A fire prevention early warning method based on multi-dimensional forest health state analysis

Abstract

This application discloses a fire early warning method based on multidimensional forest health status analysis, belonging to the field of forest fire prevention technology. The method includes: acquiring multidimensional observable data for each assessment unit in a target forest area; inputting the multidimensional observable data into a pre-constructed hybrid prediction model, outputting a comprehensive flammability risk index for each assessment unit within a specified future time period, and generating a visualized spatiotemporal distribution map on a GIS map; wherein the hybrid prediction model calculates the flammability probability of each tree species using a single-tree physical mechanism model, obtains a stand risk index using a spatial weighted aggregation algorithm, outputs a landscape-level probability through a landscape-level risk prediction network, and performs weighted fusion of three sources of uncertainty based on traditional empirical paths; and issuing early warning information based on the spatiotemporal distribution map. This application combines physical mechanisms with transfer learning to achieve rapid adaptation with few samples and accurate prediction at multiple scales, providing reliable decision support for forest fire prevention.

Description

Technical Field

[0001] This application relates to the field of intelligent sensing and early warning of ecological security, and in particular to a fire prevention and early warning method based on multidimensional forest health status analysis, as well as a computing device. Background Technology

[0002] Forest fire prevention is a crucial aspect of ecological security. Currently, forest fire early warning mainly relies on satellite thermal infrared monitoring, video smoke recognition, and manual patrols. Satellite thermal infrared monitoring detects high-temperature anomalies on the ground using polar-orbiting satellites, but the long satellite transit period and cloud cover can easily lead to missed detections. Furthermore, it can only detect fires when the fire area reaches 0.1 hectares or more, by which time the fire is already difficult to control. Forest fire video monitoring systems can automatically identify smoke and fire, but the coverage radius of a single camera is limited (less than 5 kilometers), and the false alarm rate is as high as 30% or more at night and in rainy or foggy weather. Manual patrols rely on forest rangers conducting foot patrols, with an average daily patrol radius of less than 5 kilometers, leaving many monitoring blind spots in remote mountainous areas. Essentially, all of these traditional methods only trigger alarms when the fire has already developed into open flames or smoke, by which time the fire has often spread to an uncontrollable level, failing to achieve early prevention.

[0003] To improve the timeliness of early warnings, existing technologies attempt to introduce data-driven AI models to predict fire risk by analyzing factors such as meteorology, vegetation, and topography. However, these methods face the following inherent drawbacks: First, the scarcity of data. Historical fire data represents low-probability events, and the probability of fires varies greatly across different regions and seasons. Purely data-driven models are difficult to train effectively in areas with sparse data. Traditional deep learning methods require thousands of samples to achieve usable accuracy, while most forest areas have fewer than a hundred historical fire records. Second, poor species adaptability. Different tree species have significantly different physiological characteristics: pine trees are rich in oil and highly flammable; birch trees have high water content and are relatively difficult to burn; oak trees have slow litter decomposition and accumulate a lot of surface combustibles. Existing models treat vegetation as a homogenized layer and use a single vegetation index to represent vegetation status, failing to distinguish the differences in flammability among different tree species, leading to a significant decrease in prediction accuracy in mixed forest areas. Third, a lack of mechanistic support. The flammability of trees is directly related to physical factors such as water deficit, physiological activity, internal temperature distribution, and litter accumulation. Purely data-driven models cannot characterize these mechanistic processes, and the prediction results lack biological and physical basis, have weak interpretability, and are difficult to provide scientific guidance for forest fire prevention. Fourth, the prediction timeliness is insufficient. Existing methods mostly assess the risk of the day based on meteorological data at the current moment, lacking the ability to predict the evolution of risk in future periods, resulting in insufficient advance warning and difficulty in supporting the formulation of preventive measures. While time-series prediction schemes based on interpolation or missing value processing can fill in the missing data in meteorological observations, restoring the data to a continuous sequence before applying deep learning models to predict future risks, the virtual data generated by interpolation deviates significantly from the actual meteorological evolution. This error will be passed on to subsequent prediction stages, leading to false alarms or missed alarms, and cannot fundamentally solve the problem of scarce historical fire data. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] In view of the aforementioned shortcomings and deficiencies of existing technologies, this application provides a fire early warning method based on multidimensional forest health status analysis, overcoming the limitations of existing methods such as satellite thermal infrared monitoring, video smoke recognition, and manual patrols in proactive prevention; the weaknesses of existing data-driven AI models in generalization under conditions of scarce historical fire data and difficulty in adapting to differences in the flammability of different tree species; the inability of purely data-driven models to characterize physical mechanisms such as water deficit and physiological activity in trees, and the lack of interpretability in prediction results; and the lack of ability to predict future risk evolution and insufficient early warning lead time in existing methods. Simultaneously, it overcomes the limitations of existing technologies in predicting multi-source fires. To address the shortcomings of uncertainty quantification and adaptive fusion in measurement results, and the difficulty in assessing the reliability of early warnings, this application provides a fire early warning method based on multidimensional forest health status analysis. It introduces plant physiological constraints through a single-tree physical mechanism model, enables rapid adaptation to new tree species with a small number of samples through cross-species transfer learning, achieves progressive risk assessment from single tree to forest stand to landscape through multi-scale spatial aggregation, and improves prediction accuracy through weighted fusion of three sources of uncertainty. Finally, it outputs the spatiotemporal distribution of flammability risk for each assessment unit in the next 7-30 days, achieving high-precision, long-term, and proactive prediction and early warning of forest fire risks for different tree species and at different scales.

[0006] (II) Technical Solution

[0007] To achieve the above objectives, the main technical solutions adopted in this application include:

[0008] In a first aspect, embodiments of this application provide a fire early warning method based on multidimensional forest health status analysis, the specific steps of which include:

[0009] S100. Acquire multispectral remote sensing images corresponding to each evaluation unit of the target forest area, and preprocess the multispectral remote sensing images; the evaluation unit is a grid of multiple predetermined spatial scales divided from the target forest area.

[0010] Vegetation index data, soil index data, topographic data and biological status data of each assessment unit are extracted from the preprocessed multispectral remote sensing images, and meteorological data of the corresponding locations are fused to obtain multidimensional observable data of each assessment unit.

[0011] S200. Input the multidimensional observable data of each assessment unit into the pre-built hybrid prediction model, and output the comprehensive flammability risk index of each assessment unit within a specified future time period; and generate a visualized spatiotemporal distribution map of flammability risk on a GIS map based on the comprehensive flammability risk index.

[0012] S300. Generate and release corresponding forest fire prevention early warning information based on the spatiotemporal distribution map of flammability risk.

[0013] Optionally, in some embodiments of this application, the processing of the hybrid prediction model in S200 includes:

[0014] S201. Identify the tree species type of each tree in the current evaluation unit from the biological state data in the multidimensional observable data;

[0015] S202. By querying the pre-built tree species mechanism model parameter library, determine whether each tree species type corresponds to a pre-trained dedicated mechanism model parameter; the tree species mechanism model parameter library is pre-built based on historical observation data and stores dedicated mechanism model parameters for multiple tree species.

[0016] If the tree species type corresponds to a pre-trained dedicated mechanism model parameter, then the dedicated mechanism model parameter is loaded into the single tree physical mechanism model;

[0017] If the tree species type does not have pre-trained dedicated mechanism model parameters, then the cross-species migration model is used to generate dedicated mechanism model parameters applicable to the tree species based on 10-50 calibration sample data of the tree species. The generated dedicated mechanism model parameters are then loaded into the single tree physical mechanism model and stored in the tree species mechanism model parameter library.

[0018] The meteorological and soil data corresponding to the current single tree are input into the single tree physical mechanism model after loading parameters, and the single tree flammability probability is calculated.

[0019] Summarize the individual tree flammability probability of all trees in the current assessment unit to obtain the set of individual tree flammability probabilities for each tree in the current assessment unit.

[0020] S203. Based on the set of combustibility probabilities of individual trees, and combined with the spatial location coordinates and canopy area of ​​each tree extracted from multidimensional observable data, the stand risk index of the assessment unit is calculated using a spatial weighted aggregation algorithm.

