Spatial pattern prediction of forest wildfire susceptibility in Central Yunnan Province, China based on multivariate data

Wildfires are an important disturbance factor in forest ecosystems. Assessing the probability of forest wildfires can assist in forest wildfire prevention, control, and supervision. The logistic regression model is widely used to forecast the probability, spatial patterns, and drivers of forest wildfires. This study used logistic regression to establish a spatial prediction model for forest wildfire susceptibility, which was applied to evaluate the risk of forest wildfires in Central Yunnan Province (CYP), China. A forest wildfire risk classification was implemented for CYP using forest burn scar data for 2001 to 2020 and the logistic spatial prediction model for forest wildfire susceptibility. Climate, vegetation, topographical, human activities, and location were selected as forest wildfire prediction variables. The results showed that: (1) The distributions of temperature, vegetation coverage, distance to water bodies, distance to roads, and precipitation were positively correlated with the occurrence of forest wildfires. Elevation, relative humidity, the global vegetation moisture index, wind speed, slope, latitude, and distance to residential areas were negatively correlated with the occurrence of forest wildfires. (2) The results of the logistic spatial prediction model for forest wildfire susceptibility showed a good fit to wildfire data, with an overall simulation probability of 81.6%. The optimal threshold for spatial prediction for forest wildfire susceptibility in CYP was determined to be 0.414. A significance level of a selected model variable of < 0.05 resulted in an area under the receiver operating characteristic curve (AUC) of 0.882–0.890. (3) Forest wildfire prevention efforts should focus on Southwest Yuxi City and southern Qujing City accounted for a high proportion of the areas at high risk of forest wildfires. Other localities should adjust forest wildfire prevention measures according to local conditions and strengthen existing wildfire prevention and emergency resource planning and allocation. (4) Some factors contributing to forest wildfires where different among the different areas. Forest wildfire risk factors had different degrees of impact under different spatial and temporal scales. The spatial relationships between wildfire disasters and influencing factors should be established in areas with heterogeneous environmental conditions for the selection of relevant factors.


