Forest fire prediction method different from manual recording and satellite remote sensing data

Abstract

The invention relates to the technical field of forest fire prediction, in particular to a forest fire prediction method different from manual recording and satellite remote sensing data, and the method comprises the following steps: S1, data acquisition and preprocessing, S2, model training and prediction, and S3, prediction precision detection. According to the method, data training is carried out by establishing six data models, related data of forest fires are fully acquired, and two data sources can be used for mutual evidence, so that the defects and errors of a single data source are overcome; the accuracy of the model is fully improved, then the model is distinguished, the most suitable prediction model suitable for different regions or different fire prevention periods and different seasons in the same region can be screened out, the prediction accuracy is higher, pertinent use is facilitated during follow-up prediction, mutual verification among multiple models can be achieved, and the prediction efficiency is improved. And the prediction accuracy is improved.

Description

Technical Field

[0001] The invention relates to the technical field of forest fire prediction, and in particular to a forest fire prediction method that is different from artificial records and satellite remote sensing data. Background Art

[0002] Forest fire is one of the important factors affecting forest ecosystems, directly endangering standing trees, soil, microorganisms and wildlife, playing an important role in the stability, renewal and succession of forest ecosystems, endangering the atmospheric environment and the safety of human life and property, and causing great environmental and economic losses. In recent years, due to global warming and the extension of the fire weather season, extreme fire meteorological conditions have become increasingly common around the world, increasing the flammability of vegetation and the frequency and intensity of fires. The fire situation in many regions has gradually become more severe. A large number of studies have shown that the occurrence of fires is the result of the combined influence of multiple driving factors. Therefore, it is an important aspect of forest fire management and prevention to establish a forest fire model suitable for the study area for prediction, identify the main driving factors affecting the fire and their probability of occurrence, and grasp the fire distribution law and its dynamic changes in the region, which can help effectively implement fire prevention strategies. Due to the limitations of the accumulation and acquisition of long-term historical fire record data, remote sensing data can provide observation information with multi-scale and clear temporal and spatial characteristics. It is currently a key technology for monitoring large-scale fire events and their dynamics and an important source for improving fire information databases.

[0003] Different models have differences in predicting the probability of fire, drawing fire probability maps, and evaluating fire influencing factors. The most suitable models selected in different regions are also different. Therefore, it is of great significance to select a model suitable for the study area for prediction research.

[0004] In the prior art, six commonly used technical models are often used for fire prediction. How to classify such models and fully understand their prediction accuracy, suitable forest areas or forest conditions so as to apply the corresponding models in the corresponding forest areas is the technical problem to be solved by this application. Summary of the invention

[0005] The purpose of the present invention is to solve the shortcomings of the prior art and to propose a forest fire prediction method that is different from artificial records and satellite remote sensing data.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A method for predicting forest fires that is different from manual records and satellite remote sensing data, comprising the following steps:

[0008] S1. Data acquisition and preprocessing: Obtain fire data at the location where the forest fire occurred and data on the surrounding conditions of the forest for model training;

[0009] The fire situation data of the forest fire location needs to obtain two different data: historical fire record data and MODIS remote sensing satellite fire point data;

[0010] Historical fire record data is provided by the Forestry and Grassland Bureau, which contains information such as the time, location, cause, and grade of each fire;

[0011] MODIS remote sensing satellite fire point data are processed by NASA's scientific computing facilities based on MODIS data and published in the Fire Information Resource Management System;

[0012] After the data is acquired, the following operations are performed on the data: first, combined with the land use type data with a resolution of 1km, non-forest fire points such as urban construction land and farmland are removed;

[0013] Second, select fire points with a confidence level of ≥80% to screen and extract valid fire points and their information. Since there may be multiple fire points in a fire, this study uses ArcGIS 10.8 software to remove fire points with a distance of less than 1 km within 24 hours.

[0014] The forest surrounding conditions data include three categories of climate, human factors, vegetation and terrain, totaling 54 factors; the climate factors include meteorological data; the human factors include socio-economic data and infrastructure data; the vegetation and terrain factors include terrain data and vegetation data.