[0021] S204. The forest stand risk index is used as a vegetation feature and is combined with the topographic data and meteorological data in the multidimensional observable data to form a multidimensional feature vector, which is then input into the landscape-level risk prediction network to output the landscape-level flammability risk probability.

[0022] S205. Calculate the forest health index based on the multidimensional observable data, and convert the forest health index into a traditional empirical flammability risk index through a preset mapping relationship.

[0023] S206. Calculate the cognitive uncertainty and stochastic uncertainty of the forest stand risk index, landscape-level flammability risk probability, and traditional empirical flammability risk index, respectively.

[0024] Adding the cognitive uncertainty to the corresponding random uncertainty yields the total uncertainty of the forest stand risk index, the landscape-level flammability risk probability, and the traditional experience flammability risk index.

[0025] The fusion weights are dynamically allocated based on the total uncertainty, with higher fusion weights awarded to evaluation results that have lower total uncertainty.

[0026] The forest stand risk index, landscape-level flammability risk probability, and traditional experience-based flammability risk index are weighted and summed according to the fusion weights to generate the comprehensive flammability risk index.

[0027] Optionally, in some embodiments of this application, the single-tree physical mechanism model is a mechanism model used to simulate the relationship between the physiological and ecological processes of a single tree and the evolution of its flammability, including the following sub-models:

[0028] A transpiration mechanism model based on the Penman-Monteith equation was used to simulate the water dissipation process of trees.

[0029] A photosynthesis mechanism model based on the Farquhar model was constructed to simulate the process of carbon assimilation and biomass accumulation in trees.

[0030] A physical model of tree heat conduction based on the finite element heat conduction equation is used to simulate the temperature field distribution process inside the tree.

[0031] A fuel accumulation mechanism model based on litter generation and decomposition kinetics is used to simulate the accumulation process of surface combustible load.

[0032] Optionally, in some embodiments of this application, if the tree species type does not correspond to pre-trained dedicated mechanistic model parameters in S202, then a cross-species transfer model is used to generate dedicated mechanistic model parameters suitable for the tree species based on calibration sample data of that tree species, including:

[0033] A model-agnostic meta-learning framework is adopted, with the pre-trained parameters in the prior knowledge base of the source species as the initial state of the physical mechanism model of the single tree, and an inner loop gradient update is performed on 10-50 calibration sample data of the tree species type.

[0034] The prior knowledge base of the source species is pre-constructed based on the historical observation data of the source species and stores the basic pre-training parameters of the physical mechanism model of a single tree; the source species refers to the tree species with historical observation data used for pre-training the physical mechanism model of a single tree.

[0035] Meanwhile, the calibration sample data is projected into the feature space through the prototype network, and similarity is measured with the class prototype vector generated by the prototype network based on the source species prior knowledge base, to assist in parameter adaptation.

[0036] Generate specific mechanistic model parameters adapted to the tree species type, and store the specific mechanistic model parameters in the tree species mechanistic model parameter library.

[0037] Optionally, in some embodiments of this application, the calibration sample data is augmented using a generative adversarial network to generate diverse synthetic training samples to assist in the inner loop gradient update.

[0038] Optionally, in some embodiments of this application, the stand risk index of the assessment unit is calculated using a spatial weighted aggregation algorithm in step S203, including:

[0039] Calculate the stand risk index using the following formula:

[0040] ;

[0041] Among them, w ij The spatial weight factor between the i-th tree and the j-th tree is calculated using the following formula:

[0042] ;

[0043] d ij Let be the horizontal distance between the i-th tree and the j-th tree;

[0044] This is a spatially relevant length parameter that is adaptively adjusted based on stand density.

[0045] s i Let s be the canopy area of ​​the i-th tree. j Let be the area of ​​the canopy of the j-th tree;

[0046] R i Let be the probability of flammability of the i-th tree.

[0047] n represents the total number of trees within the forest stand area.

[0048] Optionally, in some embodiments of this application, the multidimensional feature vector input to the landscape-level risk prediction network in step S204 includes:

[0049] The stand feature vector is composed of the stand risk index and stand structure parameters, wherein the stand structure parameters include average tree height, average diameter at breast height, stand density, and canopy closure extracted from multidimensional observable data.

[0050] A topographic feature vector consisting of slope, aspect, elevation, and topographic relief;

[0051] A meteorological characteristic vector consisting of temperature, humidity, wind speed, precipitation, and evaporation;

[0052] The landscape-level risk prediction network is a multilayer perceptron or convolutional neural network, and its output is the landscape-level flammability risk probability in the range of 0 to 1.

[0053] Optionally, in some embodiments of this application, the cognitive uncertainty in S206 is calculated using Monte Carlo dropout or deep ensemble methods; the random uncertainty is determined based on the observation noise statistical characteristics of the multidimensional observable data.

[0054] Optionally, in some embodiments of this application, S300 further includes:

[0055] Based on the risk level of each assessment unit in the aforementioned spatiotemporal distribution map of flammability risk, the warning level is divided according to the corresponding level of cognitive uncertainty.

[0056] Set a baseline warning trigger threshold and an uncertainty determination threshold; dynamically adjust the baseline warning trigger threshold according to seasonal changes or extreme weather events;

[0057] When the cognitive uncertainty of a certain assessment unit exceeds the uncertainty judgment threshold, the warning trigger threshold of that assessment unit is increased according to the preset rules to obtain the adjusted warning trigger threshold.

[0058] When the comprehensive flammability risk index of the assessment unit exceeds the adjusted warning trigger threshold, a warning of the corresponding level is triggered.

[0059] Optionally, in some embodiments of this application, the specified future time period in S200 is the next 7-30 days; the hybrid prediction model is based on historical multidimensional observable data sequences, performs rolling prediction through a time series prediction network, and outputs the comprehensive flammability risk index of each evaluation unit for each day in the specified future time period.

[0060] (III) Beneficial Effects

[0061] This application presents a fire early warning method based on multidimensional forest health status analysis. By introducing a single-tree physical mechanism model, it achieves a comprehensive assessment of tree flammability based on water deficit, physiological activity, heat conduction, and litter accumulation, significantly improving the interpretability of the prediction results. Secondly, by using a spatial weighted aggregation algorithm to couple the flammability probability of a single tree with the distance between trees and the canopy area, it realizes multi-scale risk aggregation from single trees to forest stands to landscapes. Based on historical data sequences, it performs rolling predictions and outputs the spatiotemporal distribution of flammability risk for each assessment unit over the next 7-30 days, significantly extending the early warning lead time from several hours in traditional methods to several weeks, allowing sufficient preparation time for preventive measures such as understory flammability cleanup and fire prevention resource allocation. Furthermore, by quantifying the uncertainty of the forest stand risk index and adaptively weighting and fusing the landscape-level flammability risk probability, a comprehensive flammability risk index is generated, effectively suppressing the uncertainty of a single model and improving the reliability and stability of the early warning results. Attached Figure Description

[0062] Figure 1 The flowchart below shows a fire early warning method based on multidimensional forest health status analysis according to this application.

[0063] Figure 2 This is a process logic diagram of a pre-built hybrid prediction model according to an embodiment of this application. Detailed Implementation

[0064] In existing technologies, forest fire early warning methods can be mainly categorized into the following two types:

[0065] The first category is early warning methods based on traditional monitoring techniques, such as satellite thermal infrared monitoring, video smoke recognition, and manual patrols. These methods are essentially reactive, triggering alarms only when a fire has developed into open flames or smoke. This results in significant warning delays and problems with large blind spots and limited coverage. Satellite monitoring is limited by transit cycles and cloud cover, video surveillance is susceptible to nighttime and rainy / foggy weather, and manual patrols cannot provide 24 / 7 monitoring, failing to meet the needs of proactive fire prevention.

[0066] The second category is fire risk prediction methods based on data-driven AI models, such as random forests, support vector machines, and deep learning models. These methods attempt to predict fire risk by analyzing factors such as meteorology, vegetation, and topography. However, their training heavily relies on a large amount of historical fire data, while fire itself is a low-probability event, and data scarcity leads to weak model generalization ability. At the same time, existing models treat vegetation as a homogeneous layer, failing to distinguish the physiological characteristics and flammability differences of different tree species, resulting in a significant drop in prediction accuracy in mixed forest areas. Furthermore, purely data-driven models lack a characterization of the physical mechanisms of fire occurrence, resulting in poor interpretability of prediction results and difficulty in providing a scientific basis for forest fire prevention.

