Forest fire prediction method

Abstract

The invention discloses a forest fire prediction method, which comprises the following steps of: 1, dividing a forest region into a plurality of sub-regions, and generating a digital map; step 2, collecting fire factors of each sub-region, wherein the fire factors comprise static fire factors and dynamic fire factors; 3, carrying out risk grading based on the fire factors, and outputting region classification information; 4, deploying a monitoring network according to the region classification information and acquiring monitoring data; 5, processing the monitoring data through a graph neural network model, and outputting a fire danger probability; and 6, when the fire danger probability exceeds a threshold value, starting an emergency response mechanism. According to the invention, by fusing static and dynamic fire factors, combining with meteorological data to dynamically correct risk classification and differentially deploy monitoring resources, the prediction precision is optimized by adopting a dual-channel graph neural network model, and a closed-loop mechanism from monitoring to prediction to emergency is established, so that the accuracy of forest fire prediction and the emergency disposal capability are improved.

Description

Technical Field

[0001] The present invention relates to the field of prediction methods, and in particular to a method for predicting forest fires. Background Art

[0002] Forest fires are highly destructive natural disasters. Their occurrence and spread are influenced by multiple factors, including vegetation, topography, weather, and human activity. They not only damage forest ecosystems and reduce biodiversity, but can also threaten the lives and property of surrounding residents and even trigger regional environmental disasters. Therefore, accurate and timely forest fire forecasting, early identification of high-risk areas, and the initiation of prevention and control measures are key to reducing fire losses.

[0003] Existing forest fire prediction technologies have problems such as reliance on a single type of fire factor, fixed regional risk classification results, and lack of targeted monitoring resource allocation. Therefore, a forest fire prediction method is proposed. Summary of the Invention

[0004] The present invention solves the problems of the prior art through the following technical solutions, and the present invention comprises the following steps: Step 1: Divide the forest area into several sub-areas and generate a digital map; Step 2: Collect the fire factors of each sub-area, which include static fire factors and dynamic fire factors; Step 3: Perform risk classification based on fire factors and output regional classification information; Step 4: Deploy the monitoring network and obtain monitoring data based on regional classification information; Step 5: Process the monitoring data through the graph neural network model and output the fire risk probability; Step 6: When the fire risk probability exceeds the threshold, activate the emergency response mechanism.

[0005] Furthermore, the static fire factors include flammable vegetation density, historical fire point density, lightning strike frequency and terrain characteristics, including slope and altitude; The dynamic fire factors include vegetation moisture content and human activity intensity distribution data.

[0006] Furthermore, step 3 includes: Step 3.1: Fusion processing is performed on the collected static fire factors (including flammable vegetation density, historical fire point density, lightning strike frequency, and terrain characteristics) and dynamic fire factors (including vegetation moisture content and human activity intensity distribution data). The attribute information of these two types of factors is encoded into a spatiotemporal feature vector that contains both temporal variation characteristics and spatial distribution characteristics. Step 3.2: Input the spatiotemporal feature vector into the clustering model, and generate the flammability classification results of the region through model calculation, which are specifically divided into three levels: high flammability, medium flammability, and low flammability; Step 3.3: Associate real-time and short-term forecast meteorological data (including temperature, relative humidity, wind speed, and precipitation probability). Based on the meteorological data, dynamically adjust and correct the classification results generated in Step 3.2 to adapt to the impact of changing meteorological conditions on regional flammability and improve the timeliness and accuracy of the classification results. Meteorological data are collected in real time through satellite remote sensing or ground-based meteorological stations.

[0007] Furthermore, the dynamic adjustment and correction includes: When the meteorological conditions corresponding to the meteorological data reach the preset danger standard, the medium and low flammability sub-areas will be upgraded to high flammability categories; Through the preset adversarial sample generation module, data on fire scenarios with superimposed extreme meteorological conditions and terrain features are simulated to obtain simulated data. Based on the simulated data, the weighted coefficients of terrain features (including slope and altitude) are optimized (the steeper the slope and the higher the altitude, the larger the weighted coefficient, which is specifically set based on measured data that shows that terrain accelerates the spread of fire). This strengthens the model's learning of the impact of terrain during the dynamic correction process and improves the accuracy of classification results in extreme scenarios.

[0008] Furthermore, the specific process of step 4 is as follows: For highly flammable sub-areas: deploy drone swarms equipped with monitoring equipment to conduct inspections, and dynamically adjust the monitoring frequency based on changes in fire risk probability, increasing the monitoring frequency when the fire risk probability increases and decreasing it when it decreases; For flammable sub-areas: deploy a ground sensor network and combine it with satellite remote sensing data to achieve coordinated ground and satellite monitoring; For low-flammability sub-areas: use satellite remote sensing monitoring to obtain regional macro-monitoring data.

[0009] Furthermore, the path planning of the drone inspection aims to achieve full coverage of the sub-area, and the path needs to avoid high-risk terrain; The inspection route is updated periodically based on changes in fire risk probability: when the fire risk probability fluctuates slightly, the route is updated at a fixed period; when the fire risk probability fluctuates beyond the preset standard range, the route is updated in real time, giving priority to covering sub-areas where the fire risk probability increases faster.