[0015] S2. Model training and prediction: Establish the corresponding prediction model, import the processed data into the model, and determine the model's selection of important driving factors;

[0016] The prediction models include logistic regression model, random forest algorithm, enhanced regression tree model, extreme gradient boosting model, support vector machine and Gompit regression model;

[0017] The data is in binomial distribution form; with the help of ArcGIS 10.8, non-fire points are created according to the ratio of fire point: random point = 1:1.5, following the principle of dual randomness in time and space, and all data are randomly divided into modeling data and test data in a ratio of 7:3 for fitting;

[0018] The random forest algorithm, enhanced regression tree model, extreme gradient boosting model and support vector machine model: the “varSelRF” package in Rstudio4.1.3 is used to perform selection calculation of feature variables of the random forest algorithm on 5 groups of training samples, and the variables that appear 3 times or more in the 5 training samples are selected; the enhanced regression tree model uses the “gbm.simplify” function in Rstudio to test the 5 training samples for abandoning unimportant variables, adjust the parameter model accuracy to the highest level and select the variables that are retained in the 5 subsets and are not abandoned 3 times or less;

[0019] The logistic regression model and Gompit regression model were fitted to 5 groups of training samples in SAS9.4 after multicollinearity diagnosis and significance test. In this study, factors with occurrence times ≥ 3 were selected to fit the full sample data.

[0020] S3. Prediction accuracy detection: Use ROC (Receiver Operating Characteristic Curve) to evaluate the accuracy of the model; ROC calculates a series of sensitivity and specificity values ​​by setting several different critical values ​​for continuous variables, and draws a curve with the false positive rate (the probability of judging a correct result as a wrong result) as the horizontal axis and the sensitivity (the probability of judging a correct result as a wrong result) as the vertical axis. The larger the AUC value, the better the fitting effect of the model.

[0021] Preferably, the meteorological data is downloaded and obtained from the National Tibetan Plateau Science Data Center; the daily average temperature, daily maximum temperature, daily minimum temperature, daily average relative humidity, daily minimum relative humidity, daily average surface temperature, daily maximum surface temperature, daily average wind speed, daily maximum wind speed, sunshine hours, daily average precipitation, and daily average air pressure are selected;

[0022] At the same time, the variable values ​​of each meteorological station at the monthly and annual scales were calculated, and the ArcGI S10.8 layer connection function was used to match the fire points and random points with the nearest meteorological stations. Navicat software was used to extract the corresponding meteorological values ​​based on the time information of the fire points and random points, and finally the meteorological element data of the fire points and non-fire points at the daily, monthly and annual scales were obtained.

[0023] Preferably, the terrain data is derived from a 30m high-resolution digital elevation model provided by a geospatial data cloud platform, and the digital elevation model is calculated and mapped using a 3D analysis tool in ArcGIS software to obtain the factors with the greatest impact on the study area, which are altitude, slope and aspect index;

[0024] The slope index calculation formula is as follows:

[0025] Aspect index=CoS((θ×2PI) / 360

[0026] Preferably, the socio-economic data includes:

[0027] First, the global population grid dataset developed by the Oak Ridge National Laboratory of the U.S. Department of Energy. Its pixel value is the average population per unit area over 24 hours, with a spatial resolution of 1 km.

[0028] Second, night light images can show the brightness distribution of the earth's surface at night, and thus indirectly show the intensity of human activities. Long-term series of night light datasets are of great reference value for forest fire research. The night light data used in this study comes from the China Long-term Series Night Light Dataset of the Global Change Data Warehouse Electronic Magazine.

[0029] Preferably, the infrastructure data is obtained through a 1:250,000 national basic geographic database in the national geographic information resource directory service system, which includes railways, highways, settlements, watchtowers, fire brigades, fire-proof roads, and fire-proof barrier factors closest to the fire point.

[0030] Preferably, the vegetation data generally refers to the percentage of the vertical projection area of ​​vegetation on the ground to the total area of ​​the statistical area, which is an important indicator for measuring the status of surface vegetation and can be used to represent the total amount of live combustibles and dead combustibles above the surface;

[0031] At present, many methods for measuring vegetation coverage using remote sensing have been developed. A more practical method is to use vegetation index to approximate vegetation coverage. The commonly used vegetation index is NDVI; the data is provided by the geospatial data cloud with a spatial resolution of 500m MODIS Normalized Difference Vegetation Index (NDVI);

[0032] The calculation formula is as follows:

[0033] FVC=(NDVI-NDVI sail ) / (NDVI veg -NDVI sail )

[0034] Among them, NDVIsoil is the NDVI value of the area with bare soil or no vegetation cover, and NDVIveg represents the NDVI value of the pixel completely covered by vegetation.

