Prediction And Evaluation of Forest Fire In Yunnan of China Based On Geographically Weighted Logistic Regression Model

13 Establishing an effective forest fire forecasting mechanism is the premise of scientific planning and management of 14 forest fires and forest fire prevention. In recent years, the forest fire prediction mechanism has been one of the key 15 areas of concern for the government forestry management departments and forestry researchers. One of them, is 16 logistic regression ( LR ). It is a relatively frequent prediction probability model used in forest fire prediction and 17 prediction in China and abroad for the past few years. However, with the gradual deepening of research, it is found 18 that the logistic regression model fails to fully consider the spatial non-stationary relationship between forest fires 19 and driving factors, which leads to poor fitting effect and low prediction accuracy of the model. But its extended 20 counterpart, the Geographically weighted logistic regression ( GWLR ) model, takes into account the spatial 21 correlation between model variables, and effectively improves the fitting ability and prediction accuracy of the 22 model. Therefore, this paper compares the ability of the logistic regression model and the geographically weighted 23 logistic regression model in terms of fitting ability and prediction accuracy in order to obtain the ability of the two 24 models to predict forest fires in Yunnan Province. In this paper, the samples were divided into 60% training samples 25 and 40% test samples, and repeated sampling was carried out 5 times for training. Variables that appeared in the 26 training model for 3 or more times were used to construct the final LR and GWLR models. Finally, the models with 27 better fitting ability and higher prediction accuracy were used to classify the fire risks in Yunnan Province. The 28 results show that the geographically weighted logistic regression model is superior to the logistic regression model 29 in terms of fitting effect and accuracy. The geographically weighted logistic regression model is more suitable for 30 the data structure of forest fires in Yunnan Province and has better prediction ability. The AUC value of the 31 geographically weighted logistic regression model is 0.902, and the prediction accuracy is 82.7 %; The AUC value 32 of logistic regression model was 0.891, and the prediction accuracy was 80.1%; Fully considering the spatial 33 heterogeneity among model variables can, to some extent, predict forest fires more accurately. The fitting of the two 34 models shows that the relative humidity, temperature, air pressure, sunshine hours, daily precipitation, wind speed, 35 and other meteorological factors; Vegetation type; terrain factor; Population road network and other human 36 activity factors become the cause of forest fires in Yunnan Province.