[0010] Furthermore, the graph neural network model in step 5 includes the following dual-channel structure: Spatial processing channel: By extracting the vegetation type, vegetation density and terrain connection characteristics of adjacent sub-areas, the vegetation continuity index representing the possibility of fire spread is calculated. This index is used to quantify the spatial correlation of fire spread between adjacent sub-areas; Time processing channel: Based on the collection timestamp of the monitoring data, time-varying weight coefficients are assigned to the monitoring data at different times. The weight of recent data is higher than that of historical data. The weight decay rule is set according to the time sensitivity of fire development to highlight the impact of real-time monitoring data on the calculation of fire risk probability.

[0011] Furthermore, the calculation formula of the vegetation continuity index is: ; Among them, d ij is the distance between adjacent sub-regions; σ is the characteristic parameter of vegetation type; V i is the comprehensive quantitative value of vegetation and terrain in sub-region i; V j is the comprehensive quantitative value of vegetation and terrain in sub-region j.

[0012] Furthermore, step 6 specifically includes the following operations: Step 6.1: Deploy mobile sensing devices (including temperature detection units, smoke concentration sensors, and vegetation status monitoring modules) in sub-areas where the fire risk probability exceeds the threshold (i.e., target sub-areas) to collect real-time supplementary data in the area. The supplementary data includes local temperature changes, smoke diffusion status, and dynamic changes in vegetation moisture content. Step 6.2: Input the collected supplementary data into the graph neural network model in step 5, and iterate the model’s time-varying weight coefficients and specific parameters of spatial correlation to optimize the real-time prediction accuracy of fire risk probability; Step 6.3: When the supplementary data meets the preset fire point confirmation criteria, that is, when the monitoring indicators reach a critical state that can determine the occurrence of a fire, the resource dispatch mechanism is immediately triggered. The dispatch content includes the deployment of firefighting forces, the allocation of emergency materials, and the demarcation of regional warning ranges.

[0013] Furthermore, the mobile sensing device includes a gas detection unit with a high temperature resistant package and a motion control module based on terrain data; High-temperature resistant packaged gas detection unit: encapsulated in a fireproof and heat-insulating shell, used to detect fire-related gas parameters such as combustible gas concentration and smoke composition in the target sub-area in real time; Motion control module based on terrain data: Built-in sub-region terrain database (including slope and obstacle distribution information), using terrain matching algorithm to plan movement paths, enabling autonomous movement or fixed-point deployment of the device in complex terrain (such as steep slopes and dense vegetation areas); The mobile sensing device is equipped with an overheating protection mechanism. The specific content of the overheating protection mechanism is: when the temperature sensor inside the device detects that the temperature exceeds the preset critical value, it automatically activates multi-level protection measures, including activating the built-in insulation layer, cutting off non-core circuits to reduce energy consumption, and sending an overheating warning signal to the monitoring center to ensure the continuity of the core detection function in a high-temperature environment.

[0014] Compared with the existing technology, the present invention has the following advantages: the forest fire prediction method generates a vector containing spatiotemporal characteristics by fusing static fire factors and dynamic fire factors, and dynamically corrects the classification results in combination with real-time and short-term meteorological data. At the same time, through extreme scenario simulation and terrain weighted enhancement model learning on the influence of terrain, the regional flammability classification is made more adaptable to environmental changes, thereby improving classification accuracy and timeliness; differentiated monitoring strategies are adopted for high, medium and low flammability sub-regions, and the monitoring frequency and drone inspection path are dynamically adjusted according to the fire risk probability to achieve precise resource allocation, focus on covering high-risk areas, improve monitoring efficiency and reduce redundant costs; the graph neural network model adopts a dual-channel structure. The spatial processing channel quantifies the spatial correlation of fire propagation in adjacent areas through vegetation continuity indicators, and the temporal processing channel highlights the influence of real-time data through time-varying weights. Combined with mobile sensing device supplementary data, the model parameters are iteratively optimized, significantly improving the accuracy of fire risk probability prediction; when the fire risk probability exceeds the threshold, the mobile sensing device quickly collects supplementary data, promptly triggering the resource scheduling mechanism, and the equipment has an overheating protection mechanism to ensure continuous operation in high-temperature environments, effectively shortening the response time and improving fire response capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is the overall structural diagram of the present invention. DETAILED DESCRIPTION

[0016] The following is a detailed description of an embodiment of the present invention. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process. However, the protection scope of the present invention is not limited to the following embodiment.

[0017] like Figure 1 As shown, this embodiment provides a technical solution: a forest fire prediction method, comprising the following steps: Step 1: Divide the forest area into several sub-areas and generate a digital map; Step 2: Collect the fire factors of each sub-area, which include static fire factors and dynamic fire factors; Step 3: Perform risk classification based on fire factors and output regional classification information; Step 4: Deploy the monitoring network and obtain monitoring data based on regional classification information; Step 5: Process the monitoring data through the graph neural network model and output the fire risk probability; Step 6: When the fire risk probability exceeds the threshold, activate the emergency response mechanism.

[0018] The static fire factors include flammable vegetation density, historical fire point density, lightning strike frequency and terrain characteristics (including slope and altitude); The dynamic fire factors include vegetation moisture content and human activity intensity distribution data.