[0035] Preferably, the AUC value is the area under the curve. The larger the area under the curve, the higher the model accuracy. An AUC value of 0.5-0.7 indicates a poor fitting effect; 0.7-0.9 indicates a fair fitting effect, and above 0.9 indicates a very good fitting effect of the model.

[0036] The Youden index (sensitivity + specificity - 1) calculated by ROC sensitivity and specificity is used to obtain the optimal critical value, that is, the optimal threshold (cut-off). This index can be used as a threshold for the forest fire prediction model to determine the probability of forest fire occurrence.

[0037] Preferably, the selection of the important driving factors includes that all variables in the model are significant at the 0.05 level, and the VIF values ​​of the variables are all less than 10.

[0038] The beneficial effects of the present invention are:

[0039] 1. This application can provide more accurate and comprehensive data support for the establishment and training of simulation by fully acquiring data. The prediction based on two data sources (including manual records and satellite remote sensing) overcomes the shortcomings and errors of a single data source. The two data sources can corroborate each other and improve the quality of simulation training;

[0040] 2. Through the setting and training of multiple models, the bias of the model can be distinguished; that is, the forest fire in which area the model is applicable, so as to effectively make predictions based on different situations and improve the accuracy of the prediction;

[0041] 3. By establishing six data models for data training and fully acquiring relevant data on forest fires, the accuracy of the model is fully improved. Then the models are differentiated to determine the optimal model for forest fire prediction in the study area. That is, the most suitable prediction model for different regions or different fire prevention periods and seasons in the same region can be screened out. The prediction accuracy is higher, so that it can be used in a targeted manner in subsequent predictions. It can also achieve mutual verification between multiple models to improve the accuracy of the prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 The present invention provides a flowchart of a forest fire prediction method that is different from manual records and satellite remote sensing data. DETAILED DESCRIPTION

[0043] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0044] Reference Figure 1 , a forest fire prediction method that is different from manual records and satellite remote sensing data, comprising the following steps:

[0045] S1. Data acquisition and preprocessing: Obtain fire data at the location where the forest fire occurred and data on the surrounding conditions of the forest for model training;

[0046] S2. Model training and prediction: Establish the corresponding prediction model, import the processed data into the model, and determine the model's selection of important driving factors;

[0047] S3. Prediction accuracy detection.

[0048] In the present invention, the fire situation data of the forest fire location, the forest surrounding situation data and the preprocessing method required in the above S1 are:

[0049] 1. Fire point data:

[0050] Two different data types are used: historical fire record data and MODIS remote sensing satellite fire point data (MCD14ML); predictions based on two data sources (including manual records and satellite remote sensing) overcome the shortcomings and errors of a single data source, and the two data sources can corroborate each other.

[0051] Historical fire record data is provided by the Forestry and Grassland Bureau and contains information such as the time, location, cause, and level of each fire.

[0052] MCD14ML is processed by NASA's Scientific Computing Facility based on MODIS data and published in the Fire Information Resource Management System (FIRMS); this product uses a contextual algorithm to detect fires with a spatial resolution of 1km pixels. It is a global monthly active fire positioning product, equipped with two satellites, Terra and Aqua. The MODIS sensor detects the geographic location, date and other information of the fire; because the satellite cannot identify the type of fire, this product includes non-forest fires such as urban and marginal fires, farmland fires, etc.

[0053] The MODIS remote sensing satellite fire point data needs to be processed as follows: 1. Combined with the land use type data with a resolution of 1 km, non-forest fire points such as urban construction land and farmland are removed; 2. Fire points with a confidence level ≥ 80% are selected to screen and extract valid fire points and their information. Since there may be multiple fire points in one fire, this study uses ArcGIS10.8 software to eliminate fire points with a distance of less than 1 km within 24 hours.

[0054] 2. Factor data:

[0055] A large number of potential driving factors were added, and a total of 54 factors were selected, which were mainly divided into three categories: climate factors, human factors, vegetation and terrain factors, and were studied as potential influencing factors of forest fires in the study area.

