Forest fire susceptibility assessment method and system based on XGBoost-SHAP interpretable model

Abstract

The application discloses a kind of forest fire-prone area assessment method and system based on XGBoost-SHAP explainable model, the method includes steps: obtaining the multi-source data of target area in preset historical period and pre-processing, obtain vegetation type spatial distribution dataset, forest vegetation is divided into independent category, construct XGBoost-based wild fire susceptibility model, generate multi-season susceptibility dynamic atlas;Build evaluation model, determine factor weight, carry out comprehensive risk assessment and spatial zoning, form the double-dimension risk atlas of fusion "occurrence probability-potential loss", support multi-season, the differentiated forest fire risk management of multiple vegetation types.Adopting the technical scheme, independent modeling according to vegetation type, overcome the limitation that traditional method is homogenized to assess heterogeneous forest area, fully consider the seasonal dynamic change of fire risk, provide fine technical support for differentiated, dynamic fire prevention resource allocation.

Description

Technical Field

[0001] This invention belongs to the field of fire risk assessment technology, and relates to a method and system for assessing forest fire susceptibility based on the XGBoost-SHAP interpretable model. Background Technology

[0002] Forest fires are among the most serious natural disasters posing a severe threat to the ecological environment, socio-economic development, and human safety globally. With the intensification of global climate change and the increasing frequency of extreme weather events, the frequency and intensity of forest fires are on the rise, making fire prevention and control an increasingly challenging task. Therefore, conducting scientific and accurate forest fire risk assessment research has become a crucial foundation for improving forest fire prevention capabilities, optimizing resource allocation, and formulating prevention and control strategies.

[0003] Currently, forest fire risk assessment research has gradually evolved from early empirical statistical models to a multi-factor comprehensive evaluation system that combines remote sensing, geographic information systems, and machine learning technologies. Scholars both domestically and internationally widely employ machine learning methods such as logistic regression, random forests, and support vector machines, combined with driving factors such as meteorology, topography, vegetation, and human activities, to construct fire susceptibility prediction models.

[0004] Existing research has propelled fire risk assessment from qualitative description to quantitative and spatial methods. However, several key issues remain to be addressed: First, most assessment models do not adequately consider the impact of vegetation heterogeneity on fire occurrence mechanisms. For example, different vegetation types, such as evergreen broad-leaved forests, deciduous broad-leaved forests, and coniferous forests, exhibit significant differences in combustible composition, moisture content, and combustion characteristics, leading to varying dominant factors and spatial patterns of fire risk. Second, many existing studies treat the entire forest area as a homogeneous whole for modeling, neglecting the classification and assessment of vegetation types. This limits the accuracy of model predictions and makes it difficult to accurately reflect the fire risk characteristics of different forest types.

[0005] Secondly, current research largely focuses on fire susceptibility assessments, lacking a systematic integration of potential fire consequences. Existing assessment systems generally overlook the potential impacts of fire on ecosystem services, population safety, and economic development, leading to risk zoning results that favor areas with a "high probability of occurrence" while potentially ignoring high-risk areas with a "moderate probability of occurrence but severe consequences."

[0006] Furthermore, most machine learning models are "black box" models, making their decision-making processes difficult to interpret and hindering the mechanistic analysis of fire factors and support for prevention and control decisions. At the same time, many studies are based on static or annual-scale data, failing to adequately consider the seasonal dynamics of fire risk, resulting in assessment results that are difficult to adapt to the fire prevention and control needs of different seasons. Summary of the Invention

[0007] The purpose of this invention is to address the aforementioned problems in existing technologies by proposing a forest fire susceptibility assessment method and system based on the XGBoost-SHAP interpretable model.

[0008] To achieve the above objectives, the basic solution of this invention is: a forest fire susceptibility assessment method based on the XGBoost-SHAP interpretable model, comprising the following steps:

[0009] S1. Obtain multi-source data of the target area within a preset historical period, and preprocess it to obtain a spatial distribution dataset of vegetation types.

[0010] S2, based on the spatial distribution dataset of vegetation types, divides forest vegetation into independent categories, including: evergreen broad-leaved forest, deciduous broad-leaved forest, and coniferous forest;

[0011] S3. For each independent category of forest vegetation, a wildfire susceptibility model based on XGBoost is constructed, and the hyperparameters of the model are tuned. The contribution and direction of each driving factor are quantified using the SHAP (Shapley addition interpretation) interpretability analysis method, and a multi-seasonal susceptibility dynamic map is generated.

[0012] S4. Construct an assessment model that comprehensively reflects ecological, social and economic vulnerability. Use the AHP-PCA-minimum relative information entropy combined weighting method to determine factor weights, and achieve dynamic assessment of vulnerability through coefficient of variation and trend analysis.

[0013] S5, based on the results of susceptibility and vulnerability assessments, conducts comprehensive risk assessments and spatial zoning, forming a two-dimensional risk map that integrates "probability of occurrence - potential loss," supporting differentiated forest fire risk management across multiple seasons and vegetation types.

[0014] The working principle and beneficial effects of this basic scheme are as follows: This technical scheme proposes a refined assessment framework for forest fire risk based on vegetation type differentiation, models different forest types separately, and breaks through the limitations of traditional methods in homogenizing heterogeneous forest areas.

[0015] In forest fire risk assessment, the SHAP interpretability framework is systematically applied to solve the "black box" problem of machine learning models, enabling quantitative analysis of the contribution and direction of influence of each driving factor. A dual-dimensional coupled hazard assessment method of "probability of occurrence - potential loss" is proposed, which integrates fire susceptibility (probability) and vulnerability (consequences), realizing a paradigm shift from single probability prediction to comprehensive risk assessment.

[0016] Furthermore, the method for preprocessing multi-source data of the target area within a preset historical period is as follows:

[0017] The original fire point product VNP14IMGN was obtained from the satellite platform NASA and FIRMS. Based on the built-in "confidence level" and "fire point type" fields, high-confidence vegetation burning points were selected. The vector boundary of the study area was used for spatial cropping to remove irrelevant and low-quality records.

[0018] Set a spatial buffer and a time window, aggregate multiple fire points that fall within the spatial buffer, treat them as a single fire event, and take the center point or the earliest recorded point of the point set formed by the aggregated multiple fire points as the representative sample of the event to ensure sample independence and avoid overfitting.

[0019] The aggregated fire points are spatially overlaid with the high-resolution vegetation type map for analysis. Each fire point is assigned a vegetation type attribute for its corresponding pixel. Specifically:

[0020] Calculate the spatial relationship between fire points and vegetation type maps, extract the vegetation type attributes of each fire point location and assign them to the fire point, providing a data foundation for subsequent modeling according to different vegetation types;

[0021] Complete the temporal dimension deconstruction and sample set organization: Based on the observation date of the fire points, divide them into corresponding years and climate seasons: spring, summer, autumn, and winter, and output a structured fire point sample database containing spatial coordinates, vegetation type, season, and year labels. The fire point sample database is used as positive samples and combined with negative samples randomly generated in the corresponding vegetation type areas to form a vegetation type spatial distribution dataset.