[0019] Step 3.1: Fusion processing is performed on the collected static fire factors (including flammable vegetation density, historical fire point density, lightning strike frequency, and terrain characteristics) and dynamic fire factors (including vegetation moisture content and human activity intensity distribution data). The attribute information of these two types of factors is encoded into a spatiotemporal feature vector that contains both temporal variation characteristics and spatial distribution characteristics. Step 3.2: Input the spatiotemporal feature vector into the clustering model, and generate the flammability classification results of the region through model calculation, which are specifically divided into three levels: high flammability, medium flammability, and low flammability; Step 3.3: Associate real-time and short-term forecast meteorological data (including temperature, relative humidity, wind speed, and precipitation probability) and dynamically adjust and correct the classification results generated in Step 3.2 based on the meteorological data to adapt to the impact of changing meteorological conditions on regional flammability and improve the timeliness and accuracy of the classification results. Meteorological data is collected in real time through satellite remote sensing or ground-based meteorological stations. Through multi-dimensional data fusion and dynamic correction, the comprehensiveness, accuracy, and timeliness of regional flammability classification are significantly improved, making the classification results more in line with actual fire risk conditions and providing a reliable basis for subsequent monitoring deployment and fire risk prediction; For example, a forest sub-area Y (area 1 square kilometer); Fusion of fire factors generates spatiotemporal feature vectors; Static fire factor: Flammable vegetation density: 85% (dominated by coniferous forests, which are highly flammable); Historical fire density: 4 times / km², with 4 fires occurring in the past 5 years, indicating a medium risk; Lightning strike frequency: 3 times / year. The average annual number of lightning strikes in this area is slightly higher than that in the surrounding areas. Topographic features: Gentle slope of 20°, medium altitude of 600 meters above sea level.

[0020] Dynamic Fire Factor: Vegetation moisture content: 22%, real-time monitoring value, below the safety threshold of 30%, flammability increases; Human activity intensity: 0.3 times / day. The frequency of hikers' activities has been relatively high recently, and slightly higher on weekends.

[0021] Fusion processing: The spatial attributes of static factors (such as the spatial distribution of slope and altitude) and the temporal changes of dynamic factors (such as the vegetation moisture content dropped from 28% to 22% in the past three days, and the human activities increased to 0.5 times / day on weekends) are encoded into spatiotemporal feature vectors. For example: [85%, 4 times / km2, 3 times / year, 20°, 600 m, 22%, 0.3 times / day, 28%→22% (3-day variation), 0.3→0.5 times / day (weekend fluctuation)]; This vector contains both spatial intrinsic properties (static) and temporal dynamic changes (dynamic), providing comprehensive features for subsequent classification.

[0022] The clustering model generates initial classification results; The above spatiotemporal feature vectors are input into a clustering model (such as K-means clustering). The model calculates based on a preset threshold (for example, a comprehensive risk score of 50-70 is classified as medium flammability): In the static factors of subregion Y, the density of flammable vegetation is high (85%) but the density of historical fire points is medium (4 times / km2), and the terrain slope is moderate (20°); Among the dynamic factors, the moisture content of vegetation is slightly low (22%) but the intensity of human activities is low (0.3 times / day). The comprehensive risk score is 62 points, which falls into the medium flammable range.

[0023] Initial classification result: Sub-area Y is classified as medium flammable.

[0024] Dynamically revise classification results by associating meteorological data; Real-time and short-term meteorological data collection, collected through ground weather stations and satellite remote sensing: Real-time temperature: 36°C, exceeding the dangerous threshold of 32°C; Relative humidity: 25%, below the dangerous threshold of 30%; Wind speed: Level 7, exceeding the dangerous threshold of Level 5, which is likely to accelerate the spread of fire; Probability of short-term precipitation: 0%, no precipitation in the next 6 hours, and the moisture content of vegetation will continue to decline.

[0025] Dynamic correction basis: The preset danger standards are temperature ≥ 32℃, humidity ≤ 30%, and wind speed ≥ level 5. The current meteorological data fully meets this standard, triggering the classification upgrade mechanism.

[0026] Correction results: The flammable category in the initial classification of sub-area Y was dynamically revised to the highly flammable category to adapt to the actual risk of a sudden increase in vegetation flammability in an environment of high temperature, low humidity, and strong wind.

[0027] The dynamic adjustment and correction includes: When the meteorological conditions corresponding to the meteorological data reach the preset hazard standards, the medium and low flammability sub-areas will be upgraded to high-risk categories; The adversarial sample generation module simulates extreme fire scenario data and assigns weighted coefficients to terrain features (including slope and altitude) (the steeper the slope and the higher the altitude, the larger the coefficient, based on measured data that terrain accelerates fire spread) to enhance the model's learning of the impact of terrain. Through a targeted dynamic correction mechanism, the regional flammability classification's ability to respond to extreme weather conditions and terrain influences is further enhanced to prevent risk underestimation. This also improves the model's adaptability to predicting fire risks in complex scenarios, making the classification results more consistent with actual fire development patterns. When meteorological conditions reach the preset danger level, sub-areas classified as medium or low flammable are directly upgraded to high-risk categories, avoiding risk omissions caused by conventional classification lagging behind sudden meteorological changes, and ensuring that high-risk areas receive timely attention. Strengthen the weight of terrain's impact on fire spread: Simulate extreme fire scenarios through the adversarial sample generation module, and assign weighted coefficients to terrain features such as slope and altitude. The steeper the slope and the higher the altitude, the larger the coefficient. This allows the model to more deeply learn the law of how terrain accelerates fire spread, improving the accuracy of risk prediction in extreme scenarios.

[0028] Continuing from the previous forest sub-area, it was initially classified as medium flammable due to the fusion of static and dynamic factors, and was later revised to high flammable due to meteorological data of high temperature, low humidity, and level 5 wind.

