Method for constructing heavy and super heavy forest fire emergency rescue channel path based on knowledge graph

Abstract

The application discloses a knowledge graph-based heavy and super heavy forest fire emergency rescue channel path construction method, and belongs to the field of forest fire rescue, and comprises the following steps: S1, multi-source heterogeneous data fusion and standardization processing; S2, forest fire dynamic risk assessment; S3, knowledge graph entity extraction; S4, emergency rescue channel path construction; S5, calculating the feasibility score of the emergency rescue channel path. The above-mentioned knowledge graph-based heavy and super heavy forest fire emergency rescue channel path construction method is adopted, multi-modal data semantic association is coupled with real-time risk, unmanned aerial vehicle fleet cooperative reconnaissance is linked with path planning linkage, fire spread simulation is fused with knowledge graph reasoning, the breakthrough from static planning to dynamic adaptation of the emergency rescue channel path is realized, and the heavy and super heavy forest fire rescue efficiency is significantly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of forest fire rescue, and in particular to a method for constructing an emergency rescue channel path for a heavy or super heavy forest fire based on a knowledge graph. BACKGROUND

[0002] Forest fires are one of the major natural disasters that threaten ecological safety and human life and property, and have the characteristics of strong suddenness, great destructiveness, and rapid spread. In recent years, under the influence of global climate change, intensified drought, and human activities, heavy or super heavy forest fires have occurred frequently, posing a severe challenge to traditional emergency rescue channel path construction technology.

[0003] Currently, the construction of a heavy or super heavy forest fire rescue channel mainly relies on the following traditional methods:

[0004] 1. Manual planning based on experience and static maps: first-line commanders manually select a rescue channel path by combining historical experience and traditional GIS maps with the general situation of the fire site. This method relies on individual cognitive levels and lacks systematic integrated analysis of dynamic risk factors such as terrain, vegetation, and weather, making it difficult to adapt to complex scenarios with rapidly changing fire conditions.

[0005] 2. Local dynamic optimization method: Some solutions attempt to combine unmanned aerial vehicle reconnaissance or fire spread models for limited path optimization, but mostly remain at the application level of single-dimensional data (such as relying only on remote sensing images or single-point weather data), lacking the ability to deeply associate multi-modal data and couple semantic-level risks.

[0006] 3. Isolated case library and rule engine: Some systems store historical emergency rescue channel path construction cases, but lack graphed knowledge association and dynamic reasoning mechanisms, and are unable to automatically extract risk evolution rules or core constraints of similar scenarios, resulting in poor adaptability and low iteration efficiency of emergency rescue channel path planning. SUMMARY

[0007] The purpose of the present application is to provide a method for constructing an emergency rescue channel path for a heavy or super heavy forest fire based on a knowledge graph, which solves the above technical problems.

[0008] To achieve the above purpose, the present application provides a method for constructing an emergency rescue channel path for a heavy or super heavy forest fire based on a knowledge graph, comprising the following steps:

[0009] S1. Multi-source heterogeneous data fusion and standardized processing: integrate remote sensing, weather, historical fire, forest fire resources, and rescue text data, and obtain a feature dataset through format unification and feature extraction;

[0010] S2. Dynamic risk assessment of forest fires: use a forest fire risk prediction model trained based on the feature dataset obtained in step S1 to predict fire risks in real time;

[0011] S3, knowledge graph entity extraction: based on the feature data set obtained in step S1 and the fire risk real-time prediction result obtained in step S2, knowledge graph entities are extracted from multi-source heterogeneous data;

[0012] S4, emergency rescue channel path construction: taking the knowledge graph obtained in step S3 as a constraint, the unmanned aerial vehicle reconnaissance trajectory is determined through a double deep Q network, and the A* algorithm is used to generate the emergency rescue channel path combined with the fire observation data on the unmanned aerial vehicle reconnaissance trajectory;

[0013] S5, verify the emergency rescue channel path by calculating the path feasibility score, and update the knowledge graph described in step S3 based on the score result and real-time fire, realize the dynamic optimization of the emergency rescue channel path.

