Forest fire prediction method and system based on artificial intelligence

Abstract

The invention relates to a forest fire prediction method and system based on artificial intelligence, and belongs to the technical field of fire prediction, and the system comprises a data collection and fusion module which is used for obtaining multi-source time sequence data of a target forest, the multi-source time series data comprises meteorological data, remote sensing image data, ground sensor data, topographic data and humanistic geographic data; the forest attribute analysis module judges whether the target forest belongs to a high-human-trace activity forest or a low-human-trace activity forest through a preset classification model; the self-adaptive fire risk prediction engine is used for calculating a comprehensive fire risk index of the target forest and generating a fire occurrence probability space-time distribution diagram; and the targeted remedial measure generation module is used for generating differentiated remedial measure plans matched with the forest attributes. According to the method, the human trace activity attributes of the forest can be accurately identified, the fire risk factor weight is adaptively adjusted for forests with different attributes, and the accuracy of comprehensive fire hazard index calculation is improved.

Description

Technical Field

[0001] This invention relates to the field of fire prediction technology, specifically to a forest fire prediction method and system based on artificial intelligence. Background Technology

[0002] Forest fires are one of the major disasters threatening ecological environment security and the safety of people's lives and property. They are characterized by their suddenness, rapid spread, and wide impact. Therefore, accurate prediction and efficient remediation of forest fires are the core requirements of forest fire prevention and control.

[0003] Current forest fire prediction technologies mostly employ uniform prediction models, failing to adequately consider the impact of varying levels of human activity in different forest areas on fire causes. Fires in forests with high human activity (such as suburban woodlands and scenic area forests) are primarily caused by human factors (e.g., careless use of fire, tourist gatherings during holidays), while fires in forests with low human activity (such as remote virgin forests) are mainly caused by natural factors (e.g., lightning strikes, extreme weather conditions). This uniform prediction model ignores these differences, leading to unreasonable risk factor weighting and low prediction accuracy.

[0004] Therefore, those skilled in the art have provided an artificial intelligence-based forest fire prediction method and system to address the problems mentioned in the background section. Summary of the Invention

[0005] The purpose of this invention is to provide an artificial intelligence-based forest fire prediction method and system that accurately identifies the human activity attributes of forests (forests with high human activity / forests with low human activity) and adaptively adjusts the weights of fire risk factors for forests with different attributes, thereby improving the accuracy of the comprehensive fire hazard index calculation and solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: An artificial intelligence-based forest fire prediction system includes: The data acquisition and fusion module is used to acquire multi-source time-series data of the target forest, including meteorological data, remote sensing image data, ground sensor data, topographic data, and human geography data. The forest attribute analysis module is used to determine whether a target forest belongs to a high-human-activity forest or a low-human-activity forest based on the road density, nighttime light index, and proportion of human-caused fire sources extracted from the remote sensing image data and a preset classification model. An adaptive fire risk prediction engine, connected to the forest attribute analysis module, is used to process the multi-source time series data by enabling the corresponding risk factor weighting strategy based on the judgment result of the forest attribute analysis module, calculate the comprehensive fire hazard index of the target forest and generate a spatiotemporal distribution map of the probability of fire occurrence. The targeted remedial measures generation module is connected to the adaptive fire risk prediction engine and is used to generate differentiated remedial measures plans that match forest attributes based on the judgment results of the forest attribute analysis module and the spatiotemporal distribution map of fire occurrence probability.

[0007] As a further aspect of the present invention: the preset classification model in the forest attribute analysis module performs the following process: Feature extraction was performed on remote sensing images of the target forest area over a historical period to obtain the road length per unit area as the road density feature. and average nighttime light intensity as a characteristic of the light index. ; By analyzing historical fire data, the proportion of human-caused fires in a target forest area within the same historical time period is statistically analyzed to obtain the characteristic of human-caused fire source proportion. ; The road density features Lighting index characteristics Characteristics of the proportion of human-caused fire sources The input is fed into a trained lightweight gradient boosting decision tree model, which outputs a binary label, which is the judgment result of high human activity forest or low human activity forest.

[0008] As a further aspect of the present invention: the adaptive fire risk prediction engine includes a first prediction sub-model and a second prediction sub-model in parallel. When the forest attribute is a high-human-activity forest, the adaptive fire risk prediction engine enables the first risk factor weighting strategy corresponding to the first prediction sub-model. When the forest attribute is a low-human-activity forest, the adaptive fire risk prediction engine enables the second risk factor weighting strategy corresponding to the second prediction sub-model.

[0009] As a further aspect of the present invention: the comprehensive fire hazard index The calculation process is as follows: First, calculate the basic hazard values ​​for each natural factor. and human factors baseline risk values ,in The aridity index is obtained by weighted fusion of temperature, humidity, wind speed from the meteorological data and vegetation aridity index extracted from the remote sensing image data. By analyzing the road density characteristics The light index features The results were obtained by logistic regression analysis of holiday markers; Then, the weights are dynamically adjusted according to the activated risk factor weighting strategy to synthesize the fire hazard index. From the formula The calculation shows that, among which The terrain complexity factor is calculated based on the terrain data. For natural factor weights, Human-made factor weights, To fix the terrain factor weights, + + =1; The weighting parameters of the first risk factor weighting strategy are set to satisfy... ; The second risk factor weighting strategy sets the weight parameters to satisfy... .