[0029] The actual terrain of this sub-area is a steep slope of 25° and a high altitude of 800 meters; Since the real-time meteorological data (high temperature above 35°C, humidity below 30%, and wind force 5) has reached the preset danger standard, the system directly upgraded the risk level from the revised high flammability category to a higher level, clarifying its high-risk attribute affected by extreme weather conditions. The adversarial sample generation module simulates the scene data of an extreme fire in the area. Combined with the measured data that terrain accelerates the spread of fire, it assigns a weighting coefficient of 0.8 to the 25° slope (the coefficient for a gentle slope is 0.3) and a weighting coefficient of 0.7 to the altitude of 800 meters (the coefficient for a low altitude is 0.2). When calculating fire risk, the model will focus on the characteristics of the terrain that cause the fire to spread rapidly upward, and predicts that the fire will spread faster than in gentle, low-altitude areas under the same meteorological conditions.

[0030] Based on this, in subsequent monitoring, the drone cluster not only increased the inspection frequency, but also adjusted the route in a targeted manner, giving priority to covering the area above the steep slope of the sub-area, and deploying firefighting forces in the buffer zone at the bottom of the slope in advance, effectively improving the efficiency of responding to the risks of superposition of extreme weather and terrain.

[0031] The specific process of step 4 is as follows: For highly flammable sub-areas: deploy drone swarms equipped with monitoring equipment to conduct inspections, and dynamically adjust the monitoring frequency based on changes in fire risk probability (increase the frequency when the fire risk probability increases, and decrease the frequency when it decreases); For flammable sub-areas: deploy a ground sensor network and combine it with satellite remote sensing data to achieve coordinated ground and satellite monitoring; For low-flammability sub-areas: Use satellite remote sensing monitoring to obtain regional macro-monitoring data; By adopting differentiated monitoring strategies for sub-areas with different flammability categories and dynamically adjusting monitoring frequency based on fire risk probability, precise allocation of monitoring resources is achieved. This ensures focused coverage of high-risk areas while avoiding resource redundancy in low-risk areas, thereby improving overall monitoring efficiency and timely response. Based on regional flammability classification, namely high flammability, medium flammability, and low flammability, different monitoring methods are matched, such as drone swarms, ground-based and satellite collaboration, and satellite remote sensing. This ensures that monitoring intensity is proportional to risk level, ensuring that high-risk areas receive more intensive and accurate monitoring data. The monitoring frequency is adjusted in real time as the fire risk probability changes (more frequent when it rises, less frequent when it falls), avoiding inadequate monitoring during high-risk periods or wasted resources during low-risk periods due to a fixed frequency, ensuring that the monitoring rhythm closely matches the risk dynamics. If it is classified as highly flammable after correction, there will be high temperature above 35°C, humidity below 30%, wind force 5, slope of 25°, and altitude of 800 meters.

[0032] This sub-area is classified as highly flammable, so a drone swarm equipped with temperature and smoke sensors is deployed for inspections. When the initial fire risk probability is 60%, the inspection frequency is set to once every two hours, covering key locations within the sub-area, such as steep slopes and densely vegetated areas. Two hours later, due to the continued high temperature, the graph neural network model predicted that the fire risk probability had risen to 85% (exceeding the 70% warning threshold). The system automatically adjusted the drone inspection frequency to once every 30 minutes, and the inspection route prioritized the top of the slope, where the fire risk probability was increasing the fastest (due to the steep slope and high altitude, fires are prone to spread quickly). If the subsequent weather conditions improve (such as the probability of precipitation increases and the fire risk probability drops to 40%), the inspection frequency will be reduced to once every 4 hours to reduce resource investment.

[0033] This differentiated and dynamically adjusted monitoring method not only ensures that regional status changes can be captured in real time during high-risk periods, but also avoids ineffective monitoring when the fire risk is reduced, significantly improving the targeted nature of fire prevention and resource utilization efficiency.

[0034] The path planning of the drone inspection aims to achieve full coverage of the sub-area, and the path needs to avoid high-risk terrain; The inspection route is updated periodically based on changes in fire risk probability: when the fire risk probability fluctuates slightly, the route is updated at a fixed interval; when the fire risk probability fluctuates beyond a preset standard, the route is updated in real time, prioritizing coverage of sub-areas with faster increases in fire risk probability. By scientifically planning and dynamically updating drone inspection routes, we ensure comprehensive monitoring coverage of highly flammable sub-areas while avoiding high-risk terrain to ensure equipment safety. Furthermore, we flexibly adjust monitoring priorities based on fluctuations in fire risk probability, further improving the relevance and efficiency of monitoring and avoiding resource waste. Path planning aims to fully cover the sub-area while avoiding high-risk terrain, such as steep slopes, dense vegetation, and areas prone to ignition. This ensures no blind spots in monitoring, reduces drone failures caused by dangerous terrain, and ensures monitoring continuity. The path update frequency is adjusted according to the degree of fluctuation in fire risk probability. When the fire risk fluctuates for hours, it is updated at a fixed period to avoid invalid adjustments. When the fluctuation is large, it is updated in real time and priority is given to areas where the fire risk is rising rapidly, ensuring that high-risk points are monitored in key areas and improving the timeliness of fire warnings.