[0014] Therefore, the present application adopts the above-mentioned knowledge graph-based heavy and super heavy forest fire emergency rescue channel path construction method, which has the beneficial effects of:

[0015] 1. Through knowledge graph unified semantic modeling, accurately associate global risk entities with multi-modal features, and improve risk quantification accuracy;

[0016] 2. Fusion of high-frequency dynamic data (unmanned aerial vehicle reconnaissance + meteorological update) and graph reasoning, realize real-time dynamic optimization of path (minute-level response to fire mutation);

[0017] 3. Build a multi-dimensional composite evaluation system (fire out-of-control rate, traffic capacity, infrared detection accuracy, etc.), ensure path safety redundancy and extreme scene adaptation.

[0018] In summary, the present application realizes entity semantic association and rule reasoning by fusing multi-source remote sensing, meteorological and text data, and constructing dynamic knowledge graph; combined with deep reinforcement learning (DRL) to optimize unmanned aerial vehicle group reconnaissance trajectory, and link fire spread simulation (FARSITE) and real-time risk coupling analysis; innovatively establish a complete closed-loop mechanism covering "data fusion-risk modeling-path generation-dynamic verification-knowledge archiving", solve the problems of multi-source heterogeneous data fragmentation, dynamic adaptability lag, safety redundancy deficiency, knowledge reuse inefficiency and collaborative communication bottleneck in traditional methods, realize the fundamental breakthrough of emergency rescue channel path from "passive experience response" to "active intelligent prediction", and significantly improve the heavy and super heavy fire rescue efficiency and safety.

[0019] The technical solutions of the present application will be further described in detail below with the help of the drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 The flowchart of the present application based on knowledge graph heavy and super heavy forest fire emergency rescue channel path construction method. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the embodiments of the present application are further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the embodiments of the present application and not to limit the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application. Examples of the embodiments are shown in the drawings, wherein the same or similar reference signs represent the same or similar elements or elements having the same or similar functions throughout.

[0022] It should be noted that the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or server comprising a series of steps or units does not have to be limited to only those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0023] The embodiments of the present application are described in detail below with reference to the drawings.

[0024] As shown in Figure 1 The knowledge graph-based super-large forest fire emergency rescue channel path construction method comprises the following steps:

[0025] S1, multi-source heterogeneous data fusion and standardization processing: integrating remote sensing, meteorological, historical fire, forest fire resources and rescue text data, obtaining feature data set through format unification and feature extraction;

[0026] Step S1 specifically comprises the following steps:

[0027] S11, multi-source heterogeneous data acquisition: acquiring remote sensing, meteorological, historical fire, forest fire resources and rescue text data T s of the target area, wherein the remote sensing data includes vegetation distribution data, fire distribution data, road and river system GIS data, the meteorological data includes monthly precipitation m P , potential evapotranspiration PEI, daily average relative temperature Hu and daily average relative humidity Te, wind speed and wind direction;

[0028] S12, data format unification and resampling: converting the collected multi-source heterogeneous data into UTM coordinates, and resampling to a unified resolution after standardization processing to obtain a standardized data set;

[0029] S13, feature extraction: extracting vegetation features V C and meteorological features NC , fire characteristics I C , terrain characteristics T f , text remote sensing correlation characteristics TRS and forest fire resource characteristics F r , and after min-max standardization, the feature data set F = {V C , N C , I C , T f , TRS, F r} is obtained.

[0030] In step S13, the vegetation characteristics V C include the normalized vegetation index NDVI, the vegetation moisture index GVMI, the fuel load classification L c and the vegetation coverage Ve;

[0031] The normalized vegetation index NDVI is expressed as follows:

[0032]

[0033] In the formula, NIR represents the near-infrared band reflectivity of the remote sensing satellite; R R represents the red band reflectivity of the remote sensing satellite;

[0034] The vegetation moisture index is expressed as follows:

[0035]

[0036] In the formula, SWIR represents the short-wave infrared band reflectivity of the remote sensing satellite;

[0037] The fuel load classification L c is expressed as follows:

[0038]

[0039] In the formula, l b , m b and h b represent low flammability, medium flammability and high flammability, respectively; F l represents the fuel load, with the unit of t / hm 2 ;

[0040] The vegetation coverage Ve is expressed as follows:

[0041]

[0042] In the formula, NDVI min and NDVI max represent the minimum and maximum values of the normalized vegetation index in the target area, respectively;

[0043] Meteorological characteristics N C including standardized precipitation evapotranspiration index SPEI, humidity-temperature index HTI, wind speed W S and wind direction θ d , wherein the standardized precipitation evapotranspiration index is expressed as follows:

[0044]