Kunming 650500, China 7 4* Center for Geospatial Information Engineering and Technology of Yunnan Province, Kunming 650500, China 8 5* Key Laboratory of Resources and Environmental Remote Sensing for Universities in Yunnan, Kunming 650500, 9 China 10 *Corresponding author(s). E-mail(s): go2happiness@163.com; yjh2ok@163.com 11 12 Abstract 13 Establishing an effective forest fire forecasting mechanism is the premise of scientific planning and management of 14 forest fires and forest fire prevention. In recent years, the forest fire prediction mechanism has been one of the key 15 areas of concern for the government forestry management departments and forestry researchers. One of them, is 16 logistic regression ( LR ). It is a relatively frequent prediction probability model used in forest fire prediction and 17 prediction in China and abroad for the past few years. However, with the gradual deepening of research, it is found 18 that the logistic regression model fails to fully consider the spatial non-stationary relationship between forest fires 19 and driving factors, which leads to poor fitting effect and low prediction accuracy of the model. But its extended 20 counterpart, the Geographically weighted logistic regression ( GWLR ) model, takes into account the spatial 21 correlation between model variables, and effectively improves the fitting ability and prediction accuracy of the 22 model. Therefore, this paper compares the ability of the logistic regression model and the geographically weighted 23 logistic regression model in terms of fitting ability and prediction accuracy in order to obtain the ability of the two 24 models to predict forest fires in Yunnan Province. In this paper, the samples were divided into 60% training samples 25 and 40% test samples, and repeated sampling was carried out 5 times for training. Variables that appeared in the 26 training model for 3 or more times were used to construct the final LR and GWLR models. Finally, the models with 27 better fitting ability and higher prediction accuracy were used to classify the fire risks in Yunnan Province. The 28 results show that the geographically weighted logistic regression model is superior to the logistic regression model 29 in terms of fitting effect and accuracy. The geographically weighted logistic regression model is more suitable for 30 the data structure of forest fires in Yunnan Province and has better prediction ability. The AUC value of the 31 geographically weighted logistic regression model is 0.902, and the prediction accuracy is 82.7 %; The AUC value 32 of logistic regression model was 0.891, and the prediction accuracy was 80.1%; Fully considering the spatial 33 heterogeneity among model variables can, to some extent, predict forest fires more accurately. The fitting of the two 34 models shows that the relative humidity, temperature, air pressure, sunshine hours, daily precipitation, wind speed, 35 and other meteorological factors; Vegetation type; terrain factor; Population density, road network and other human 36 activity factors become the cause of forest fires in Yunnan Province. . Every year, a large area of forests around the world are burned, which not only seriously destroys the balance 45 of the ecosystem, but also consumes a huge human effort and material resources, often even causing fatalities (Wen et al. 46 2019, Xu et al. 2003). The establishment of an effective forest fire prediction system is a key step and an important basis 47 for the establishment of a forest fire defense system, which is of great significance to the protection of forest ecological 48 resources, the maintenance of the balance and sustainable development of the global ecosystem, and the development and 49 protection of human economy (Fischer et al. 2016, NorTh et al. 2015, Zhang et al. 2019. Therefore, in recent years, forest 50 fire prediction has been the focus of many scholars and the forest fire management department of the government at home 51 and abroad. 52 Since the 1920s, western European countries began to study and forecast forest fires, among which Canada, the 53 United States and the former Soviet Union achieved rapid development in the early forest fire prediction and forecast 54 research (Shu &Tian 1997). Earlier papers, based their models on historical forest fire data to establish forest fire 55 prediction model. Poisson regression model and negative binomial regression model were first used in forest fire 56 prediction, and these two models were gradually applied in forest fire research in many countries and regions (Boubeta et  57 al. 2015, Crosby 1954, Cunningham &Martell 1973, Dayananda 1977, Wotton &Martell 2005. China was relatively late 58 to join this endeavor, which was mainly after the founding of New China in the 1950s. In terms of methods and models, 59 it more or less borrowed that of the former Soviet Union and European countries, which it gradually developed. In the 60 subsequent research on forest fires in China, Shu Lifu et al. divided forest fire prediction and prediction into fire risk 61 weather forecast, forest fire occurrence forecast and forest fire behavior forecast, and also divided forest fire prediction 62 and prediction methods into empirical method, mathematical statistics method and physical method (Shu et al. 1998 that the spatial relationship between the dependent variable and the independent variable is stable, that is, the model 86 parameter is a constant in the whole study area, and a parameter is applied to the whole study area. However, the spatial 87 relationship between the occurrence of forest fires and the driving factors is not stable, so the global logistic regression 88 model cannot solve the spatial instability relationship between the two. But geographically weighted regression (GWR) 89 model, which can solve the problem of spatial nonstationarity, is beneficial to reduce the differences between the model 90 and to some extent, improve the accuracy of the model. Based on this, foreign scholars began to expand geographically 91 weighted regression model that is geographically weighted logistic regression model (GWLR) applied in the prediction Logistic regression model applicable to any region? And should the fitting accuracy and prediction accuracy be higher 98 than the traditional global logistic regression model? Based on the forest fire data of Yunnan Province from 2010 to 2020, 99 this study combined the driving factors of forest fire such as topography, meteorology, vegetation and population density. 100 The applicability of global logistic regression model and local geographically weighted logistic regression model to forest 101 fire prediction in Yunnan Province and their respective prediction ability were discussed from the aspects of model fitting 102 ability and prediction accuracy. 103

Data source and processing 126
The main data of this study include Yunnan Province 2010-2020 satellite monitoring forest fire data points, Land 127 use, vegetation type data, population density data, residential area data, road network data and daily meteorological data 128 of Yunnan Province from 2010 to 2020, Yunnan Digital Elevation Model ( DEM ) terrain data. 129

Forest fire data 130
The forest fire data of Yunnan Province comes from the VIIRS375 meters active fire products from NASA, including 131 the latitudes and longitudes of the fire point, the date and time of the fire point, the confidence of the fire point and the 132 type of fire point. Through comprehensive fire point confidence and use of fire point and land use type data overlay 133 analysis of two conditions to screen out the final actual forest fire data, a total of 4021 items are obtained.