[0022] By filtering based on confidence level and fire point type, low-quality data was eliminated, ensuring the reliability of the modeling samples. By setting spatial buffers and time windows for fire point aggregation, the problem of multiple records of the same fire caused by satellite revisits was effectively overcome, ensuring sample independence, avoiding model overfitting, and laying a solid data foundation for subsequent high-precision modeling.

[0023] Furthermore, for each independent category of forest vegetation, a wildfire susceptibility model based on XGBoost was constructed, and Bayesian optimization was used for hyperparameter tuning. The specific steps are as follows:

[0024] For each vegetation type, pre-defined driving factors are extracted from the spatial distribution dataset of vegetation types, and multicollinearity diagnosis and screening are performed to ensure the independence of input factors.

[0025] The gradient boosting method is used to progressively optimize the prediction error. This is achieved by generating a series of weak learners and weighting their predictions to form a strong learner. The following parameters are selected for tuning: n_estimators (number of weak learners), max_depth (maximum tree depth), learning_rate (learning rate), subsample (sample sampling ratio), and colsample_bytree (feature / column sampling ratio).

[0026] The XGBoost model is hyperparameter tuned using the scikit-optimize framework. Hyperparameters and their ranges are defined. Several points and their corresponding objective function values ​​are obtained through initial sampling. A surrogate model is constructed using a Gaussian process to approximate the objective function. Specifically:

[0027] Before starting the formal Bayesian optimization iteration, a random search method is used to randomly select several sets of parameter combinations from the hyperparameter space, and the model is trained separately to obtain its performance index, i.e. the objective function, as the data basis for building the initial surrogate model.

[0028] Using existing hyperparameter combinations and corresponding model accuracy as observation data, the similarity between different parameter groups is calculated through the objective function. The objective function value of any unknown point in the entire hyperparameter space is predicted probabilistically, and its expected value (mean) and uncertainty (variance) are output.

[0029] The next sampling point is selected by the sampling function, its target value is evaluated and the surrogate model is updated until the optimal value of the objective function is found as the optimal model parameter.

[0030] By employing Bayesian optimization algorithms to automatically optimize key parameters (such as n_estimators, max_depth, etc.) of the XGBoost model, the globally optimal parameter combination can be found with fewer iterations compared to traditional grid search or random search, thus improving parameter tuning efficiency and model performance.

[0031] Furthermore, the following steps were taken to quantify the contribution and direction of influence of each factor using the SHAP interpretability analysis method to generate a multi-seasonal susceptibility dynamic map:

[0032] The trained BO-XGBoost is interpreted using TreeExplainer from the SHAP library. The SHAP value of each driving factor for each sample is calculated. The contribution of the factors is ranked by the mean absolute SHAP value. The direction of influence is determined by the positive or negative SHAP value. A SHAP value > 0 indicates that increasing the value of the driving factor will increase the probability of fire, while a SHAP value < 0 indicates that it will decrease the probability of fire.

[0033] The data was split into spring, summer, autumn and winter, and then input into the wildfire susceptibility model to obtain continuous probability values. In ArcGIS, the probability values ​​were reclassified into five susceptibility levels using the natural discontinuity method: extremely low susceptibility, low susceptibility, medium susceptibility, high susceptibility and extremely high susceptibility. Four susceptibility spatial distribution maps for spring, summer, autumn and winter were generated, and their rationality was verified by overlaying historical fire points. The spatiotemporal dynamic change patterns of susceptibility areas were analyzed.

[0034] Introducing the SHAP model interpretation tool, SHAP predicted value y pre for:

[0035] y pre =y base +f(x i1 )+f(x i2 )+....+f(x in ),

[0036] Where, x ik y is the k-th feature of the i-th sample; pre y represents the model's predicted value for this sample. base This serves as the baseline for the entire model;

[0037] Using b1-b12 to represent the driving factors of each participating model, we analyze models for different vegetation types to enhance the interpretability and transparency of the models. Specifically:

[0038] By comparing the bar lengths of each factor's code b1-b12 and creating SHAP contribution bar charts for each vegetation type, the core driving factors of evergreen broad-leaved forests were identified. The atlas is a spatially visualized distribution map generated in ArcGIS by classifying the wildfire susceptibility probability values ​​predicted by the model using the natural discontinuity method.

[0039] By introducing the SHAP tool and calculating the contribution of each feature to the prediction results, the specific direction and magnitude of the influence of driving factors such as temperature and rainfall on the fire risk of different vegetation types are quantified and analyzed, making the complex "black box" model decision-making process transparent. By drawing SHAP contribution bar charts, the core driving factors of fire occurrence in different forest areas (such as evergreen broad-leaved forests) can be intuitively identified, greatly enhancing the interpretability of the model.

[0040] Furthermore, for each vegetation type, predefined driving factors are extracted from the spatial distribution dataset of vegetation types, and multicollinearity diagnosis and screening are performed to ensure the independence of input factors. The method is as follows:

[0041] Driving factors include climate factors, topographic factors, vegetation factors, and population factors.

[0042] From the spatial distribution dataset of vegetation types, 13 key driving factors in four categories were extracted: temperature, precipitation, potential evapotranspiration, elevation, slope, topographic humidity index, water flow intensity index, aspect, curvature, soil classification, normalized vegetation index (NDVI), ecological environment quality, and population density.

[0043] Spatial registration, clipping, and resampling of all raster data were completed in the ArcGIS platform. For the spatial distribution data of each vegetation type, the variance inflation factor (VIF) and tolerance (TOL) of each driving factor were calculated. Multicollinearity diagnosis and factor screening were performed based on the criteria of VIF < 10 and TOL > 0.1 to ensure that the feature variables input into the machine learning model are independent, stable, and reliable, laying a solid data foundation for subsequent high-precision modeling.

[0044] ,

[0045] Among them, TOL is an indicator that measures the degree of linear correlation between a certain independent variable and other independent variables, reflecting the degree of independence of the independent variable from the dependent variable; The coefficient of determination for the regression model is denoted by ; the VIF value is the reciprocal of TOL.

[0046] By integrating 13 key driving factors across four categories—climate, topography, vegetation, and human activities—the study comprehensively depicts the complex environment influencing forest fire occurrence.

[0047] Furthermore, an assessment model comprehensively reflecting ecological, social, and economic vulnerability is constructed. The AHP-PCA-minimum relative information entropy combined weighting method is used to determine factor weights, and dynamic vulnerability assessment is achieved through coefficient of variation and trend analysis. The specific steps are as follows:

[0048] Integrate vegetation dynamics data, social vulnerability data, and economic vulnerability data from the spatial distribution dataset of vegetation types;

[0049] All integrated data are spatially clipped, resampled to a uniform resolution and coordinate system transformation to construct a spatially aligned, temporally consistent, multidimensional standardized dataset.

[0050] Trend and stability analysis was performed on vegetation dynamics data and economic vulnerability data with time series characteristics in multidimensional standardized datasets, including calculating the coefficient of variation to assess the stability of their temporal fluctuations and performing linear trend analysis to identify their direction of change.