[0067] Therefore, this application provides a fire early warning method based on multidimensional forest health status analysis. It aims to construct a hybrid prediction model that integrates single tree physical mechanism model, cross-species transfer learning, multi-scale spatial aggregation and uncertainty weighted fusion. This model can achieve multi-scale progressive risk prediction from single tree to forest stand to landscape. It can quickly adapt to new tree species with only a small number of samples and output the spatiotemporal distribution of flammability risk for the next 7-30 days. It effectively overcomes the problems of existing technologies such as delayed early warning, scarce data, poor species adaptability and lack of interpretability of prediction results.

[0068] To better understand the above technical solutions, exemplary embodiments of this application will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application can be understood more clearly and thoroughly, and that the scope of this application can be fully conveyed to those skilled in the art.

[0069] Example 1

[0070] Figure 1 The flowchart below shows a fire early warning method based on multidimensional forest health status analysis according to this application. Figure 1 As shown, the fire early warning method based on multidimensional forest health status analysis includes:

[0071] S100. Acquire multispectral remote sensing images corresponding to each evaluation unit of the target forest area, and preprocess the multispectral remote sensing images; the evaluation unit is a grid of multiple predetermined spatial scales divided from the target forest area.

[0072] Vegetation index data, soil index data, topographic data and biological status data of each assessment unit are extracted from the preprocessed multispectral remote sensing images, and meteorological data of the corresponding locations are fused to obtain multidimensional observable data of each assessment unit.

[0073] S200. Input the multidimensional observable data of each assessment unit into the pre-built hybrid prediction model, and output the comprehensive flammability risk index of each assessment unit within a specified future time period; and generate a visualized spatiotemporal distribution map of flammability risk on a GIS map based on the comprehensive flammability risk index.

[0074] S300. Generate and release corresponding forest fire prevention early warning information based on the spatiotemporal distribution map of flammability risk.

[0075] This embodiment achieves comprehensive coverage of the entire area and independent risk assessment for each grid by dividing the target forest area into grids of predetermined spatial scales as assessment units, thus overcoming the limitations of traditional manual patrols in terms of coverage. Through the technical concept of predicting forest fires before they occur, it outputs a flammability risk index for a specified future time period, shifting from traditional post-fire detection to pre-fire prediction. This provides ample preparation time for fire prevention decisions and preventative measures (such as clearing flammable materials from under the forest canopy and allocating resources). By generating a visualized spatiotemporal distribution map of the comprehensive flammability risk index on a GIS map, and issuing early warning information based on this map, fire prevention personnel can intuitively and quickly grasp high-risk areas, improving the practicality and operability of early warning information.

[0076] Example 2

[0077] Figure 1 The flowchart below shows a fire early warning method based on multidimensional forest health status analysis according to this application. Figure 1 As shown, the method includes:

[0078] S100. Acquire multispectral remote sensing images corresponding to each evaluation unit of the target forest area, and perform phased preprocessing on the multispectral remote sensing images; the evaluation unit is a grid of multiple predetermined spatial scales divided from the target forest area.

[0079] Vegetation index data, soil index data, topographic data, and biological status data of each assessment unit are extracted from the preprocessed multispectral remote sensing images, and meteorological data of the corresponding locations are fused together to obtain multidimensional observable data for each assessment unit.

[0080] Specifically, firstly, multispectral remote sensing images of each assessment unit in the target forest area are acquired. These images can be collected in real time by satellite remote sensing platforms (such as Sentinel-2, Landsat, etc.). Simultaneously, meteorological data such as temperature, humidity, and wind speed are collected in real time through a network of ground meteorological monitoring stations deployed within the target forest area, and soil moisture and temperature changes are monitored in real time through a soil sensor network.

[0081] The assessment unit is a grid of multiple predetermined spatial scales divided from the target forest area, preferably a 1km×1km grid, to achieve full coverage monitoring of the target forest area without blind spots.

[0082] The preprocessing of the acquired multispectral remote sensing images includes two stages:

[0083] The first stage is the professional preprocessing of remote sensing images, which involves performing standard remote sensing image processing procedures such as radiometric calibration, atmospheric correction, geometric correction, cloud cover detection, band registration, and topographic correction. This process eliminates the effects of systematic errors, atmospheric scattering / absorption interference, topographic distortion, and cloud cover, restoring the true reflective characteristics of ground objects and providing a high-precision image foundation for subsequent data extraction.

[0084] The second stage is multidimensional data cleaning and standardization. This involves cleaning, normalizing, standardizing, handling missing values, and detecting and removing outliers from the vegetation index data, soil index data, topographic data, and biological status data extracted from the corrected remote sensing images, as well as the fused meteorological data. This ensures that the multidimensional observable data input into the model is noise-free, unbiased, and dimensionally consistent.

[0085] The following multidimensional observable data were extracted from each evaluation unit from the preprocessed remote sensing image:

[0086] Vegetation index data, including Normalized Difference Vegetation Index (NDVI), Soil-Adjusted Vegetation Index (SAVI), Enhanced Vegetation Index (EVI), Leaf Area Index (LAI), Normalized Burn Index (NBR), etc., are extracted through corresponding band operations of multispectral remote sensing images and are used to characterize vegetation coverage, growth status, and vitality level.

[0087] Soil index data, including soil moisture, pH value, organic matter content, and soil temperature, are obtained by fusing remote sensing inversion with monitoring data from a ground-based soil sensor network.

[0088] Topographic data, including slope, aspect, elevation, and topographic relief, is extracted from the digital elevation model (DEM) and used to characterize the impact of topography on fire spread.

[0089] Biological status data: including tree species distribution, tree age structure, pest and disease index, canopy area, diameter at breast height, tree height, and branch and leaf density, are obtained by integrating high-resolution remote sensing image interpretation, forest resource survey data and sample plot survey data;

[0090] Simultaneously, by integrating meteorological data such as temperature, humidity, wind speed, precipitation, evaporation, and solar radiation collected by a network of ground meteorological stations, as well as soil moisture and temperature data monitored by a soil sensor network, multidimensional observable data for each assessment unit is obtained. This multidimensional observable data combines information from multiple sources, including satellite remote sensing, ground meteorological monitoring, and soil sensor networks, achieving organic fusion of multi-source data and providing comprehensive and accurate input for subsequent risk assessment.

[0091] S200. Input the multidimensional observable data of each assessment unit into the pre-constructed hybrid prediction model, and output the comprehensive flammability risk index of each assessment unit within a specified future time period; and generate a visualized spatiotemporal distribution map of flammability risk on a GIS map based on the comprehensive flammability risk index.

[0092] Figure 2 The processing logic diagram for the pre-built hybrid prediction model is as follows: Figure 2 As shown, the processing steps of this hybrid prediction model include:

[0093] S201. Identify the tree species type of each tree in the current evaluation unit from the biological state data in the multidimensional observable data;

[0094] Specifically, by combining biological state data from multidimensional observable data with tree species feature recognition algorithms and forest resource survey data, the tree species type of each tree in the current assessment unit can be identified. Common tree species such as pine, oak, birch, and fir, as well as rare and unique tree species, can be accurately identified, laying the foundation for loading parameters for subsequent tree species-specific mechanism models.

[0095] S202. By querying the pre-constructed tree species mechanism model parameter library, determine whether each tree species type corresponds to a pre-trained dedicated mechanism model parameter; the tree species mechanism model parameter library is pre-constructed based on historical observation data and stores dedicated mechanism model parameters for multiple tree species, including tree species-specific parameters such as maximum stomatal conductance, litter production rate, decomposition rate, and moisture regulation coefficient, as well as environmental response parameters such as temperature sensitivity and humidity response coefficient. The parameter range is determined based on a large amount of field measurement data.