[0056] 1. Meteorological data:

[0057] Meteorological data were obtained from the National Tibetan Plateau Science Data Center. Twelve factors were selected, including daily average temperature, daily maximum temperature, daily minimum temperature, daily average relative humidity, daily minimum relative humidity, daily average surface temperature, daily maximum surface temperature, daily average wind speed, daily maximum wind speed, sunshine hours, daily average precipitation, and daily average air pressure (see Table 1).

[0058] At the same time, the monthly and annual variable values ​​of each meteorological station were calculated. The ArcGIS 10.8 layer connection function was used to match the fire points and random points with the nearest meteorological station. The corresponding meteorological values ​​were extracted based on the time information of the fire points and random points using Navicat software. Finally, the meteorological element data of the fire points and non-fire points were obtained on the daily, monthly and annual scales.

[0059] 2. Terrain data:

[0060] The impact of topography on fire is a combination of multiple topographic factors. Topographic differences have a significant impact on vegetation composition and the spatial distribution of combustibles, thus shaping different fire environments and affecting the occurrence and spread of fires. Therefore, it is considered an important fire environment factor.

[0061] The terrain data used comes from the 30m high-resolution digital elevation model (DEM) provided by the geospatial data cloud platform. The 3D analysis tool in ArcGIS software is used to calculate and draw the digital elevation model (DEM). The factors with the greatest influence in the study area are altitude, slope and aspect (aspect index);

[0062] The slope index calculation formula is as follows:

[0063] Aspect index=cos((θ×2PI) / 360

[0064] 3. Socio-economic data:

[0065] LandScan is a global population grid dataset developed by the Oak Ridge National Laboratory of the U.S. Department of Energy. Its pixel value is the average population per unit area in 24 hours, and the spatial resolution is 1 km.

[0066] Night light images can show the brightness distribution of the earth's surface at night, and thus indirectly show the intensity of human activities. Long-term series of night light datasets are of great reference value for forest fire research. The night light data used in this study comes from the China Long-term Series Night Light Dataset of the Global Change Data Warehouse Electronic Magazine.

[0067] 4. Infrastructure data:

[0068] The infrastructure data were retrieved from the 1:250,000 national basic geographic database in the national geographic information resource directory service system, including factors such as railways, highways, residential areas, watchtowers, fire brigades, fire-prevention roads, and fire barriers closest to the fire point.

[0069] 5. Vegetation data:

[0070] Vegetation coverage usually refers to the percentage of the vertical projection area of ​​vegetation (including leaves, stems, and branches) on the ground to the total area of ​​the statistical area. It is an important indicator for measuring the status of surface vegetation and can be used to represent the total amount of live and dead combustibles above the surface. At present, many methods have been developed to measure vegetation coverage using remote sensing. A more practical method is to use vegetation index to approximate vegetation coverage. The commonly used vegetation index is NDVI. The data is provided by the geospatial data cloud, MODIS Normalized Difference Vegetation Index (NDVI) with a spatial resolution of 500m.

[0071] The calculation formula is as follows:

[0072] FVC=(NDVI-NDVI sail ) / (NDVI veg -NDVI sail )

[0073] Among them, NDVIsoil is the NDVI value of the area with bare soil or no vegetation cover, and NDVIveg represents the NDVI value of the pixel completely covered by vegetation.

[0074] In the present invention, the corresponding prediction model established in S2 includes:

[0075] 1. Logistic regression model:

[0076] Logistic regression model (LR) is the most commonly used forest fire prediction model at home and abroad. It is one of the generalized linear regression models. This model is usually combined with remote sensing technology to predict forest fires. This study uses the LR model to predict the probability of forest fires. Assuming the probability of forest fires (y = 1) is P, and the probability of no forest fires (y = 0) is 1-P, the probability of forest fires and the calculation formulas of each variable are as follows:

[0077]

[0078] After transformation, the probability prediction formula of forest fire occurrence can be obtained as follows:

[0079]

[0080] In the formula: P is the probability of forest fire occurrence, β0 is a constant, m is the number of covariates, β1, β2, βm are the logistic model regression coefficients of the independent variables, and x1, x2, xm are the independent variables that affect the occurrence of forest fire.

[0081] 2. Random Forest Algorithm:

[0082] Random Forest (RF) is a machine learning algorithm based on the idea of ​​ensemble learning. It improves the accuracy and stability of the model by constructing multiple decision trees and combining their prediction results. The core lies in integrating multiple decision trees (DecisionTree), constructing a training set for each decision tree by autonomously extracting samples with replacement. A single decision tree is a weak classifier, and Random Forest combines the classification results of multiple decision trees into a strong classifier. Each tree is independent, and the features in the tree are randomly selected, which can reduce the variance of the model. In this study, whether a forest fire occurs is a simple binary classification problem, that is, classification is performed based on the majority of the voting results of all decision trees.