[0035] For example, in the flammable sub-area, the fire risk probability has risen to 85% under conditions of high temperature above 35°C, humidity below 30%, force 5 wind, steep slope of 25°, and altitude of 800 meters; The drone inspection route planning for this area is as follows: The initial path aims to fully cover the sub-area while avoiding direct flights over 25° steep slopes. Steep slopes are high-risk terrain, fires can spread rapidly along the slopes, and drones flying there are easily affected by airflow and lose control. The path is designed to detour along the flat area at the edge of the steep slope, and indirect monitoring of the steep slope area is achieved through multi-angle shooting, ensuring complete coverage while avoiding exposing equipment to high-risk environments.

[0036] When the fire risk probability fluctuates slightly, such as a slow decrease from 85% to 82% with a fluctuation of 3% within 2 hours, the route is updated at a fixed period (every 2 hours) to maintain regular coverage of the entire area; If it is detected that the fire risk probability at the top of the slope in the area suddenly increases due to a sudden increase in wind speed, such as from level 5 to level 7, the fire risk probability will soar from 82% to 95% within 30 minutes, a fluctuation of 13%, which is a large fluctuation. The system will immediately update the route in real time, suspend inspections of other stable areas, and prioritize the concentration of drone clusters to cover the top of the slope and the surrounding areas where the fire risk is increasing the fastest. Through high-frequency, close-range monitoring, it will capture details such as temperature changes and smoke diffusion in the area, providing accurate data for fire risk prediction and emergency response; This path planning and update mechanism not only ensures that there are no blind spots in monitoring and that the equipment is safe, but also can quickly focus on key points when risks suddenly change, significantly improving the effectiveness of monitoring in highly flammable areas and the preemptive nature of emergency response.

[0037] The graph neural network model in step 5 contains the following dual-channel structure: Spatial processing channel: By extracting the vegetation type, vegetation density and terrain connection characteristics of adjacent sub-areas, the vegetation continuity index representing the possibility of fire spread is calculated. This index is used to quantify the spatial correlation of fire spread between adjacent sub-areas; Time processing channel: Based on the acquisition timestamp of the monitoring data, time-varying weight coefficients are assigned to the monitoring data at different times (recent data has a higher weight than historical data, and the weight decay rule is set according to the time sensitivity of fire development) to highlight the impact of real-time monitoring data on fire risk probability calculations; The dual-channel structure of the graph neural network model (spatial processing channel + temporal processing channel) accurately captures the spatial correlation and temporal dynamics of fire risks, significantly improving the accuracy and real-time performance of fire risk probability predictions and providing a more scientific basis for fire early warning and spread trend assessment. The spatial processing channel quantifies the possibility of fire spread between adjacent sub-areas by calculating vegetation continuity indicators, such as the impact of vegetation type, density, and terrain connection on fire spread, avoiding isolated analysis of the risks of a single area and predicting in advance the path and speed of fire spread from high-risk areas to the surrounding areas.

[0038] Highlight the impact of real-time data and adapt to the time sensitivity of fire development: The time processing channel uses a time-varying weight coefficient, with recent data having a higher weight, allowing the model to pay more attention to the latest monitoring data. For example, temperature and smoke data from 10 minutes ago are more valuable than those from 2 hours ago. This solves the interference of historical data on real-time risks and makes the prediction results more in line with the current fire development status.

[0039] The calculation formula of the vegetation continuity index is: ; Among them, d ij is the distance between adjacent sub-regions, which is calculated based on the digital map generated in step 1 by extracting the spatial coordinate information of adjacent sub-regions and using the Euclidean distance formula; σ is a vegetation type characteristic parameter, which is based on the vegetation type data in the static fire factor collected in step 2 (obtained through field vegetation surveys and satellite remote sensing image interpretation). The specific value is set according to the differences in vegetation types between adjacent sub-regions (such as the difference in flammability between coniferous forests and broad-leaved forests, and the similarity of vegetation communities). The greater the difference in vegetation types, the larger the σ value. V i is the comprehensive quantitative value of vegetation and terrain in sub-region i, and its calculation parameters and acquisition method are as follows: Basic parameters: the density of flammable vegetation in sub-area i is obtained through field sampling and vegetation cover inversion from remote sensing images; the moisture content of vegetation in sub-area i is obtained through real-time monitoring by ground sensors and inversion of remote sensing data; Terrain connection characteristics: Based on the terrain connection data of sub-area i and adjacent sub-area j collected in step 2, including the slope and altitude of sub-area i, the slope difference between sub-area i and sub-area j, and the continuity of altitude transition, obtained through digital elevation model (DEM) and topographic mapping data; Weighting coefficient: This is set based on the measured impact of the topographic connection between sub-areas i and j on fire spread (e.g., the increase in fire spread from sub-area i to j for every 5° increase in slope, and the correlation between the continuity of the elevation transition between sub-areas i and j and the probability of fire spread). The closer the topographic connection (e.g., the smaller the slope difference between sub-areas i and j, and the gentler the elevation transition), the greater the weighting coefficient. The comprehensive quantitative value is calculated by weighted summation of the basic parameters of sub-region i and the corresponding weighted coefficients, which is used to reflect the synergistic effect of vegetation density, moisture content and terrain connection in sub-region i on the spread of fire to adjacent areas; V j is the comprehensive quantitative value of vegetation and terrain in sub-region j, and its calculation parameters and acquisition method are as follows: Basic parameters: The density of flammable vegetation in sub-area j is obtained through field sampling and vegetation cover inversion from remote sensing images; the moisture content of vegetation in sub-area j is obtained through real-time monitoring by ground sensors and inversion of remote sensing data; Terrain connection characteristics: Based on the terrain connection data of sub-area j and adjacent sub-area i collected in step 2, including the slope and altitude of sub-area j, its slope difference with sub-area i, and the continuity of altitude transition, obtained through digital elevation model (DEM) and topographic mapping data; Weighting coefficient: This is set based on the measured impact of the topographic connection between sub-regions j and i on fire spread (e.g., the increase in fire spread to i for every 5° increase in slope in sub-region j, and the correlation between the continuity of the altitude transition between sub-regions j and i and the probability of fire spread). The closer the topographic connection (e.g., the smaller the difference in slope between sub-regions j and i, and the gentler the altitude transition), the greater the weighting coefficient. The comprehensive quantitative value is calculated by weighted summation of the basic parameters of sub-region j and the corresponding weighted coefficients, which is used to reflect the synergistic effect of vegetation density, moisture content and terrain connection in sub-region j on the spread of fire to adjacent areas; For example, in the highly flammable sub-area mentioned in the previous example, the temperature is above 35°C, the humidity is below 30%, the wind speed is level 7, the slope is 25°, and the altitude is 800 meters. The fire risk probability at the top of the slope soars from 82% to 95% within 30 minutes.