[0096] If the tree species type corresponds to a pre-trained dedicated mechanism model parameter, then the dedicated mechanism model parameter is loaded into the single tree physical mechanism model;

[0097] Optionally, the physical mechanism model for a single tree includes specific parameters for different tree species, as shown in Table 1:

[0098]

[0099] Table 1

[0100] As shown in Table 1, the parameter ranges for different tree species in the single-tree physical mechanism model reflect the species specificity of the model. By setting differentiated parameters such as maximum stomatal conductance and litter production rate, the model can accurately characterize the physiological characteristics and flammability differences of different tree species such as pine (highly flammable) and birch (lowly flammable), providing a reasonable parameter initialization benchmark for subsequent cross-species transfer learning and effectively improving the prediction accuracy of mixed forest areas.

[0101] If the tree species type does not have a pre-trained dedicated mechanism model parameter (i.e., a new tree species), then a cross-species migration model is used to generate dedicated mechanism model parameters applicable to the tree species based on 10-50 calibration sample data of the tree species. The generated dedicated mechanism model parameters are then loaded into the physical mechanism model of a single tree and stored in the tree species mechanism model parameter library to realize the dynamic updating and expansion of the parameter library.

[0102] Specifically, the calibration sample data serves as input data for adapting cross-species migration models. It includes measured data on tree physiology and ecology obtained through field surveys, as well as empirical anchor data derived from forestry experts' knowledge or long-term observation experience, which are used to guide the rapid adaptation of model parameters to the target tree species.

[0103] For ease of understanding, if there are pre-trained dedicated mechanism model parameters corresponding to a tree species type, then the tree species type is defined as a known tree species, and the dedicated mechanism model parameters are directly loaded into the physical mechanism model of a single tree.

[0104] If no pre-trained dedicated mechanistic model parameters correspond to a particular tree species type, then the tree species type is defined as a new tree species, and the cross-species transfer learning process is initiated. Using the cross-species transfer model, based on calibration sample data (10-50 samples) for that tree species, dedicated mechanistic model parameters suitable for that tree species are generated. After loading the parameters, the meteorological data (temperature, humidity, wind speed, etc.) and soil data corresponding to the current individual tree are input into the individual tree physical mechanism model to calculate the individual tree flammability probability. The individual tree flammability probability is a value in the range of 0-1, with a higher value indicating a stronger flammability of the individual tree.

[0105] Repeat the above process for all trees in the current assessment unit, and summarize the individual tree flammability probability of all trees in the current assessment unit to obtain the set of individual tree flammability probabilities for each tree in the current assessment unit.

[0106] Among them, the single-tree physical mechanism model is a mechanism model used to simulate the relationship between the physiological and ecological processes of a single tree and its flammability evolution. It is constructed based on the basic principles of plant physiology and physics and includes four core sub-models. These sub-models are coupled with each other to comprehensively depict the evolutionary process of a tree from its physiological state to its flammability.

[0107] A transpiration mechanism model based on the Penman-Monteith equation is used to simulate the water dissipation process of trees, calculate the water stress index of trees, and characterize the water deficit state of trees; its core calculation formula is:

[0108] Evaporation rate = 1.6 × stomatal conductance × vapor pressure difference / (101.3 × (1 + 0.66 × wind speed));

[0109] Wherein, vapor pressure difference = saturated water vapor pressure - actual water vapor pressure;

[0110] Saturated water vapor pressure = 0.6108 × exp(17.27 × temperature / (temperature + 237.3));

[0111] Actual water vapor pressure = saturated water vapor pressure × humidity / 100;

[0112] Stomatal conductance = maximum stomatal conductance × (1 - exp(-0.01 × solar radiation)) × exp(-0.1 × vapor pressure difference);

[0113] A photosynthetic mechanism model based on the Farquhar model was constructed to simulate the carbon assimilation and biomass accumulation process in trees, assess the plant vigor index of trees, and characterize the physiological activity of trees.

[0114] A tree heat conduction physical model based on the finite element heat conduction equation is used to simulate the temperature field distribution process inside the tree, calculate the heat stress index and heat accumulation index of the tree, and characterize the thermal state of the tree.

[0115] A fuel accumulation mechanism model based on litter generation and decomposition kinetics is used to simulate the accumulation process of surface combustible load and characterize the accumulation state of flammable materials around trees; its core calculation formula is:

[0116] Fuel accumulation rate = Litter generation rate - Decomposition rate × Fuel quantity × (1 - Moisture content / 100).

[0117] This provides a solid biological and physical basis for assessing the flammability of individual trees, significantly improving the interpretability of the prediction results.

[0118] For new tree species not yet included in the database, a cross-species transfer learning method combining a model-agnostic meta-learning (MAML) framework and a prototype network is employed. This method can quickly generate dedicated mechanistic model parameters with only 10-50 calibration samples, specifically including:

[0119] A model-independent meta-learning (MAML) framework is adopted, using pre-trained parameters from the source species prior knowledge base as the initial state of the single-tree physical mechanism model. The source species prior knowledge base is pre-constructed based on historical observation data of source species (such as European red pine, British oak, European beech, spruce, etc., which have rich historical observation data), and stores the basic pre-trained parameters and class prototype vectors of the single-tree physical mechanism model.

[0120] Perform an inner loop gradient update on 10-50 calibration sample data of this tree species to quickly adapt to the physiological characteristics of the target tree species.

[0121] Simultaneously, after calculating the similarity vectors between the target tree species and each source species through the prototype network, the similarity vectors are smoothed and expanded by combining them with a pre-constructed similarity matrix in the prior knowledge base of the source species (which records the functional trait similarity between the source species). For example, if the target tree species has a high similarity to pine trees, and the similarity matrix shows that pine trees are highly similar to fir trees, the association weight between the target tree species and fir trees can be appropriately increased. Thus, in the subsequent inner loop gradient update, not only are the parameters of the most similar source species used as the benchmark, but the parameter adjustment experience of multiple related source species is also comprehensively referenced, making the parameters of the generated dedicated mechanism model more robust and generalizable.

[0122] To more accurately assist in parameter adaptation, this embodiment pre-calculates the functional trait similarity among source species, as shown in Table 2:

[0123]

[0124] Table 2

[0125] Based on the above traits, class prototype vectors for each source species are generated using a prototype network, and a similarity matrix is ​​constructed, as shown in Table 3:

[0126]

[0127] Table 3

[0128] After generating specialized mechanistic model parameters adapted to the tree species type, these parameters are stored in a tree species mechanistic model parameter library for future use. This effectively solves the technical challenges of scarce historical fire data and difficulty in quantifying differences in flammability among different tree species, significantly reducing the data costs of deploying and applying the technology in new areas and with new tree species.

[0129] Optionally, for scenarios with limited calibration sample data, data augmentation can be performed using generative adversarial networks (GANs) to generate diverse synthetic training samples. These synthetic samples are based on the physiological and physical constraints of trees to ensure their rationality and are used to assist in the inner loop gradient update. This addresses the problem of insufficient adaptation accuracy caused by a small number of calibration samples and improves the stability and accuracy of transfer learning.

[0130] To verify the effectiveness of cross-species transfer learning, this embodiment compared the prediction accuracy under different sample sizes, and the results are shown in Table 4:

[0131]

[0132] Table 4

[0133] The table visually demonstrates the significant technical advantages of this invention under conditions of limited samples. When the sample size is extremely scarce, this application improves the accuracy of predicting the flammability probability of a single tree through transfer learning, effectively solving the core pain point of insufficient data for new tree species and new forest areas.

[0134] S203. Based on the set of individual tree flammability probabilities, and combined with the spatial coordinates and canopy area of ​​each tree extracted from multidimensional observable data, a spatial weighted aggregation algorithm is used to calculate the stand risk index of the assessment unit. The stand risk index is a value in the range of 0-1, representing the overall flammability risk level of the stand within the assessment unit, achieving risk aggregation from the individual tree level to the forest level.

[0135] Specifically, the formula for calculating the forest stand risk index is:

[0136] ;

[0137] Among them, w ij The spatial weight factor between the i-th tree and the j-th tree is calculated using the following formula:

[0138] ;

[0139] d ij Let be the horizontal distance between the i-th tree and the j-th tree; The spatially relevant length parameter is adaptively adjusted based on stand density; s i Let s be the canopy area of ​​the i-th tree. j Let R be the canopy area of ​​the j-th tree; i Let be the probability of flammability of the i-th tree; n is the total number of trees in the forest stand area.

