A method for evaluating seasonal wild fire proneness

Abstract

The present application relates to the technical field of prediction, in particular to a seasonal wild fire susceptibility evaluation method, first, historical wild fire samples are collected and screened, then a CF wild fire susceptibility model of the whole season is constructed by using a CF algorithm, based on which a negative sample selection area is demarcated and positive and negative samples are integrated; secondly, the positive and negative samples in different periods are divided into training set and test set, an annual and different seasonal wild fire susceptibility prediction model is constructed based on an XGBoost model, and the rationality and prediction performance thereof are compared and analyzed by using multiple evaluation indexes; finally, the SHAP global interpretability method is used to analyze the difference in the influence degree of factors in the susceptibility model of different seasons, the probability of wild fire occurrence in different seasons and possible reasons are understood, so that corresponding preventive measures are taken in advance, so as to reduce the occurrence rate and harm of wild fire, and overcome the problems that the existing machine learning method has poor model interpretability and it is difficult to determine the dominant factors when predicting the wild fire susceptibility of each season.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of prediction technology, and in particular to a seasonal wild fire susceptibility evaluation method. BACKGROUND

[0002] Wild fire disaster refers to a phenomenon that, in natural environments such as forest, grassland and desert, due to the comprehensive action of various factors, fire breaks out and spreads in a certain range. Once wild fire breaks out, due to its fierce fire, fast spreading speed and wide range, it not only brings serious threat to people's life and property safety, but also leads to a large amount of vegetation being burned, affects the balance of ecological system, even causes natural disasters such as mudslides, and has adverse effects on soil, water and air quality, and further affects people's health and quality of life. Related researches have shown that the frequency of wild fire disasters in different seasons is different. In southern China, wild fire mainly occurs in spring and winter, while the frequency of wild fire disasters in summer and autumn is relatively low. Understanding the seasonal regularity of wild fire can help the present application to predict the occurrence season and location of wild fire, and help to study the relationship between wild fire and environmental factors such as climate, vegetation and other factors, so as to better understand the formation mechanism and development trend of seasonal wild fire.

[0003] In recent years, machine learning methods have been proven to be effective methods for natural disaster assessment. The wide use of machine learning methods in disaster susceptibility is based on three factors: (1) it can be used to train and build disaster susceptibility prediction models based on historical disaster data, for predicting future disaster events and providing guidance for emergency rescue. (2) it can be combined with technologies such as geographic information system (GIS) to conduct spatial analysis and modeling of disaster-prone areas to assess disaster risk and develop disaster prevention measures. (3) it can extract factors that have important influence on disaster susceptibility, such as meteorological factors, topographical factors, human activities, etc., so as to help the present application to better understand the mechanism and regularity of disaster occurrence. Due to the complexity of machine learning "black box" models, it is often difficult to explain the decision-making process and reasons behind the model, so it is difficult to make people believe the reliability of the model. This will affect the application range and credibility of the model. Therefore, it is very important and necessary to build a reasonable and high-performance seasonal wild fire susceptibility model and analyze the occurrence regularity of wild fire susceptibility in different seasons, so as to develop accurate and effective prevention and response measures. SUMMARY

[0004] The present application aims to provide a seasonal wild fire susceptibility evaluation method, which uses CF and XGBoost models to build a reasonable and high-performance wild fire susceptibility model, and solves the problems of poor model interpretability and difficulty in determining the dominant factors when using existing machine learning methods to predict wild fire susceptibility in different seasons.