134
In the establishment of forest fire probability model in Yunnan Province, a certain proportion of non-fire points are 135 needed to participate in the modeling. According to previous research experience, the proportion of fire points and non-136 fire points is set to 1 : 1, and random creation of non-fire points must follow two rules : non-fire points must fall on the 137 area of forest land use type. Non-fire points must be random in time and space. Finally, 8042 forest fire point data ( sum 138 of fire point and non-fire point ) in Yunnan Province are formed. In this paper, the total number of samples is randomly 139 divided into 60 % ( 4826 forest fire sample data ) training samples to construct LR model and GWLR model and 40 % ( 140 3216 forest fire sample data ) test samples to test the goodness of fit of the two models. At the same time, in order to 141 avoid the influence of random distribution samples on the construction of the model, the samples were randomly divided 142 into five groups and the experiment was repeated five times. Finally, the significant variables that appear at least three 143 times or more in five experiments were selected as the full sample data for model fitting. This study will use spss25.0, 144 spssau and MGWR2.2 software to fit the logistic regression model and the geographically weighted logistic regression 145 model, and use ArcGIS10.2 software to visualize the parameter estimation coefficient and coefficient t test value of the 146 geographically weighted logistic regression model. 147

Meteorological data 148
Meteorological data mainly come from the National Meteorological Science Data Center ( http://data.cma.cn/), ), 149 which includes 125 national meteorological stations in Yunnan Province and 27 surrounding provinces, including daily 150 maximum temperature, daily minimum temperature and daily average temperature. Daily maximum surface temperature, 151 daily minimum surface temperature and daily average surface temperature. Sunshine hours. 24 hours precipitation. Daily 152 maximum pressure, daily minimum pressure, daily average pressure. The data of meteorological factors such as daily 153 maximum wind speed, wind direction of daily maximum wind speed, daily average relative humidity, and minimum 154 relative humidity of Eyre. By creating a Tyson polygon of meteorological station in ArcGIS10.2, the 8042 forest fire 155 sample points were matched with the meteorological station points, and the sample points were corresponding to the 156 particular meteorological station points based on the spatial position. Finally, Python was used to match the sample points 157 with the meteorological data according to the meteorological station points and dates. 158

Topography and vegetation data 159
Terrain data mainly comes from geospatial data cloud ( http://www.gscloud.cn/), ). Terrain data mainly includes 160 digital elevation data, aspect and slope. Mainly through the elevation data obtained to calculate the slope and aspect data. 161 The vegetation data used in the study is the vegetation type data of Yunnan Province, which is derived from the geospatial 162 data cloud. According to the national classification standard, the secondary class of vegetation type data in Yunnan 163 Province is reclassified into the first class by ArcGIS10. 2, mainly including coniferous forest, broadleaf forest, shrub 164 forest, grass, meadow, alpine vegetation, cultivated vegetation, coniferous and broad-leaved mixed forest. 165

The data of humanities 166
The humanistic data in this study mainly include the population density data, road network data, railway network 167 data and residential area data of Yunnan Province from 2010 to 2020. The vector data points of each year ' s fire samples 168 and the corresponding population density raster data are superimposed and analyzed by ArcGIS10.2, and the population 169 density values of each forest fire point are extracted by the point tool of value extraction. The residential area, highway 170 network and railway network data are mainly calculated by using the nearest neighbor analysis tool in ArcGIS10.2 to 171 calculate the nearest distance between each sample point and the residential area, railway and highway. 172

Data normalization 173
Since a wide range of data with different sources is involved in the study, the dimensions and grades used are also 174 different, this will decrease the fitting degree of the model and make it impractical when the model is fitted later. So in 175 order to eliminate the dimension between the data, the level of difference, and the difference between the data level, the 176 numerical problems caused by the data must be normalized. However, the methods used by different data are different, 177 including temperature, air pressure, ground temperature, sunshine hours, precipitation, wind direction, wind speed, and 178 other meteorological factors. Population density, the distance between the fire and the highway, the distance from the 179 railway, the distance from the residential area and other social infrastructure and human data. The vegetation type data 180 and elevation data will be normalized by Formula ( 1 ). The slope data will be normalized by Formula ( 2 ). The relative 181 humidity data will be normalized by Formula ( 3 ). 182 In the formula, Xi represents the value after normalization, x represents the value before normalization, xmax and 184 Xmin represent the maximum and minimum values in this set of data. 185

= sin
(2) 186 In the formula, xθ represents the normalized slope value, while θ represents the slope value.
In the formula, xσ represents the normalized relative humidity, while σ represents the original relative humidity. 189