[0045] In the formula, σ P-PET represents the standard deviation of historical monthly precipitation m P and potential evapotranspiration PEI;

[0046] The humidity-temperature index HTI is expressed as follows:

[0047] HTI = 0.6 × (1 - Te / 100) + 0.4 × (Hu / 35);

[0048] Forest fire characteristics I C including forest fire density D t , fire radiation power FPR, and the shortest straight line distance d between real-time fire points and road entities Fire-Road , wherein the forest fire density D t is expressed as follows:

[0049]

[0050] In the formula, ΔN i represents the number of newly added fire points in the target area within a time period Δt; S target represents the area of the target area;

[0051] Topographic characteristics T f including slope, and the slope Sl is expressed as follows:

[0052]

[0053] In the formula, and respectively represent the elevation change rates in x and y directions, with units of °;

[0054] The shortest straight line distance d between real-time fire points and road entities is expressed as follows: Fire-Road

[0055]

[0056] In the formula, (x Fire , y Fire ) represents the UTM coordinates of the fire point; (x Road , y Road ) represents the UTM coordinates of the road node; ​

[0057] Forest fire fighting resource characteristics F r Including response radius r i and maximum water supply distance D max-water ;

[0058] The steps for obtaining the text remote sensing association feature TRS are as follows:

[0059] The first step is to use the BERT-base model to analyze the text data T on forest fire prevention resources and rescue. s Perform text semantic encoding to obtain text vectors:

[0060] Vec(T s ) = BERT(s tokenized );

[0061] In the formula, Vec(T) s ) represents the semantic vector of forest fire fighting resources and rescue text data; BERT(s tokenized This indicates the use of the BERT-based model for semantic encoding.

[0062] Step 2: Remote sensing image feature extraction: A CNN network is used to extract infrared image features (R) from the remote sensing images. i Output image features:

[0063] Feat(r i ) = CNN(i n );

[0064] In the formula, Feat(R) i ) represents infrared image feature R i Feature vectors; CNN(i n ) represents the pixel value i normalized to [0,1]. n Feature extraction is performed using a CNN network;

[0065] Step 3: Extracting Text Remote Sensing Association Features (TRS):

[0066]

[0067] In the formula, Sim(T) s ,R i ) represents the cosine similarity between the text vector and the image feature; Sim(·) represents the cosine similarity function.

[0068] S2. Dynamic risk assessment of forest fires: Real-time prediction of fire risk using a forest fire risk prediction model trained on the feature dataset obtained in step S1.

[0069] Step S2 specifically includes the following steps:

[0070] S21, constructing a forest fire risk prediction model based on an XGBoost model;

[0071] S22, determining optimal initialization hyperparameters of the forest fire risk prediction model through a TPE algorithm of an Optuna framework;

[0072] S23, dividing the feature data set F into a training set and a validation set, first inputting the training set into the forest fire risk prediction model for determining the optimal initialization hyperparameters for training until the area under the curve AUC of the set validation set reaches a set value, determining that the forest fire risk prediction model converges, and stopping training;

[0073] S24, collecting real-time multi-source heterogeneous data of a target area and inputting the trained forest fire risk prediction model to output a real-time predicted fire risk value risk, and dividing the fire risk value risk into grades to obtain a fire risk grade F l ;

[0074] S3, knowledge graph entity extraction: based on the feature data set obtained in step S1 and the fire risk real-time prediction result obtained in step S2, extracting knowledge graph entities from multi-source heterogeneous data;

[0075] Step S3 specifically includes the following steps:

[0076] S31, extracting knowledge graph entities: dividing the target area into grids, and associating the fire risk real-time prediction result, vegetation features V C , meteorological features N C , forest fire features I C , terrain features T f , and text remote sensing associated features TRS to the corresponding grid;

[0077] S32, entity extraction result verification: verifying the consistency of the entity knowledge graph spatial range and the spatial consistency through the spatial connection tool of Arc GIS to obtain a knowledge graph entity set;

[0078] S33, triple relationship extraction: based on the knowledge graph entity set, extracting the semantic relationship between entities through spatial analysis and business rules to form a subject-relation-object triple, obtaining a knowledge graph triple set, wherein the knowledge graph triple set includes the spatial position relationship and interaction association relationship between entities and the support or restriction association of the corresponding emergency rescue channel path construction;

[0079] S34, based on the knowledge graph triple set, constructing a structured knowledge graph and storing it to form a semantic network supporting path construction reasoning.