[0005] To achieve the above object, the application provides a seasonal wild fire susceptibility evaluation method, comprising the following steps:

[0006] Collect historical wild fire data in a specific area, and conduct general investigation and verification on wild fire samples through various channels, and screen out unqualified and repeated sample points to optimize the wild fire sample data set;

[0007] According to the geographical and climatic characteristics of the area and the distribution of wild fire samples in the area, the wild fire influence factors are selected from static and dynamic aspects, and the coordinate system and evaluation unit size are unified;

[0008] The wild fire sample data and dynamic influence factors are divided into five periods of the whole year, spring, summer, autumn and winter, and the data matching is conducted in combination with the static influence factor data;

[0009] Based on the distribution of the whole year wild fire samples in the area on the static factor and dynamic factor layer of the whole year period, the CF value corresponding to each factor in different attribute intervals is calculated by using the certainty coefficient algorithm;

[0010] Based on the CF value calculation results of each factor, the sampling range of negative samples is determined, and the same number of negative samples as the wild fire samples in the whole year, spring, summer, autumn and winter five periods are randomly selected, and the data matching is conducted with the original data of the corresponding factor in the period;

[0011] Based on the number of positive and negative samples in different periods, the sample division of training set and test set is conducted, the wild fire susceptibility model is constructed based on the XGBoost algorithm using the training set, and the rationality and prediction performance of the model are verified using the test set;

[0012] Using multiple evaluation indexes, the differences of wild fire susceptibility model evaluation results based on annual and seasonal are analyzed, the advantages of seasonal evaluation in wild fire susceptibility prediction performance are discussed, and the wild fire susceptibility grades in each period are divided and warned;

[0013] Using the SHAP global explainability method, the influence of seasonal factors on wild fire susceptibility in different seasons is discussed, and through analyzing the interaction and dependence relationship between the characteristics of different seasons, the distribution rule of high wild fire susceptibility area in different seasons and the difference of factor action are revealed.

[0014] Optionally, the channels for obtaining historical wild fire data include relevant government departments, mainstream media, remote sensing images and products, and the optimization principles of the wild fire sample data set are good credibility, occurrence in vegetation coverage area and different wild fire sites in each period.

[0015] Optionally, the static influence factor is a non-seasonal factor, and the selection dimensions include soil topography, surface hydrological environment and human activities; the dynamic influence factor is a seasonal factor, and the selection dimensions include meteorological climate and vegetation coverage.

[0016] Optionally, the time range covered by the year is from January to December, the time range covered by spring is from March to May, the time range covered by summer is from June to August, the time range covered by autumn is from September to November, and the time range covered by winter is from December to February.

[0017] Optionally, the reason for using the certainty factor algorithm to calculate the CF value corresponding to each factor in different attribute intervals is that the meanings and dimensions of different factors are greatly different, the certainty factor algorithm can standardize the factor data, and the numerical change of the factor can be converted into a positive relationship with the occurrence of wild fire, reflecting the sensitivity of wild fire to each factor.

[0018] Optionally, based on the calculation results of the CF value of each factor, the process of delimiting the sampling range of the negative sample includes the following steps:

[0019] Replace the original values in different attribute intervals of each factor with the corresponding CF values;

[0020] Weighted superposition calculation is performed on the CF layers of each factor to obtain a CF-based wild fire susceptibility model;

[0021] The region with CF value greater than 0 in the CF wild fire susceptibility model is removed, and a 500m buffer region centered on the historical wild fire is removed to obtain the final negative sample selection range;

[0022] Randomly select the same number of negative samples as the number of wild fire samples in different periods within the negative sample selection range, keep the negative sample spacing at 1000m, ensure that the samples are uniformly distributed in the entire region, and have different environmental properties.

[0023] Optionally, in the process of dividing the training set and the test set, first, the positive and negative sample sets are randomly shuffled, and then 70% of the samples are randomly selected from the shuffled sample sets for model training and construction, and the remaining 30% are used for model performance verification.

[0024] Optionally, the index for evaluating the wild fire susceptibility is divided into a zoning rationality index and a prediction performance index, the zoning rationality index includes the area ratio of high and very high susceptibility areas based on annual and seasonal wild fire susceptibility results, the number of wild fire samples contained in the areas, and the frequency ratio of each susceptibility level; the prediction performance index includes the receiver operating characteristic curve, the fire point detection rate, the false alarm rate, the overall accuracy, the Kappa coefficient and the root mean square error based on the confusion matrix.