[0051] The final weights of each driving factor in the forest fire vulnerability assessment were determined by adopting a weight determination strategy of "coordination between subject and object and information fusion".

[0052] A linear weighted summation model is used as the core method for calculating the vulnerability index. The vulnerability index (B) is:

[0053] That is, the evaluation model;

[0054] in, Due to wildfire vulnerability, This represents the value of the i-th indicator factor, where i is the sequence number of the time parameter. This represents the combined weight of the i-th indicator factor. This represents the total number of years of research.

[0055] By integrating multi-source vulnerability data, including vegetation dynamics, social and economic data, and employing a linear weighted summation model, a comprehensive quantitative assessment of the potential consequences of fires was achieved.

[0056] Furthermore, for a given raster cell, the coefficient of variation (CV) of a certain factor in its time series is:

[0057] ,

[0058] in, Total number of years of research; It is a time series; These are the variable values ​​at the corresponding time points; The average value of the variable;

[0059] For each pixel's time-series data, the slope of its trend, slop, is calculated using univariate linear regression:

[0060] ,

[0061] When slop>0, it indicates that the factor value corresponding to the pixel is on an upward trend, i.e., an improving trend; when slop<0, it indicates that the factor value is on a downward trend, i.e., a deteriorating trend; when slop≈0, it indicates that there is no obvious trend of change.

[0062] By clarifying the calculation formulas for the coefficient of variation (CV) and the trend slope, a precise and quantifiable mathematical method is provided for dynamic vulnerability assessment.

[0063] Furthermore, adopting a weighting strategy of "coordination between subject and object, and information fusion," the final weights of each influencing factor in the forest fire vulnerability assessment are determined as follows:

[0064] Using the Analytic Hierarchy Process (AHP), a judgment matrix is ​​constructed, and a subjective weight vector reflecting management cognition and regional characteristics is calculated, specifically:

[0065] A three-tiered structure of objectives, criteria, and solutions is established. Based on management cognition and regional characteristics, a 1-9 scale method is used to compare and score factors at the same level to construct a judgment matrix. Preliminary weights are calculated through normalization, and finally, a consistency test (CR<0.1) is performed to ensure logical rationality. The output is a subjective weight vector that reflects subjective experience.

[0066] Principal component analysis (PCA) was used to perform statistical analysis on the standardized multidimensional dataset. Principal components were extracted based on variance contributions, and objective weight vectors based on the intrinsic structure of the data were calculated. Specifically:

[0067] The raw data of the nine vulnerability factors (EVI, GDP, NPP, NX, DTR, RTR, RS, DTF, UL) were constructed into a sample matrix and standardized to eliminate the influence of units.

[0068] Calculate the covariance matrix of the standardized data, i.e., the correlation coefficient matrix, and perform eigenvalue decomposition on it to obtain 9 eigenvalues: λ1≥λ2≥...≥λ9, and their corresponding eigenvectors;

[0069] Calculate the variance contribution rate (e.g., PC1 contribution rate 0.316) and cumulative contribution rate of each principal component based on the eigenvalues, and select the top k principal components (the top 5, with a cumulative contribution rate of 0.852) with a cumulative contribution rate exceeding 85%.

[0070] The eigenvectors are multiplied by the square roots of their eigenvalues ​​to construct the loading matrix, which reflects the correlation between the original factors and the principal components.

[0071] The weights of each principal component contribution rate are used to perform a weighted summation of the absolute values ​​of the loadings to obtain the comprehensive score coefficient for each factor.

[0072] Finally, the overall score coefficient is normalized;

[0073] By introducing the principle of minimum relative information entropy from information theory as an optimization criterion, a mathematical model is established to fuse and calculate the subjective and objective weight vectors, and the comprehensive weight vector with the smallest deviation from both is obtained.

[0074] ,

[0075] ,

[0076] ,

[0077] in, This indicates the final weight of each evaluation indicator; and These represent the weights determined by the AHP (Analog-Hybrid Analysis) method and the weights determined by the PCA (Principal Component Analysis) method, respectively.

[0078] The weight determination strategy of "subject-object collaboration and information fusion" is adopted to ensure the rationality of the weights. Principal component analysis (PCA) is based on the inherent statistical laws of data to objectively reflect the contribution of factors, avoid the arbitrariness of purely subjective weighting, and realize the scientific integration and optimization of subjective experience and objective data.

[0079] Furthermore, based on the susceptibility and vulnerability assessment results, a comprehensive hazard assessment and spatial zoning are conducted to form a two-dimensional risk map integrating "probability of occurrence - potential loss," supporting differentiated forest fire risk management across multiple seasons and vegetation types. The specific methods are as follows:

[0080] By spatial overlaying and hierarchical zoning of geographic information, risk maps with clear spatial orientation and seasonal differences are generated, revealing the distribution of high-risk areas for different vegetation types in different seasons:

[0081] R = S × V,

[0082] in, Wildfire hazard in the target area; Wildfire susceptibility in the target area; Wildfire vulnerability in the target area;

[0083] The continuous R values ​​in the ArcGIS platform (Geographic Information Processing Platform) are divided into five levels using the natural discontinuity method: very low, low, medium, high, and very high. The higher the level, the greater the probability of fire occurrence and the more severe the potential losses in the area. The rationality of the classification is verified by frequency ratio.

[0084] By spatially coupling the results of susceptibility and vulnerability assessments, a comprehensive risk map integrating "probability of occurrence" and "potential loss" is generated, which intuitively reveals the spatial distribution pattern of risk.

[0085] The present invention also provides a forest fire susceptibility assessment system based on the XGBoost-SHAP interpretable model, including a data acquisition unit and a processing unit. The data acquisition unit is used to acquire multi-source data of the target area within a preset historical period.

[0086] The input terminal of the processing unit is connected to the output terminal of the data acquisition unit. The processing unit executes the method described in this invention to obtain the forest fire susceptibility assessment results.

[0087] This system integrates complex data processing, model building, risk calculation, and zoning processes through the collaborative work of data acquisition and processing units, achieving full automation from data input to risk assessment result output. Attached Figure Description

[0088] Figure 1This is a flowchart illustrating the forest fire susceptibility assessment method based on the XGBoost-SHAP interpretable model of this invention.

[0089] Figure 2 This is a wildfire hazard assessment result diagram for different vegetation types based on the forest fire susceptibility assessment method of the present invention, which is based on the XGBoost-SHAP interpretable model.

