire risk profile, outputting fire threat prediction results with explicit spatial location and future evolution paths. Based on the spatial distribution pattern of this fire threat prediction result, targeted prevention and control measures with hierarchical differences are derived in reverse.

[0023] The structural adaptive analysis engine consists of a feature decoupling module, a correlation identification module, a weight adjustment module, and a profile reconstruction module connected sequentially. The feature decoupling module receives a standardized feature field and separates the intertwined signals from three dimensions—abnormal heat accumulation state, continuous water loss rate of surface combustibles, and potential activity level of near-surface turbulent motion—into independent feature vectors. The correlation identification module calculates the Pearson correlation coefficient between each independent feature vector to identify the dominant coupling mode in the current spatiotemporal scenario. The weight adjustment module dynamically allocates the contribution weights of the three feature vectors according to the dominant coupling mode. The profile reconstruction module reassembles the weighted feature vectors into a three-dimensional data cube according to spatial location, which is the multidimensional fire risk situation profile.

[0024] See Figure 2 This is a 3D visualization of a multidimensional fire risk profile, clearly presenting the core information of a four-dimensional data cube: spatial location, energy state, combustible material dryness, and turbulence state. The red / orange points in the diagram are mainly concentrated in areas with high spatial and energy state dimensions, while the combustible material and turbulence dimensions are also at a medium-to-high level. This corresponds to a high-incidence fire scenario involving the coupling of "heat anomalies, combustible material dryness, and active turbulence." The blue points are mostly distributed in areas with lower energy state dimensions. Even with differences in spatial location and combustible material conditions, insufficient energy accumulation significantly reduces the overall risk, demonstrating the dominant role of heat anomalies in the coupling mode. The distribution of risk points is not linear, indicating that the model achieves dynamic adjustment and convergence of the three variables through a two-way information transmission link, ultimately outputting stable risk assessment results and fully presenting the final outcome of the structural adaptive analysis engine.

[0025] In one embodiment of the present invention, see [reference] Figure 3 The data assimilation process receives and integrates periodic meteorological observation records, surface thermal infrared image sequences, and soil moisture sampling values ​​at different depths for the target forest area. Temporal variation curves of wind speed vectors, relative humidity, and precipitation are extracted from the periodic meteorological observation records. These curves are recorded in a spatially distributed network of meteorological stations with hourly or daily time steps. Pixel-level brightness and temperature data from the surface thermal infrared image sequences are converted into true surface temperatures through radiometric calibration and emissivity correction. A variability distribution map of the surface temperature field between two or more consecutive images is calculated, reflecting the intensity of surface temperature variation over time and its spatial distribution characteristics. Soil moisture sampling values ​​at different depths (surface, shallow, and deep) are correlated with corresponding geographic location information. A three-dimensional raster data representation of soil moisture content is generated using a Kriging spatial interpolation algorithm. The raster data has a uniform resolution on the horizontal plane and distinguishes different depth layers vertically.

[0026] In some embodiments, an energy-moisture exchange balance equation is established that synchronously constrains the variability distribution maps of wind speed vectors, relative humidity, and surface temperature fields. Solving the energy-moisture exchange balance equation yields a quantitative characterization of the abnormal heat accumulation state. The energy-moisture exchange balance equation expresses the coupling relationship between surface energy budget and water transport processes, and its discretized form can be expressed as:

[0027] in: A quantitative indicator representing an abnormal accumulation of heat; The rate of change of time for each pixel in the variability distribution map representing the surface temperature field; Relative humidity of the air; This is the wind speed vector; It is the spatial Laplace operator of the Earth's surface temperature field, characterizing the degree of planar diffusion of temperature; coefficients , and These are weighted parameters calibrated based on land surface type and climate regionalization. The solution process is performed independently on each spatial grid cell, ultimately outputting a spatially continuous [database / structure]. Value raster layer. Taking the temporal variation curve of precipitation as input, a physical process-based water transport model is driven. The water transport model combines three-dimensional raster data of soil moisture content to simulate the spatial distribution of the continuous water loss rate of surface combustibles.