[0025] Optionally, the SHAP global explainability method includes the explanation of the influence factor action direction strength and the explanation of the influence factor dependency, for the explanation of the influence factor action direction strength, the influence degree of different features on the wild fire susceptibility prediction result is displayed; the features which have positive influence on the prediction result and the features which have negative influence on the prediction result are understood, and the influence strength of each feature is understood, and then it is determined which factors are most critical for the prediction of wild fire susceptibility; for the explanation of the influence factor dependency, the influence of the interaction relationship between different features on the prediction result is displayed, the contribution degree of the interaction effect between different features on the prediction result is understood, and the correlation between the features how to affect the prediction result is understood, the behavior of the wild fire susceptibility prediction model is understood, so as to improve the prediction accuracy of the model.

[0026] The present application provides a seasonal wild fire susceptibility evaluation method, first, a variety of channels are used to collect historical wild fire samples, and the wild fire samples are screened and optimized to obtain high reliability and rich historical wild fire samples, and the samples are divided into five periods of full season, spring, summer, autumn and winter. According to the regional environmental characteristics and the dispersion of wild fire samples, static and dynamic wild fire influence factors in different periods are selected. Then, the CF algorithm is used to construct a CF wild fire susceptibility model for the whole season, based on which the negative sample selection area is determined, and the same number of negative samples as the positive samples in each period are selected from the negative samples, and the positive and negative samples are integrated. Secondly, the training set and test set are divided according to the number of positive and negative samples in different periods, the XGBoost model is used to construct the wild fire susceptibility prediction model of different seasons and years, and based on the same zoning, the susceptibility level is divided into five susceptibility levels of extremely low, low, medium, high and extremely high. The rationality and prediction performance are compared and analyzed by using multiple evaluation indexes, and the advantages of the seasonal wild fire susceptibility evaluation method are discussed. Finally, the SHAP global explainability method is used to analyze the difference in the influence degree of factors in the susceptibility model of different seasons, understand the probability of wild fire occurrence in different seasons and the possible reasons, so as to take corresponding preventive measures in advance, so as to reduce the occurrence rate and harm of wild fire. BRIEF DESCRIPTION OF DRAWINGS

[0027] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, briefly introduce the drawings needed to be used in the embodiments or prior art description, obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creating labor.

[0028] Figure 1 It is a flowchart of a seasonal wild fire susceptibility evaluation method of the present application.

[0029] Figure 2is a flowchart of selecting negative samples based on the CF method in the present application.

[0030] Figure 3 is a flowchart of constructing wild fire susceptibility models in different periods by using the XGBoost model in the present application.

[0031] Figure 4 is a flowchart of analyzing wild fire susceptibility in different seasons by using the SHAP global explanation method in the present application. DETAILED DESCRIPTION

[0032] Embodiments of the present application are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numbers represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the accompanying drawings are exemplary and are intended to explain the present application, and cannot be understood as a limitation of the present application.

[0033] Referring to Figure 1 , the present application provides a seasonal wild fire susceptibility evaluation method, comprising the following steps:

[0034] S1: Collecting historical wild fire data in a specific area, and conducting general investigation and verification on the wild fire samples through various channels, screening out unqualified and repeated sample points, and optimizing the wild fire sample data set;

[0035] S2: According to the geographical and climatic characteristics of the region, the dispersion of wild fire samples in the region, selecting wild fire influence factors from static and dynamic aspects, and unifying the coordinate system and the size of the evaluation unit;

[0036] S3: Dividing the wild fire sample data and dynamic influence factors into five periods of the whole year, spring, summer, autumn and winter, and jointly matching the data with the static influence factor data;

[0037] S4: Based on the dispersion of the whole year wild fire samples in the region on the static factor and dynamic factor layer of the whole year period, the CF value corresponding to each factor in different attribute intervals is calculated by using the certainty coefficient algorithm;

[0038] S5: Based on the CF value calculation results of each factor, the sampling range of negative samples is determined, and the same number of negative samples as the wild fire samples in the whole year, spring, summer, autumn and winter five periods are randomly selected, and the data is matched with the original data of the corresponding factor;