[0080] S4, emergency rescue passage path construction: taking the knowledge graph obtained in step S3 as a constraint, determining a UAV reconnaissance trajectory through a double deep Q network, and generating an emergency rescue passage path by using an A* algorithm in combination with fire observation data on the UAV reconnaissance trajectory;

[0081] Step S4 specifically includes the following steps:

[0082] S41, attribute a passage cost attribute to a corresponding grid of a target region:

[0083] C(i,j) = road cost + 0.3ω1×risk+ω2×Sl+ω3×ρ v ;

[0084] In the formula, C(i,j) represents the comprehensive passage cost of grid i to grid j; road cost represents the passage opening cost; ω1, ω2, and ω3 all represent weight coefficients, and ω1+ω2+ω3=1; risk represents the fire risk value of the current grid; ρ v represents the vegetation density of the current grid, ρ v = Ve×k veg , k veg represents the vegetation type coefficient;

[0085] S42, determine a UAV group trajectory based on a double deep Q network;

[0086] S421, define the state space s = (x UVA ,y UVA ,risk,d Fire-Road ,comm) of the double deep Q network; wherein (x UVA ,y UVA ) represents the coordinates of the UAV; omm represents the communication connection state; the action space s = (East, South, West, North, Hover), East, South, West, North, and Hover respectively represent east, south, west, north, and hovering;

[0087] Fuse the FRZI risk index FRZI and the path cost reward function r:

[0088] r = ω α ·FRZI - ω β ·d Fire-Road - ω γ ·collision - ω δ ·comm loss - ω ∈ ·P block ;

[0089] wherein,

[0090]

[0091] P block =ω W ×wind factor +ω s ×slope factor +ω f ×fuel factor ;

[0092] In the formula, ω α , ω β , ω γ , ω δ and ω ∈ all represent weight coefficients, and ω α +ω β +ω γ +ω δ +ω ∈ =1; collision represents a UAV collision penalty term; comm loss represents a communication interruption penalty; P block represents a fire out-of-control probability of the grid; ω W , ω s and ω f all represent weight coefficients, ω W +ω s +ω f =1, wind factor represents a wind speed factor between [0,1], and W s_min and W s_max respectively represent a minimum reference wind speed and a maximum reference wind speed; slope factor represents a slope factor between [0,1]; fuel factor represents a combustible factor between [0,1]; L i,j represents a path length; w0, w1, w2 and w3 all represent weight coefficients, and w0+w1+w2+w3=1;

[0093] S422, train the double-depth Q network defined in step S421 until the training step number reaches a set number of times, customize the training, and output the double-depth Q network;

[0094] S423, input the real-time state space s into the double-depth Q network to obtain an optimal reconnaissance trajectory of the UAV group and fire observation data corresponding to the trajectory;

[0095] S43. Generate emergency rescue route: Based on the fire observation data corresponding to the trajectory obtained in step S423 and the road entities in the knowledge graph, combined with the FARSITE model to predict the risk of the fire spreading out of control across the road, the A* algorithm is used to generate the initial emergency rescue route:

[0096] Cost(n) = g(n) + h(n) + λ·P block (n);

[0097] in,

[0098]

[0099] h(n) = |x n -x goal |+|y n -y goal |;

[0100] In the formula, Cost(n) represents the cost function for establishing the initial emergency rescue route; g(n) represents the actual cost from the starting point to node n; k represents the number of grids traversed from the starting point to node n; and road cost (i) represents the cost of opening a channel within grid i; risk cost (i) represents the fire risk cost of grid i; h(n) represents the heuristic function; (x n ,y n (x) represents the UTM coordinates of node n; goal ,y goal ) represents the endpoint coordinates; λ represents the weighting coefficient for the risk of fire spiraling out of control; P block (n) represents the probability of fire getting out of control at node n;

[0101] S44. By combining the road entity connectivity in the knowledge graph, interrupted road sections are eliminated, and the points of passage are fitted with Bézier curves to achieve smooth path processing, thereby obtaining the final emergency rescue channel path.

[0102] S5. Verify the emergency rescue route by calculating the route feasibility score, and dynamically optimize the emergency rescue route based on the score results and the knowledge graph described in step S3 of real-time fire update.