[0028] The water transport model employs stratified soil hydrodynamic equations to describe the water transport process within the soil-vegetation-atmosphere continuum. The model input includes hourly precipitation time series, soil volumetric water content at each depth layer in the 3D raster data, and the corresponding surface temperature of the raster. The output is the diurnal variation rate of the surface combustible water content. The 3D raster data is divided into independent computational units according to horizontal resolution. Each computational unit corresponds to a set of soil volumetric water contents at different depths. Hourly precipitation is input into the stratified hydrodynamic equation of the corresponding computational unit according to the time step. Combined with surface temperature, vegetation transpiration and soil evaporation water consumption are calculated. The variation rate of surface combustible water content in each computational unit is obtained through differential iteration. Finally, the results of all computational units are stitched together to form a spatially distributed raster layer of surface combustible water loss rate. Utilizing the pulsating characteristics of wind speed vectors and the spatial heterogeneity of the surface temperature field, a microscale turbulence parameterization scheme is constructed. This scheme outputs a classification of the potential activity level of near-surface turbulent motion. The pulsation characteristics of wind speed vectors are obtained by calculating the turbulence intensity and standard deviation of the observation sequence. The spatial heterogeneity of the surface temperature field is quantified by calculating the standard deviation of the surface temperature in the local area. The microscale turbulence parameterization scheme integrates these two factors and divides the calculation results into multiple discrete activity levels according to a pre-set threshold range.

[0029] Optionally, when separating variables from periodic meteorological observation records, the original time-series data can be filtered and smoothed to eliminate high-frequency noise. When calculating the variability distribution map of the surface temperature field, time-series image registration techniques can be used to ensure a one-to-one correspondence of spatial pixels. When generating three-dimensional raster data of soil moisture content, a digital elevation model can be introduced as a covariate in the spatial interpolation process to improve accuracy in complex terrain areas. Optionally, the energy-water exchange balance equation can be solved using iterative numerical methods until the residuals converge. The water transport model can be driven by integration operations with a daily time step. The hierarchical classification results of the microscale turbulence parameterization scheme can be stored as integer values, with larger values ​​representing higher potential activity of near-surface turbulent motion.

[0030] In one embodiment of the present invention, see [reference] Figure 4The construction of a multidimensional fire risk profile begins with the analysis of quantitative characterization data on anomalous heat accumulation. Local peak regions are identified in the raster map representing these anomalous heat accumulation, and continuous pixels with values ​​exceeding the surrounding background threshold are aggregated into independent peak region entities. For each identified local peak region, the geometric center coordinates are extracted, the direction of the maximum temperature gradient within the region is calculated as the temperature gradient direction, and the average rate of energy accumulation within the region is statistically analyzed as the energy accumulation rate. The spatial correspondence between the spatial distribution map of the continuous water loss rate of surface combustibles and the pre-input topographic relief digital elevation model layer and vegetation type classification layer is analyzed. Through overlay analysis and statistical induction, sub-regions with different combustible dryness sensitivity levels are delineated; for example, areas with high water loss rates and coniferous forest or shrubland vegetation are classified as high-sensitivity areas.

[0031] In some embodiments, the classification results of the potential activity level of near-surface turbulent motion are mapped from a two-dimensional plane to a three-dimensional spatial regular grid containing elevation information. Based on topographic elevation and atmospheric boundary layer height parameters, interpolation or assignment is performed in the vertical direction to form a three-dimensional structure of the turbulent energy field. A coupled feedback mechanism is designed, using energy accumulation rate, combustible dryness sensitivity level, and the three-dimensional structure of the turbulent energy field as the core variables of the dynamic system, constructing a two-way information transmission link: when the energy accumulation rate increases, the activity weight of the corresponding region in the three-dimensional structure of the turbulent energy field is simultaneously increased; when the activity of a certain region in the three-dimensional structure of the turbulent energy field increases, the rate of decrease in the combustible dryness sensitivity level of that region is simultaneously accelerated; when the combustible dryness sensitivity level decreases, the calculation weight of the energy accumulation rate is simultaneously enhanced. In each iteration, the contribution weights of the three variables are adjusted according to preset rules. The iteration stops when the relative changes of the three variables are all less than 0.01, outputting a stable multidimensional fire risk profile. The profile is a multidimensional data cube whose dimensions include spatial location, energy state, combustible dryness state, and turbulence state.