[0039] S6: Based on the number of positive and negative samples in different periods, the sample division of training set and test set is carried out, the wild fire susceptibility model is constructed by using the training set based on the XGBoost algorithm, and the rationality and prediction performance of the model are verified by using the test set;

[0040] S7: Adopt multiple evaluation indexes to analyze the differences in the evaluation results of the annual and seasonal wildfire susceptibility models, explore the advantages of seasonal evaluation in the performance of wildfire susceptibility prediction, and divide the wildfire susceptibility levels in each period for early warning;

[0041] S8: Use the SHAP global explainability method to explore the influence of seasonal factors on wildfire susceptibility in different seasons, and reveal the distribution patterns and factor action differences of high wildfire susceptibility areas in different seasons by analyzing the interactions and dependencies between the characteristics of different seasons.

[0042] The application will be described in detail in combination with the implementation steps as follows:

[0043] In step S1, after determining the research area, the wildfire sample data in a specific historical period is collected based on multiple channels, and the sample is checked to optimize the sample data set. Specifically:

[0044] S1.1, the historical wildfire data can be obtained through channels such as reports from mainstream media at home and abroad, fire reports and investigation records provided by government agencies, remote sensing image data and related wildfire product datasets, and literature queries at home and abroad.

[0045] S1.2, the most important historical wildfire data source for wildfire susceptibility research is remote sensing wildfire products. Compared with other data sources, it has the advantages of stronger real-time performance, wider data coverage, higher data resolution, and more convenient data processing and analysis. Common remote sensing wildfire products include MODIS fire point and hot spot data, VIIRS fire point data, and Landsat high-resolution remote sensing images.

[0046] S1.3, taking VIIRS fire point data as an example, the steps for optimizing wildfire sample selection are as follows:

[0047] (a) According to its "Type" field, only samples with the attribute "presumed vegetation fire" are retained; according to its "Confidence" field, unqualified wildfire samples with the attribute "Low" are removed.

[0048] (b) Based on high-resolution land cover type data products, only samples in areas with vegetation cover such as farmland, forest land, and grassland are retained.

[0049] (c) Use the "Euclidean distance" analysis tool to remove duplicate samples within the same evaluation unit in the same period.

[0050] In step S2, according to the geographical and climatic characteristics of the region and the distribution of wildfire samples in the region, wildfire impact factors are selected from static (non-seasonal) and dynamic (seasonal) aspects, and the coordinate system and evaluation unit size are unified. Specifically:

[0051] S2.1, the non-seasonal factors are mainly selected from the aspects of soil topography, surface hydrological environment, human activities, etc. Specifically, there are soil type, elevation, slope, slope direction, curvature, terrain wetness index (TWI), stream power index (SPI), distance from water system, distance from road, distance from residential area, population density, surface coverage type, etc.

[0052] S2.2, the seasonal factors are selected from the aspects of meteorological climate, vegetation coverage, etc. Specifically, there are normalized difference vegetation index (NDVI), leaf area index (LAI), aridity index (NDBSI), humidity index (WET), period total rainfall, average wind speed, average air temperature, period total solar radiation, average potential evapotranspiration, etc.

[0053] S2.3, the evaluation unit size refers to the image grid size, and the optimal grid unit size is selected according to the empirical formula, and the calculation formula is as follows:

[0054] G s =7.49+0.0006×S-2.0×10 -9 ×S 2 +2.9×10 -15 ×S 3 (1)

[0055] In the formula: G s is the appropriate grid size, and S is the denominator of the map digital scale. For example, for a scale of 1:50000, it is calculated that the optimal evaluation unit is 30m x 30m grid unit.

[0056] In step S3, first, the historical wild fire samples are divided into different sample sets according to the whole year (1-12 months), spring (3-5 months), summer (6-8 months), autumn (9-11 months) and winter (12-2 months). Correspondingly, the dynamic factors are collected and processed according to the same 5 periods as the standard, and all the static factors are matched to the wild fire samples in the corresponding period.