[0090] Figure 3 This is an example of the SHAP feature importance analysis diagram of the forest fire susceptibility assessment method based on the XGBoost-SHAP interpretable model of this invention. Detailed Implementation

[0091] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0092] In the description of this invention, it should be understood that the terms "longitudinal", "lateral", "up", "down", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0093] In the description of this invention, unless otherwise specified and limited, it should be noted that the terms "installation", "connection" and "linking" should be interpreted broadly. For example, they can refer to mechanical or electrical connections, or internal connections between two components. They can be direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0094] This invention discloses a forest fire susceptibility assessment method based on the XGBoost-SHAP interpretable model. It overcomes the shortcomings of existing technologies, such as model homogenization, single dimension, weak interpretability, and static assessment. This method can distinguish vegetation types, couple susceptibility and vulnerability, possess interpretability, and support seasonal dynamic assessment, thus achieving more scientific, refined, and practical fire risk zoning and management. Figure 1 As shown, the forest fire susceptibility assessment method based on the XGBoost-SHAP interpretable model includes the following steps:

[0095] S1. Obtain multi-source data of the target area within a preset historical period, and preprocess it to obtain a spatial distribution dataset of vegetation types.

[0096] S2, based on the spatial distribution dataset of vegetation types, divides forest vegetation into independent categories, including: evergreen broad-leaved forest, deciduous broad-leaved forest, and coniferous forest;

[0097] S3. For each independent category of forest vegetation, a wildfire susceptibility model based on XGBoost is constructed, and the hyperparameters of the model are tuned. The SHAP (Shapley addition interpretation) interpretability analysis method is used to quantify the contribution and direction of influence of each driving factor (climate, vegetation, topography, population, etc.) and generate a multi-seasonal susceptibility dynamic map.

[0098] S4. Construct an assessment model that comprehensively reflects ecological, social and economic vulnerability. Use the AHP-PCA-minimum relative information entropy combined weighting method to determine factor weights, and achieve dynamic assessment of vulnerability through coefficient of variation and trend analysis.

[0099] S5, based on the results of susceptibility and vulnerability assessments, conducts comprehensive risk assessments and spatial zoning, forming a two-dimensional risk map that integrates "probability of occurrence - potential loss," supporting differentiated forest fire risk management across multiple seasons and vegetation types.

[0100] In a preferred embodiment of the present invention, the method for preprocessing multi-source data of a target area within a preset historical period is as follows:

[0101] The original fire point product VNP14IMGN was obtained from the satellite platform NASA and FIRMS. Based on the built-in "confidence level" and "fire point type" fields, high-confidence vegetation burning points were selected. The vector boundary of the study area was used for spatial cropping to remove irrelevant and low-quality records.

[0102] To overcome the problem of multiple recordings of the same fire caused by satellite revisit, a spatial buffer (creating a circular buffer for each fire point to identify and aggregate adjacent fire points) and a time window are set up. Multiple fire points falling within the spatial buffer are aggregated and treated as a single fire event. The center point or the earliest recorded point of the point set formed by the aggregated multiple fire points is taken as the representative sample of the event to ensure sample independence and avoid overfitting.

[0103] The aggregated fire points are spatially overlaid with a high-resolution vegetation type map, and each fire point is assigned a vegetation type attribute to its corresponding pixel, such as evergreen broad-leaved forest or deciduous broad-leaved forest. This step is the core foundation supporting subsequent "vegetation type-specific modeling," enabling the model to learn the unique fire risk patterns of different vegetation systems. Specifically:

[0104] The spatial relationship between fire points and vegetation type maps (referring to the location of forest fire ignition points, derived from VIIRS satellite fire point data) is calculated. The vegetation type attribute of each fire point's location is extracted and assigned to that fire point, providing a data foundation for subsequent modeling according to different vegetation types. Specifically:

[0105] Raw fire point data is extracted from satellite fire point product data, and filtered based on fire point confidence level and fire point type, retaining fire point samples that meet preset conditions. The filtering conditions are as follows:

[0106] ,

[0107] Among them, confidence j Let j be the confidence level of the j-th fire point. To pre-set the reliability threshold, type j This is a fire point type identifier.

[0108] Spatial constraints: To eliminate redundant fire points resulting from multiple observations of the same fire event, fire points are aggregated based on spatial distance and time windows.

[0109] ,

[0110] ,

[0111] Where, d ij Let be the spatial distance between fire point i and fire point j, and D be the spatial distance threshold. , Let i be the coordinates of the fire point. , Let J be the coordinates of the fire point.

[0112] Time constraints: ,

[0113] Among them, t i , t j , where represents the time of fire occurrence, and T is the time window threshold.

[0114] Aggregation determination:

[0115] ,

[0116] Fire points that meet the above conditions belong to the same fire event.

[0117] Fire incident statement:

[0118] ,

[0119] Representative fire point calculation:

[0120] ,

[0121] The aggregated fire points are spatially overlaid with multi-source raster data for analysis, and the attribute values ​​of each driving factor are extracted as follows:

[0122] ,

[0123] in, Rm represents the value of the k-th fire point on the m-th driving factor, and Rm is the corresponding raster data function.

[0124] The objective function is:

[0125] ,

[0126] Where n represents the number of samples, y i For true labels (0 or 1). To predict the probability of a fire occurring using the model.

[0127] Complete the temporal dimension deconstruction and sample set organization: Based on the observation date of fire points, they are divided into corresponding years and climate seasons: spring, summer, autumn, and winter. A structured fire point sample database containing spatial coordinates, vegetation type, season, and year labels is output. This fire point sample database serves as the positive sample and is combined with randomly generated negative samples within the corresponding vegetation type areas (specifically, locations randomly generated within the study area that have never experienced fires historically. The model learns the characteristics of these locations to understand where fires are unlikely to occur). Together, they form a vegetation type spatial distribution dataset.

[0128] In a preferred embodiment of the present invention, a wildfire susceptibility model based on XGBoost is constructed for each independent category of forest vegetation, and Bayesian optimization is used for hyperparameter tuning. The specific steps are as follows:

[0129] For each vegetation type, pre-defined driving factors are extracted from the spatial distribution dataset of vegetation types, and multicollinearity diagnosis and screening are performed to ensure the independence of input factors.

[0130] The gradient boosting method is used to progressively optimize the prediction error. This is achieved by generating a series of weak learners and weighting their predictions to form a strong learner. The following parameters are selected for tuning: n_estimators (number of weak learners), max_depth (maximum tree depth), learning_rate (learning rate), subsample (sample sampling ratio), and colsample_bytree (feature / column sampling ratio).

[0131] The XGBoost model is hyperparameter tuned using the scikit-optimize framework. Hyperparameters and their ranges are defined. Several points and their corresponding objective function values ​​are obtained through initial sampling. A surrogate model is constructed using a Gaussian process to approximate the objective function. Specifically:

[0132] Before starting the formal Bayesian optimization iteration, a random search method is used to randomly select several sets of parameter combinations from the hyperparameter space, and the model is trained separately to obtain its performance index, i.e., the objective function (an index used to measure the quality of the model), as the data basis for building the initial surrogate model.

[0133] Using existing hyperparameter combinations and corresponding model accuracy as observation data, the similarity between different parameter groups is calculated through the objective function. The objective function value of any unknown point in the entire hyperparameter space is predicted probabilistically, and its expected value (mean) and uncertainty (variance) are output.