[0032] It is understandable that when performing scenario matching and extrapolation based on event patterns in a historical fire case database, the database is searched using combinations of driving factors extracted from a multidimensional fire risk profile. This retrieves historical cases similar in terms of heat accumulation patterns, combustible dryness spatial patterns, and turbulence conditions. The ignition point coordinates, initial spread velocity after the fire, and final burnt shape polygon are extracted for each matching historical case. A quantitative comparison is then performed between the set of local peak regions and the set of sub-regions with different combustible dryness sensitivity levels in the current multidimensional fire risk profile and the corresponding features of each retrieved historical case. A feature similarity matrix is ​​calculated, and the elements in the similarity matrix... Indicates the first The current potential risk area and the first The comprehensive similarity between historical cases is calculated using a formula that takes into account differences in spatial patterns, numerical values, and structural features.

[0033] Optionally, calculate the overall similarity. The formula can be expressed as:

[0034] in: Indicates the current number The geometric center of each local peak region and its historical position The Euclidean distance between the coordinates of the fire ignition points in each case; It is the current number Normalized values ​​of combustible dryness sensitivity levels for each sub-region; It is the first in history Normalized values ​​of the dryness of representative combustibles at the time of each case; It is the current number The range of peak regions and historical levels The intersection-union ratio between the over-exposed polygonal shapes in each case; It is a distance attenuation parameter; These are weighting coefficients corresponding to spatial distance, dryness difference, and shape similarity, respectively. .

[0035] Based on the calculated similarity matrix, one or more historical cases with the highest similarity are assigned as reference templates for each currently identified potential risk area. Using the initial spread rate and final fire shape recorded in the reference templates as the simulation baseline, and combining the currently acquired real-time wind speed vector data with the constructed 3D structure of the turbulent energy field, a dynamic simulation model of fire behavior based on cellular automata or physical equations is driven to predict multiple possible development paths for each potential risk area under various meteorological scenarios. The predicted fire spread paths of all potential risk areas are merged, and the terrain barrier effects obtained from digital terrain model analysis, such as the inhibitory effects of ridgelines and rivers on fire spread, are overlaid to generate the final fire threat prediction result. The fire threat prediction result is presented in the form of a series of time-slice raster maps, with each raster cell labeled with the probability value of fire occurrence in that future time slice.

[0036] In some embodiments, the feature comparison process can incorporate a dynamic time warping algorithm to match the similarity of time series features. The dynamic simulation model of fire behavior can combine the Lagrange particle diffusion method to simulate the occurrence and propagation probability of flying fires. When generating fire threat prediction results, the Monte Carlo ensemble method can be used to fuse the probabilities of multiple possible development paths. Optionally, the event patterns in the historical fire case database can include meteorological reanalysis data, satellite observation fire point sequences, and post-disaster on-site investigation reports. Before calculating the similarity matrix, all feature vectors can be standardized to eliminate the influence of dimensions. When assigning reference templates, a similarity threshold can be set; assignment is only performed when the highest similarity exceeds this threshold, otherwise it is marked as a new risk pattern. The terrain barrier effect can be quantified as a reduction factor for the spread rate and applied in the dynamic simulation process of fire behavior.

[0037] In one embodiment of the present invention, a targeted prevention and control measure configuration scheme is derived by reverse deduction based on the spatial distribution pattern of fire threat prediction results, and the evolution process of the raster data sequence of fire hazard range probability under different time slices in the fire threat prediction results is analyzed. By calculating the fire hazard range probability difference of the same spatial grid cell between adjacent time slices, the core transmission channel with the fastest increase in fire hazard range probability and the high-probability stable area with a continuously higher fire hazard range probability than the high threshold are identified. The core transmission channel is manifested as a strip or linear corridor formed by the continuous spatial distribution of high probability difference pixels, and the high-probability stable area is manifested as a concentrated contiguous area where the probability value remains at a high level in multiple consecutive time slices.