[0057] In step S4, based on the distribution of the whole year wild fire samples in the static factor layer and the dynamic factor layer of the whole year period in the region, the CF value corresponding to the different attribute intervals of each factor is calculated by using the CF algorithm.

[0058] Specifically, the reason for using the CF algorithm to process the factor data is that the meaning and dimension of different factors have great differences, and the CF algorithm can standardize the factor data, and the numerical change of the factor can be converted into a positive relationship with the occurrence of wild fire, reflecting the sensitivity of wild fire occurrence to each factor. The CF value calculation formula is

[0059]

[0060] where CF ij represents the certainty factor of the occurrence of wildfires in the jth secondary classification interval of the ith factor, and its value range is [-1, 1]. When CF>0, the greater the value, the greater the possibility of the occurrence of wildfires; when CF<0, the smaller the value, the smaller the possibility of the occurrence of wildfires. PP s is the prior probability of the occurrence of wildfires in the entire study area, which is represented by the ratio of the total number of wildfires in the entire study area to the total number of grids in the study area. PP ij is the conditional probability of the occurrence of wildfires in the jth secondary classification interval of the ith adjustment factor, which is usually represented by the ratio of the number of wildfires in the factor classification to the number of grids in the factor classification in the study.

[0061] In step S5, based on the calculation results of the CF values of each factor, the sampling range of negative samples (non-wildfire points) is delineated, and the same number of negative samples as the number of wildfire samples in the whole year, spring, summer, autumn and winter is randomly selected from the sampling range and matched with the original data of the corresponding factor in the period. The process of selecting negative samples based on the CF method is shown in Figure 2 , and specifically:

[0062] S5.1, replace the original values in different attribute intervals of each factor with the corresponding CF values;

[0063] S5.2, calculate the CF of each factor by weighting and superposition to obtain the wildfire-prone area model based on CF, and the factor weight calculation formula based on CF is shown in formula (3), and the weighting and superposition calculation formula is shown in formula (4).

[0064] W i = CF i_max -CF i_min (3)

[0065]

[0066] where W i is the CF weight of the ith factor; CF i_max and CF i_min are the maximum and minimum values of the CF values corresponding to different attribute intervals of the ith factor; CF is the CF value obtained by weighting and superposition of all factors; and n is the number of factors.

[0067] S5.3, taking CF value as 0 as the demarcation line, eliminating the area CF > 0 in the regional CF wild fire prone area model, that is, extracting the area not prone to wild fire disasters in the area based on CF; eliminating the 500 radius buffer area centered on the historical wild fire sample to eliminate the area with similar geographical environment as the historical wild fire sample; finally obtaining the regional negative sample selection area.

[0068] S5.4, in the negative sample selection area, taking 1000m distance as the sampling interval, randomly selecting the same number of modeling negative samples as the number of wild fires in different periods, ensuring that the samples are uniformly distributed in the entire area and have different environmental attributes.

[0069] In step S6, based on the number of positive and negative samples in different periods, the sample division of the training set and the test set is carried out, the XGBoost algorithm is used to construct the wild fire prone area model based on the training set, and the rationality and prediction performance of the model are verified based on the test set. The process of using XGBoost model to construct wild fire prone area model in different periods is as shown in Figure 3 , specifically:

[0070] S6.1, divide the training set and the test set, first randomly shuffle the positive and negative sample set, and randomly select 70% of the samples for model training and construction, and the remaining 30% for model performance verification.

[0071] S6.2, XGBoost is an ensemble learning algorithm based on decision tree, and the prediction formula is as shown in formula (5). XGBoost uses sorting by feature value, supports parallel computing and other techniques to speed up training, and has very high training and prediction efficiency. In addition, the model has good robustness to noise and outliers, and has very high prediction performance, which performs well on various data sets, usually better than other machine learning algorithms. In general, XGBoost has the advantages of high efficiency, robustness, good prediction performance, strong robustness and feature selection, and is very suitable for the construction of wild fire prone area model in different periods.