[0010] As a further aspect of the present invention: the targeted remedial measure generation module includes: The High-Human-Activity Forest Remediation Strategy Unit is activated when the forest attribute is a high-human-activity forest. The remediation plan generated by this unit includes dynamic evacuation route planning. The dynamic evacuation route planning takes the high-risk areas in the spatiotemporal distribution map of the fire probability as the threat source, the population gathering points in the population heat map of the human geography data as the starting point, and the safe area as the ending point to construct a time-varying route planning network. It also uses an improved A* algorithm that considers the cost of fire spread speed and route congestion to calculate multiple backup evacuation routes and their expected safe time windows for each population gathering point. The low-human-activity forest remediation strategy unit is activated when the forest attribute is low-human-activity forest. The remediation measures plan generated by this unit includes the optimal layout scheme of firebreaks. The optimal layout scheme of firebreaks is based on the spatiotemporal distribution map of fire occurrence probability, the slope data in the topographic data, and the vegetation type data extracted from the remote sensing image data. It constructs a cost-benefit optimization model of firebreaks and solves it through a genetic algorithm, outputting the recommended geographical location and priority construction level of the firebreak group.

[0011] As a further aspect of the present invention: the plan generated by the targeted remedial measures generation module also includes: For forests with high human activity, a scheduling plan is generated to strengthen the inspection of man-made fire control points in high-risk areas; For forests with low human activity, generate aerial operation plans for areas with high lightning strike incidence.

[0012] As a further aspect of the present invention, it also includes a feedback learning module for continuously optimizing the adaptive fire risk prediction engine and the targeted remedial measures generation module. The feedback learning module collects data on actual fire incidents, remedial measures taken, and their actual effects. It compares this data with the system's previous predictions and contingency plans to generate an effect evaluation signal. Using the aforementioned effect evaluation signal, the risk factor weighting strategy and the remedial measure generation logic are fine-tuned online through a reinforcement learning algorithm.

[0013] As a further aspect of the present invention, the adaptive fire risk prediction engine also utilizes transfer learning technology to share the underlying feature extraction network parameters between the first prediction sub-model and the second prediction sub-model, but retains their respective independent risk factor weighted fusion layers.

[0014] As a further aspect of the present invention: when the feedback learning module detects a long-term change in the level of human activity in a certain forest area, the system activates a model switching mechanism, so that the adaptive fire risk prediction engine smoothly transitions from the currently enabled first risk factor weighting strategy or second risk factor weighting strategy to the corresponding second risk factor weighting strategy or first risk factor weighting strategy.

[0015] This application also discloses an artificial intelligence-based forest fire prediction method, which employs an artificial intelligence-based forest fire prediction system and includes the following steps: Step 1: Acquire multi-source time-series data of the target forest, including meteorological data, remote sensing image data, ground sensor data, topographic data, and human geography data; Step 2: Based on the road density, nighttime light index, and proportion of human-caused fire sources extracted from the remote sensing image data, determine whether the target forest belongs to a high-human-activity forest or a low-human-activity forest using a preset classification model. Step 3: Based on the multi-source time series data, select the corresponding risk factor weighting strategy according to the judgment result of Step 2, and use the corresponding risk factor weighting strategy to process the multi-source time series data, calculate the comprehensive fire hazard index of the target forest and generate a spatiotemporal distribution map of the probability of fire occurrence. Step 4: Based on the judgment results of Step 2 and the spatiotemporal distribution map of the probability of fire occurrence, generate a differentiated remedial measures plan that matches the forest attributes.

[0016] Compared with the prior art, the beneficial effects of the present invention are: High prediction accuracy: The forest attribute analysis module accurately identifies the attributes of human activities in the forest and adaptively adjusts the weights of risk factors. This solves the problem of inaccurate prediction caused by the neglect of attribute differences in existing unified prediction models. By comprehensively considering multiple factors such as nature, human activities, and terrain, the comprehensiveness and accuracy of fire risk prediction are further improved.

[0017] The remedial measures are highly targeted: Remedial measures centered on personnel safety and fire isolation are formulated to address the different fire prevention needs of forests with high and low human activity. Dynamic evacuation route planning ensures the safety and flexibility of personnel evacuation, and the optimal layout scheme of firebreaks achieves a balance between cost and benefit, solving the problem that existing technical remedial measures are highly versatile but not very effective.

[0018] The system is highly adaptable: the feedback learning module continuously optimizes the model and strategy through feedback from actual data, the transfer learning technology improves the generalization ability and data utilization of the model, and the model switching mechanism adapts to the long-term changes in the attributes of human activity in the forest, enabling the system to run stably for a long time and maintain good performance.

[0019] Highly practical: The system integrates readily available data from multiple sources and adopts mature and efficient algorithm models. The generated prediction results and remedial measures are intuitive and operable, providing scientific decision support for forest fire prevention and control departments and reducing fire losses. Attached Figure Description

[0020] Figure 1 This is a structural block diagram of an artificial intelligence-based forest fire prediction system; Figure 2 This is a flowchart of an artificial intelligence-based forest fire prediction method. Detailed Implementation

[0021] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0022] As mentioned in the background section of this application, research has found that existing forest fire prediction technologies mostly employ uniform prediction models, failing to fully consider the impact of varying levels of human activity in different forest areas on the causes of fires. Fires in forests with high human activity (such as suburban woodlands and scenic area forests) are primarily caused by human factors (such as careless use of fire and tourist gatherings during holidays), while fires in forests with low human activity (such as remote virgin forests) are mainly caused by natural factors (such as lightning strikes and extreme weather conditions). The uniform prediction model ignores these attribute differences, leading to unreasonable risk factor weighting and low prediction accuracy, thus exhibiting certain shortcomings.

[0023] To address the aforementioned shortcomings, this application discloses an artificial intelligence-based forest fire prediction method and system, which can accurately identify the human activity attributes of forests (forests with high human activity / forests with low human activity) and adaptively adjust the weights of fire risk factors for forests with different attributes, thereby improving the accuracy of the comprehensive fire hazard index calculation.