[0038] In some embodiments, for the identified core transmission channels, the spatial orientation, width variation, and spread resistance of the core transmission channels are calculated. The spatial orientation is obtained by extracting the central axis of the core transmission channel and calculating its azimuth angle. The width variation is obtained by performing profile analysis perpendicular to the central axis to obtain a width index representing the probability difference distribution. The spread resistance is comprehensively estimated by overlaying terrain slope data and vegetation flammability rating data. Based on these calculated parameters, a dynamically variable width physical buffer zone planning route is designed. The width of the physical buffer zone planning route is positively correlated with the rate of increase in the fire hazard probability at the corresponding location of the core transmission channel. The design of the dynamically variable width physical buffer zone planning route is achieved as follows: Continuous centerline vector data of the core transmission channel and land cover type classification raster data within a preset buffer zone on both sides of the channel centerline are obtained. A series of sampling points are set at fixed intervals along the continuous centerline. At each sampling point, the fire hazard probability value and its spatial gradient value for the corresponding geographical location in the fire threat prediction results for a specific prediction time period are read. Based on the fire hazard probability value and its spatial gradient value, a preset width mapping function is used to calculate the baseline width of the physical buffer zone at the sampling point.

[0039] In some embodiments, for areas identified as having a high probability of stability, the combustible load, water accessibility, and transportation accessibility within these areas are assessed. Combustible load is estimated based on remotely sensed leaf area index, vegetation type, and forest age data; water accessibility is measured by calculating the shortest surface path distance from the area to the nearest natural or artificial water source; and transportation accessibility is measured by calculating the shortest time distance from the area to the existing forest road network. Based on the assessment results, a tiered water storage point deployment plan and site selection recommendations for forward fire station locations are generated. The tiered water storage point deployment plan specifies the location, capacity level, and construction priority of new water storage facilities, while the site selection recommendations for forward fire station locations specify the location of the station, the recommended equipment type, and personnel size.

[0040] Optionally, a spatial overlay analysis is performed between the vector map layer of the dynamically variable width physical firebreak planning route and the spatial point layer of the tiered water storage point layout scheme. The spatial overlay analysis identifies overlapping areas where the physical firebreak and water storage point locations overlap, as well as high-risk areas not covered by either the planned physical firebreak route or the water storage points. For overlapping areas, the intensity of fire prevention measures is redistributed and optimized; for example, the density of water storage points can be appropriately reduced or the width of the firebreak can be optimized in the overlapping areas. For areas without measures, suggestions for the placement of supplementary lookout posts or emergency material storage points are generated. Finally, a structured, targeted prevention and control measure configuration plan document is generated, containing detailed information on measure types, geographical coordinates of implementation, intensity levels, and recommended priority of implementation, in the form of tables and map attachments.

[0041] In one embodiment of the invention, the model self-evolution stage is automatically triggered upon completion of each risk assessment task. This stage records all types of raw data used in the risk assessment task, the intermediately generated standardized feature fields, the multi-dimensional fire risk profile, and the final output fire threat prediction results. These records, along with the unique identifier of the risk assessment task and the task execution timestamp, are stored in the assessment task metadata database. The system continuously monitors the actual fire situation in the target forest area using methods including satellite remote sensing hotspot monitoring networks, ground lookout reports, and aerial patrol information. Once a real fire point is detected, the system immediately captures the occurrence time, accurate coordinates, and initial spread trajectory obtained through satellite image analysis or ground reports.

[0042] In some embodiments, the occurrence time and coordinates of the actual fire point are compared retrospectively with all fire threat prediction results stored in the historical assessment task metadata database to find prediction records where fire risk exists at the corresponding time and predicted location. The retrospective comparison employs a spatiotemporal convolutional matching algorithm. The steps of the spatiotemporal convolutional matching algorithm are as follows: a four-dimensional search space is established centered on the geographical coordinates of the actual fire point. The four dimensions of the four-dimensional search space include longitude, latitude, a time window, and a prediction probability threshold. The longitude and latitude dimensions define the spatial search range, the time window defines the time length for retrospectively searching historical prediction results, and the prediction probability threshold defines the minimum probability of fire risk range among the fire threat prediction results considered valid predictions. Within the four-dimensional search space, a spatiotemporal convolution kernel of a preset scale is slidable. The spatiotemporal convolution kernel is a multi-dimensional array defining a weight distribution.