[0134] The next sampling point is selected through a sampling function (a decision rule that guides where to sample next; based on the mean and variance of the surrogate model output, it calculates the "potential value" of each point in the entire hyperparameter space and selects the point with the highest potential value as the next set of parameters to be tried; for example, when adjusting learning_rate z, EI will select a hidden better solution with a lower mean but higher uncertainty) to evaluate its target value and update the surrogate model until the optimal value of the objective function is found as the optimal model parameters.

[0135] To assess wildfire susceptibility in the study area, non-fire hotspots were first randomly selected within the area at a 1:1 ratio, and the validation and training sets were then divided at a 3:7 ratio. A wildfire susceptibility model based on an extreme gradient boosting model was then optimized using the scikit-optimize framework. The model's performance was validated using four methods: model accuracy, ROC-AUC, Recall, F1, and Kappa. Finally, wildfire susceptibility results were obtained based on the model, and wildfire-susceptible regions were delineated and the model results visualized.

[0136] In a preferred embodiment of the present invention, the steps for quantifying the contribution and influence direction of each factor using the SHAP interpretability analysis method to generate a multi-seasonal susceptibility dynamic map are as follows:

[0137] The TreeExplainer library of the SHAP library is used to interpret the trained wildfire susceptibility model (i.e., the BO-XGBoost model). The SHAP value of each driving factor for each sample is calculated. The contribution of the factors is ranked by the mean absolute SHAP value. The direction of influence is judged by the positive or negative SHAP value. A SHAP value > 0 indicates that increasing the value of the driving factor will increase the probability of fire, and < 0 indicates that it will decrease the probability of fire.

[0138] The research data was split into spring, summer, autumn and winter, and then input into the wildfire susceptibility model to obtain continuous probability values. In ArcGIS, the probability values ​​were reclassified into five susceptibility levels using the natural discontinuity method: extremely low susceptibility, low susceptibility, medium susceptibility, high susceptibility and extremely high susceptibility. Four susceptibility spatial distribution maps for spring, summer, autumn and winter were generated, and their rationality was verified by overlaying historical fire points. The spatiotemporal dynamic change patterns of susceptibility areas were analyzed.

[0139] Introducing the SHAP model interpretation tool, SHAP predicted value y pre for:

[0140] y pre =y base +f(x i1 )+f(x i2 )+....+f(x in ),

[0141] Where, x ik y is the k-th feature of the i-th sample; pre y represents the model's predicted value for this sample. base This serves as the baseline for the entire model;

[0142] Due to image area limitations, b1-b12 are used to represent the driving factors of each participating model, such as... Figure 2 As shown, analysis of models for different vegetation types enhances the interpretability and transparency of the models. This not only helps in understanding the decision-making basis of the models but also identifies invalid features or potential overfitting problems, providing a scientific basis for further model optimization. Specifically:

[0143] By drawing SHAP contribution bar charts for each vegetation type and comparing the bar lengths of each factor code b1-b12, the core driving factors of evergreen broad-leaved forests were identified.

[0144] In a preferred embodiment of the present invention, for each vegetation type, a preset driving factor is extracted from the spatial distribution dataset of vegetation types, and multicollinearity diagnosis and screening are performed to ensure the independence of the input factors. The method is as follows:

[0145] Driving factors include climate factors, topographic factors, vegetation factors, and population factors.

[0146] From the spatial distribution dataset of vegetation types, 13 key driving factors in four categories were extracted: temperature, precipitation, potential evapotranspiration, elevation, slope, topographic humidity index, water flow intensity index, aspect, curvature, soil classification, normalized vegetation index (NDVI), ecological environment quality, and population density.

[0147] Spatial registration, clipping, and resampling of all raster data were completed in the ArcGIS platform. For the spatial distribution data of each vegetation type, the variance inflation factor (VIF) and tolerance (TOL) of each driving factor were calculated. Multicollinearity diagnosis and factor screening were performed based on the criteria of VIF < 10 and TOL > 0.1 to ensure that the feature variables input into the machine learning model are independent, stable, and reliable, laying a solid data foundation for subsequent high-precision modeling.

[0148] ,

[0149] Among them, TOL is an indicator that measures the degree of linear correlation between a certain independent variable and other independent variables, reflecting the degree of independence of the independent variable from the dependent variable; The coefficient of determination for the regression model is denoted by ; the VIF value is the reciprocal of TOL.

[0150] In a preferred embodiment of the present invention, an assessment model that comprehensively reflects ecological, social, and economic vulnerability is constructed. The factor weights are determined using the AHP-PCA-minimum relative information entropy combined weighting method, and dynamic vulnerability assessment is achieved through coefficient of variation and trend analysis. The specific steps are as follows:

[0151] Vegetation dynamics data, including multi-year series enhanced vegetation index (EVI) and net primary productivity (NPP) from remote sensing satellites such as MODIS.

[0152] Social vulnerability data includes census-based proportions of vulnerable populations (RS) and female populations (NX), road accessibility based on transportation networks (RTR), and fire station accessibility based on public facility locations (DTF).

[0153] Economic vulnerability data include the Nighttime Light Index (UL) derived from nighttime light imagery and Gross Domestic Product (GDP) derived from statistical yearbooks.

[0154] Integrate vegetation dynamics data, social vulnerability data, and economic vulnerability data from the spatial distribution dataset of vegetation types;

[0155] All integrated data are spatially clipped, resampled to a uniform resolution, and transformed to a coordinate system to construct a spatially aligned, temporally consistent, multidimensional standardized dataset, laying the data foundation for subsequent dynamic analysis and fusion evaluation.

[0156] Trend and stability analysis was performed on vegetation dynamics data and economic vulnerability data with time series characteristics in multidimensional standardized datasets, including calculating the coefficient of variation to assess the stability of their temporal fluctuations and performing linear trend analysis to identify their direction of change.

[0157] The final weights of each driving factor in the forest fire vulnerability assessment were determined by adopting a weight determination strategy of "coordination between subject and object and information fusion".

[0158] A linear weighted summation model is used as the core method for calculating the vulnerability index. The vulnerability index (B) is:

[0159] That is, the evaluation model;

[0160] in, Due to wildfire vulnerability, This represents the value of the i-th indicator factor, where i is the sequence number of the time parameter. This represents the combined weight of the i-th indicator factor. This represents the total number of years of research.

[0161] To achieve dynamic assessment of forest fire vulnerability, this invention performs trend and stability analysis on vegetation dynamic data with time-series characteristics (such as Enhanced Vegetation Index (EVI) and Net Primary Productivity (NPP)) and economic vulnerability data (such as Nighttime Light Index (UL)).

[0162] The coefficient of variation (CV) is used to quantify the volatility and stability of each factor over time. A higher CV indicates greater volatility and instability of the factor over time; conversely, a lower CV indicates relative stability.

[0163] In a preferred embodiment of the present invention, for a given raster pixel, the coefficient of variation (CV) of a certain factor in its time series is:

[0164] ,

[0165] in, Total number of years of research; It is a time series; These are the variable values ​​at the corresponding time points; The average value of the variable;

[0166] In a GIS or remote sensing processing platform, the mean and standard deviation of the time series data for each pixel are calculated, and then the coefficient of variation (CV) is obtained, generating a "spatial distribution map of the coefficient of variation" that reflects the stability of fluctuations across the entire region.