[0103] Step S5 specifically includes the following steps:

[0104] S51. Route Feasibility Verification:

[0105] F S =F ars +R us +A s +R tr +D vrs ;

[0106] wherein,

[0107]

[0108]

[0109] wherein, F S represents the path feasibility total score; F ars , R us , A s , R tr and D vrs respectively represent the fire risk avoidance rate score, the passage utilization rate score, the traffic capacity adaptation degree score, the real-time risk matching degree score and the unmanned aerial vehicle verification rate score; d min represents the minimum distance from the constructed path to the fire area boundary; d min,max and d min,min respectively represent the maximum safety distance and the minimum reasonable distance of the target area; L road and L total represent the length of the constructed path in the bare soil zone (no combustible material) and the total length of the constructed path; L qual represents the length of the road in the constructed path that meets the forest fire resources passing through; represents the average fire risk value on the constructed path; r target represents the target fire risk tolerance value; r min and r max respectively represent the global minimum value and the maximum value of the fire risk value of the target area; L verify represents the length of the path verified by the unmanned aerial vehicle aerial photography;

[0110] S52, judge whether the path feasibility total score F S is not less than the threshold set by the expert, if yes, output the final emergency rescue passage path, otherwise, based on the path feasibility total score F S and the real-time fire change, return to step S3, dynamically update and optimize the final emergency rescue passage path through the knowledge graph.

[0111] Simulation experiment

[0112] Experimental scene setting

[0113] Target area: select a typical high mountain forest area (about 100km 2 ), including complex terrain (slope 0°-45°), diverse vegetation (60% of coniferous forest, 30% of broad-leaved forest, 10% of shrub forest) and hierarchical road network (2 main roads and 15 rural roads).

[0114] Fire scene: simulate the crown fire scene of heavy fire, the initial fire point is located in the center of the region, the fire spreads dynamically with wind speed (3-12 m / s), slope, and the fire changes are updated every 10 minutes.

[0115] The above experimental scene is constructed for emergency rescue channel path by using traditional method 1 (manual construction of emergency rescue channel path based on static GIS (relying on historical experience, without considering dynamic fire), traditional method 2 (local dynamic construction of emergency rescue channel path by single unmanned aerial vehicle (without knowledge graph semantic association)) and the method described in the application.

[0116] Table 1 comparison of core indicators

[0117]

[0118] As can be seen from Table 1, the response speed of the method described in the application to the fire mutation is 47% faster than that of the traditional method 2, and compared with the traditional method 1, the fire out-of-control risk can be reduced by 72%, the channel length can be shortened by 30% (compared with the traditional method 1) under the premise of ensuring safety, and the high-risk area coverage rate is increased by 31.4%, balancing the rescue efficiency and risk control, thereby proving the effectiveness of the application.

[0119] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the application and not to limit them, although the application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that: it can still modify or equivalently replace the technical solutions of the application, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the application.

Claims

1. A method for constructing emergency rescue routes for major forest fires based on knowledge graphs, characterized by: Includes the following steps: S1. Multi-source heterogeneous data fusion and standardization processing: Integrate remote sensing, meteorological, historical fire, forest fire resources and rescue text data, and obtain feature datasets through format unification and feature extraction; S2. Dynamic risk assessment of forest fires: Real-time prediction of fire risk using a forest fire risk prediction model trained on the feature dataset obtained in step S1. S3. Knowledge Graph Entity Extraction: Based on the feature dataset obtained in step S1 and the real-time fire risk prediction results obtained in step S2, knowledge graph entities are extracted from multi-source heterogeneous data. S4. Emergency Rescue Channel Path Construction: Using the knowledge graph obtained in step S3 as constraints, a dual-depth Q-network is used to determine the UAV reconnaissance trajectory, and fire observation data on the UAV reconnaissance trajectory is combined with... Algorithm generates emergency rescue route; S5. Verify the emergency rescue route by calculating the route feasibility score, and realize the dynamic optimization of the emergency rescue route based on the score results and the knowledge graph described in step S3 of real-time fire update.

2. The method for constructing emergency rescue routes for major forest fires based on knowledge graphs according to claim 1, characterized in that: Step S1 specifically includes the following steps: S11. Multi-source heterogeneous data acquisition: Acquire remote sensing, meteorological, historical fire, forest fire resources and rescue text data of the target area. The remote sensing data includes vegetation distribution data, fire distribution data, and GIS data of roads and river systems; the meteorological data includes monthly precipitation. , potential evapotranspiration Daily average relative temperature Daily average relative humidity Wind speed and wind direction; S12. Data format unification and resampling: The collected multi-source heterogeneous data is converted to UTM coordinates and resampled to a uniform resolution after standardization to obtain a standardized dataset. S13. Feature Extraction: Extracting vegetation features from standardized datasets. Meteorological characteristics Forest fire characteristics Topographic features Textual remote sensing correlation features and characteristics of forest fire fighting resources After undergoing min-max standardization, the feature dataset is obtained. .