[0072]

[0073] where y i represents the predicted value of the i-th sample; x i represents the feature vector of the i-th sample; K is the number of trees; f k is the k-th tree; F is the set of available trees.

[0074] S6.3, Adjust the hyperparameters of the XGBoost model, mainly including 'n_estimators': the number of trees in the model, that is, the number of iterations.'max_depth': the maximum depth of each tree, the value of this parameter usually affects the degree of overfitting and underfitting of the model. 'learning_rate': learning rate, which represents the step size of each weight update. A smaller learning rate can make the model more stable, but it requires more iterations.'subsample': the proportion of subsamples used to construct each tree. 'colsample_bytree': the feature sampling ratio of each tree used to construct each tree. 'gamma': controls the minimum loss reduction value of the tree's leaf node splitting. Increasing this parameter can prevent overfitting.'min_child_weight': controls the minimum sample weight and of the tree's leaf node. Reducing this parameter can prevent overfitting.

[0075] S6.4, Based on all the feature maps in the region, use the trained XGBoost model to make predictions and obtain the entire region's wildfire susceptibility prediction model.

[0076] In step S7, multiple evaluation indicators are used to analyze the differences in the evaluation results of the annual and seasonal wildfire susceptibility models, explore the advantages of seasonal evaluation in the performance of wildfire susceptibility prediction, and divide the wildfire susceptibility levels of each period for early warning. Specifically:

[0077] S7.1, The wildfire susceptibility zoning level is divided into five susceptibility levels: very low (0-20%), low (20-40%), medium (40-60%), high (60-80%), and very high (80-100%).

[0078] S7.2, Zoning rationality indicators include the area proportion of high and very high susceptibility areas based on annual and seasonal wildfire susceptibility results, and the number of wildfire samples contained in the area, the frequency ratio of each susceptibility level, and the frequency ratio calculation formula is as follows:

[0079]

[0080] where FR i represents the frequency ratio of the i-th susceptibility level; N i is the number of wildfires in the i-th susceptibility level; S i is the number of grids in the susceptibility interval; N and S represent the number of wildfires and the number of grids in the entire region, respectively.

[0081] S7.3, the performance indicators include the receiver operating characteristic (ROC) curve, the fire point detection rate, the false alarm rate, the overall accuracy, the Kappa coefficient (Kappa'), and the root mean square error (RMSE) based on the confusion matrix, and the calculation formula is as follows:

[0082]

[0083]

[0084]

[0085]

[0086]

[0087] Wherein, TP and FN are the number of positive sample prediction correct and prediction error respectively; TN and FP are the number of negative sample prediction correct and prediction error respectively. N is the number of samples; Y pre_i And Y i Are the predicted value and actual value of the ith sample.

[0088] In step S8, the SHAP global explainability method is used to reveal the distribution rule of the high-risk area of wild fire in different seasons and the difference in the role of factors. The process of using the SHAP global explainability method to analyze the wild fire risk in different seasons is as shown in Figure 4 , and specifically:

[0089] S8.1, for the wild fire risk model in different seasons, the feature analysis and dependency analysis of the SHAP global explainability method are the same, both of which are to explain the contribution degree of the features in the model to the prediction of the model and the interaction relationship between the features. Specifically:

[0090] (1) Feature analysis: through feature analysis, the present application can determine the contribution degree of each feature to the prediction result of the model, and whether there is a difference in the importance of the features in different seasons. For example, in spring, there may be more vegetation growth and rainfall, so the vegetation-related features may have a greater impact on the prediction result of the wild fire risk in spring. In summer, temperature and humidity and the like may be more important.

[0091] (2) Dependency analysis: through dependency analysis, the present application can determine the interaction relationship between each feature and other features, and understand whether there is a difference in the relationship between the features in different seasons. For example, in spring, there may be a greater interaction between vegetation and rainfall, while in summer, there may be a more significant interaction between temperature and humidity.