[0140] This formula uses an exponential decay function to characterize the law that the mutual influence between trees decreases with increasing distance. At the same time, it uses the product of canopy areas to reflect the greater contribution of large trees to the overall risk of the forest stand. Finally, it obtains a forest stand risk index that reflects the overall flammability level of the assessment unit, realizing accurate risk aggregation from the micro-level of individual trees to the macro-level of forest stands, and providing key technical support for refined risk assessment.

[0141] S204. The forest stand risk index is used as a vegetation feature and is combined with the topographic data and meteorological data in the multidimensional observable data to form a multidimensional feature vector, which is then input into the landscape-level risk prediction network to output the landscape-level flammability risk probability.

[0142] Specifically, the multidimensional feature vector includes: a stand feature vector composed of the stand risk index and stand structure parameters, wherein the stand structure parameters include average tree height, average diameter at breast height, stand density, and canopy closure extracted from multidimensional observable data;

[0143] A topographic feature vector consisting of slope, aspect, elevation, and topographic relief;

[0144] A meteorological characteristic vector consisting of temperature, humidity, wind speed, precipitation, and evaporation.

[0145] The landscape-level risk prediction network is a multilayer perceptron (MLP) or convolutional neural network (CNN), which outputs a landscape-level flammability risk probability in the range of 0 to 1, representing the fire risk level on a larger spatial scale. It realizes progressive risk prediction from the stand scale to the landscape scale, and significantly improves the accuracy of risk assessment on a larger spatial scale.

[0146] The preferred choice is a multilayer perceptron, whose network architecture is as follows:

[0147] Input layer: Fusion of forest stand feature vector (64-dimensional), topographic feature vector (32-dimensional), and meteorological feature vector (48-dimensional), with a total input dimension of 144 dimensions;

[0148] Hidden layer 1: 128 neurons are set, ReLU activation function is used, and batch normalization (BN) and Dropout (0.2) are added to prevent overfitting;

[0149] Hidden layer 2: 64 neurons are used, employing the ReLU activation function, with batch normalization (BN) and Dropout (0.2; );

[0150] Output layer: Set up 1 neuron, use the Sigmoid activation function, and output the landscape-level flammability risk probability in the range of 0-1;

[0151] Loss function: Mean Squared Error (MSE);

[0152] Optimizer: Adam, learning rate 0.0001.

[0153] S205. A forest health index is calculated based on the multidimensional observable data, and then converted into a traditional empirical flammability risk index through a preset mapping relationship. The traditional empirical flammability risk index is a value within the range of 0-1, constructed based on traditional forestry assessment methods, and forms a three-source complement with the stand risk index (mechanism path) and the landscape-level flammability risk probability (data path). Specifically, the forest health calculation formula is as follows:

[0154] Health index = 0.4 × vegetation health + 0.3 × soil health + 0.3 × weather suitability;

[0155] This formula is used to integrate multi-source observation data into a comprehensive forest health index, with a value range of 0-1. The higher the value, the better the forest health and the lower the fire risk.

[0156] Wherein: Vegetation health = 0.3 × NDVI + 0.2 × SAVI + 0.2 × EVI + 0.15 × LAI + 0.15 × NBR;

[0157] NDVI is the Normalized Difference Vegetation Index, SAVI is the Soil-Regulated Vegetation Index, EVI is the Enhanced Vegetation Index, LAI is the Leaf Area Index, and NBR is the Normalized Burning Index.

[0158] Soil health = 0.4 × soil moisture + 0.3 × soil pH + 0.3 × organic matter content;

[0159] Meteorological suitability = 0.25 × Temperature suitability + 0.25 × Humidity suitability + 0.2 × Wind speed suitability + 0.15 × Precipitation suitability + 0.15 × Evaporation suitability.

[0160] The values ​​for each of the above sub-indicators range from 0 to 1, and can be obtained in the following ways:

[0161] Remote sensing inversion: directly calculated from multispectral remote sensing images (such as NDVI, SAVI, etc.);

[0162] Empirical formulas: These are derived from meteorological observation data through a normalization function (e.g., temperature suitability).

[0163] Ground monitoring: Data is acquired through sensor networks (such as soil moisture and pH value).

[0164] Expert knowledge base: A mapping table built based on historical statistics or expert experience.

[0165] Then, the forest health index is converted into a traditional experience-based flammability risk index through a preset mapping relationship. The mapping relationship can be implemented using one of three methods: forestry expert experience lookup table method, neural network mapping method, or linear mapping method.

[0166] The traditional experience-based flammability risk index, as a third type of risk assessment result, is used to integrate with the forest stand risk index (mechanistic path) and the landscape-level flammability risk probability (data path) to generate a comprehensive flammability risk index.

[0167] This step, based on the forest health index calculation formula, integrates multi-source observation data into a comprehensive forest health index, fully inheriting the long-accumulated experience and knowledge in the forestry field. Through a pre-defined mapping relationship, the forest health index is transformed into a traditional empirical flammability risk index, serving as a third-category risk assessment result. This complements the mechanistic and data pathways, ensuring that this application retains the reliability of traditional forestry experience while incorporating the efficiency of modern AI technology.

[0168] S206. Calculate the cognitive uncertainty and stochastic uncertainty of the forest stand risk index, landscape-level flammability risk probability, and traditional empirical flammability risk index, respectively.

[0169] Specifically, cognitive uncertainty is calculated using Monte Carlo dropout or deep ensemble methods, reflecting the model's grasp of the prediction results. For the forest stand risk index (derived from single-tree model aggregation), Monte Carlo dropout can be used to randomly drop neurons multiple times during model inference, and the variance of the prediction results is calculated as cognitive uncertainty. For the landscape-level flammability risk probability (derived from neural networks), deep ensemble methods can be used to train multiple independent networks, and the consistency of the outputs of each network is calculated as cognitive uncertainty.

[0170] The random uncertainty reflects the noise level of the input data itself. It is determined based on the statistical characteristics of the measurement error, spatiotemporal sampling error, and other characteristics of each observed variable in the multidimensional observable data, for example, by taking the mean of the normalized measurement error of each variable.

[0171] The cognitive uncertainty is added to the corresponding random uncertainty to obtain the total uncertainty of each risk index; the fusion weight is dynamically allocated according to the total uncertainty, and the lower the total uncertainty, the higher the fusion weight is obtained.

[0172] Specifically, the uncertainty levels are classified according to the criteria in Table 5:

[0173]

[0174] Table 5

[0175] Based on the above uncertainty level classification results, the corresponding fusion strategy is executed to generate a comprehensive flammability risk index:

[0176] For low-uncertainty evaluation units, the fusion weight is calculated according to the standard inverse variance weighting formula;

[0177] For assessment units with moderate uncertainty, a penalty coefficient of 0.9 is applied to models with higher uncertainty based on standard inverse variance weighting to further reduce their weight contribution.

[0178] For the high uncertainty assessment unit, a conservative weighting strategy is adopted: if the historical performance of the stand risk index (mechanism path) is better than the other two types of models, its weight is set to 0.6, and the remaining weights of the other two types of models are distributed according to the inverse variance weighting, with the remaining weights being 0.4.

[0179] For the assessment unit with extremely high uncertainty, the risk assessment result with the lowest total uncertainty is selected as the comprehensive flammability risk index, and the outputs of the other two models are discarded.

[0180] The formula for weighting the inverse standard variance is as follows:

[0181] ;

[0182] Where k represents one of three risk assessment results: forest stand risk index, landscape-level flammability risk probability, and traditional experience-based flammability risk index. The fusion weight of the risk assessment results for the k-th type is... For the total uncertainty of the risk assessment result of type k, U stand Corresponding to the total uncertainty of the stand risk index, U landscape Corresponding to the total uncertainty of the stand risk index, U expert This corresponds to the total uncertainty of the traditional empirical flammability risk index.

[0183] This formula ensures that the sum of all weights is 1, and that the weight of each risk index is inversely proportional to its total uncertainty.

[0184] Finally, the calculated fusion weights are used to weight and sum the results of the three risk assessments to generate the comprehensive flammability risk index for this assessment unit:

[0185]

[0186] in, To comprehensively assess the flammability risk index, R is the forest stand risk index. landscape For landscape-level flammability risk probability, R expert The traditional experience-based flammability risk index.