Introduction
As an important part of the terrestrial ecosystem, forests play a vital role in the dynamic balance of natural landscapes. However, forest ecosystems are extremely vulnerable to interference from various factors, among which fire is one of the most important (Daubenmire 1968;Vogl 1974). Forest fires strongly affect the vegetation and environment of an ecosystem (Flannigan et al. 2000), therefore, the scientific study and evaluation of fire risk play an important role in fire monitoring, prevention, and control, as well as in developing emergency resource planning and allocation measures. The assessment of fire risk involves the generation of a qualitative or quantitative index of risk through the analysis of environmental factors that may affect the occurrence and spread of fire. These estimates of risk are then categorized into different grades to describe changes to fire risk in an area (Chuvieco et al 2010). The assessment of the grade of forest fire risk is an important basis for forest wildfire prevention.
Fire risk assessment and early warning methods began in the late 1920s (Li 2016a), which have been continuously improved with the development of science and technology. The rise of remote sensing technology provides broad data sources for forestry fire risk assessment (Roy et al. 2008;Alonso-Canas and Chuvieco 2015;Giglio et al. 2018;Roteta 2019). Also, with the increase in available geographical element data, the factors for constructing a fire risk index are becoming more and more comprehensive, and many factors, such as terrain, vegetation, and cultural facilities are also taken into account (Shi 2009;Sun and Zhang 2011;Chas-Amil et al. 2015;Tian et al. 2016;). In some studies, various machine learning methods, such as random forests, deep learning, and neural networks, have been applied to forest fire risk assessment and model building (Cao et al. 2017;Zhang et al. 2019a, b;Sevinca et al. 2020). Although among these methods, machine learning has been shown to provide better model parameterization, it tends to give predicted results without interpretation (Gunning and Aha 2019), and thus, it fails to provide a clear relationship between forest fires and drivers. Jaiswal et al. (2002) classified forest fire hazard areas in Deja, India based on satellite data but considering only a few driving factors. Zhang et al. (2018Zhang et al. ( , 2019a used mathematical statistics and cluster analysis to evaluate the grades of forest fire risk and established fire risk zoning for townships and for Chifeng City in the Bahrain right flag of Inner Mongolia. Similarly, their study considered a limited number of variables and did not consider their mutual influences. Johnston et al. (2020) established a forest fire risk assessment method suitable for Canada based on the possibility of fire occurrence, exposure, and vulnerability and their potential impacts (Woo et al. 2017;You et al. 2017;Molaudzi and Adelabu 2018). Ye et al.(2017) Used weights-of-evidence to explore the spatial distribution of forest wildfire points in Yunnan Province from 2007 to 2013. These previous studies focused more on the direct effects of anthropogenic impacts on fire risk with less focus on the indirect impacts of the ecology.
In summary, studies conducted both in China and globally have made good progress in the spatial prediction model for forest wildfire susceptibility. Forest fire risk assessment at any spatial scale can be analyzed based on short-term or long-term driving factors. However, most of the previous studies only considered climate factors, with a short research period, and used a simple spatial weighted overlay method to classify forest fire risk, resulting in low prediction accuracy. Therefore, in the study, we selected 15 fire risk indicators related to climate, vegetation, terrain, human and location that have a greater impact on wildfire, and expanded the study period to build a forest wildfire sensitivity prediction model. Due to sufficient sunshine, high temperature and dry air, Yunnan Province is prone to forest wildfires during the dry season, which is the province with the largest number of forest fires in China (Su et al. 2015;Yang et al. 2015;He et al. 2017). Central Yunnan Province (Kunming City, Yuxi City, Chuxiong Yi Autonomous Prefecture and Qujing City) contains rich forest resources. Through forest wildfire risk assessment, it is of great significance for forest management in this area. Therefore, the study takes CYP as the study area, collects the forest wildfire data during the wildfire prevention period from 2001 to 2020 (December to June of the next year), constructs the spatial prediction model for forest wildfire susceptibility by using climate, vegetation, terrain, human and location factors, evaluates the driving factors of forest wildfire. Factors that were highly correlated and contributed to model bias were eliminated. Logistic regression was then used to construct a spatial prediction model for forest fire susceptibility. Finally, the model was applied to assessing the long-term probability of forest wildfires in the CYP based on comprehensive factors and forest wildfire risk zoning.

Study area
The CYP incorporates the Kunming, Yuxi, Qujing, and the Chuxiong Yi Autono-mous Prefecture in central Yunnan (Fig. 1), consisting of 42 counties (urban areas) and a total area accounting for 24% of the entire province ). The CYP is located in the low-latitude plateau region, showing the characteristics of monsoon climate. Its annual average temperature is 18.61 °C, the climate is warm, the temperature difference between the four seasons is low, and the mean precipitation is 817.83 mm (Li. 2016b). In addition, the dry season and rainy season in central Yunnan are distinct, and the annual precipitation distribution is uneven. Generally, the rainy season is defined from May to October, and the dry season is defined from November to April (Zheng et al. 2017).
The CYP is rich in forest resources, with a forest coverage rate of nearly 47% (Long et al. 2017). The dominant tree species are coniferous forests composed of flammable tree species, including Pinus, Picea, and Eucalyptus (Chen et al. 2012). The special terrain, climate, diversity of vegetation types, and the complexity of fire causes make forest wildfires frequent in central Yunnan, and wildfires mainly occur in winter and spring from December to May, causing immeasurable losses to society. Therefore, we collected data on climate, humanities, topography, and vegetation, and analyzed the probability of forest wildfires in the study area, hoping to provide a reference.

Data resources and data preprocessing
Forest wildfires can occur extremely rapidly and different factors affect the occurrence and spatial distribution of forest wildfires (Tian et al. 2014;Bisquert et al. 2014). The present study selected variables falling in the categories of topography, vegetation, climatology, human activities, and location as wildfire risk factors. These factors were combined with forest burn scar data for the analysis of the correlation between impact factors and the occurrence of forest wildfires. The independent variable data is normalized, and the aspect and vegetation type are used as classification variables. Resampling the resolution of all data to 500 m, unifying all factors into Lambert projection, and using statistical tests to eliminate multicollinearity between selected predictors. Table 1 summarizes the data used in the current study.