[0024] The following will describe in detail, with reference to the accompanying drawings, how the solution of this application solves the above-mentioned technical problems.

[0025] Please see Figure 1 In this embodiment of the invention, an artificial intelligence-based forest fire prediction system includes: a data acquisition and fusion module for acquiring multi-source time-series data of a target forest, including meteorological data, remote sensing image data, ground sensor data, topographic data, and human geography data; a forest attribute analysis module for determining whether the target forest belongs to a high-human-activity forest or a low-human-activity forest based on road density, nighttime light index, and the proportion of human-caused fire sources in historical fire point data extracted from remote sensing image data, using a preset classification model; an adaptive fire risk prediction engine connected to the forest attribute analysis module for processing the multi-source time-series data using a corresponding risk factor weighting strategy based on the judgment results of the forest attribute analysis module, calculating the comprehensive fire hazard index of the target forest, and generating a spatiotemporal distribution map of fire occurrence probability; and a targeted remedial measures generation module connected to the adaptive fire risk prediction engine for generating differentiated remedial measures plans matching the forest attributes based on the judgment results of the forest attribute analysis module and the spatiotemporal distribution map of fire occurrence probability. This application accurately identifies forest human activity attributes through a forest attribute analysis module and adaptively adjusts risk factor weights, solving the problem of inaccurate predictions caused by existing unified prediction models ignoring attribute differences. By comprehensively considering multiple factors such as natural, human, and topographical factors, it further improves the comprehensiveness and accuracy of fire risk prediction. For forests with high and low human activity, it formulates remedial measures centered on personnel safety and fire isolation, respectively, to address the different fire prevention needs. Dynamic evacuation route planning ensures the safety and flexibility of personnel evacuation, and the optimal layout scheme for firebreaks achieves a balance between cost and benefit, solving the problem that existing remedial measures are highly versatile but ineffective.

[0026] In this embodiment, the preset classification model in the forest attribute analysis module performs the following process: feature extraction is performed on remote sensing images of the target forest area within a historical time period to obtain the road length per unit area as the road density feature. and average nighttime light intensity as a characteristic of the light index. By analyzing historical fire data, the proportion of human-caused fires in a target forest area within the same historical time period is statistically analyzed to obtain the characteristic of human-caused fire source proportion. Road density characteristics Lighting index characteristics Characteristics of the proportion of human-caused fire sources The input is fed into a trained lightweight gradient boosting decision tree model, which outputs a binary label indicating whether the forest is characterized as having high or low human activity. By selecting three core features directly related to forest human activity—road density, nighttime light index, and the proportion of human-caused fire sources—the model achieves accurate characterization of forest attributes. These features are highly complementary and have clear physical meanings. The lightweight gradient boosting decision tree model balances classification accuracy with processing efficiency, quickly outputting reliable binary classification results. This provides precise attribute data for subsequent adaptive fire risk prediction and differentiated remedial measures, preventing prediction strategies and remedial solutions from failing due to biased forest attribute assessments.

[0027] Example: Assuming the target forest region Y is selected, and the historical time period is set to July 1, 2023 - June 30, 2024 (one year), the specific output process for the binary classification labels is as follows: 1. Feature extraction: Road density characteristics Semantic segmentation of roads was performed on 365-day high-resolution remote sensing images of region Y, yielding a total forest area of ​​S=120km². 2 The total road length L = 36km, calculated as follows =S / L=36 / 120=0.3; Nighttime light index characteristics After preprocessing the nighttime remote sensing images, including radiometric calibration and atmospheric correction, the average nighttime light intensity of region Y is statistically analyzed to obtain... =0.21 (value range 0-1, higher value indicates more concentrated human activity at night). Characteristics of the proportion of human-caused fire sources : Retrieve fire records for region Y from historical fire data. The total number of fires is N=8, of which N were caused by human factors (such as agricultural waste fires, fires at remote work sites). h =2 times, calculated to get =N h / N=2 / 8=0.25.

[0028] Model reasoning: =0.3、 =0.21、 =0.25 Input to the trained lightweight gradient boosting decision tree model. The model outputs a binary classification label "0", and the decision region Y is a low-human-activity forest. It should be noted that when <0.4 and <0.3 and When <0.3, output label 0; when >0.6 and >0.6 and When the value is greater than 0.6, the output label is 1. The lightweight gradient boosting decision tree model is a highly efficient and optimized version of the gradient boosting decision tree. Its core is to work together through multiple simple decision trees, with each tree focusing on correcting the prediction error of the previous tree. At the same time, with the help of lightweight design such as data / feature sampling and optimized tree growth, it can significantly improve training and inference speed and reduce memory consumption without reducing prediction accuracy. It can accurately capture non-linear relationships between features, and is not sensitive to noisy data. It has a low deployment threshold and is particularly suitable for scenarios such as forest human activity classification that require fast batch processing and have moderate computing power requirements.

[0029] Results Verification: Region Y is actually a remote primary forest reserve with only a few forest fire prevention patrol roads. The average number of people entering the area is less than 5,000 per year. Human factors account for only 22% of historical fires, which is highly consistent with the model classification results. This proves that the classification method can accurately capture the core differences in human activities in the forest. At the same time, the model only takes 0.5 seconds to process the feature data of Region Y, which can quickly adapt to the batch judgment of attributes in a large area of ​​forest, meeting the real-time requirements of fire prevention and control.