[0092] S8.2, Global interpretation of the model using SHAP to determine which factors have the greatest impact on the model's prediction results and identify the features that play a dominant role in different seasons. Calculate the SHAP value of each static and dynamic factor for each season to determine which factors have the greatest impact on wild fire susceptibility in different seasons.

[0093] S8.2, Global interpretation of the model using SHAP to determine the direction of the effect of all factors on the prediction results. For example, in summer, wind speed has a negative relationship with the occurrence of wild fires, while in winter, wind speed has a positive relationship with the occurrence of wild fires.

[0094] S8.3, Global interpretation of the model using SHAP to determine the dependence of single factors on wild fire susceptibility prediction results and judge the attribute interval of factors prone to wild fires; determine the influence of the interaction between double factors on wild fire susceptibility prediction results. Specifically:

[0095] (1) First, use the SHAP method to globally interpret the model to understand the overall contribution of each factor to the prediction results of wild fire susceptibility. By calculating the SHAP value of each factor, the dependence between a single factor and the prediction results of wild fire susceptibility can be determined. (2) Second, determine the attribute interval of a single factor to determine which attribute value range is more likely to cause the occurrence of wild fires. By observing the trend of SHAP value changes with attribute values, the attribute interval of factors prone to wild fires can be determined.

[0096] (3) Finally, the present application considers the influence of the interaction between double factors on the prediction results of wild fire susceptibility. By analyzing the SHAP value changes under different factor combinations, it can be identified which factors have interactions and understand the influence of these interactions on the prediction results of wild fire susceptibility.

[0097] At the same time, in order to more intuitively illustrate the execution process of the evaluation method, the following is an auxiliary description of an evaluation example:

[0098] In evaluating wild fire susceptibility, the present application selected 18 evaluation factors, including elevation, normalized difference vegetation index (NDVI), annual average rainfall, and average temperature. First, SHAP analysis was used to determine the influence of each factor on the prediction results of wild fire susceptibility. By sorting the SHAP values of the model output, the present application found that NDVI and annual average rainfall were the most important factors affecting wild fire susceptibility. Second, by analyzing the relationship between wild fire susceptibility and these factors, the NDVI range and threshold interval of annual average rainfall prone to wild fires were determined.

[0099] Then, the influence of the interaction between the two factors on the prediction result of the wild fire susceptibility is further studied. By drawing the interaction graph between the factors and analyzing the corresponding SHAP value, it is revealed that the influence degree of some factors on the wild fire susceptibility may be mutually enhanced or weakened under certain combinations. For example, when the elevation is lower and the NDVI is higher, the wild fire susceptibility may present a higher trend.

[0100] Through the execution process of the above evaluation method, the prediction result of the model can be comprehensively explained, the relationship between the single factor and the factors is revealed, and a deeper understanding and guidance for preventing and managing the wild fire disaster is provided.

[0101] The above disclosure is only one preferred embodiment of the present application, and of course cannot limit the scope of the rights of the present application. Those skilled in the art can understand that all or part of the above-mentioned embodiments are implemented, and equivalent changes made according to the claims of the present application still belong to the scope covered by the present application.