[0040] Spatial processing channel: Calculate the vegetation continuity index of the slope top sub-area (A) and the adjacent slope middle sub-area (B). Assume that the distance between A and B is d ij =50 meters, vegetation characteristic parameter σ=100, vegetation and terrain comprehensive quantitative value V of A i =0.8, the comprehensive quantitative value of vegetation and terrain of B is V j =0.7, substituting into the formula we get: ; This indicator shows that due to the close distance between A and B and the flammable vegetation conditions, the spatial correlation of fire spreading from A to B is strong (the higher the indicator value, the stronger the correlation). Based on this, the model predicts that if a fire occurs in A, it will spread quickly to B.

[0041] Temporal processing channel: The model assigns time-varying weights to monitoring data at different times. For example, real-time data within 30 minutes (wind speed increased from level 5 to level 7, and slopetop temperature increased from 35°C to 38°C) is weighted 0.9, while data from two hours ago (wind speed level 5, temperature 35°C) is weighted 0.3. Because recent data carries a higher weight, the model can keenly capture the impact of sudden increases in wind speed and temperature on fire risk, accurately calculating the real-time change in the fire risk probability in area A, from 82% to 95%, rather than letting historical data drag down the prediction accuracy.

[0042] Through dual-channel collaboration, the model not only clarifies the possible spread path of the fire (A to B), but also accurately reflects the surge in real-time risks, providing an accurate basis for subsequent emergency responses (such as prioritizing the deployment of fire isolation zones between A and B and dispatching firefighting forces to the top area of ​​the slope), significantly improving the practicality of fire risk prediction.

[0043] The step 6 specifically includes the following operations: Step 6.1: Deploy mobile sensing devices (including temperature detection units, smoke concentration sensors, and vegetation status monitoring modules) in sub-areas where the fire risk probability exceeds the threshold (i.e., target sub-areas) to collect real-time supplementary data in the area. The supplementary data includes local temperature changes, smoke diffusion status, and dynamic changes in vegetation moisture content. Step 6.2: Input the collected supplementary data into the graph neural network model in step 5, and iterate the model’s time-varying weight coefficients and specific parameters of spatial correlation to optimize the real-time prediction accuracy of fire risk probability; Step 6.3: When the supplementary data meets the preset fire point confirmation criteria (i.e., the monitoring indicators reach a critical state that can be used to determine the occurrence of a fire), the resource dispatch mechanism is immediately triggered. The dispatch content includes the deployment of firefighting forces, the allocation of emergency supplies, and the demarcation of regional warning areas; Through a closed-loop mechanism from supplementary monitoring to model optimization to emergency triggering, the real-time and accuracy of fire risk warnings and the efficiency of emergency response are significantly improved, ensuring that in high-risk scenarios, the fire status can be quickly identified, the prediction results can be optimized, and resources can be dispatched in a timely manner to minimize fire losses.

[0044] Mobile sensing devices are deployed in target sub-areas where fire risk exceeds the threshold to collect micro data such as local temperature changes and smoke diffusion, to make up for the neglect of local details by large-scale monitoring (such as drones or satellites), and provide a more detailed basis for risk judgment.

[0045] The supplementary data is input into the graph neural network model, and the specific parameters of time-varying weights and spatial correlations are iteratively updated, so that the model can adapt to environmental changes in real time, such as sudden increases in local temperature and accelerated smoke diffusion, to avoid prediction bias caused by data lag.

[0046] Quickly trigger emergency response and shorten handling time: When supplementary data confirms the occurrence of a fire, the resource dispatch mechanism is immediately activated to achieve the precise deployment of firefighting forces and supplies, significantly improving the timeliness and pertinence of the emergency response.

[0047] For example, in highly flammable sub-areas (high temperature above 35°C, humidity below 30%, wind force 7, slope of 25°, altitude of 800 meters, the fire risk probability of sub-area A on the top of the slope has reached 95%, exceeding the threshold of 90%).

[0048] A mobile sensing device (including a high-temperature gas detection unit and a terrain-adaptive motion control module) was deployed in target sub-area A. It autonomously moved along a 25-degree steep slope to the core area at the top, collecting supplemental data in real time: the local temperature soared from 38°C to 45°C (near the ignition point of vegetation) within 10 minutes, smoke concentrations rose from 0.1mg / m³ to 0.8mg / m³ (far exceeding normal levels), and vegetation moisture content dropped from 15% to 10% (extremely dry).