[0083] 3. Enhanced regression tree model:

[0084] Boosted regression tree (BRT) is a machine learning algorithm based on classification regression tree, and is an integrated method for fitting statistical models. It combines the advantages of regression tree and boosted tree algorithms, and can improve prediction accuracy by generating robust models through random sampling and self-learning. The advantage of this algorithm is that it can calculate the relative importance of variables and capture complex nonlinear relationships. During the calculation process, a certain amount of data is randomly sampled multiple times to analyze the effect of independent variables on dependent variables. The remaining data is used to verify the results, and the generated multiple regression trees are averaged and output. It can calculate the relationship between an independent variable and a dependent variable when other independent variables are taken as the average or remain unchanged, so as to derive the influence of the independent variable on the dependent variable.

[0085] 4. Extreme Gradient Boosting Model:

[0086] eXtremeGradientBoosting (XGBoost) was promoted by Chen and Guestrin in 2016. It is an efficient and scalable distributed gradient boosting decision tree (GradientBoostingDecisionTree, GBDT) machine learning ensemble algorithm; the model has three important advantages: regularized objective function for better generalization; gradient tree boosting for additive training; shrinkage and column downsampling for preventing overfitting.

[0087] 5. Support Vector Machine

[0088] Support vector machine (SVM) is a new method of data mining in recent years. It is a supervised machine learning algorithm. The basic idea is to construct a hyperplane in the sample input space or feature space so that the distance between the hyperplane and the two types of sample sets is maximized, and the data points are divided into two different categories in the best way. It has good generalization ability and adapts to high-dimensional and small sample data. By selecting appropriate kernel functions and parameters, SVM can produce a classification model with high accuracy. It has unique advantages in dealing with problems such as small data volume and nonlinearity, and to a large extent overcomes problems such as the curse of dimensionality and over-learning.

[0089] 6. Gompit regression model (hereinafter referred to as GR):

[0090] Used to describe the relationship between a binary variable and one or more explanatory variables; Gompit regression follows a cumulative positive value distribution (Gompertz distribution), and its cumulative probability function is:

[0091]

[0092] Where: P(Y=1) represents the probability that the corresponding individual chooses 1 for a given x; F(z) is the extreme value distribution function; Y is converted into probability through the extreme value distribution function;

[0093] ln[-ln(1-P)]=z=β0+β1x1+β2x2+...+β m x m

[0094] In the formula: P is the probability of forest fire; m is the number of explanatory variables; β1, β2, ... βm are the correlation coefficients of each explanatory variable; x1, x2, ... xm are the explanatory variables that affect the occurrence of forest fire; z is linearly related to several predictor variables x.

[0095] In the present invention, the data in S2 is processed because the data is required to be in the form of binomial distribution in the process of building the model. Therefore, this application creates a certain proportion of random points (i.e., random points where no fire has occurred). At present, there is no unified standard for the creation ratio of random points at home and abroad. In order to ensure that the data will not be over-dispersed, based on previous applications, ArcGIS 10.8 is used to create non-fire points in accordance with the ratio of fire point: random point = 1:1.5, following the dual random principle of time and space, and all data are randomly divided into modeling data and test data in a ratio of 7:3 for fitting.

[0096] In the present invention, the important driving factors of the judgment model in S2 are because different fire influencing factors have different degrees of influence on the model performance; RF, BRT, XGBoost and SVM models: the "varSelRF" package in Rstudio4.1.3 is used to select and calculate the characteristic variables of the random forest algorithm for 5 groups of training samples, and the variables that appear 3 times or more in the 5 training samples are selected; the BRT model uses the "gbm.simplify" function in Rstudio to abandon the unimportant variable test for 5 training samples, adjust the parameter model accuracy to achieve the highest and select the variables that have not been abandoned 3 times or less in the 5 subsets; LR and GR models: after using multicollinearity diagnosis and significance test, 5 groups of training samples are fitted in SAS9.4, and this application selects factors with a number of occurrences ≥3 times to enter the fitting of the full sample data; all variables included in the final model are significant at the 0.05 level, and the VIF values ​​of the variables are all less than 10.