Claims

1. A method of seasonal wildland fire susceptibility assessment, characterized by, The method comprises the following steps: Collect historical wildfire data in a specific area and verify wildfire samples through various channels, screen out unqualified and repeated sample points, and optimize the wildfire sample data set; According to the geographical and climatic characteristics of the region and the distribution of wildfire samples in the region, select wildfire impact factors from static and dynamic aspects, unify the coordinate system and evaluation unit size; Divide the wildfire sample data and dynamic impact factor into five periods: whole year, spring, summer, autumn and winter, and match the data with static impact factor data; Based on the distribution of the whole year wildfire sample in the region and the dynamic factor layer of the whole year period, use the certainty factor algorithm to calculate the CF value corresponding to each factor in different attribute intervals; Based on the CF value calculation results of each factor, the sampling range of negative samples is determined, and the same number of negative samples as the number of wildfire samples in the whole year, spring, summer, autumn and winter periods are randomly selected, and the data are matched with the corresponding period factor original data; Based on the CF value calculation results of each factor, the sampling range of negative samples is determined, which comprises the following steps: Replace the original values in different attribute intervals of each factor with the corresponding CF values; Weighted superposition calculation is performed on the CF layers of each factor to obtain the CF-based wildfire susceptibility model; Remove the regions with CF values greater than 0 in the CF wildfire susceptibility model and remove the 500m buffer area centered on the historical wildfire to obtain the final negative sample selection range; Randomly select negative samples in the negative sample selection range, keep the negative sample spacing at 1000m, and ensure that the samples are evenly distributed in the entire region and have different environmental properties; Based on the number of positive and negative samples in different periods, divide the training set and test set, use the training set to build a wildfire susceptibility model based on the XGBoost algorithm, and use the test set to verify the rationality and prediction performance of the model; Use multiple evaluation indexes to analyze the differences in the evaluation results of the annual and seasonal wildfire susceptibility models, explore the advantages of seasonal evaluation in the prediction performance of wildfire susceptibility, and divide the wildfire susceptibility levels in each period for early warning; Use SHAP global interpretability method to explore the influence of seasonal factors on wildfire susceptibility in different seasons, and analyze the interaction and dependence between different seasonal characteristics to reveal the distribution rule of high wildfire susceptibility area and the difference in factor action in different seasons.

2. The seasonal wildfire susceptibility evaluation method of claim 1, wherein the channels for obtaining historical wildfire data include relevant government departments, mainstream media, remote sensing images and products, and the optimization principles of the wildfire sample data set are good credibility, occurrence in vegetation coverage area and different wildfire locations in each period.

3. The seasonal wildfire susceptibility evaluation method of claim 2, wherein the static impact factor is a non-seasonal factor, and the selection dimension includes soil topography, surface hydrological environment and human activity, and the dynamic impact factor is a seasonal factor, which can be selected from meteorological climate and vegetation coverage dimension. ​ ​ 4. The seasonal wild fire susceptibility evaluation method according to claim 3, wherein the annual time range is 1-12 months, the spring time range is 3-5 months, the summer time range is 6-8 months, the autumn time range is 9-11 months, and the winter time range is 12-2 months.

5. The seasonal wild fire susceptibility evaluation method according to claim 4, wherein the reason for calculating the CF value corresponding to each factor in different attribute intervals by using the certainty factor algorithm is that the meanings and dimensions of different factors are greatly different, the factor data can be standardized by using the certainty factor algorithm, and the numerical value change of the factor can be converted into a positive relationship with the occurrence of wild fire, reflecting the sensitivity of wild fire occurrence to each factor.

6. The seasonal wild fire susceptibility evaluation method according to claim 5, wherein in the process of dividing the training set and the test set, the positive and negative sample sets are first randomly shuffled, and then 70% of the samples are randomly selected from the shuffled sample sets for model training and construction, and the remaining 30% are used for model performance verification.

7. The seasonal wild fire susceptibility evaluation method according to claim 6, wherein the indexes for evaluating wild fire susceptibility are divided into zoning rationality indexes and prediction performance indexes, the zoning rationality indexes include the area proportion of high and very high susceptibility zones based on annual and seasonal wild fire susceptibility results, the number of wild fire samples contained in the zones, and the frequency ratio of each susceptibility level, and the prediction performance indexes include the receiver operating characteristic curve, the fire point detection rate, the false alarm rate, the overall accuracy, the Kappa coefficient and the root mean square error based on the confusion matrix.

8. The seasonal wild fire susceptibility evaluation method according to claim 7, wherein the SHAP global explainability method includes the explanation of the influence factor action direction and strength and the explanation of the influence factor dependency, for the explanation of the influence factor action direction and strength, the influence degree of different features on the wild fire susceptibility prediction result is displayed, and for the explanation of the influence factor dependency, the influence of the interaction relationship between different features on the prediction result is displayed. ​ ​ ​ ​ ​ ​

Images