3. The method for constructing emergency rescue routes for major forest fires based on knowledge graphs according to claim 2, characterized in that: In step S13, vegetation characteristics Including Normalized Difference Vegetation Index Vegetation humidity index Combustible material load classification and vegetation coverage ; Among them, the normalized vegetation index The expression is as follows: ； In the formula, This indicates the near-infrared reflectance of remote sensing satellites. This indicates the reflectivity of the red band of a remote sensing satellite. The expression for the vegetation humidity index is as follows: ； In the formula, This indicates the shortwave infrared reflectance of remote sensing satellites; Combustible material load classification The expression is as follows: ； In the formula, , and These represent low flammability, medium flammability, and high flammability, respectively. This indicates the combustible material load, measured in t / hm². 2 ; vegetation coverage The expression is as follows: ； In the formula, and These represent the minimum and maximum values ​​of the normalized vegetation index within the target area, respectively. meteorological characteristics Including Standardized Precipitation Evapotranspiration Index Humidity-Temperature Index Wind speed and wind direction The standardized precipitation evapotranspiration index is expressed as follows: ； In the formula, Indicates historical monthly precipitation and potential evaporation Standard deviation; Humidity-Temperature Index The expression is as follows: ； Forest fire characteristics Including forest fire density Fire radiation power And the shortest straight-line distance between the real-time fire point and the road entity. Forest fire density The expression is as follows: ； In the formula, Indicates the time period within the target area. The number of newly added fire points within; Indicates the area of ​​the target region; Topographic features Including slope, and slope The expression is as follows: ； In the formula, and These represent the rates of elevation change in the x and y directions, respectively, in degrees (°). The shortest straight-line distance between the real-time fire point and the road entity The expression is as follows: ； In the formula, UTM coordinates representing the fire point; Represents the UTM coordinates of the road node; Characteristics of forest fire fighting resources Including response radius and maximum water supply distance ; Text remote sensing association features The steps to obtain it are as follows: The first step is to use the BERT-base model to analyze forest fire fighting resources and rescue text data. Perform text semantic encoding to obtain text vectors: ； In the formula, Semantic vectors representing textual data on forest fire fighting resources and rescue; This indicates that the semantic encoding operation is performed using the BERT-base model; Step 2: Remote sensing image feature extraction: A CNN network is used to extract infrared image features from the remote sensing images. Output image features: ； In the formula, Indicating infrared image features eigenvectors; Indicates normalization to pixel values Feature extraction is performed using a CNN network; Step 3: Extracting text remote sensing association features : ； In the formula, Represents the cosine similarity between text vectors and image features; This represents the cosine similarity function.

4. The method for constructing emergency rescue routes for major forest fires based on knowledge graphs according to claim 3, characterized in that: Step S2 specifically includes the following steps: S21. Construct a forest fire risk prediction model based on the XGBoost model: S22. Determine the optimal initialization hyperparameters of the forest fire risk prediction model using the TPE algorithm within the Optuna framework; S23, Transfer the feature dataset The model is divided into a training set and a validation set. First, the training set is input into the forest fire risk prediction model with the optimal initial hyperparameters determined for training. The training continues until the area under the validation curve (AUC) of the validation set reaches the set value. At this point, the forest fire risk prediction model is considered to have converged and training is stopped. S24. Collect real-time multi-source heterogeneous data of the target area and input it into the trained forest fire risk prediction model, outputting the real-time predicted fire risk value. and fire risk values Fire risk level is determined by classifying the levels. .