[0187] By dynamically adjusting the uncertainty level classification and fusion strategy, the system can adaptively adjust the fusion method according to the prediction confidence of different evaluation units. In the low uncertainty region, it can give full play to the advantages of multi-source fusion, and in the high uncertainty region, it can adopt a conservative strategy to avoid misjudgment, which significantly improves the reliability and robustness of the evaluation results.

[0188] Furthermore, steps S201 to S206 above are repeated for all assessment units within the target forest area to obtain the comprehensive flammability risk index of each assessment unit at the current moment.

[0189] The hybrid prediction model is based on historical multidimensional observable data sequences (preferably time series data from the past 24 hours). It performs rolling predictions through an LSTM time series prediction network and a multi-head attention mechanism. Each time the multidimensional observable data sequence from the past 24 hours is input, the comprehensive flammability risk index for the next 24 hours is output. The model iteratively generates prediction sequences for the next 7-30 days, thus achieving long-term risk evolution forecasting.

[0190] The long-term rolling prediction in the hybrid prediction model is implemented based on an LSTM time-series prediction network and a multi-head attention mechanism, wherein:

[0191] LSTM layer: The input sequence length is 30 days (historical data), the hidden state dimension is 128, and it captures the temporal features and long-term dependencies of multidimensional observable data;

[0192] Multi-head attention mechanism: Eight attention heads are set up to give higher weights to key features in historical time series data (such as high temperature, drought, strong wind, etc.) to enhance the representation ability of important features;

[0193] Fully connected layer: Maps the output features of LSTM and attention mechanism to a comprehensive flammability risk index sequence for the next 7-30 days.

[0194] The risk index of all assessment units for each future day is spatially visualized on a GIS map, forming a spatiotemporal distribution map of flammability risk over a specified future time period. This distribution map is presented in the form of a heat map, with different colors representing different risk levels (0-0.3 for low risk, 0.3-0.6 for medium risk, and 0.6-1.0 for high risk). It also supports timeline scrolling, allowing users to view the spatial distribution and evolution trend of risk for each future day, intuitively displaying the spatiotemporal characteristics of high-risk areas.

[0195] This embodiment significantly extends the early warning lead time from several hours in traditional methods to 7-30 days in the future, allowing ample preparation time for preventative measures such as clearing combustibles under forest cover and optimizing the allocation of fire prevention resources.

[0196] S300. Generate and release corresponding forest fire prevention early warning information based on the spatiotemporal distribution map of flammability risk.

[0197] Meanwhile, in the early warning issuance phase, the early warning strategy is further dynamically adjusted based on the results of the cognitive uncertainty level classification:

[0198] Based on the risk level of each assessment unit in the aforementioned spatiotemporal distribution map of flammability risk, the warning level is divided according to the corresponding level of cognitive uncertainty.

[0199] Based on the comprehensive flammability risk index level of each assessment unit in the aforementioned flammability risk spatiotemporal distribution map, and combined with the corresponding cognitive uncertainty level, three warning levels are divided into red, orange, and yellow; among them, high risk (≥0.6) corresponds to red warning, medium risk (0.3-0.6) corresponds to orange warning, and low risk (0-0.3) corresponds to yellow warning.

[0200] Set a baseline warning trigger threshold and an uncertainty judgment threshold; dynamically adjust the baseline warning trigger threshold according to seasonal changes or extreme weather events (such as high temperature, drought, strong wind, continuous no precipitation, etc.), such as appropriately lowering the baseline warning trigger threshold during the drought season to improve warning sensitivity.

[0201] When the total uncertainty of an assessment unit exceeds the uncertainty threshold, the warning trigger threshold for that assessment unit is increased according to preset rules (e.g., the warning threshold is increased by 10% for every 0.1 increase in uncertainty), resulting in the adjusted warning trigger threshold. The higher the level of uncertainty, the greater the increase in the threshold.

[0202] Differentiated early warning trigger thresholds are used for assessment units with different levels of uncertainty:

[0203] Low uncertainty assessment unit: adopts standard early warning trigger threshold;

[0204] Medium uncertainty assessment unit: Early warning trigger threshold increased by 10%;

[0205] High uncertainty assessment unit: Early warning trigger threshold increased by 20%;

[0206] Extremely high uncertainty assessment unit: The warning trigger threshold is increased by 50%, or the warning is temporarily not issued and only a notification message is pushed.

[0207] When the comprehensive flammability risk index of the assessment unit exceeds the adjusted warning trigger threshold, a warning of the corresponding level (such as red warning, orange warning, or yellow warning) is triggered. The warning information includes core contents such as risk level, location of high-risk area, duration of risk, prevention and control suggestions, and level of uncertainty. It is pushed to relevant entities such as forest fire prevention management departments, forest rangers, and forest residents through various means such as SMS, forest fire command platform push, warning broadcast, and GIS system pop-up window.

[0208] Based on the spatiotemporal distribution map of flammability risks, it provides precise emergency decision support for forest fire prevention work, including key areas for clearing combustibles under the forest canopy, the direction for optimizing the allocation of fire prevention resources, and the key deployment points for patrol personnel, thereby improving the pertinence and efficiency of forest fire prevention work.

[0209] This dynamic threshold adjustment mechanism effectively avoids false alarms caused by insufficient model confidence, while making the early warning rules more scientific and flexible, adapting to fire prevention needs under different time and space conditions.

[0210] By incorporating uncertainty levels into the early warning issuance process, the early warning rules become more scientific and flexible, effectively avoiding false alarms caused by insufficient model confidence. At the same time, it ensures that early warnings can be issued in a timely and accurate manner in high-risk, high-certainty areas, significantly improving the practicality and reliability of the early warning system.

[0211] In summary, the fire early warning method based on multidimensional forest health status analysis proposed in this application realizes multi-scale progressive risk prediction from individual trees to forest stands and then to the landscape by constructing a hybrid prediction model that integrates a single tree physical mechanism model, cross-species transfer learning, spatial weighted aggregation, landscape-level risk prediction network, traditional experience path and uncertainty perception weighted fusion. This application requires only 10-50 samples of target tree species to quickly adapt to new species, effectively solving the industry challenges of scarce historical fire data and significant differences in the physiological characteristics of different tree species. Compared with traditional deep learning methods, the adaptation time for new tree species is shortened from several months to several days. It can accurately distinguish the differences in flammability among different tree species such as pine, oak, birch, and fir, significantly improving the prediction accuracy of mixed forests and rare tree species forests. By providing biological and physical basis for the prediction results through a physical mechanism model, the interpretability is significantly improved. Through weighted fusion of three sources of uncertainty, the complementary advantages of the mechanism path, data path, and experience path are fully utilized, greatly improving the reliability and robustness of the prediction results. Through gridded parallel computing and GIS visualization, the early warning lead time is extended from several hours to 7-30 days in the future. Through a dynamic threshold adjustment mechanism, the early warning rules are made more scientific and flexible, effectively avoiding false alarms. This method has a clear and standardized operation process, a high degree of standardization in data collection and processing, low model deployment and migration costs, and can be adapted to diverse forest ecosystems under different geographical and climatic conditions, providing practical technical support and scientific decision-making basis for forest fire prevention and management.

[0212] Example 3

[0213] This embodiment takes the 2,800-hectare mixed coniferous and broad-leaved forest area in Changbai Mountain as the target forest area. The terrain of this area is mainly low and medium mountains, with an altitude range of 500-1,800m. The climate is a temperate continental monsoon climate, with warm and rainy summers and dry and windy autumns, which is the peak season for forest fires.

[0214] The region has a complex tree species structure, mainly including common tree species such as oak and birch, as well as rare and unrecorded tree species such as European fir (without pre-trained mechanism model parameters); the stand density is uneven and the mixed ratio varies greatly. Traditional models cannot distinguish the differences in the flammability of tree species and lack the ability to adapt to new tree species, resulting in a prediction accuracy of less than 60%.

[0215] First, pre-configure the hardware and data sources:

[0216] Data acquisition equipment: Landsat-9 satellite multispectral remote sensing imagery (spatial resolution 30m), 8 automatic weather monitoring stations in the region (monitoring indicators: temperature, humidity, wind speed, precipitation, evaporation, temporal resolution 1 hour), 15 soil sensors (monitoring indicators: soil moisture, temperature, pH value, deployment depth 10cm), and UAVs (used for field measurement data collection in new tree species sample plots).