[0049] The above supplementary data is input into the graph neural network model, and the model immediately iteratively updates the parameters: the time-varying weight is tilted towards high temperature and high smoke data within 10 minutes (the weight is increased from 0.9 to 0.95), and the spatial correlation parameters between sub-area A and adjacent area B are adjusted (due to the sudden increase in local wind speed, the vegetation continuity index is increased from 0.061 to 0.085), so that the accuracy of fire risk probability prediction is further optimized, from 95% to 98%, and the probability of fire is extremely high.

[0050] When the supplementary data shows that the smoke concentration is 0.8mg / m 3(reaching the critical value for fire point confirmation) and the temperature reaching 45°C (exceeding the ignition point of vegetation), the system immediately triggers the resource scheduling mechanism: dispatching a fire helicopter 3 kilometers away to prioritize the top of the slope, deploying fire extinguishing bombs and protective equipment from nearby reserves, and at the same time demarcating a warning area with a radius of 1 kilometer centered on sub-area A, prohibiting people from entering.

[0051] Through this process, supplementary data makes model predictions more accurate, and emergency response quickly transitions from early warning to actual combat, effectively shortening the time from risk identification to resource availability, and gaining a critical window for early firefighting.

[0052] The mobile sensing device includes a gas detection unit with a high temperature resistant package and a motion control module based on terrain data; High-temperature resistant packaged gas detection unit: encapsulated in a fireproof and heat-insulating shell, used to detect fire-related gas parameters such as combustible gas concentration and smoke composition in the target sub-area in real time; Motion control module based on terrain data: Built-in sub-region terrain database (including slope and obstacle distribution information), using terrain matching algorithm to plan movement paths, enabling autonomous movement or fixed-point deployment of the device in complex terrain (such as steep slopes and dense vegetation areas); The mobile sensor device is equipped with an overheat protection mechanism. Specifically, when the temperature sensor inside the device detects that the temperature exceeds a preset critical value, it automatically activates a multi-level protection measure, including activating the built-in thermal insulation layer, disconnecting non-core circuits to reduce energy consumption, and sending an overheat warning signal to the monitoring center to ensure the continuity of core detection functions in high-temperature environments. By equipping mobile sensing devices with specialized functional modules (high-temperature resistant packaging, terrain-adaptive motion control, and overheat protection), they can operate stably in extreme environments such as high temperatures and complex terrains, continuously collecting accurate fire-related data. This provides reliable support for fire risk identification and emergency response, and avoids monitoring interruptions or data distortion caused by equipment failure. The high-temperature-resistant encapsulated gas detection unit adopts a fireproof and heat-insulating shell, which can stably detect key parameters such as combustible gas concentration and smoke composition in high-temperature and smoky environments, avoiding equipment damage due to interruption of data collection.

[0053] The motion control module based on terrain data allows the device to move or stay fixed autonomously on complex terrains such as steep slopes and densely vegetated areas, ensuring that the monitoring points cover core areas of fire risk (such as slope tops and areas with dense flammable vegetation).

[0054] The overheat protection mechanism automatically activates multi-level protection when the device temperature exceeds the limit, ensuring that core detection functions are not interrupted and that data collection is continuous and effective in high-temperature environments. The mobile sensing device is a high-temperature fire monitoring robot with terrain adaptive capabilities; The working status and advantages of the robot are as follows: High-temperature-resistant encapsulated gas detection unit: The robot's shell is made of fireproof and heat-insulating materials. Even in a high temperature of 45°C and a smoky environment, the internal gas detection unit can still operate stably, detecting in real time the combustible gas concentration in sub-area A reaching 1200ppm (far exceeding the safety threshold) and a large amount of carbon monoxide in the smoke (concentration 300ppm), providing key data for confirming the fire point.

[0055] Motion control module based on terrain data: The robot has a built-in terrain database of sub-area A, which includes the distribution of obstacles at the 25° steep slope and the edge of the slope top. It plans the path through the terrain matching algorithm, autonomously avoids obstacles such as protruding rocks, and gradually moves along the steep and gentle slope sections to the core monitoring point at the top of the slope and fixes it, ensuring close monitoring of the areas with the highest fire risk.

[0056] Overheat protection mechanism: When the robot's internal temperature sensor detects that the temperature rises to 60°C due to high ambient temperature and continuous operation of the equipment, exceeding the preset critical value of 55°C, the protection measures are immediately activated: the built-in insulation layer is activated to reduce the internal temperature rise rate, non-core lighting circuits are cut off to reduce energy consumption, and an overheat warning signal is sent to the monitoring center. At this time, the core gas detection and temperature monitoring functions are still operating normally, and the combustible gas concentration of 1300ppm and the smoke concentration of 0.9mg / m3 are continuously transmitted to the background. 3 Real-time data from the fire department provides continuous support for the deployment of firefighting forces.

[0057] Through these functions, the fire monitoring robot can still work stably in extreme environments, ensuring uninterrupted data collection and accurate monitoring points, providing reliable guarantees for fire confirmation and emergency resource dispatch.

[0058] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features being referred to. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0059] In the description of this specification, the reference terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" mean that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and features of different embodiments or examples without contradiction.