[0043] It is understandable that the spatiotemporal convolution kernel assigns higher weights to recent times in the time dimension, meaning that historical predictions closer to the actual fire occurrence time receive higher weights; in the spatial dimension, the spatiotemporal convolution kernel uses Gaussian decay weights centered on the coordinates of the actual fire point. For each probability grid cell of a historical fire threat prediction result falling within the four-dimensional search space, the matching response value between the probability grid cell and the current spatiotemporal convolution kernel is calculated. The matching response value comprehensively considers spatial distance, time difference, and the probability value itself. The calculation formula integrates spatial response With time response :

[0044] in: This represents the response components that decay Gaussian based on spatial distance; This represents the response components that decay linearly or exponentially based on the time difference; It is the weighting coefficient for the fusion of spatial and temporal components; It is the probability value of the probability grid cell. The threshold function, when When the probability is below the predicted probability threshold ,otherwise When a matching response value is found by calculating a sliding convolution within the search space. A successful prediction backtracking is defined as when the number of raster cells exceeds a set threshold. All successful backtracking matching events are recorded, and each matching event is associated with corresponding assessment task metadata, including the multidimensional fire risk profile data and input data version used to generate the prediction result, for subsequent model evolution analysis. See Table 1.

[0045] Table 1: Backtracking Matching Results Real fire point number Time of occurrence Fire point coordinates (longitude, latitude) Matching historical prediction record number Match response value Estimated record generation time F2023-045 2023-08-1514:30 118.752°E, 32.105°N EP20230814_1200_001 0.872 2023-08-1412:00 F2023-045 2023-08-1514:30 118.752°E, 32.105°N EP20230813_1800_005 0.641 2023-08-1318:00 If a matching prediction record is found, all types of raw data and intermediate states used in generating the prediction record are extracted from the assessment task metadata database. This extracted data is then bundled with the actual development data of real fire points and stored as a new positive sample case in the historical fire case database. Cluster analysis is periodically performed on all cases in the historical fire case database to update the feature weights and similarity calculation rules used for scenario matching. The cluster analysis employs an unsupervised machine learning algorithm, grouping cases based on their multidimensional fire risk profile feature vectors, and recalculating or adjusting the parameters in the feature weights and similarity calculation rules during the feature comparison process based on the grouping results.

[0046] In some embodiments, the captured initial spread trajectory can be used to correct the fire spread parameters in the reference template. The length of the time window can be set as the effective forecast duration of the fire threat prediction result. The prediction probability threshold can be dynamically adjusted based on statistical analysis of historical prediction performance. Optionally, the scale of the spatiotemporal convolution kernel can be configured according to the terrain complexity and data spatial resolution. When recording matching events, the probability value and time difference of the matching location can be saved simultaneously as additional information for evaluating prediction accuracy. The triggering of periodic cluster analysis can be set to the number of new cases in the historical fire case database reaching a certain threshold or a fixed time period. The feature weights can be updated using a redistribution method based on the cluster center distance.

[0047] See Figure 5 This is a bar chart comparing the weight adjustments of fire risk characteristics, clearly showing the changes in the weight allocation of the five major risk characteristics in the forestry fire risk assessment model before and after its self-evolution. After the model evolution, the overall weight allocation focuses more on the two most direct fire driving factors: heat anomalies and combustible dryness, while slightly increasing the influence of topography, demonstrating its ability to self-optimize based on real fire point data. The direction of the weight adjustment is highly consistent with the scientific understanding of forestry fires, proving the effectiveness and rationality of the model's self-evolution mechanism. By regularly generating such weight comparison charts, the technical team can track the model's evolution trend over the long term and promptly identify any anomalies or deviations in weight allocation from business logic.

[0048] In one embodiment of the present invention, the preprocessing stage for data quality enhancement and completion is performed before receiving and integrating various types of raw data. The preprocessing stage performs outlier detection on the input periodic meteorological observation records. Outlier detection employs statistical distribution-based methods, such as calculating the absolute median difference of each meteorological element's time series, and marking observation data exceeding a preset threshold as outliers. The spatiotemporal correlation of neighboring stations is used to imputate the data marked as missing or outliers. Leveraging the high correlation between spatially adjacent meteorological stations with similar topography and climate in their time series, substitute values ​​are generated using linear regression or Kriging interpolation methods to form a continuous and complete meteorological data sequence. Cloud mask removal and atmospheric correction are performed on the surface thermal infrared image sequence. Cloud mask removal identifies and removes pixels covered by clouds by utilizing the difference in reflectance characteristics between the thermal infrared band and the visible light band. Atmospheric correction uses an atmospheric radiative transfer model to eliminate the influence of atmospheric absorption and scattering on the surface thermal radiation signal. The terrain correction model is used to eliminate radiation distortion caused by slope and aspect. Based on the digital elevation model, the terrain correction model calculates the solar incidence angle and sensor observation angle for each pixel, performs cosine correction on the surface temperature, and generates an accurate surface temperature product.