[0097] In the present invention, the prediction accuracy detection of S3 is because after the forest fire prediction model is established, evaluating the adaptability and accuracy of the prediction model is an important criterion for testing the model fitting effect; the present application uses the receiver operating characteristic curve (ROC) to evaluate the model accuracy; ROC calculates a series of sensitivity and specificity values ​​by setting several different critical values ​​for continuous variables, and draws a curve with the false positive rate (the probability of judging an actual wrong as correct) as the horizontal coordinate and the sensitivity (the probability of judging an actual correct as wrong) as the vertical coordinate. The larger the area under the curve, the higher the model accuracy; the larger the AUC value, the better the fitting effect of the model; when the AUC is 0.5-0.7, it means the fitting effect is poor, and 0 .7-0.9 indicates that the fitting effect is general, and above 0.9 indicates that the fitting effect of the model is very good; and the Youden index (sensitivity + specificity - 1) calculated by the sensitivity and specificity of ROC is used to obtain the optimal critical value, that is, the optimal threshold (cut-off). This index can be used as the threshold for the forest fire prediction model to determine the probability of forest fire occurrence; by establishing six data models for data training, and fully obtaining relevant data on forest fires, the accuracy of the model is fully improved, and then the models are distinguished to determine the optimal model for forest fire prediction in the study area, that is, the most suitable prediction model for different regions or different fire prevention periods and different seasons in the same region can be screened out, with a higher prediction accuracy, so that it can be used in a targeted manner in subsequent predictions, and it can also achieve mutual verification between multiple models to improve the accuracy of predictions.

[0098] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A forest fire prediction method different from artificial records and satellite remote sensing data, characterized in that: The following steps are involved: S1. Data acquisition and preprocessing: Obtain fire data at the location where the forest fire occurred and data on the surrounding conditions of the forest for model training; The fire situation data of the forest fire location needs to obtain two different data: historical fire record data and MODIS remote sensing satellite fire point data; Historical fire record data is provided by the Forestry and Grassland Bureau, which contains information such as the time, location, cause, and grade of each fire; MODIS remote sensing satellite fire point data are processed by NASA's scientific computing facilities based on MODIS data and published in the Fire Information Resource Management System; After the data is acquired, the following operations are performed on the data: first, combined with the land use type data with a resolution of 1km, non-forest fire points such as urban construction land and farmland are removed; Second, select fire points with a confidence level of ≥80% to screen and extract valid fire points and their information. Since there may be multiple fire points in a fire, this study uses ArcGIS 10.8 software to remove fire points with a distance of less than 1 km within 24 hours. The forest surrounding conditions data include three categories of climate, human factors, vegetation and terrain, totaling 54 factors; the climate factors include meteorological data; the human factors include socio-economic data and infrastructure data; the vegetation and terrain factors include terrain data and vegetation data. S2. Model training and prediction: Establish the corresponding prediction model, import the processed data into the model, and determine the model's selection of important driving factors; The prediction models include logistic regression model, random forest algorithm, enhanced regression tree model, extreme gradient boosting model, support vector machine and Gompit regression model; The data is in binomial distribution form; with the help of ArcGIS 10.8, non-fire points are created according to the ratio of fire point: random point = 1:1.5, following the principle of dual randomness in time and space, and all data are randomly divided into modeling data and test data in a ratio of 7:3 for fitting; The random forest algorithm, enhanced regression tree model, extreme gradient boosting model and support vector machine model: the "varSelRF" package in Rstudio4.1.3 was used to perform selection calculation of feature variables of the random forest algorithm on 5 groups of training samples, and variables that appeared 3 times or more in the 5 training samples were selected; the enhanced regression tree model used the "gbm.simplify" function in Rstudio to test the 5 training samples for abandoning unimportant variables, adjusted the parameter model accuracy to achieve the highest and selected the variables that were not abandoned 3 times or less in the 5 subsets; The logistic regression model and Gompit regression model were fitted to 5 groups of training samples in SAS9.4 after multicollinearity diagnosis and significance test. In this study, factors with occurrence times ≥ 3 were selected to fit the full sample data. S3. Prediction accuracy detection: Use ROC (Receiver Operating Characteristic Curve) to evaluate model accuracy; ROC calculates a series of sensitivity and specificity values ​​by setting several different critical values ​​for continuous variables, and draws a curve with the false positive rate (the probability of judging a correct result as wrong) as the horizontal axis and the sensitivity (the probability of judging a correct result as wrong) as the vertical axis. The larger the AUC value, the better the fitting effect of the model.