[0217] Basic data: forest resource survey data (tree species distribution, tree age structure, diameter at breast height / tree height statistics), digital elevation model (DEM, used to extract topographic data), forestry expert experience database (including growth characteristics of tree species such as fir, and criteria for judging flammability period);

[0218] Computing equipment: A cloud server cluster (8-core CPU, 32GB memory, GPU accelerator card) deployed in the forest fire prevention command center, storing tree species mechanism model parameter library and source species prior knowledge base, and running hybrid prediction model.

[0219] The target forest area was divided into 1km×1km grid assessment units, totaling 28 assessment units. Each unit independently performed data processing, model inference, and risk prediction to ensure the precision of risk assessment.

[0220] S100. Acquire multispectral remote sensing images corresponding to each evaluation unit of the target forest area, and preprocess the multispectral remote sensing images; the evaluation unit is a grid of multiple predetermined spatial scales divided from the target forest area.

[0221] Specifically, the study acquired nearly 30 days of Landsat-9 satellite multispectral remote sensing imagery of the target area, covering seven bands including blue, green, red, and near-infrared; collected nearly 30 days of real-time and historical meteorological data from eight meteorological monitoring stations, including daily average temperature (5-22℃), relative humidity (45%-80%), and wind speed (1-6m / s); collected real-time soil sensor monitoring data (soil moisture 15%-35%, pH 5.5-7.2) and laboratory-tested organic matter content data (1.5%-3.8%); acquired high-resolution imagery through low-altitude drone photography (flight altitude 100m), and combined with forest resource survey data, identified the tree species distribution in each assessment unit (oak 42%, birch 38%, European fir 12%, other tree species 8%); and collected calibration sample data (age, diameter at breast height, crown width, leaf water content, etc.) of European fir through field measurements (setting up 20 20m×20m quadrats).

[0222] Then, the above multispectral remote sensing images are preprocessed;

[0223] The first stage (professional preprocessing of remote sensing images) involves performing radiometric calibration (eliminating sensor errors), atmospheric correction (removing atmospheric scattering / absorption interference), geometric correction (error ≤ 1 pixel), cloud cover detection (removing image areas with cloud coverage exceeding 10% and supplementing with interpolation from adjacent temporal images), and terrain correction (eliminating brightness distortion caused by terrain undulations based on DEM) on Landsat-9 images to ensure the accuracy of image data.

[0224] Phase Two (Multidimensional Data Cleaning and Standardization):

[0225] Data extraction: Extract vegetation indices (NDVI, SAVI, EVI) and topographic data (slope, aspect, elevation, topographic relief) from preprocessed images; interpret the spatial coordinates of trees and canopy area from UAV images;

[0226] Data cleaning: using 3 Outliers in meteorological and soil data (such as invalid data caused by extreme high temperatures or sensor malfunctions) are removed in principle.

[0227] Standardization processing: Perform normalization processing on all multidimensional data (mapped to the 0-1 interval) to ensure uniformity of units;

[0228] Missing value imputation: Kriging interpolation was used to impute a small number of missing soil moisture and meteorological data to ensure data integrity.

[0229] The final result is multidimensional observable data for each evaluation unit, covering 25 feature dimensions across 5 major categories: vegetation, soil, topography, meteorology, and biological status.

[0230] S200. Input the multidimensional observable data of each assessment unit into the pre-built hybrid prediction model, and output the comprehensive flammability risk index of each assessment unit within a specified future time period; and generate a visualized spatiotemporal distribution map of flammability risk on a GIS map based on the comprehensive flammability risk index.

[0231] (1) Construct a prior knowledge base for source species: Select common tree species with abundant data, such as European red pine, British oak, and spruce, as source species, and construct a knowledge base based on their historical observation data (meteorological, soil, and tree structure data of the past 5 years);

[0232] (2) Acquisition of fir calibration samples: 50 calibration samples containing measured data and expert experience anchor data were obtained and standardized for adaptation;

[0233] Measured data: Tree age (8-35 years), diameter at breast height (12-45cm), crown width (2.5-8m), leaf water content (45%-65%), and bark thickness (0.8-2.2cm) measured in the sample plots.

[0234] Expert experience anchor data: Qualitative data provided by forestry experts, such as "fir trees enter the flammable period after 15 years of age" and "flammability increases significantly when leaf moisture content is below 50%", are quantified into numerical characteristics (e.g., flammable period is marked as 1, non-flammable period is marked as 0).

[0235] (3) Since European fir is not included in the tree species mechanism model parameter library, the cross-species transfer learning process is initiated: the spruce pre-training parameters are loaded into the initial state, the similarity is calculated through the prototype network (the similarity between fir and spruce is 0.82), the inner loop gradient update is performed in combination with the MAML framework, and with the assistance of GAN data augmentation, the fir-specific mechanism model parameters are generated and stored in the parameter library within 48 hours.

[0236] The inference process of the hybrid prediction model includes:

[0237] Single tree flammability probability calculation: Pre-trained parameters were loaded for common tree species (oak and birch), and special parameters generated by cross-species migration were loaded for fir. The flammability probability of each tree was calculated through a single tree physical mechanism model. The results showed that the flammability probability of fir was concentrated in the range of 0.22-0.76, and the probability of individuals with an age of more than 15 years and a leaf water content of less than 50% exceeded 0.5.

[0238] Multi-scale risk aggregation: The stand risk index is obtained by using a spatial weighted aggregation algorithm. The stand, topographic and meteorological feature vectors are spliced ​​together and input into the landscape-level prediction network to output the landscape-level flammability risk probability.

[0239] Three-source integration and long-term prediction: Calculate the uncertainty of forest stand risk index, landscape-level probability and traditional experience index, dynamically allocate weights to generate a comprehensive flammability risk index, output the risk sequence for the next 20 days through LSTM rolling prediction, and generate a GIS visualization spatiotemporal distribution map.

[0240] S300. Generate and release corresponding forest fire prevention early warning information based on the spatiotemporal distribution map of flammability risk.

[0241] Dynamically adjust early warning thresholds: lower the baseline threshold by 10% based on the windy characteristics of autumn, and raise the threshold by 20% for units with high uncertainty;

[0242] Warning determination and issuance: Five units triggered orange alerts, and two high-altitude units with concentrated fir trees triggered red alerts. Warning information including risk areas, duration and prevention and control recommendations was pushed out.

[0243] Targeted prevention and control: Focus on clearing combustibles under fir forests, deploying drones to enhance monitoring, and setting up firebreaks.

[0244] This embodiment fully verifies the core practical value and innovative advantages of the present invention in the typical scenario of Changbai Mountain mixed coniferous and broad-leaved forest, which has a complex tree species structure, contains rare tree species that have not been pre-trained, and has a large altitude range: Through a cross-species transfer learning architecture, only a small number of calibration samples are needed to quickly generate mechanistic model parameters for new tree species, effectively solving the industry pain point of poor adaptability of traditional methods to new and niche tree species, and significantly expanding the scope of application of the technology; relying on single-tree-level mechanistic modeling, multi-scale risk aggregation, and uncertainty-weighted fusion, it breaks through the limitations of vegetation homogenization processing in traditional models, accurately characterizes the differences in flammability of different tree species, and greatly improves the matching degree between the prediction results and the actual risk level. The prevention and control measures formulated here are more targeted, effectively transforming technological advantages into practical fire prevention results. Through long-term rolling forecasting technology, the early warning lead time is extended from several hours in traditional methods to tens of days. At the same time, the efficient automation of the new tree species adaptation process greatly reduces the time for technology deployment and the cost of data collection, and reduces reliance on manual intervention. With the combination of physical mechanism models and cross-species transfer learning, the prediction results are supported by both biological and physical mechanisms. Furthermore, the bias of a single model is reduced through multi-source data fusion and uncertainty quantification. It maintains stable prediction performance in areas with variable climate conditions, and has strong environmental adaptability and anti-interference ability, laying a solid foundation for subsequent large-scale promotion and application.