[0049] In some embodiments, for soil moisture sampling values ​​at different depth levels, the spatial distribution uniformity of the soil moisture sampling values ​​is examined, a spatial density distribution map of the sampling points within the study area is calculated, and sparse sampling areas with sampling point density below a set threshold are identified. In these sparse sampling areas, a surrogate model based on the topographic moisture index is introduced to simulate and increase data density. The topographic moisture index uses a digital elevation model to calculate topographic slope and runoff accumulation area to characterize the potential spatial distribution trend of soil moisture. The surrogate model is established based on the regression relationship between soil moisture sampling values ​​in densely sampled areas and the corresponding topographic moisture index. This relationship is applied to sparsely sampled areas to generate simulated soil moisture values. The processed meteorological data sequences, surface temperature products, and the increased soil moisture data are uniformly resampled to the same spatial resolution and map projection coordinate system to complete the data preparation.

[0050] It is understandable that outlier detection methods based on absolute median difference are suitable for a time series dataset of a meteorological element. First, calculate the median. Then calculate the absolute deviation of each data point from the median. Next, the median of these absolute deviations is calculated, i.e., the absolute median difference. Any satisfaction data points Considered outliers, among which It is an adjustable scaling factor, typically correlated with the expected distribution of the data. For labeled outliers or original missing values, neighboring site data is utilized. Interpolation can be used to establish, for example... The linear relationship, where the coefficients and It was obtained by fitting historical synchronous observation data from two stations during normal periods.

[0051] See Figure 6 This is a heatmap showing the standardization level of multi-source data features. It clearly demonstrates the standardization effect of five types of raw data, including meteorological observations and thermal infrared imagery, when transformed into five types of fire risk features, such as heat anomalies and combustible material dehydration. This can directly guide multi-source data fusion strategies, such as prioritizing the quality of meteorological observations and topographic data, as they contribute the most to the standardization of core risk features. For combinations with low standardization levels (e.g., meteorological → vegetation type), it is advisable to consider introducing additional data sources or optimizing algorithms to improve the transformation effect. The heatmap reveals the matching efficiency of different data-feature pairs, helping technical teams identify weaknesses in the model and thus iterate in a targeted manner.

[0052] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A forestry fire risk assessment method based on multi-source data fusion, characterized in that, The method includes: Receive and integrate various types of raw data about the target forest area, wherein the various types of raw data at least cover periodic meteorological observation records, surface thermal infrared image sequences, and soil moisture sampling values ​​at different depth levels; Data assimilation processing is performed on the integrated raw data of various types to generate a set of standardized feature fields with a unified spatiotemporal reference. The standardized feature fields characterize the state of abnormal heat accumulation, the rate of continuous water loss of surface combustibles, and the potential activity of near-surface turbulent motion. Using a structural adaptive analysis engine, the interrelated dimensions in the standardized feature field are dynamically decoupled and recombined to construct a multidimensional fire risk profile. The multidimensional fire risk profile is used to indicate the combination of internal driving factors and external expansion conditions of forest fires. Based on event patterns in the historical fire case database, scenario matching and extrapolation calculations are performed on the multidimensional fire risk profile to output fire threat prediction results with explicit spatial location and future evolution path. Based on the spatial distribution pattern of the fire threat prediction results, a targeted prevention and control measure configuration scheme with hierarchical differences is derived in reverse.