2. A forest fire prediction method according to claim 1 that is different from artificial records and satellite remote sensing data, characterized in that: The meteorological data are obtained by downloading from the National Tibetan Plateau Science Data Center; the daily average temperature, daily maximum temperature, daily minimum temperature, daily average relative humidity, daily minimum relative humidity, daily average surface temperature, daily maximum surface temperature, daily average wind speed, daily maximum wind speed, sunshine hours, daily average precipitation, and daily average air pressure are selected; At the same time, the variable values ​​of each meteorological station at the monthly and annual scales were calculated, and the ArcGIS10.8 layer connection function was used to match the fire points and random points with the nearest meteorological stations. Navicat software was used to extract the corresponding meteorological values ​​based on the time information of the fire points and random points, and finally the daily, monthly and annual scale meteorological element data of the fire points and non-fire points were obtained.

3. The method for predicting forest fires according to claim 1, which is different from artificial records and satellite remote sensing data, is characterized in that: The terrain data is derived from the 30m high-resolution digital elevation model provided by the geospatial data cloud platform. The digital elevation model is calculated and drawn using the 3D analysis tool in ArcGIS software. The factors with the greatest influence in the study area are altitude, slope and aspect index. The slope index calculation formula is as follows: Aspect index=cos((θ×2PI) / 360.

4. The method for predicting forest fires according to claim 1, which is different from artificial records and satellite remote sensing data, is characterized in that: The socio-economic data include: First, the global population grid dataset developed by the Oak Ridge National Laboratory of the U.S. Department of Energy. Its pixel value is the average population per unit area over 24 hours, with a spatial resolution of 1 km. Second, night light images can show the brightness distribution of the earth's surface at night, and thus indirectly show the intensity of human activities. Long-term series of night light datasets are of great reference value for forest fire research. The night light data used in this study comes from the China Long-term Series Night Light Dataset of the Global Change Data Warehouse Electronic Magazine.

5. The method for predicting forest fires according to claim 1, which is different from artificial records and satellite remote sensing data, is characterized in that: The infrastructure data is obtained through the 1:250,000 national basic geographic database in the national geographic information resource directory service system, which includes the railways, roads, settlements, watchtowers, fire brigades, fire-proof roads, and fire-proof barriers closest to the fire point.

6. The method for predicting forest fires according to claim 1, which is different from artificial records and satellite remote sensing data, is characterized in that: The vegetation data usually refers to the percentage of the vertical projection area of ​​vegetation on the ground to the total area of ​​the statistical area. It is an important indicator for measuring the status of surface vegetation and can be used to indicate the total amount of live and dead combustibles above the surface. At present, many methods for measuring vegetation coverage using remote sensing have been developed. A more practical method is to use vegetation index to approximate vegetation coverage. The commonly used vegetation index is NDVI; the data is provided by the geospatial data cloud with a spatial resolution of 500m MODIS Normalized Difference Vegetation Index (NDVI); The calculation formula is as follows: FVC=(NDVI-NDVI sail ) / (NDVI veg -NDVI sail ) Among them, NDVIsoil is the NDVI value of the area with bare soil or no vegetation cover, and NDVIveg represents the NDVI value of the pixel completely covered by vegetation.

7. The method for predicting forest fires according to claim 1, which is different from artificial records and satellite remote sensing data, is characterized in that: The AUC value is the area under the curve. The larger the area under the curve, the higher the model accuracy. An AUC value of 0.5-0.7 indicates a poor fitting effect; 0.7-0.9 indicates a fair fitting effect, and 0.9 or above indicates a very good fitting effect of the model. The Youden index (sensitivity + specificity - 1) calculated by ROC sensitivity and specificity is used to obtain the optimal critical value, that is, the optimal threshold (cut-off). This index can be used as a threshold for the forest fire prediction model to determine the probability of forest fire occurrence.

8. The method for predicting forest fires according to claim 1, which is different from artificial records and satellite remote sensing data, is characterized in that: The selection of the important driving factors includes that all variables in the model are significant at the 0.05 level, and the VIF values ​​of the variables are all less than 10.

Images