[0030] In this embodiment, the adaptive fire risk prediction engine includes a first prediction sub-model and a second prediction sub-model in parallel. When the forest attribute is a high-human-activity forest, the adaptive fire risk prediction engine activates the first risk factor weighting strategy corresponding to the first prediction sub-model; when the forest attribute is a low-human-activity forest, the adaptive fire risk prediction engine activates the second risk factor weighting strategy corresponding to the second prediction sub-model. By setting up parallel first and second prediction sub-models with corresponding different risk factor weighting strategies, accurate adaptation between fire risk prediction and forest human activity attributes is achieved. This avoids the neglect of core risk factors of forests with different attributes by a unified prediction model, significantly improving the targeting and accuracy of risk prediction. At the same time, the parallel architecture can quickly respond to the forest attribute judgment results, ensuring prediction efficiency and providing accurate and efficient risk basis for the generation of subsequent differentiated remedial measures.

[0031] In this embodiment, the comprehensive fire hazard index The calculation process is as follows: First, calculate the basic hazard values ​​of natural factors respectively. and human factors baseline risk values ,in It is obtained by weighted fusion of temperature, humidity, wind speed from meteorological data and vegetation aridity index extracted from remote sensing image data. By analyzing road density characteristics Lighting index characteristics The fire hazard index is obtained by performing logistic regression on holiday markers; then, the weights are dynamically adjusted according to the activated risk factor weighting strategy to obtain the comprehensive fire hazard index. From the formula The calculation shows that, among which The terrain complexity factor is calculated based on terrain data. For natural factor weights, Human-made factor weights, To fix the terrain factor weights, + + =1; The weighting parameters of the first risk factor weighting strategy satisfy... The weighting parameters of the second risk factor weighting strategy are set to meet the following requirements. By clarifying the calculation logic of three core risk elements—natural factors, human factors, and topographic factors—a multi-dimensional and comprehensive quantification of fire risk was achieved. Furthermore, by dynamically adjusting the α and β weights to adapt to different forest attributes, the calculation of the comprehensive fire hazard index is both mathematically rigorously supported and accurately matches the dominant risk factors of different forests. This avoids misjudgments caused by single dimensions or fixed weights, significantly improving the scientific rigor and accuracy of fire risk assessment and providing a reliable quantitative basis for generating spatiotemporal distribution maps of fire probability.

[0032] Example: Two typical forest areas are selected for verification. Area M is a suburban scenic forest (classified as a forest with high human activity by the classification model), and area N is a remote primary forest (classified as a forest with low human activity). The calculation process is as follows: 1. Basic parameter settings: Natural Factor Weighted Fusion Weights: Temperature Weight w T =0.3, humidity weight w H =−0.2 (higher humidity, lower risk), wind speed weight w V =0.2, vegetation aridity index weight w VDDI =0.3; The weights for calculating terrain complexity factors are: slope weight a = 0.4, aspect score weight b = 0.3, and elevation undulation weight c = 0.3. The overall index weight constraint is α+β+γ=1, and the fixed terrain factor weight is γ=0.2.

[0033] 2. Basic risk value of natural factors calculate: Region M: Temperature T after normalization of meteorological data norm =0.6, humidity H norm =0.4, wind speed V norm =0.3, Vegetation Aridity Index (VDDI) extracted from remote sensing imagery norm =0.5, calculated according to the weighted fusion formula: ; Region N: Temperature T norm =0.5, humidity H norm =0.3, wind speed V norm =0.4, Vegetation Aridity Index (VDDI) norm =0.6, calculated as follows: .

[0034] 3. Baseline hazard values ​​for human factors calculate: Area M: Road density =0.8, Nighttime Light Index =0.7, Holiday sign H hol =1 (prediction period includes holidays), obtained by fitting a logistic regression model: ; Area N: Road density =0.1, Nighttime Light Index =0.2, Holiday sign H hol =0 (prediction period excludes holidays), calculated as follows: .

[0035] 4. Terrain complexity factor calculate: Area M: Slope S norm =0.3, Aspect Rating A score =0.7 (eastward slope), elevation relief E norm =0.2, calculated as follows: ; Region N: Slope S norm =0.6, Aspect Rating A score =0.7 (south-facing slope), elevation relief E norm =0.8, calculated as follows: .

[0036] 5. Overall Fire Hazard Index calculate: Region M (High Population): Activate the first risk factor weighted strategy (β>α), setting α=0.3 and β=0.5, and substitute into the formula: ; Region N (low population density): Activate the second risk factor weighting strategy (α>β), setting α=0.5 and β=0.3, and substitute into the formula: =0.5×0.38+0.3×0.495+0.2×0.69=0.19+0.1485+0.138=0.4765.

[0037] 6. Comparative verification (if factor segmentation or fixed weights are ignored): Assuming that natural and human factors are simply averaged without distinguishing attributes (α=0.4, β=0.4, γ=0.2), then the region M... This is 17.9% lower than the above calculation result (0.593), failing to reflect the dominant role of human risk in high-population areas; Region N The result was underestimated by 5.1% compared to the calculated result (0.4765), which weakened the impact of natural risks in low-population areas.