2. The forestry fire risk assessment method based on multi-source data fusion according to claim 1, characterized in that, The process of performing data assimilation on the integrated original data of various types generates a set of standardized feature fields with a unified spatiotemporal reference. The specific implementation method is as follows: The temporal variation curves of wind speed vector, relative humidity, and precipitation were separated from the periodic meteorological observation records. The pixel-level brightness temperature in the surface thermal infrared image sequence is converted into the actual surface temperature, and the variability distribution map of the surface temperature field between consecutive images is calculated. By associating soil moisture sampling values ​​at different depth levels with their corresponding geographical locations, three-dimensional raster data of soil moisture content is generated through spatial interpolation. An energy-moisture exchange balance equation is established that synchronously constrains the wind speed vector, air relative humidity, and the variability distribution map of the surface temperature field. Solving the energy-moisture exchange balance equation yields a quantitative characterization of the abnormal heat accumulation state. The time-series variation curve of the precipitation is used as input to drive a water transport model. The water transport model combines the three-dimensional grid data of soil moisture content to simulate the spatial distribution of the continuous water loss rate of surface combustibles. By utilizing the pulsating characteristics of wind speed vectors and the spatial heterogeneity of the surface temperature field, a microscale turbulence parameterization scheme is constructed, which outputs a classification of the potential activity level of near-surface turbulent motion.

3. The forestry fire risk assessment method based on multi-source data fusion according to claim 2, characterized in that, The method employs a structural adaptive analysis engine to dynamically decouple and recombine the interrelated dimensions in the standardized feature field, constructing a multidimensional fire risk profile, including the following steps: Identify local peak regions in the quantitative characterization of the abnormal heat accumulation state, and extract the geometric center, temperature gradient direction, and energy accumulation rate of each peak region; The spatial distribution of the continuous water loss rate of the surface combustibles was analyzed in relation to the topographic relief and vegetation type layers, and sub-regions with different combustible dryness sensitivity levels were divided. The classification results of the potential activity level of the near-surface turbulent motion are mapped onto a three-dimensional spatial grid to form a three-dimensional structure of the turbulent energy field; Design a coupling feedback mechanism that enables bidirectional information flow among the energy accumulation rate, the combustible material drying sensitivity level, and the three-dimensional structure of the turbulent energy field. Through the aforementioned coupling feedback mechanism, the contribution weights of each element are iteratively adjusted until the system state converges, thereby generating a stable multidimensional fire risk profile that reflects the interaction of multiple factors.

4. The forestry fire risk assessment method based on multi-source data fusion according to claim 3, characterized in that, The event patterns based on the historical fire case database are used to perform scenario matching and extrapolation calculations on the multidimensional fire risk profile, outputting fire threat prediction results with explicit spatial location and future evolution path, executed sequentially: Retrieve historical cases from the historical fire case database that are similar to the current multidimensional fire risk profile in terms of the combination of driving factors, and extract the coordinates of the ignition point, the initial spread rate and the final fire shape of the historical cases; The local peak regions and sub-regions with different combustible dryness sensitivity levels in the current situation profile are compared with the retrieved historical cases to calculate a similarity matrix; Based on the similarity matrix, one or more of the most similar historical cases are assigned as reference templates for each currently identified potential risk area; Using the initial spread rate and final burned shape of the reference template as a baseline, and combining the currently acquired real-time wind speed vector with the three-dimensional structure of the turbulent energy field, dynamic simulation of fire behavior is performed to predict multiple development paths for each potential risk area. By integrating the predicted paths of all potential risk areas and overlaying the terrain barrier effect, the fire threat prediction results are generated, which are labeled with the probability of fire hazard range under different time slices.

5. The forestry fire risk assessment method based on multi-source data fusion according to claim 4, characterized in that, Based on the spatial distribution pattern of the fire threat prediction results, a targeted prevention and control measure configuration scheme with hierarchical differences is derived in reverse, further including: The evolution of the fire hazard range probability under different time slices in the fire threat prediction results was analyzed to identify the core transmission channel with the fastest increase in fire hazard range probability and the areas with a high probability of stable maintenance. For the core transmission channel, its spatial orientation, width variation and spread resistance are calculated, and a dynamically variable width physical isolation zone is planned accordingly. The width of the planned route is positively correlated with the rate of increase of the probability of fire hazard range. For areas with a high probability of stability, assess their internal combustible material load, water accessibility, and transportation accessibility to generate a tiered water storage point layout plan and site selection recommendations for forward fire station points. The spatial overlay analysis of the planned route of the dynamically variable width physical isolation zone and the layout scheme of the graded water storage points is performed to identify the overlapping areas and blank areas of the measures, and the intensity of the measures is redistributed. The final document is a configuration plan for the targeted prevention and control measures, which includes the type of measure, the location of implementation, the intensity level, and the priority of execution.