Multicollinearity test 190
Multicollinearity means that there is a certain degree or high correlation between explanatory variables in the linear 191 regression model, which can lead to the loss of significance of variable significance test and a failure in the model 192 prediction function. Therefore, the main purpose of the multicollinearity test is to determine the relevant driving factor 193 model variables that lead to forest fires and determine the independent variables that eventually enter the model. 194 Therefore, when the formula contains multiple independent variables and the correlation between variables has to be 195 tested, multiple collinearity tests should be conducted on independent variables to exclude the factors with significant 196 collinearity. In this paper, the variance inflation factor ( VIF ) is used to test the multiple collinearity of the driving factors 197 of forest fire occurrence applied in the study. The multiple collinearity test method used in this study is the variance 198 inflation coefficient. The calculation formula is as follows : 199 This coefficient is usually judged using 10 as the critical value. When VIF<10, there is no multicollinearity. When 201 10<=VIF<100, there is strong multicollinearity. When VIF>=100, severe multicollinearity exists (Chang et  Among them， 214 LR model obtained through logical transformation is as follows:： 216 Where, P is the probability of forest fire; β0, β1, β2,... ,βn is the parameter estimation coefficient of the model, which Among them, 231 ) ...

Multicollinearity test results 260
According to the results of multicollinearity diagnosis, VIF of seven meteorological variables, including daily mean 261 pressure, daily minimum pressure, daily mean air temperature, daily minimum air temperature, daily mean surface 262 temperature, daily minimum surface temperature, daily mean wind speed and so on, are all greater than 10, indicating 263 that they all have a collinearity relationship. After removing these eight variables, what remains is average relative 264 humidity, minimum relative humidity, sunshine time, daily highest temperature, daily maximum pressure, precipitation, 265 24 hours a day, the biggest manners, the daily maximum surface temperature, population density, slope, slope to fire 266 recently, altitude, vegetation type, settlement to zero distance, railway, highway to point nearest to the point in recent 267 distance 16 independent variables (

Logistic model fitting results 270
In this paper, the stepwise regression method and the Logistic model are used to fit and calculate the five groups of 271 training samples. After fitting, five different subsets of characteristic variables are obtained ( Nearest distance from road to fire point Nearest distance from railway to fire point Note :√ indicates that the variable appears in the model, and × indicates that the variable is not in the model. 277 Table 2 shows that 11 independent variables, such as daily average relative humidity, daily minimum relative 278 humidity, sunshine hours, daily maximum temperature, daily maximum pressure, daily maximum surface temperature, 279 population density, altitude, the nearest distance between residential area and fire point, the nearest distance between road 280 and fire point, and the nearest distance between railway and fire point, will enter the final full sample model fitting stage 281 ( Table 3 ). Among them, 8 variables, such as daily average relative humidity, daily minimum relative humidity, sunshine 282 hours, daily maximum pressure, altitude, daily maximum surface temperature, the nearest distance between residential 283 area and fire point, and the nearest distance between railway and fire point, appear in 5 sample experiments. The highest 284 temperature appeared four times a day. Population density and the nearest zero distance from road to fire occurred three 285 times. In addition, the five variables of 24 -hour precipitation, maximum wind speed, slope, aspect and vegetation type 286 will not enter the final full-sample fitting data, because they appear less than three times in the five sample experiments, 287 among which 24 -hour precipitation once, maximum wind speed and slope twice, aspect and vegetation type 0 times. 288 It can be seen from Table 3 that five variables, including sunshine duration, daily maximum pressure, daily maximum 290 temperature, the nearest distance from residential area to fire point, and the nearest distance from railway to residential 291 area, are positively correlated with the occurrence of forest fire. In addition, six variables, including average relative 292 humidity, minimum relative humidity, daily maximum surface temperature, population density, altitude, and the nearest 293 distance from highway to fire point, are negatively correlated with the occurrence of forest fire. Finally, using a p value 294 of less that 0.001. it is determined that all these 11 variables have a significant impact on the occurrence of forest fire. 295