[0167] The CV value is divided into five levels—"extremely low volatility," "low volatility," "medium volatility," "high volatility," and "extremely high volatility"—using the natural discontinuity method. Figure 3 As shown in the figure, low-fluctuation areas represent long-term stability of vegetation cover or socio-economic status, usually corresponding to mature ecosystems or steady-state structures; while high-fluctuation areas indicate that environmental or economic conditions are easily disturbed, and may contain higher potential risks or uncertainties in resilience.

[0168] Linear regression trend analysis is used to identify the direction of change (improvement or degradation) of each factor over a long time scale and its significance, thereby capturing the dynamic evolution trend of vulnerability.

[0169] For each pixel's time-series data, the slope of its trend, slop, is calculated using univariate linear regression:

[0170] ,

[0171] Linear fitting was performed on the time series data of each pixel within the study area to calculate the trend slope of each pixel, generating a "spatial distribution map of change trend". When slop > 0, it indicates that the factor value corresponding to that pixel is on an upward trend, i.e., an improving trend (such as improved vegetation growth and enhanced economic activity); when slop < 0, it indicates that the factor value is on a downward trend, i.e., a deteriorating trend (such as reduced vegetation cover and declining economic activity); when slop ≈ 0, it indicates no obvious change trend.

[0172] In a preferred embodiment of the present invention, a weighting determination strategy of "subject-object collaboration and information fusion" is adopted to determine the final weights of each influencing factor in the forest fire vulnerability assessment. The method is as follows:

[0173] Using the Analytic Hierarchy Process (AHP), a judgment matrix is ​​constructed, and a subjective weight vector reflecting management cognition and regional characteristics is calculated, specifically:

[0174] A three-tiered structure of objectives, criteria, and solutions is established: Objective layer, evaluating the overall objective of wildfire vulnerability; Criteria layer, evaluating dimensions: vegetation socio-economic factors; Solution layer, evaluating indicators: EVI and NPP. Based on management understanding and regional characteristics, a 1-9 scale is used to construct a judgment matrix by comparing and scoring factors pairwise within the same tier (a tool in AHP for quantifying pairwise comparisons, where 1 represents equal importance, 3, 5, 7, and 9 represent increasing importance, and 2, 4, 6, and 8 represent the median value. The judgment matrix is ​​constructed using this scale, and the weights of each indicator are calculated after a consistency check). Preliminary weights are calculated through normalization, and finally, a consistency check (CR < 0.1) is performed to ensure logical rationality, outputting a subjective weight vector reflecting subjective experience.

[0175] Principal component analysis (PCA) was used to perform statistical analysis on the standardized multidimensional dataset. Principal components were extracted based on variance contributions, and objective weight vectors based on the intrinsic structure of the data were calculated. Specifically:

[0176] The raw data of the nine vulnerability factors (EVI, GDP, NPP, NX, DTR, RTR, RS, DTF, UL) were constructed into a sample matrix and standardized to eliminate the influence of units.

[0177] Calculate the covariance matrix of the standardized data, i.e., the correlation coefficient matrix, and perform eigenvalue decomposition on it to obtain 9 eigenvalues: λ1≥λ2≥...≥λ9, and their corresponding eigenvectors;

[0178] Calculate the variance contribution rate (e.g., PC1 contribution rate 0.316) and cumulative contribution rate of each principal component based on the eigenvalues, and select the top k principal components (the top 5, with a cumulative contribution rate of 0.852) with a cumulative contribution rate exceeding 85%.

[0179] The eigenvectors are multiplied by the square roots of their eigenvalues ​​to construct the loading matrix, which reflects the correlation between the original factors and the principal components.

[0180] The weights of each principal component contribution rate are used to perform a weighted summation of the absolute values ​​of the loadings to obtain the comprehensive score coefficient for each factor.

[0181] Finally, the overall score coefficient is normalized;

[0182] By introducing the principle of minimum relative information entropy from information theory as the optimization criterion, a mathematical model is established to fuse and calculate the subjective and objective weight vectors, and the comprehensive weight vector minF, which has the smallest deviation from both, is obtained:

[0183] ,

[0184] ,

[0185] ,

[0186] in, This indicates the final weight of each evaluation indicator; and Let m represent the weights determined by the AHP (Analog-Hybrid Analysis) method and the PCA (Principal Component Analysis) method, respectively. Let m be the time series, i be the i-th factor, and minF be the minimum value of the comprehensive weight vector under the constraints.

[0187] In a preferred embodiment of the present invention, based on the susceptibility and vulnerability assessment results, a comprehensive hazard assessment and spatial zoning are performed to form a two-dimensional risk map integrating "probability of occurrence - potential loss," supporting differentiated forest fire risk management across multiple seasons and vegetation types. The specific method is as follows:

[0188] By spatial overlaying and hierarchical zoning of geographic information, risk maps with clear spatial orientation and seasonal differences are generated, revealing the distribution of high-risk areas for different vegetation types in different seasons:

[0189] R = S × V,

[0190] in, Wildfire hazard in the target area; Wildfire susceptibility in the target area; Wildfire vulnerability in the target area;

[0191] The continuous R values ​​were divided into five levels in ArcGIS (a geographic information processing platform used for data processing, spatial analysis, model result visualization, etc.) using the natural discontinuity method: very low, low, medium, high, and very high (R is not a fixed global threshold; the five levels are divided using the natural discontinuity method and change dynamically with different vegetation types, seasons, and evaluation objects). The higher the level, the greater the probability of fire occurrence and the more severe the potential losses in the area. The rationality of the classification was verified by frequency ratio.

[0192] This approach represents a leap from simple probability assessment to comprehensive risk management, and provides a direct and scientific basis for subsequent differentiated fire prevention resource allocation, key area monitoring and early warning, and the formulation of seasonal prevention and control strategies, thereby enhancing the accuracy and systematic nature of forest fire risk prevention and control.

[0193] The present invention also provides a forest fire susceptibility assessment system based on the XGBoost-SHAP interpretable model, including a data acquisition unit and a processing unit. The data acquisition unit is used to acquire multi-source data of the target area within a preset historical period.

[0194] The input terminal of the processing unit is electrically connected to the output terminal of the data acquisition unit. The processing unit executes the method described in this invention to obtain the forest fire susceptibility assessment results.

[0195] This invention, through integrated innovation of technical solutions, effectively solves the inherent defects of existing forest fire risk assessment methods in terms of relevance, comprehensiveness, and interpretability. The results produced have higher spatial accuracy, richer risk dimensions, and clearer decision-making guidance value, providing solid technical support for improving the modernization, refinement, and intelligence of regional forest fire prevention and control.