6. The forestry fire risk assessment method based on multi-source data fusion according to claim 5, characterized in that, The design of dynamically variable width physical isolation strips is implemented in the following ways: Obtain the continuous centerline of the core transmission channel, as well as the surface cover type data within a preset range on both sides of the channel; Along the continuous centerline, sampling points are set at fixed intervals. At each sampling point, the probability value of the fire risk range and its gradient of the corresponding location in the fire threat prediction result are read. Based on the fire hazard range probability value and its gradient, the reference width of the physical isolation zone at the sampling point is calculated using a preset width mapping function. Based on the land cover type data at the sampling points, the baseline width is corrected: if it is flammable vegetation, the width is increased; if it is a natural barrier zone, the width is decreased. Connect the corrected width values ​​of all sampling points to form a dynamically variable width physical isolation strip planning line with a smooth width change.

7. The forestry fire risk assessment method based on multi-source data fusion according to claim 4, characterized in that, The method also includes a model self-evolution phase, which is automatically triggered after each risk assessment task is completed, and includes: Record all the various types of raw data used in this risk assessment task, the intermediately generated standardized feature fields, the multidimensional fire risk situation profile, and the final output fire threat prediction results; Continuously monitor the actual fire situation in the target forest area, and once a real fire point is detected, immediately capture its occurrence time, accurate coordinates, and initial spread trajectory; The occurrence time and coordinates of the actual fire points are compared with all the fire threat prediction results output in the historical assessment tasks to find the prediction records where there is fire risk at the corresponding time and location. If a matching prediction record is found, all the data and intermediate states that the prediction record relied on when it was generated are extracted, and it is bundled with the actual development data of the real fire point and stored as a new positive sample case in the historical fire case library. Periodically perform cluster analysis on all cases in the historical fire case database and update the feature weights and similarity calculation rules used for scenario matching.

8. The forestry fire risk assessment method based on multi-source data fusion according to claim 7, characterized in that, The step of retrospectively comparing the occurrence time and coordinates of the actual fire point with all the fire threat prediction results output in the historical assessment tasks is carried out using a spatiotemporal convolutional matching algorithm, the steps of which are as follows: A four-dimensional search space is established with the coordinates of the actual fire point as the center. The four dimensions include longitude, latitude, time window, and prediction probability threshold. Within the four-dimensional search space, a spatiotemporal convolution kernel of a preset size is slid, which is given high weight in the near time dimension and Gaussian decay in the spatial dimension. For each probability grid in the fire threat prediction result, calculate its matching response value with the current spatiotemporal convolution kernel. The matching response value takes into account spatial distance, time difference and probability value itself. When a grid cell with a matching response value exceeding a set threshold is found in the search space, it is determined to be a successful prediction backtracking. Record all successful backtracking matching events and associate them with the corresponding evaluation task metadata for subsequent model evolution analysis.

9. The forestry fire risk assessment method based on multi-source data fusion according to claim 1, characterized in that, Before receiving and integrating various types of raw data about the target forest area, a preprocessing stage for data quality enhancement and completion is also included, which performs the following: Outlier detection is performed on the input periodic meteorological observation records, and the missing or abnormal data is interpolated using the spatiotemporal correlation of neighboring stations to form a continuous and complete meteorological data sequence. The surface thermal infrared image sequence is subjected to cloud mask removal and atmospheric correction, and a terrain correction model is used to eliminate radiation distortion caused by slope and aspect, generating an accurate surface temperature product. For the soil moisture sampling values ​​at different depth levels, their spatial distribution uniformity is checked, and a surrogate model based on the topographic moisture index is introduced to simulate sparsely sampled areas in order to increase data density. The processed meteorological data sequences, surface temperature products, and enhanced soil moisture data are uniformly resampled to the same spatial resolution and map projection coordinate system to complete the data preparation work.

10. A forestry fire risk assessment system based on multi-source data fusion, characterized in that, The system includes a processor and a memory that are interconnected, the memory storing a computer program, and the processor being configured to, when executing the computer program, implement the forestry fire risk assessment method based on multi-source data fusion as described in any one of claims 1 to 9.

Images