5. The method for constructing emergency rescue routes for major forest fires based on knowledge graphs according to claim 4, characterized in that: Step S3 specifically includes the following steps: S31. Extract knowledge graph entities: Divide the target area into grids and extract real-time fire risk prediction results and vegetation characteristics. Meteorological characteristics Forest fire characteristics Topographic features and text remote sensing association features Associated with the corresponding grid; S32. Entity extraction result verification: Using ArcGIS's spatial connectivity tool, verify the consistency of the spatial range and spatial consistency of the entity knowledge graph to obtain the entity set of the knowledge graph; S33. Triple Relationship Extraction: Based on the knowledge graph entity set, semantic relationships between entities are extracted through spatial analysis and business rules to form subject-relationship-object triples, resulting in a knowledge graph triple set. The knowledge graph triple set includes spatial location relationships and interaction relationships between entities, as well as supporting or restricting relationships between entities on the construction of emergency rescue channel paths. S34. Based on the knowledge graph triple set, construct and store a structured knowledge graph to form a semantic network that supports reasoning for constructing emergency rescue channel paths.

6. The method for constructing emergency rescue routes for major forest fires based on knowledge graphs according to claim 5, characterized in that: Step S4 specifically includes the following steps: S41. Assign the toll cost attribute to the corresponding grid in the target area: ； In the formula, Represents a grid To grid The overall cost of passage; This indicates the cost of setting up the channel; , and Both represent weighting coefficients, and ; This indicates the fire risk value of the current grid. This indicates the current vegetation density of the grid. , Indicates the vegetation type coefficient; S42. Determining the trajectory of a drone swarm based on a dual-depth Q-network; S421. Define the state space of a dual-depth Q-network. ;in, Indicates the coordinates of the drone; Indicates the communication connection status; action space , These represent east, south, west, north, and hovering, respectively. Fusion Risk Index With path cost reward function : ； in, ； ； In the formula, , , , and Both represent weighting coefficients, and ; This indicates the drone collision penalty. Indicates a penalty for communication interruption; This indicates the probability of a fire going out of control within the grid. , and All represent weighting coefficients. , Indicates in The wind speed factor between, and , and These represent the minimum reference wind speed and the maximum reference wind speed, respectively. Indicates in The slope factor between; Indicates in The combustible factors between them; Indicates the path length; , , and Both represent weighting coefficients, and ; S422, Training the dual-depth Q-network defined in step S421 until the number of training steps reaches the set number, customized training, and outputting the dual-depth Q-network; S423, Transfer the real-time state space Inputting a dual-depth Q-network yields the optimal reconnaissance trajectory of the UAV swarm and the corresponding fire observation data. S43. Generate emergency rescue route: Based on the fire observation data corresponding to the optimal reconnaissance trajectory obtained in step S423 and the road entities in the knowledge graph, combined with the FARSITE model to predict the risk of fire spreading out of control across the road, the A* algorithm is used to generate the initial emergency rescue route: ； in, ； ； In the formula, This represents the cost function for constructing the initial emergency rescue route; Indicates the distance from the starting point to the node. The actual cost; Indicates the distance from the starting point to the node. The number of grids traversed; Represents a grid Internal road infrastructure costs; Represents a grid The cost of fire risk; Represents a heuristic function; Represents a node UTM coordinates; Indicates the coordinates of the endpoint; Weighting coefficients representing the risk of fire getting out of control; Represents a node The probability of a fire getting out of control; S44. By combining the road entity connectivity in the knowledge graph, interrupted road sections are eliminated, and the path is smoothed by fitting the route points with a Bézier curve, thus obtaining the final emergency rescue channel path.

7. The method for constructing emergency rescue routes for major forest fires based on knowledge graphs according to claim 6, characterized in that: Step S5 specifically includes the following steps: S51. Route Feasibility Verification: ； in, ； ； ； ； ； In the formula, This represents the total score for path feasibility; , , , and These represent the scores for fire risk avoidance rate, road utilization rate, traffic capacity adaptability, real-time risk matching, and drone verification rate, respectively. This represents the minimum distance from the constructed path to the boundary of the fire zone; and These represent the maximum safe distance and the minimum reasonable distance to the target area, respectively. and This represents the length of the native soil zone within the constructed path and the total length of the constructed path; This indicates the length of the constructed path that satisfies the requirement for forest fire fighting resources to pass through; This represents the average fire risk value along the constructed path; Indicates the target fire risk tolerance value; and These represent the global minimum and maximum values ​​of the fire risk value for the target area, respectively. This indicates the verified path length for drone aerial photography; S52, Determine the total score for path feasibility If the value is not less than the threshold set by experts, output the final emergency rescue route; otherwise, output the route based on the total feasibility score. Based on real-time changes in the fire situation, return to step S3 and dynamically update and optimize the final emergency rescue route using the knowledge graph.

Images