Burn scar and random point data
The present study obtained the MCD64A1 data from the official NASA website (nasa.gov) (Giglio et al. 2018). The MRT (MODIS Reprojection Tool) was used to preprocess the fire area data of Yunnan Central from 2001 to 2020, including batch projection conversion and mosaic, cropping. The maximum value was then synthesized and a binary dependent variable (0 is non-fire, 1 is fire) grid map was generated. Finally, Overlay the land cover data with a resolution of 30 meters in 2015, from the Resources and Environment Data Center of the Chinese Academy of Sciences on the fire point data layer from 2001 to 2020, and eliminate the fire data in nonvegetation areas to obtain effective burn scar data (Fig. 2a).
The classification of forest fires requires the construction of a certain proportion of nonburn scars. The number and spatial distribution of non-fire areas are important for the construction and evaluation of forest fire index models. The present study generated non-burn scar data by creating random points under the data management module of ArcGIS software according to the ratio of burn scars and non-burn scars 1:1.5 (Fig. 2b).
The present study statistically analyzed the burned area in central Yunnan (Fig. 3). The total fire area monitored from 2001 to 2020 was 3191.25km 2 , accounting for 3.25% of the total area in central Yunnan. The monitoring results showed relatively little change in burned areas among the different years, except for spike of 506.25 km 2 in 2010. An area of 777.70 km 2 has been burned more than once over the past 20 years, and one area of 0.26 km 2 has experienced 14 periods of repeated burning (Fig. 4). The spatial distribution of

Topography
Topography is an important factor affecting the occurrence and spread of forest fires by influencing the type and distribution of vegetation (Rothermel 1991;Maingi and Henry 2007). Past studies have shown that precipitation, temperature, evaporation, and humidity change with changing elevation, thereby changing the probability of forest fires (Guo et al. 2016a, b). The degree of exposure of the land surface to solar radiation is affected by the aspect, resulting in differences in forest fire among different aspects (Hu 2005;Zhang et al. 2014), With consideration of the effect of topographic factors on the risk of forest fires, data were obtained from the geospatial data cloud platform and digital elevation model (DEM) data were analyzed using ArcGIS to extract elevation (DEM) (Fig. 5a), slope (SLOPE) (Fig. 5b), and aspect (ASPECT) (Fig. 5c).

Vegetation
Forest vegetation can affect the risk of forest fire according to forest type, water content, and other properties (Minnich and Bahre 1995;Pew and Larsen 2001), The vegetation factors considered in the present study mainly included fractional vegetation cover (FVC) (Fig. 5d), the global vegetation moisture index ( Fig. 5e) and vegetation type (Fig. 5f). Data for the FVC were extracted using the Normalized Difference Vegetation Index (NDVI) pixel dichotomy model (Liang 2011): In Eq. (1), NDVI max represents the NDVI value of a pixel in a region completely covered by vegetation and NDVI min represents the NDVI value of a pixel in an area covered by bare land or without vegetation. NDVI min and NDVI max are generally assigned the 5% and 95% cumulative probability of NDVI.
The global vegetation moisture index generally refers to the ratio of water content of the vegetation canopy to its dry matter. Three expressions of the index are typically adopted, namely the leaf water content, relative water content, and equivalent water depth (Xu and Zhang 2020). The present study used the Global Vegetation Moisture Index (GVMI) (Ceccato et al. 2002) instead: NIR and SWIR represent the reflectance values of bands 2 and 6 in the MOD09A1 product, respectively.

Climate data
Climatic factors are generally considered to be drivers of forest fires (Turco et al. 2013) and past studies have considered the comprehensive relationships between the risk of forest fire and climatology parameters, such as temperature, precipitation, evaporation, wind speed, and relative humidity, as well as various combinations of these parameters (Sevinca et al. 2020). The present study obtained data for daily average temperature (°C), 24h cumulative precipitation (mm), daily average wind speed (m s −1 ), and daily average relative humidity (%) from the Resource and Environment Science and Data Center. The data were processed using ArcGIS and the spatial distributions of annual average precipitation (PCP) (Fig. 5g), annual average air temperature (AT) (Fig. 5h), annual average wind speed (WS) (Fig. 5i), and annual average relative humidity (RH) (Fig. 5j) were established.