[0038] In this embodiment, the targeted remedial measures generation module includes: a high-human-activity forest remedial strategy unit, which is activated when the forest attribute is high-human-activity forest, and the remedial measures plan generated by it includes dynamic evacuation route planning; the dynamic evacuation route planning uses high-risk areas in the spatiotemporal distribution map of fire occurrence probability as the threat source, population gathering points in the population heat map of human geography data as the starting point, and safe areas as the ending point to construct a time-varying route planning network, and uses an improved A* algorithm that considers the cost of fire spread speed and route congestion cost to calculate multiple backup evacuation routes and their expected safe time windows for each population gathering point; and a low-human-activity forest remedial strategy unit, which is activated when the forest attribute is low-human-activity forest, and the remedial measures plan generated by it includes optimal barrier zone layout scheme; the optimal barrier zone layout scheme is based on the spatiotemporal distribution map of fire occurrence probability, slope data in terrain data, and vegetation type data extracted from remote sensing image data to construct a barrier zone cost-benefit optimization model, and solves it through a genetic algorithm to output the recommended geographical location and priority construction level of the firebreak group. By dividing forest remediation strategy units into high-human-activity and low-human-activity forest remediation strategy units, precise matching of remediation measures with forest attributes and fire risk distribution is achieved. The high-human-activity forest remediation strategy unit focuses on dynamic evacuation route planning, and improves the A* algorithm to take into account the impact of fire spread and route congestion, ensuring the safety and flexibility of personnel evacuation. The low-human-activity forest remediation strategy unit focuses on the optimal layout of firebreaks, and achieves a balance between efficient resource utilization and fire-stopping effect through cost-benefit optimization models and genetic algorithms, avoiding the problems of insufficient targeting and poor effect of general remediation measures, and significantly improving the scientific nature and practicality of fire remediation.

[0039] Example: Using the same regions M (suburban scenic forests, forests with high human activity) and N (remote primary forests, forests with low human activity), the following implementation will be carried out: 1. Region M (High-human-activity forest): Dynamic evacuation route planning; Basic parameter settings: Threat Sources: Spatiotemporal Distribution Map of Fire Occurrence Probability High-risk areas with a risk level of ≥0.6 (concentrated on the west side of the scenic area entrance and around the camping base); Starting point: Three population gathering points in human geography data (A1: scenic area entrance, gathering of approximately 300 people; A2: camping base, gathering of approximately 150 people; A3: edge of suburban residential area, gathering of approximately 200 people). End point: Safe area (open area along the provincial highway away from the forest, marked as point S); Improved A* algorithm cost function: Where g(n) is the actual distance from the starting point to the current node (unit: km), h(n) is the estimated Manhattan distance from the current node to the ending point (unit: km), and f s c(n) represents the time cost of the fire spreading to the current path segment (unit: h), c(n) represents the path congestion cost (calculated based on population density and road width, with a value range of [0,1]), and the weights w1=0.4 and w2=0.3.

[0040] Path calculation results: Generate 3 alternative routes for A1 (scenic area entrance): Route 1: A1 → Main Road → Point S, g(n) = 2.5km, h(n) = 2.0km, f s Given c(n) = 1.8h and c(n) = 0.3, we calculate f(n) = 2.5 + 2.0 + 0.4 × 1.8 + 0.3 × 0.3 = 5.21. The expected safe time window is the next 2.5 hours. Route 2: A1 → Branch Road → Rural Road → Point S, g(n) = 3.2km, h(n) = 2.8km, f s Given c(n) = 2.5h and c(n) = 0.1, we calculate f(n) = 3.2 + 2.8 + 0.4 × 2.5 + 0.3 × 0.1 = 6.83. The expected safe time window is the next 3.5 hours. Route 3: A1 → Forest Emergency Passage → Point S, g(n) = 2.8km, h(n) = 2.3km, f s Given c(n) = 2.1h and c(n) = 0.2, we calculate f(n) = 2.8 + 2.3 + 0.4 × 2.1 + 0.3 × 0.2 = 5.92. The estimated safe time window is the next 3.0 hours. Similarly, calculate three alternative paths for A2 and A3 respectively, and mark the path length, estimated travel time and safety window for each path.

[0041] 2. Region N (Low-human activity forest): Optimal barrier layout scheme; Basic parameter settings: Input data: Spatiotemporal distribution map of fire occurrence probability (high-risk areas are concentrated on south-facing slopes). ≥0.5), topographic data (slope 15°-30°, average slope 22°), and vegetation type data extracted from remote sensing images (mainly coniferous forests with high flammability). Cost-benefit optimization model: Objective function: max(E−C), where E is the barrier band blocking benefit (E=k1P) block +k2S protect P block S represents the probability of preventing the spread of fire. protectThe protected forest area is k1=0.6, k2=0.4; C is the construction cost (C=m1L+m2D, L is the length of the barrier strip, D is the construction difficulty coefficient, D=1.2 when coniferous forest + slope>20°, m1=0.5, m2=0.3). Constraints: The barrier zone must be ≥30m wide and ≤5km apart, avoiding the core ecological protection area.

[0042] Results of the genetic algorithm solution: Firebreak barrier system plan: 2 primary firebreaks (priority construction), 3 secondary firebreaks, details as follows: Level 1 barrier zone 1: Deployed along the northern boundary of the high-risk area, with a length of 4.2km, a width of 35m, E=0.85, C=0.32, and E−C=0.53; Level 1 barrier zone 2: It traverses the central part of the high-risk area, with a length of 3.8km, a width of 30m, E=0.82, C=0.30, and E−C=0.52; Secondary barrier zone: Deployed around the perimeter of high-risk areas, with a total length of 12.5km, a width of 30m, and an average E−C=0.41.

[0043] In this embodiment, the contingency plan generated by the targeted remedial measures generation module also includes: for forests with high human activity, a scheduling plan for strengthening inspections of man-made fire control points in high-risk areas; and for forests with low human activity, an aerial operation plan for areas prone to lightning strikes. This setting precisely supplements the core remedial measures for forests with high and low human activity. In high-human-activity areas, the focus is on controlling man-made fire sources in high-risk areas, curbing man-made fires in advance through strengthened inspections and scheduling. In low-human-activity areas, the focus is on lightning strikes, the main natural fire source, improving monitoring and emergency response efficiency through aerial operation planning. This makes the remedial measures plan more comprehensive and better suited to the risk characteristics of the two types of forests, extending from "post-event remediation" to "pre-event prevention," further reducing the probability of fire occurrence and the risk of fire spread.