GWLR model fitting results 296
Through the application of variance expansion factor VIF test before the daily average relative humidity, daily 297 minimum relative humidity, sunshine hours, daily maximum temperature, daily maximum pressure, 24-hour precipitation, 298 maximum wind, daily maximum surface temperature, population density, slope, aspect, altitude, vegetation type, 299 residential area to the nearest fire distance, road to the nearest zero distance, railway to the nearest fire distance and other 300 16 independent variables ( table 1 ) into the GWLR model fitting. The stability test of the sample assumes that the 301 dependent variable and the independent variable have spatial stability characteristics. After fitting the five groups of 302 training samples, the spatial non-stationarity of the spatial relationship between the dependent variable and the 303 independent variable is tested. The main discriminant principle is that the quartile range of the estimation coefficient of 304 an independent variable is greater than the ± 1 standard deviation range of the estimation coefficient of the independent that the independent variable has significant spatial nonstationarity. The same principle is that if the variable has at least 307 three or more spatial nonstationarity in five training samples, it enters the whole sample data fitting stage of the GWLR 308 model. The test results show that the nearest distance from the residential area to the fire point in all the variables is 309 spatially stable in the five training samples groups. The sunshine hours and the nearest distance from the railway to the 310 fire point are otherwise spatially stable Only one time. The other samples are all spatially non-stationary variables in the 311 five intermediate models ( Table 4 ), which are all included in the full sample fitting data of GWLR. 312 Note : √denotes that the variable is nonstationary in space, × denotes that the variable is stationary in space 314 From the fitting results of the GWLR model parameters ( Table 5 ), it can be seen that among all the variables, the 315 daily maximum temperature, 24 -hour precipitation, vegetation type, slope direction, and the nearest distance from the 316 railway to the fire point show positive or negative correlation changes in the whole region. The three variables, namely, 317 sunshine duration, daily maximum pressure and maximum wind speed were positively correlated with the occurrence of 318 forest fire. Seven variables, including average relative humidity, minimum relative humidity, daily maximum surface 319 temperature, population density, slope, altitude, and the nearest distance from highway to fire point, were negatively 320 correlated with the occurrence of forest fire in the study area. In order to better reflect the changes of variable parameter 321 estimation coefficients in the study area, Kriging interpolation is used to visualize the variable coefficients in space ( 322 Figure 2 ). Figure 2 mainly shows the change of the estimated GWLR model coefficient in local space, indicating that 323 the variables have a certain spatial heterogeneity, and the spatial relationship between the independent variables and the 324 occurrence of forest fire is a complex spatial non-stationary relationship. 325  In addition, in order to better represent the significant spatial heterogeneity estimated GWLR model coefficient in 338 local space,, the t-test value is also spatially visualized through Kriging interpolation ( Fig. 3 ). Fig. 3 mainly shows that 339 the significance of the estimated coefficients of model variables also has a strong spatial heterogeneity in the study area. 340 According to Table 5 and Fig. 3, seven variables, including average relative humidity, minimum relative humidity, daily 341 maximum surface temperature, population density, slope, altitude, and the nearest distance from highway to fire point, 342 were negatively correlated with the occurrence of forest fire in the whole study area and the correlation was statistically 343 significant. Among them, four variables, including altitude, 24 -hour precipitation, daily maximum surface temperature, 344 and daily minimum relative humidity, were significantly negatively correlated with the occurrence of forest fire in the 345 western part of the study area, Dehong Prefecture, Baoshan City, Diqing Prefecture, Lincang City, Nujiang Prefecture, 346 Pu ' er City, Dali Prefecture, Xishuangbanna Prefecture, and Lijiang City. The three variables of sunshine hours, daily 347 maximum pressure and maximum wind speed were positively correlated with the occurrence of forest fire in the whole 348 study area, and they were mainly positively correlated in Wenshan, Qujing, Zhaotong, Honghe and Yuxi. The slope 349 aspect, daily maximum temperature, vegetation type and the nearest distance from railway to fire point are positively and 350 negatively correlated in the study.

Model evaluation 365
The area under the ROC curve ( AUC ) is an important indicator for judging the accuracy and fitting effect of the 366 model. At present, it is widely used in China and abroad. Studies have shown that the AUC range is generally between 367 0.5 and 1.0. When 0.5 < AUC≤0.6, the accuracy is weak, general when 0.6 < AUC≤0.7, moderate when 0.7 < AUC≤0.8, 368 high when 0.8 < AUC≤0.9 and extremely high when AUC > 0.9. While when AUC is 0.5, it indicates that the method 369 has no diagnostic value, that is, it also indicates that the model has low accuracy, poor fitting effect, no practical 370 significance, or is not applicable in the research area. Table 6 shows the fitting effect and prediction accuracy of the 371 logistic regression model and the geographically weighted logistic regression model in the fitting process of five sample 372 groups and full samples.