[0060] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A method for predicting forest fires, characterized in that: The following steps are involved: Step 1: Divide the forest area into several sub-areas and generate a digital map; Step 2: Collect the fire factors of each sub-area, which include static fire factors and dynamic fire factors; Step 3: Perform risk classification based on fire factors and output regional classification information; Step 4: Deploy the monitoring network and obtain monitoring data based on regional classification information; Step 5: Process the monitoring data through the graph neural network model and output the fire risk probability; Step 6: When the fire risk probability exceeds the threshold, activate the emergency response mechanism.

2. A forest fire prediction method according to claim 1, characterized in that: The static fire factors include flammable vegetation density, historical fire point density, lightning strike frequency and terrain characteristics; The dynamic fire factors include vegetation moisture content and human activity intensity distribution data.

3. A forest fire prediction method according to claim 2, characterized in that: The step 3 comprises: Step 3.1: Fuse the collected static fire factors and dynamic fire factors, and encode the attribute information of the two types of factors into a spatiotemporal feature vector that contains both temporal variation characteristics and spatial distribution characteristics; Step 3.2: Input the spatiotemporal feature vector into the clustering model, and generate the flammability classification results of the region through model calculation, which are specifically divided into three levels: high flammability, medium flammability, and low flammability; Step 3.3: Associate real-time and short-term forecast meteorological data, and dynamically adjust and correct the classification results generated in Step 3.2 based on the meteorological data to adapt to the impact of changing meteorological conditions on regional flammability. Meteorological data are collected in real time through satellite remote sensing or ground-based meteorological stations.

4. A forest fire prediction method according to claim 3, characterized in that: The dynamic adjustment and correction includes: When the meteorological conditions corresponding to the meteorological data reach the preset danger standard, the medium and low flammability sub-areas will be upgraded to high flammability categories; The preset adversarial sample generation module is used to simulate the data of fire scenes in which extreme meteorological conditions and terrain features are superimposed, and the simulated data is obtained. The weighted coefficients of the terrain features are optimized based on the simulated data.

5. A forest fire prediction method according to claim 3, characterized in that: The specific process of step 4 is as follows: For highly flammable sub-areas: deploy drone swarms equipped with monitoring equipment for inspections, and dynamically adjust the monitoring frequency based on changes in fire risk probability; For flammable sub-areas: deploy a ground sensor network and combine it with satellite remote sensing data to achieve coordinated ground and satellite monitoring; For low-flammability sub-areas: use satellite remote sensing monitoring to obtain regional macro-monitoring data.

6. A forest fire prediction method according to claim 5, characterized in that: The above configuration of the drone swarm equipped with monitoring equipment during the inspection process is to plan the route of the drone inspection to achieve full coverage of the sub-area, and the route needs to avoid high-risk terrain; The inspection route is updated periodically based on changes in fire risk probability: when the fire risk probability fluctuates slightly, the route is updated at a fixed period; when the fire risk probability fluctuates beyond the preset standard amplitude, the route is updated in real time, giving priority to covering sub-areas where the fire risk probability increases rapidly.

7. The forest fire prediction method according to claim 1, characterized in that: The graph neural network model in step 5 contains the following dual-channel structure: Spatial processing channel: By extracting the vegetation type, vegetation density and terrain connection characteristics of adjacent sub-areas, the vegetation continuity index representing the possibility of fire spread is calculated. This index is used to quantify the spatial correlation of fire spread between adjacent sub-areas; Time processing channel: Based on the acquisition timestamp of the monitoring data, time-varying weight coefficients are assigned to the monitoring data at different times.

8. A forest fire prediction method according to claim 7, characterized in that: The calculation formula of the vegetation continuity index is: ； Among them, d ij is the distance between adjacent sub-regions; σ is the characteristic parameter of vegetation type; V i is the comprehensive quantitative value of vegetation and terrain in sub-region i, V j is the comprehensive quantitative value of vegetation and terrain in sub-region j.

9. The forest fire prediction method according to claim 1, characterized in that: The step 6 specifically includes the following operations: Step 6.1: Deploy mobile sensors in sub-areas where the fire risk probability exceeds the threshold to collect real-time supplementary data. The supplementary data includes local temperature changes, smoke diffusion status, and dynamic changes in vegetation moisture content. Step 6.2: Input the collected supplementary data into the graph neural network model in step 5, and update the model's time-varying weight coefficients and specific parameters of spatial correlation through data iteration; Step 6.3: When the supplementary data meets the preset fire point confirmation criteria, the resource dispatch mechanism is immediately triggered. The dispatch content includes the deployment of firefighting forces, the allocation of emergency materials, and the demarcation of regional warning ranges.

10. A forest fire prediction method according to claim 9, characterized in that: The mobile sensing device includes a gas detection unit in a high temperature resistant package and a motion control module based on terrain data; High-temperature resistant packaged gas detection unit: encapsulated in a fireproof and heat-insulating shell, used to detect combustible gas concentration, smoke composition and fire-related gas parameters in the target sub-area in real time; Motion control module based on terrain data: Built-in sub-region terrain database, using terrain matching algorithm to plan movement paths, enabling autonomous movement or fixed-point deployment of the device in complex terrain; The mobile sensor device is equipped with an overheating protection mechanism. The specific content of the overheating protection mechanism is: when the temperature sensor inside the device detects that the temperature exceeds the preset critical value, it automatically activates multi-level protection measures, including activating the built-in insulation layer, cutting off non-core circuits and sending an overheating warning signal to the monitoring center.

Images