[0044] Example: Using region M (suburban scenic forest, forest with high human activity) and region N (remote primeval forest, forest with low human activity), execute as follows: 1. Area M (High-human-activity forest): Enhanced patrol and dispatch plan for high-risk area man-made fire source control points; Basic parameter settings: Control point identification: Based on the spatiotemporal distribution map of fire occurrence probability ( Based on high-risk areas (≥0.6) and human geography data, five key human-caused fire source control points were identified (P1: smoking areas in scenic spots, P2: fire-using points in camping bases, P3: sacrificial activity sites in suburban areas, P4: construction sites on the edge of forests, P5: rest areas along roads). Inspection resources: 4 inspection personnel and 2 inspection vehicles are allocated, with an inspection radius covering a 1km radius around the control point; Scheduling optimization objective: Maximize inspection coverage efficiency η. , where t i Let s be the inspection duration of the i-th control point (in hours). i Risk weight for the i-th control point (based on The values ​​for P1-P5 are set to 0.9, 0.85, 0.8, 0.75, and 0.7 respectively. T is the total inspection time (8 hours per day), and R is the inspection resource input coefficient (personnel + vehicles, with a value of 1.2).

[0045] Scheduling scheme generation: Inspection frequency: High-risk control points (P1, P2) are inspected once every 2 hours, with each inspection lasting 0.5 hours; medium-risk control points (P3, P4, P5) are inspected once every 4 hours, with each inspection lasting 0.3 hours. Inspection route: The route is optimized using a genetic algorithm and is planned as “Inspection Base → P1 → P2 → P3 → P4 → P5 → Inspection Base”, with a total route length of 28km and a single trip time of 2.5h. Three rounds of full-coverage inspections are completed daily. Temporary control measures: During holidays (such as weekends and statutory holidays), temporary checkpoints will be set up at P1 and P2, with one additional patrol officer, and it will be prohibited to bring any source of fire into the vicinity of the control point within 500m.

[0046] 2. Area N (Low-human activity forest): Aerial operation planning for high-risk lightning strike areas; Basic parameter settings: Identification of high-incidence areas of lightning strikes: Based on historical lightning strike data and topographic data, three core high-incidence areas of lightning strikes were identified (Q1: south-facing coniferous forest belt, Q2: ridgeline above 1200m altitude, Q3: concentrated area of ​​isolated tall trees), with a total area of ​​18km². Aerial operation resources: 1 helicopter (cruising speed 180km / h, endurance 4h), 3 sets of portable lightning monitoring sensors (monitoring radius 5km). Operational planning objectives: Lightning strike monitoring coverage θ≥95%, emergency response time t≤30min.

[0047] Aerial operation plan generation: Patrol route: Design a circular patrol route of "base → Q1 → Q2 → Q3 → Q1 → base" with a total length of 120km. Each patrol takes 40 minutes. Patrols will be carried out once a day at 9:00 AM, 12:00 PM, 3:00 PM, and 6:00 PM to cover the high-risk periods for lightning strikes. Monitoring equipment deployment: One set of lightning monitoring sensors is deployed at each of the core locations Q1, Q2, and Q3. The data is transmitted to the ground command center in real time. When lightning activity is detected, the helicopter will immediately go to the target area for patrol. Emergency standby: During the peak lightning season (June-August), helicopters are on standby at the base 24 hours a day, equipped with fire extinguishing bombs and rescue equipment. Upon receiving a lightning warning or sensor alarm, they will arrive at the target area within 30 minutes.

[0048] In this embodiment, a feedback learning module is also included to continuously optimize the adaptive fire risk prediction engine and the targeted remedial measures generation module. The feedback learning module collects data on actual fire events, remedial measures taken, and their actual effects. It compares this data with the system's previous predictions and plans to generate an effect evaluation signal. Using the effect evaluation signal, the risk factor weighting strategy and remedial measures generation logic are fine-tuned online through a reinforcement learning algorithm. For example, the effect evaluation signal is labeled as S, with a value range of (0-1). The core of reinforcement learning is trial-and-error learning of "agent-environment interaction + reward / punishment," with the goal of finding the optimal strategy that maximizes the reward. Reinforcement learning rules can be set: if S < 0.8 (poor effect, triggering punishment), the human factor weight β (core of high-human-occupancy areas) is increased, and the frequency of inspections is increased; if S ≥ 0.8 (good effect, triggering reward), the parameters remain unchanged. By constructing a closed-loop iterative mechanism of "prediction-practice-evaluation-optimization" through a feedback learning module, evaluation signals are generated using actual fire events, remedial measures, and effect data. Through reinforcement learning, the risk factor weighting strategy and remedial measure generation logic are fine-tuned online. This can dynamically correct the deviation between the initial system model and the actual scenario, adapt to the dynamic changes in forest fire risk, avoid system performance degradation due to long-term operation or scenario changes, continuously improve the accuracy of the prediction engine and the practical effectiveness of remedial measures, and ensure the long-term stable application value of the system.