[0196] This invention proposes a refined assessment framework for forest fire risk based on vegetation type differentiation. It models different forest types separately, overcoming the limitations of traditional methods that homogenize heterogeneous forest areas. By systematically integrating fire susceptibility (probability) and vulnerability (consequences), it achieves a paradigm shift from single-probability prediction to comprehensive risk assessment.

[0197] In forest fire risk assessment, the SHAP interpretability framework was systematically applied to solve the "black box" problem of machine learning models and achieve quantitative analysis of the contribution and influence direction of each driving factor.

[0198] The specific embodiments described herein are merely illustrative examples of the present invention. Those skilled in the art can make various modifications or additions to the described embodiments or use similar methods to substitute them, without departing from the technology of the present invention or exceeding the scope defined by the appended claims.

Claims

1. A method for assessing forest fire susceptibility based on an XGBoost-SHAP interpretable model, characterized in that, Includes the following steps: S1. Obtain multi-source data of the target area within a preset historical period, and perform preprocessing to obtain a spatial distribution dataset of vegetation types. S2, based on the spatial distribution dataset of vegetation types, divides forest vegetation into independent categories, including: evergreen broad-leaved forest, deciduous broad-leaved forest, and coniferous forest; S3. For each independent category of forest vegetation, a wildfire susceptibility model based on XGBoost is constructed, and the hyperparameters of the model are tuned. The contribution and direction of each driving factor are quantified using the SHAP (Shapley addition interpretation) interpretability analysis method, and a multi-seasonal susceptibility dynamic map is generated. S4. Construct an assessment model that comprehensively reflects ecological, social and economic vulnerability. Use the AHP-PCA-minimum relative information entropy combined weighting method to determine factor weights, and achieve dynamic assessment of vulnerability through coefficient of variation and trend analysis. S5, based on the results of susceptibility and vulnerability assessments, conducts comprehensive hazard assessments and spatial zoning, forming a two-dimensional risk map that integrates the probability of occurrence and potential loss, supporting differentiated forest fire risk management across multiple seasons and vegetation types.

2. The forest fire susceptibility assessment method based on the XGBoost-SHAP interpretable model according to claim 1, characterized in that, The method for preprocessing multi-source data of the target area within a preset historical period is as follows: The raw fire point product VNP14IMGN was obtained from the satellite platform NASA and FIRMS. Based on the built-in "confidence level" and "fire point type" fields, high-confidence vegetation burning points were selected. The vector boundary of the study area was used for spatial cropping to remove irrelevant and low-quality records. Set a spatial buffer and a time window, aggregate multiple fire points that fall within the spatial buffer, treat them as a single fire event, and take the center point or the earliest recorded point of the point set formed by the aggregated multiple fire points as the representative sample of the event to ensure sample independence and avoid overfitting. The aggregated fire points are spatially overlaid with the high-resolution vegetation type map for analysis. Each fire point is assigned a vegetation type attribute for its corresponding pixel. Specifically: Calculate the spatial relationship between fire points and vegetation type map, extract the vegetation type attribute of each fire point location and assign it to the fire point; Complete the temporal dimension deconstruction and sample set organization: Based on the observation date of the fire points, divide them into corresponding years and climate seasons: spring, summer, autumn, and winter, and output a structured fire point sample database containing spatial coordinates, vegetation type, season, and year labels. The fire point sample database is used as positive samples and combined with negative samples randomly generated in the corresponding vegetation type areas to form a vegetation type spatial distribution dataset.

3. The forest fire susceptibility assessment method based on the XGBoost-SHAP interpretable model according to claim 1, characterized in that, For each independent category of forest vegetation, a wildfire susceptibility model based on XGBoost was constructed, and hyperparameters were tuned using Bayesian optimization. The specific steps are as follows: For each vegetation type, pre-defined driving factors are extracted from the spatial distribution dataset of vegetation types, and multicollinearity diagnosis and screening are performed to ensure the independence of input factors. The gradient boosting method is used to progressively optimize the prediction error. This is achieved by generating a series of weak learners and weighting their predictions to form a strong learner. The following parameters are selected for tuning: n_estimators (number of weak learners), max_depth (maximum tree depth), learning_rate (learning rate), subsample (sample sampling ratio), and colsample_bytree (feature / column sampling ratio). The XGBoost model is hyperparameter tuned using the scikit-optimize framework. Hyperparameters and their ranges are defined. Several points and their corresponding objective function values ​​are obtained through initial sampling. A surrogate model is constructed using a Gaussian process to approximate the objective function. Specifically: Before starting the formal Bayesian optimization iteration, a random search method is used to randomly select several sets of parameter combinations from the hyperparameter space, and the model is trained separately to obtain its performance index, i.e. the objective function, as the data basis for building the initial surrogate model. Using existing hyperparameter combinations and corresponding model accuracy as observation data, the similarity between different parameter groups is calculated through the objective function. The objective function value of any unknown point in the entire hyperparameter space is predicted probabilistically, and its expected value (mean) and uncertainty (variance) are output. The next sampling point is selected by the sampling function, its target value is evaluated and the surrogate model is updated until the optimal value of the objective function is found as the optimal model parameter.

4. The forest fire susceptibility assessment method based on the XGBoost-SHAP interpretable model according to claim 3, characterized in that, The steps for generating a multi-seasonal susceptibility dynamic map by quantifying the contribution and direction of influence of each factor using the SHAP interpretability analysis method are as follows: The trained wildfire susceptibility model is interpreted using TreeExplainer from the SHAP library. The SHAP value of each driving factor for each sample is calculated. The contribution of the factors is ranked by the mean absolute SHAP value. The direction of influence is determined by the positive or negative SHAP value. A SHAP value > 0 indicates that increasing the value of the driving factor will increase the probability of fire, while a SHAP value < 0 indicates that it will decrease the probability of fire. The research data was split into spring, summer, autumn and winter, and then input into the wildfire susceptibility model to obtain continuous probability values. In ArcGIS, the probability values ​​were reclassified into five susceptibility levels using the natural discontinuity method: extremely low susceptibility, low susceptibility, medium susceptibility, high susceptibility and extremely high susceptibility. Four susceptibility spatial distribution maps for spring, summer, autumn and winter were generated, and their rationality was verified by overlaying historical fire points. The spatiotemporal dynamic change patterns of susceptibility areas were analyzed. Introducing the SHAP model explanation tool, SHAP predicted value y pre is: y pre =y base +f(x i1 )+f(x i2 )+....+f(x in ), Where, x ik y is the k-th feature of the i-th sample; pre y represents the model's predicted value for this sample. base This serves as the baseline for the entire model; Using b1-b12 to represent the driving factors of each participating model, we analyze models for different vegetation types to enhance the interpretability and transparency of the models. Specifically: By drawing SHAP contribution bar charts for each vegetation type and comparing the bar lengths of each factor code b1-b12, the core driving factors of evergreen broad-leaved forests were identified.