Human activities
Human made infrastructure and some socioeconomic factors have a certain impact on the occurrence of forest fires (Cardille et al. 2001). Forest fires occur more often in urban forest junctions in which human activities are more frequent (Gordon et al. 1996). Therefore, the present study selected the distance from residential areas, roads, and water bodies as the variables of the forest fire prediction model. The Euclidean distance in ArcGIS was applied to generate ROAD (Fig. 5k), FAC (Fig. 5l), and RIVER ( Fig. 5m) with a resolution of 500 m.

Location of fires
An examination of the data for forest fires in Central Yunnan from 2001 to 2020 showed that there was an uneven distribution of fires in Central Yunnan. The fires were mainly concentrated in Yuxi City and southern Qujing City. Clear differences in the area of forest fires were evident at different longitudes and latitudes, particularly in areas at a low latitude. Therefore, the present study selected the longitude and latitude of each fire as the initial reference factor for establishing the model.

Methods
The present study used the forest wildfire dataset during the fire prevention period for the CYP from 2001 to 2020. Fifteen factors were selected from the climatological, vegetation, topography, and socio-economic categories as forest wildfire predictor variables. Fig. 6 shows a flow diagram of the research process followed.

Logistic model
The factors affecting the occurrence of forest fires are multi-dimensional. The occurrence of forest wildfires has complex nonparametric relationships with many environmental factors and there are typically differences in the factors influencing fires in different regions. Therefore, a fire forecasting model should be flexible. The nonlinear fitting ability of the logistic regression (LR) model can be used to simulate and analyze the categorical dependent variables with only two classification results. The model predicted value is between [0, 1] representing the forecast probability (Bar Massada et al. 2013). The use of the LR model for forest fire risk assessment has been widely recognized and is used both in China and abroad (Saefuddin et al. 2012). The logistic regression (LR) model was used in the present study to analyze the occurrence of forest fires in the CYP. The occurrence of forest fire was regarded as the dependent variable in the model. Forest fire occurs and does not occur for y = 1 and y = 0, respectively. The probability of forest fire occurrence (y = 1) was set as P, whereas the probability of forest fire not occurring (y = 0) was set as 1 − P. The ratio of forest fire occurrence to non-occurrence was equal to P/(1 − P) (Chang et al. 2013). Therefore, taking in (P/(1 − P)) as the dependent variable of the model, after logistic transformation, the probability evaluation formula of forest fire occurrence can be obtained as: In Eq. (3), P is the probability of forest fire, b0, b1, b2... bn are the correlation coefficients of the dependent variables in the binomial logistic regression, n is the number of independent variables, x1, x2, and xn are the occurrence values of forest fires for the various influencing factors, such as slope, elevation, vegetation type, daily average precipitation, distance from residential areas, and roads. In logistic regression, P/(1 − P) is the ratio of predicted probability P and is converted into the ratio form. The logarithm (Logit transformation) of the ratio of P, ln (P/(1−P)), is the predicted probability. This transformation removes the lower limit of the model: the range of P before and after transformation is [0, 1] and (− ∞, +∞), respectively. Since this transformation also reenforces the influence of the independent variable on the dependent variable, it is suitable for the fitting of probabilistic prediction models (Wang 2020).

Extraction of training and test samples
The present study adopted the occurrence of forest fire as the dependent variable and fire risk factors as the independent variables. The sampling of certain non-burn and burn areas was required in the modeling. There may be spatial autocorrelation between the selected independent variables, which will affect the fitting accuracy of the model. The current study randomly chose fire locations for the extraction of various forest fire factors. This ensured the robustness of the model and the validity of statistical inference (Dimitris et al. 2018). The data were randomly subsampled into model training and verification datasets (3) P(y = 1) = e (b0+b1x1+b2x2+⋯+bnxn) 1 + e (b0+b1x1+b2x2+⋯+bnxn) Fig. 6 Flow diagram representing the research process followed in the current study. Please refer to Table 1 for definitions of variable codes containing 80% and 20% of the data, respectively. At the same time, the training dataset was further randomly subsampled into model training subsamples and model test subsamples containing 70% and 30% of the original training dataset, respectively. This was performed to reduce the impact of random sampling on the selection of model variables. This process was repeated five times to generate five data subsamples. The logistic model was used for model fitting to generate five intermediate models. The characteristic variables that appeared three or more times in the experimental results were selected for the fitting analysis of the entire sample dataset, following which the optimal model was generated.