Forest fire risk probability distribution and fire risk classification in Yunnan Province 395
According to the comparison results of fitting ability and prediction accuracy of the LR model and the GWLR model, 396 and finally based on the prediction results of geographically weighted logistic regression model, the spatial distribution 397 of forest fire probability in Yunnan Province was interpolated and analyzed by using the empirical Bellekin interpolation 398 tool in ArcGIS10.2, and the spatial distribution map of forest fire probability in Yunnan Province was obtained ( Fig. 4a  399 ). According to the default threshold 0.5 of the GWLR model and the optimal cut-off value ( Cut-off ) of the prediction 400 probability of forest fires in Yunnan Province calculated by Haden index, the fire risk classification of Yunnan Province 401 was carried out; low fire risk level area for GWLR model predicted probability value ( P ) < 0.50 , medium fire risk level 402 area for 0.50 ≤ P < 0.660 , and high fire risk level area for P ≥ 0.660 ( Figure 4b ). According to Fig. 4, it can be seen 403 that the spatial distribution of forest fire risk probability and fire risk level in Yunnan Province is very obvious. Among 404 them, the extremely high-risk areas and high-risk areas are mainly distributed in Honghe Prefecture, Wenshan Prefecture 405 and Lijiang Prefecture. Other southern and northwestern in Yunnan Province, Nujiang and Dali prefectures are also 406 scattered. Secondly, middle-risk areas are mainly distributed in the south, northwest and central Yunnan. Areas with low 407 and extremely low risk levels are mainly distributed in Diqing Prefecture, Zhaotong City, Baoshan City and Chuxiong 408 City. From the overall layout of Yunnan Province, the probability of forest fires in Yunnan Province is mainly in the 409 southeast, south and northwest of the disaster, and the probability of occurrence in the northwest and central regions is 410 low, and the fire risk level is low.  4 .Discussion 415 In this paper, the traditional global logistic regression model and the local geographically weighted logistic 416 regression model are used to analyze the forest fire data of Yunnan Province from 2010 to 2020. In the comparative 417 analysis of the fitting ability and prediction accuracy of the two models, it is found that the meteorological factors such 418 as daily average relative humidity, daily maximum pressure, daily maximum temperature, sunshine duration, daily 419 maximum surface temperature, and the factors such as altitude, population density, distance from highway to fire point, 420 and distance from railway to fire point are the main explanatory variables of the two models, which also shows that these 421 driving factors constitute the main contributors to forest fire in Yunnan Province. In addition, through research, it is 422 found that the traditional global model assumes that the explanatory variables of the model are stable in space, while the 423 local geographically weighted logistic regression model considers the spatial non-stationary relationship of the model 424 variables, that is, the explanatory variables of each fire have corresponding parameter coefficients, rather than using the 425 parameter value from the global model applied to all prediction for different regions. Therefore, to a certain extent, the 426 local model should have better prediction accuracy and model fitting rate than the global model. The results of this study 427 show that the local model considering the spatial relationship between forest fire drivers does predict the occurrence of 428 fires more accurately. Compared with the global logistic regression model, the local geographically weighted logistic 429 regression model can explain the spatial relationship of model variables better, and has higher accuracy and fitting ability. 430 Therefore, because similar environmental variables may have different contributions to forest fires in different regions, it 431 is necessary to consider the complexity and heterogeneity of the regional space in future studies on forest fires and fire 432 risk assessment. 433 By comparison, it is found that the difference in the prediction accuracy and fitting ability between the global model 434 and the local model in Yunnan is not large. Since the geographically weighted extended model geographically weighted 435 logistic regression model is affected by the selection of kernel function and bandwidth in the process of fitting, in our 436 paper, the adaptive double square kernel function and bandwidth used in the geographically weighted logistic regression 437 model according to the previous fitting of geographically weighted logistic regression model are applied. However, it is 438 well known that the kernel function and bandwidth have a significant impact on the fitting and prediction of the model 439 and the spatial distribution of the model parameter estimation coefficient. Therefore, one of the limitations in this paper 440 is choosing the best kernel function and bandwidth of the geographically weighted logistic model. 441 According to the fire risk zoning map of Yunnan Province, the forest fires in Yunnan Province are mainly distributed 442 in Wenshan Prefecture and Xishuangbanna Prefecture in the south and southeast. Lijiang and Nujiang prefectures in the 443