[0245] Example 4

[0246] Finally, this application also proposes a computing device, which includes a processor and a memory. The memory stores a computer program, and the processor executes the instructions stored in the memory so that the computer device performs the fire early warning method based on multidimensional forest health status analysis described in the above embodiments.

[0247] Furthermore, it should be noted that in the description of this application, 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 technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0248] In this application, unless otherwise expressly specified and limited, "above" or "below" the second feature can mean that the first and second features are in direct contact, or that they are in indirect contact through an intermediate medium. Furthermore, "above," "on top of," and "over" the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0249] In the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0250] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make modifications, alterations, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A fire prevention and early warning method based on multidimensional forest health status analysis, characterized in that, include: S100. Acquire multispectral remote sensing images corresponding to each evaluation unit of the target forest area, and preprocess the multispectral remote sensing images; the evaluation unit is a grid of multiple predetermined spatial scales divided from the target forest area. Vegetation index data, soil index data, topographic data and biological status data of each assessment unit are extracted from the preprocessed multispectral remote sensing images, and meteorological data of the corresponding locations are fused to obtain multidimensional observable data of each assessment unit. S200. Input the multidimensional observable data of each assessment unit into the pre-built hybrid prediction model, and output the comprehensive flammability risk index of each assessment unit within a specified future time period. Based on the comprehensive flammability risk index, a visual spatiotemporal distribution map of flammability risk is generated on the GIS map; S300. Generate and release corresponding forest fire prevention early warning information based on the spatiotemporal distribution map of flammability risk.

2. The method according to claim 1, characterized in that, The processing steps of the hybrid prediction model in S200 include: S201. Identify the tree species type of each tree in the current evaluation unit from the biological state data in the multidimensional observable data; S202. By querying the pre-built tree species mechanism model parameter library, determine whether each tree species type corresponds to a pre-trained dedicated mechanism model parameter; the tree species mechanism model parameter library is pre-built based on historical observation data and stores dedicated mechanism model parameters for multiple tree species. If the tree species type corresponds to a pre-trained dedicated mechanism model parameter, then the dedicated mechanism model parameter is loaded into the single tree physical mechanism model; If the tree species type does not have pre-trained dedicated mechanism model parameters, then the cross-species migration model is used to generate dedicated mechanism model parameters applicable to the tree species based on 10-50 calibration sample data of the tree species. The generated dedicated mechanism model parameters are then loaded into the single tree physical mechanism model and stored in the tree species mechanism model parameter library. The meteorological and soil data corresponding to the current single tree are input into the single tree physical mechanism model after loading parameters, and the single tree flammability probability is calculated. Summarize the individual tree flammability probability of all trees in the current assessment unit to obtain the set of individual tree flammability probabilities for each tree in the current assessment unit. S203. Based on the set of combustibility probabilities of individual trees, and combined with the spatial location coordinates and canopy area of ​​each tree extracted from multidimensional observable data, the stand risk index of the assessment unit is calculated using a spatial weighted aggregation algorithm. S204. The forest stand risk index is used as a vegetation feature and is combined with the topographic data and meteorological data in the multidimensional observable data to form a multidimensional feature vector, which is then input into the landscape-level risk prediction network to output the landscape-level flammability risk probability. S205. Calculate the forest health index based on the multidimensional observable data, and convert the forest health index into a traditional empirical flammability risk index through a preset mapping relationship. S206. Calculate the cognitive uncertainty and stochastic uncertainty of the forest stand risk index, landscape-level flammability risk probability, and traditional empirical flammability risk index, respectively. Adding the cognitive uncertainty to the corresponding random uncertainty yields the total uncertainty of the forest stand risk index, the landscape-level flammability risk probability, and the traditional experience flammability risk index. The fusion weights are dynamically allocated based on the total uncertainty, with higher fusion weights awarded to evaluation results that have lower total uncertainty. The forest stand risk index, landscape-level flammability risk probability, and traditional experience-based flammability risk index are weighted and summed according to the fusion weights to generate the comprehensive flammability risk index.

3. The method according to claim 2, characterized in that, The single-tree physical mechanism model is a mechanism model used to simulate the relationship between the physiological and ecological processes of a single tree and its flammability evolution, including the following sub-models: A transpiration mechanism model based on the Penman-Monteith equation was used to simulate the water dissipation process of trees. A photosynthesis mechanism model based on the Farquhar model was constructed to simulate the process of carbon assimilation and biomass accumulation in trees. A physical model of tree heat conduction based on the finite element heat conduction equation is used to simulate the temperature field distribution process inside the tree. A fuel accumulation mechanism model based on litter generation and decomposition kinetics is used to simulate the accumulation process of surface combustible load.

4. The method according to claim 2, characterized in that, If, in step S202, the tree species type does not have pre-trained dedicated mechanistic model parameters, then a cross-species transfer model is used to generate dedicated mechanistic model parameters applicable to that tree species based on calibration sample data, including: A model-agnostic meta-learning framework is adopted, with the pre-trained parameters in the prior knowledge base of the source species as the initial state of the physical mechanism model of the single tree, and an inner loop gradient update is performed on 10-50 calibration sample data of the tree species type. The prior knowledge base of the source species is pre-constructed based on the historical observation data of the source species and stores the basic pre-training parameters of the physical mechanism model of a single tree; the source species refers to the tree species with historical observation data used for pre-training the physical mechanism model of a single tree. Meanwhile, the calibration sample data is projected into the feature space through the prototype network, and similarity is measured with the class prototype vector generated by the prototype network based on the source species prior knowledge base, to assist in parameter adaptation. Generate specific mechanistic model parameters adapted to the tree species type, and store the specific mechanistic model parameters in the tree species mechanistic model parameter library.

5. The method according to claim 4, characterized in that, The calibration sample data is augmented using a generative adversarial network to generate diverse synthetic training samples, which are used to assist in the inner loop gradient update.

6. The method according to claim 2, characterized in that, The stand risk index of the assessment unit is calculated using a spatial weighted aggregation algorithm in S203, including: Calculate the stand risk index using the following formula: ； Among them, w ij The spatial weight factor between the i-th tree and the j-th tree is calculated using the following formula: ； d ij Let be the horizontal distance between the i-th tree and the j-th tree; This is a spatially relevant length parameter that is adaptively adjusted based on stand density. s i Let s be the canopy area of ​​the i-th tree. j Let be the area of ​​the canopy of the j-th tree; R i Let be the probability of flammability of the i-th tree. n represents the total number of trees within the forest stand area.

7. The method according to claim 2, characterized in that, The multidimensional feature vector input to the landscape-level risk prediction network in S204 includes: The stand feature vector is composed of the stand risk index and stand structure parameters, wherein the stand structure parameters include average tree height, average diameter at breast height, stand density, and canopy closure extracted from multidimensional observable data. A topographic feature vector consisting of slope, aspect, elevation, and topographic relief; A meteorological characteristic vector consisting of temperature, humidity, wind speed, precipitation, and evaporation; The landscape-level risk prediction network is a multilayer perceptron or convolutional neural network, and its output is the landscape-level flammability risk probability in the range of 0 to 1.

8. The method according to claim 2, characterized in that, The cognitive uncertainty mentioned in S206 is calculated using Monte Carlo dropout or deep ensemble methods; the random uncertainty is determined based on the observation noise statistical characteristics of the multidimensional observable data.

9. The method according to claim 1, characterized in that, The S300 also includes: Based on the risk level of each assessment unit in the aforementioned spatiotemporal distribution map of flammability risk, the warning level is divided according to the corresponding level of cognitive uncertainty. Set a baseline warning trigger threshold and an uncertainty determination threshold; dynamically adjust the baseline warning trigger threshold according to seasonal changes or extreme weather events; When the cognitive uncertainty of a certain assessment unit exceeds the uncertainty judgment threshold, the warning trigger threshold of that assessment unit is increased according to the preset rules to obtain the adjusted warning trigger threshold. When the comprehensive flammability risk index of the assessment unit exceeds the adjusted warning trigger threshold, a warning of the corresponding level is triggered.

10. The method according to claim 1, characterized in that, The specified future time period in S200 is the next 7-30 days; the hybrid prediction model is based on historical multidimensional observable data sequences, performs rolling prediction through a time series prediction network, and outputs the comprehensive flammability risk index of each evaluation unit for each day in the specified future time period.

Images