[0049] In this embodiment, the adaptive fire risk prediction engine also utilizes transfer learning technology to share the underlying feature extraction network parameters between the first and second prediction sub-models, while retaining their respective independent risk factor weighted fusion layers. The core value of transfer learning is to avoid repetitive work by leveraging "common knowledge" across scenarios, while preserving the "specific needs" of the target domain to ensure adaptability. Simply put, it's like having "high-achieving, experienced student M" teach "general learning methods" (such as how to analyze questions and extract key information) to "student N with weak foundations and little practice." N doesn't need to learn from scratch; they only need to focus on refining "answering strategies suitable for themselves" to quickly improve their grades—this is the core role of transfer learning in this application. Corresponding scenario to this application: Source domain (sufficient data): Forests with high human activity (such as region M) have accumulated a large amount of multi-source time-series data and fire labels, which can fully train the general ability of "multi-source data low-level feature extraction"; Target domain (data scarcity): Low-human-activity forests (such as region N), with limited data and rare fire events, resulting in poor performance when training models alone; Common knowledge: The underlying feature extraction logic of multi-source data (such as the temporal trend of meteorological data and the spatial texture features of remote sensing images; whether it is a highly populated area or a sparsely populated forest, the basic processing logic of these data is consistent). Specific requirements: Risk factor weighted fusion strategy for different forests (high human activity requires β>α, low human activity requires α>β, this part must be independent and cannot be transferred).

[0050] By using transfer learning techniques, two prediction sub-models can share the underlying feature extraction network parameters, which reduces the computational cost and data dependence caused by repeated training. In particular, it solves the model training problem in data-scarce areas such as forests with low human activity. By retaining independent risk factor weighted fusion layers, it ensures that the weighting strategies corresponding to different forest attributes are not lost. At the same time, the sub-models can learn from each other's effective underlying features (such as common representations of multi-source data), which improves the model's generalization ability and training efficiency, shortens the iteration cycle, and makes fire risk prediction more practical for engineering and adaptable to different scenarios while ensuring accuracy and adaptability.

[0051] In this embodiment, when the feedback learning module detects a long-term change in the level of human activity in a forest area, the system activates a model switching mechanism. This smoothly transitions the adaptive fire risk prediction engine from the currently enabled first-risk-factor weighted strategy or second-risk-factor weighted strategy to the corresponding second-risk-factor weighted strategy or first-risk-factor weighted strategy. This setting avoids a sharp drop in prediction accuracy and a discontinuity in remedial measures caused by abrupt weight changes. Simultaneously, it relies on the monitoring data from the feedback learning module to ensure the accuracy of the switching timing. Combined with the shared underlying feature parameters from transfer learning, the model can quickly adapt to new attributes without retraining. This preserves the optimization results of the original model while responding promptly to dynamic changes in forest attributes, ensuring that the system maintains stable and accurate prediction and remedial effects during attribute transition periods, thus improving the system's long-term dynamic adaptability.

[0052] like Figure 2 As shown, this application also discloses an artificial intelligence-based forest fire prediction method, which employs an artificial intelligence-based forest fire prediction system and includes the following steps: Step 1: Acquire multi-source time-series data of the target forest. The multi-source time-series data includes meteorological data, remote sensing image data, ground sensor data, topographic data, and human geography data. Step 2: Based on the road density, nighttime light index, and proportion of human-caused fire sources extracted from remote sensing image data, determine whether the target forest belongs to a high-human-activity forest or a low-human-activity forest using a preset classification model. Step 3: Based on multi-source time series data, select the corresponding risk factor weighting strategy according to the judgment results of Step 2, and use the corresponding risk factor weighting strategy to process the multi-source time series data, calculate the comprehensive fire hazard index of the target forest and generate a spatiotemporal distribution map of fire occurrence probability. Step 4: Based on the judgment results of Step 2 and the spatiotemporal distribution map of fire occurrence probability, generate a differentiated remedial measure plan that matches the forest attributes.

[0053] This invention accurately identifies forest human activity attributes through a forest attribute analysis module and adaptively adjusts risk factor weights, solving the problem of inaccurate predictions caused by existing unified prediction models ignoring attribute differences. By comprehensively considering natural, human, and topographical factors, it further improves the comprehensiveness and accuracy of fire risk prediction. For forests with high and low human activity, different remedial measures are formulated, focusing on personnel safety and fire isolation, respectively. Dynamic evacuation route planning ensures the safety and flexibility of personnel evacuation, and the optimal deployment scheme of firebreaks achieves a balance between cost and benefit, addressing the problem of existing remedial measures being highly generalized but ineffective. Furthermore, the feedback learning module continuously optimizes the model and strategies through actual data feedback. Transfer learning technology improves the model's generalization ability and data utilization. The model switching mechanism adapts to long-term changes in forest human activity attributes, enabling the system to operate stably and maintain good performance over the long term. The system integrates readily available multi-source data and employs mature and efficient algorithm models. The generated prediction results and remedial measures are intuitive and operable, providing scientific decision support for forest fire prevention and control departments and reducing fire losses.

[0054] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A forest fire prediction system based on artificial intelligence, characterized in that, include: The data acquisition and fusion module is used to acquire multi-source time-series data of the target forest, including meteorological data, remote sensing image data, ground sensor data, topographic data, and human geography data. The forest attribute analysis module is used to determine whether a target forest belongs to a high-human-activity forest or a low-human-activity forest based on the road density, nighttime light index, and proportion of human-caused fire sources extracted from the remote sensing image data and a preset classification model. An adaptive fire risk prediction engine, connected to the forest attribute analysis module, is used to process the multi-source time series data by enabling the corresponding risk factor weighting strategy based on the judgment result of the forest attribute analysis module, calculate the comprehensive fire hazard index of the target forest and generate a spatiotemporal distribution map of the probability of fire occurrence. The targeted remedial measures generation module is connected to the adaptive fire risk prediction engine and is used to generate differentiated remedial measures plans that match forest attributes based on the judgment results of the forest attribute analysis module and the spatiotemporal distribution map of fire occurrence probability.