5. The forest fire susceptibility assessment method based on the XGBoost-SHAP interpretable model according to claim 3, characterized in that, For each vegetation type, predefined driving factors are extracted from the spatial distribution dataset of vegetation types, and multicollinearity diagnosis and screening are performed to ensure the independence of input factors. The method is as follows: Driving factors include climate factors, topographic factors, vegetation factors, and population factors. From the spatial distribution dataset of vegetation types, 13 key driving factors in four categories were extracted: temperature, precipitation, potential evapotranspiration, elevation, slope, topographic humidity index, water flow intensity index, aspect, curvature, soil classification, normalized vegetation index (NDVI), ecological environment quality, and population density. In the ArcGIS platform, spatial registration, clipping, and resampling of all raster data were completed. For the spatial distribution data of each vegetation type, the variance inflation factor (VIF) and tolerance (TOL) of each driving factor were calculated. Multicollinearity diagnosis and factor screening were performed based on the criteria of VIF < 10 and TOL > 0.

1. ， Among them, TOL is an indicator that measures the degree of linear correlation between a certain independent variable and other independent variables, reflecting the degree of independence of the independent variable from the dependent variable; The coefficient of determination for the regression model is denoted by ; the VIF value is the reciprocal of TOL.

6. The forest fire susceptibility assessment method based on the XGBoost-SHAP interpretable model according to claim 5, characterized in that, A comprehensive assessment model reflecting ecological, social, and economic vulnerability is constructed. The AHP-PCA-minimum relative information entropy combined weighting method is used to determine factor weights, and dynamic vulnerability assessment is achieved through coefficient of variation and trend analysis. The specific steps are as follows: Integrate vegetation dynamics data, social vulnerability data, and economic vulnerability data from the spatial distribution dataset of vegetation types; All integrated data are spatially clipped, resampled to a uniform resolution and coordinate system transformation to construct a spatially aligned, temporally consistent, multidimensional standardized dataset. Trend and stability analysis was performed on vegetation dynamics data and economic vulnerability data with time series characteristics in multidimensional standardized datasets, including calculating the coefficient of variation to assess the stability of their temporal fluctuations and performing linear trend analysis to identify their direction of change. The final weights of each driving factor in the forest fire vulnerability assessment were determined by adopting a weight determination strategy of "coordination between subject and object and information fusion". A linear weighted summation model is used as the core method for calculating the vulnerability index. The vulnerability index (B) is: ， in, Due to wildfire vulnerability, This represents the value of the i-th indicator factor, where i is the sequence number of the time parameter. This represents the combined weight of the i-th indicator factor. This represents the total number of years of research.

7. The forest fire susceptibility assessment method based on the XGBoost-SHAP interpretable model according to claim 6, characterized in that, For a given raster cell, the coefficient of variation (CV) of a certain factor in its time series is: ， in, Total number of years of research; It is a time series; These are the variable values ​​at the corresponding time points; The average value of the variable; For each pixel's time-series data, the slope of its trend, slop, is calculated using univariate linear regression: ， When slop>0, it indicates that the factor value corresponding to the pixel is on an upward trend, i.e., an improving trend; when slop<0, it indicates that the factor value is on a downward trend, i.e., a deteriorating trend; when slop≈0, it indicates that there is no obvious trend of change.

8. The forest fire susceptibility assessment method based on the XGBoost-SHAP interpretable model according to claim 6, characterized in that, The final weights of each influencing factor in forest fire vulnerability assessment are determined using a "coordinated approach between subject and object, and information fusion" weighting strategy, as follows: Using the Analytic Hierarchy Process (AHP), a judgment matrix is ​​constructed, and a subjective weight vector reflecting management cognition and regional characteristics is calculated, specifically as follows: A three-tiered structure of objectives, criteria, and options is established: the objective tier evaluates the overall wildfire vulnerability; the criteria tier evaluates the socio-economic aspects of vegetation; and the options tier evaluates the enhanced vegetation index (EVI) and net primary productivity (NPP). Based on management understanding and regional characteristics, a 1-9 scale is used to construct a judgment matrix by comparing and scoring factors at the same tier pairwise. Preliminary weights are calculated through normalization, and finally, a consistency test (CR < 0.1) is performed to ensure logical rationality. The output is a subjective weight vector reflecting subjective experience. Principal component analysis (PCA) was used to perform statistical analysis on the standardized multidimensional dataset. Principal components were extracted based on variance contributions, and objective weight vectors based on the intrinsic structure of the data were calculated. Specifically: The raw data of the nine vulnerability factors (EVI, GDP, NPP, NX, DTR, RTR, RS, DTF, UL) were constructed into a sample matrix and then standardized. Calculate the covariance matrix of the standardized data, i.e., the correlation coefficient matrix, and perform eigenvalue decomposition on it to obtain 9 eigenvalues: λ1≥λ2≥...≥λ9, and their corresponding eigenvectors; Calculate the variance contribution rate and cumulative contribution rate of each principal component based on the eigenvalues, and select the top k principal components with a cumulative contribution rate exceeding 85%. The eigenvectors are multiplied by the square roots of their eigenvalues ​​to construct the loading matrix, which reflects the correlation between the original factors and the principal components. The weights of each principal component contribution rate are used to perform a weighted summation of the absolute values ​​of the loadings to obtain the comprehensive score coefficient for each factor. Finally, the overall score coefficient is normalized; By introducing the principle of minimum relative information entropy from information theory as the optimization criterion, a mathematical model is established to fuse and calculate the subjective and objective weight vectors, and the comprehensive weight vector minF, which has the smallest deviation from both, is obtained: ， ， ， in, This indicates the final weight of each evaluation indicator; and denoted by AHP (Analog-Hybrid Hierarchical Analysis) and PCA (Principal Component Analysis), respectively; m is the time series, i is the i-th factor, and minF is the minimum value of the comprehensive weight vector obtained under constraints.

9. The forest fire susceptibility assessment method based on the XGBoost-SHAP interpretable model according to claim 6, characterized in that, Based on the susceptibility and vulnerability assessment results, a comprehensive hazard assessment and spatial zoning are conducted to form a two-dimensional risk map integrating "probability of occurrence - potential loss," supporting differentiated forest fire risk management across multiple seasons and vegetation types. The specific methods are as follows: By spatial overlaying and hierarchical zoning of geographic information, risk maps with clear spatial orientation and seasonal differences are generated, revealing the distribution of high-risk areas for different vegetation types in different seasons: R = S × V, in, Wildfire hazard in the target area; Wildfire susceptibility in the target area; Wildfire vulnerability in the target area; The continuous R values ​​in the ArcGIS platform (Geographic Information Processing Platform) are divided into five levels using the natural discontinuity method: very low, low, medium, high, and very high. The higher the level, the greater the probability of fire occurrence and the more severe the potential losses in the area. The rationality of the classification is verified by frequency ratio.

10. A forest fire susceptibility assessment system based on an XGBoost-SHAP interpretable model, characterized in that, It includes a data acquisition unit and a processing unit, wherein the data acquisition unit is used to acquire multi-source data of the target area within a preset historical period; The input terminal of the processing unit is connected to the output terminal of the data acquisition unit, and the processing unit executes the method described in any one of claims 1-9 to obtain the forest fire susceptibility assessment results.

Images