2. The forest fire prediction system based on artificial intelligence according to claim 1, characterized in that, The preset classification model in the forest attribute analysis module performs the following process: Feature extraction was performed on remote sensing images of the target forest area over a historical period to obtain the road length per unit area as the road density feature. and average nighttime light intensity as a characteristic of the light index. ; By analyzing historical fire data, the proportion of human-caused fires in a target forest area within the same historical time period is statistically analyzed to obtain the characteristic of human-caused fire source proportion. ; The road density features Lighting index characteristics Characteristics of the proportion of human-caused fire sources The input is fed into a trained lightweight gradient boosting decision tree model, which outputs a binary label, which is the judgment result of high human activity forest or low human activity forest.

3. The forest fire prediction system based on artificial intelligence according to claim 2, characterized in that, The adaptive fire risk prediction engine includes a first prediction sub-model and a second prediction sub-model in parallel. When the forest attribute is a high-human-activity forest, the adaptive fire risk prediction engine enables the first risk factor weighting strategy corresponding to the first prediction sub-model. When the forest attribute is a low-human-activity forest, the adaptive fire risk prediction engine enables the second risk factor weighting strategy corresponding to the second prediction sub-model.

4. The forest fire prediction system based on artificial intelligence according to claim 3, characterized in that, The comprehensive fire hazard index The calculation process is as follows: First, calculate the basic hazard values ​​for each natural factor. and human factors baseline risk values ,in The aridity index is obtained by weighted fusion of temperature, humidity, wind speed from the meteorological data and vegetation aridity index extracted from the remote sensing image data. By analyzing the road density characteristics The light index features The results were obtained by logistic regression analysis of holiday markers; Then, the weights are dynamically adjusted according to the activated risk factor weighting strategy to synthesize the fire hazard index. From the formula The calculation shows that, among which The terrain complexity factor is calculated based on the terrain data. For natural factor weights, Human-made factor weights, To fix the terrain factor weights, + + =1; The weighting parameters of the first risk factor weighting strategy are set to satisfy... ; The second risk factor weighting strategy sets the weight parameters to satisfy... .

5. A forest fire prediction system based on artificial intelligence according to claim 4, characterized in that, The targeted remedial measures generation module includes: The High-Human-Activity Forest Remediation Strategy Unit is activated when the forest attribute is a high-human-activity forest. The remediation plan generated by this unit includes dynamic evacuation route planning. The dynamic evacuation route planning takes the high-risk areas in the spatiotemporal distribution map of the fire probability as the threat source, the population gathering points in the population heat map of the human geography data as the starting point, and the safe area as the ending point to construct a time-varying route planning network. It also uses an improved A* algorithm that considers the cost of fire spread speed and route congestion to calculate multiple backup evacuation routes and their expected safe time windows for each population gathering point. The low-human-activity forest remediation strategy unit is activated when the forest attribute is low-human-activity forest. The remediation measures plan generated by this unit includes the optimal layout scheme of firebreaks. The optimal layout scheme of firebreaks is based on the spatiotemporal distribution map of fire occurrence probability, the slope data in the topographic data, and the vegetation type data extracted from the remote sensing image data. It constructs a cost-benefit optimization model of firebreaks and solves it through a genetic algorithm, outputting the recommended geographical location and priority construction level of the firebreak group.

6. A forest fire prediction system based on artificial intelligence according to claim 5, characterized in that, The contingency plan generated by the targeted remedial measures generation module also includes: For forests with high human activity, a scheduling plan is generated to strengthen the inspection of man-made fire control points in high-risk areas; For forests with low human activity, generate aerial operation plans for areas with high lightning strike incidence.

7. A forest fire prediction system based on artificial intelligence according to claim 6, characterized in that, It also includes a feedback learning module for continuously optimizing the adaptive fire risk prediction engine and the targeted remedial measures generation module; The feedback learning module collects data on actual fire incidents, remedial measures taken, and their actual effects. It compares this data with the system's previous predictions and contingency plans to generate an effect evaluation signal. Using the aforementioned effect evaluation signal, the risk factor weighting strategy and the remedial measure generation logic are fine-tuned online through a reinforcement learning algorithm.

8. A forest fire prediction system based on artificial intelligence according to claim 7, characterized in that, The adaptive fire risk prediction engine also utilizes transfer learning technology to share the underlying feature extraction network parameters between the first prediction sub-model and the second prediction sub-model, but retains their respective independent risk factor weighted fusion layers.

9. A forest fire prediction system based on artificial intelligence according to claim 8, characterized in that, When the feedback learning module detects a long-term change in the level of human activity in a certain forest area, the system activates the model switching mechanism, so that the adaptive fire risk prediction engine can smoothly transition from the currently enabled first risk factor weighting strategy or second risk factor weighting strategy to the other corresponding second risk factor weighting strategy or first risk factor weighting strategy.

10. A forest fire prediction method based on artificial intelligence, characterized in that, The forest fire prediction system based on artificial intelligence as described in any one of claims 1-9 includes the following steps: Step 1: Acquire multi-source time-series data of the target forest, including meteorological data, remote sensing image data, ground sensor data, topographic data, and human geography data; Step 2: Based on the road density, nighttime light index, and proportion of human-caused fire sources extracted from the remote sensing image data, determine whether the target forest belongs to a high-human-activity forest or a low-human-activity forest using a preset classification model. Step 3: Based on the multi-source time series data, select the corresponding risk factor weighting strategy according to the judgment result of Step 2, and use the corresponding risk factor weighting strategy to process the multi-source time series data, calculate the comprehensive fire hazard index of the target forest and generate a spatiotemporal distribution map of the probability of fire occurrence. Step 4: Based on the judgment results of Step 2 and the spatiotemporal distribution map of the probability of fire occurrence, generate a differentiated remedial measures plan that matches the forest attributes.

Images