Multicollinearity test of fire risk factors
Interactions between independent variables can lead to instability of the linear regression model, resulting in increased deviation of the estimated results and reduced accuracy of model prediction. Therefore, a multicollinearity test should be conducted on the selected model factors when constructing the model. Many methods currently exist to overcome the challenge of multicollinearity. The present study used the variance inflation factor (VIF) diagnostic method (Ma 2019;Wang 2017) to assess the multicollinearity of independent variables in the model. The value of VIF is positively related to the collinearity of the model. In general, a VIF = 5 is standard, whereas a VIF ≥ 10 indicates severe collinearity. Based on this general guide, independent variables with significant collinearity can be gradually eliminated. A 5 ≤ VIF < 10 indicates moderate collinearity and that significant collinearity among independent variables should be resolved, whereas 2 < VIF < 5 indicates mild collinearity which can be ignored (Krebs et al. 2012).

Receiver operating characteristic curve test
The receiver operating characteristic curve (ROC) is a comprehensive indicator that reflects the sensitivity and specificity of continuous variables. The area under the curve (AUC), can be used to analyze the prediction accuracy of the model, with the value of the AUC positively related to the diagnostic accuracy of the model. The AUC ranges from 0.5 to 1, with an AUC > 0.8 indicating good model predictive ability (Deng et al. 2012). The ROC curve inspection method has become widely used in prediction of forest fire risk Li et al. 2021). The present study used the area under the ROC curve (AUC) to analyze and evaluate the accuracy of the forest fire prediction model.

Multicollinearity test
The present study used ArcGIS to merge the attributes of each layer of forest wildfire driving variables in the CYP with those of burning scars and random points. The forest wildfire driving factors were extracted according to the coordinates of burning scars and random points, and SPSS was used to conduct statistical analyses of the integrated data. The variance inflation factor (VIF) in SPSS was used to assess multicollinearity among independent and dependent variables and independent variables with significant collinearity were eliminated.
According to the research results, all variables passed the multiple common linearity test. Therefore, the factors driving forest wildfires that passed the multicollinearity test were used as explanatory variables to establish a forest wildfire model for the CYP.

Significance test
SPSS software was used to test the significance of the independent variables in the five samples. The logistic regression model was applied to the 15 variables that passed the multicollinearity test to fit five training samples. Table 2 shows the results of the fitting analysis. Table 2 shows that among the five intermediate models, daily average wind speed, daily average precipitation, daily average temperature, daily average relative humidity, elevation, slope, fractional vegetation cover, global vegetation moisture index, distance to residential areas, distance to water systems, distance to roads, and latitude all appeared more than three times. The above variables could then be used to perform binomial logistic regression on the full sample dataset to eliminate the aspect, vegetation types, and longitude that do not meet the model requirements.

ROC curve test
MATLAB software was used to construct the ROC curve, and the ROC curve was used to test the goodness of fit of the model and to calculate the optimal critical value for forest Table 2 Results of fitting independent variables in the logistic model. Please refer to Table 1 for definitions of variable codes * Note: " + " and " − "indicate that the factor appeared and did not appear in the model, respectively wildfires in the CYP. Fig. 7 represents the ROC curve of the five sub-samples and fullsample dataset. As shown in Table 3, the area under the ROC curve (AUC) of the five sub-samples and the full sample ranged from 0.882 to 0.890, with the significance level P less than 0.001. The model showed a high prediction accuracy and statistical significance. Therefore, the model could be used to predict the occurrence of forest wildfires in the CYP and the critical value of wildfire occurrence was 0.414.

Establishment of the probability model of forest wildfire occurrence
The extracted training sample data were imported into the SPSS software for binomial classification logistic regression to obtain the parameters of the forest wildfire risk probability model (Table 4).  According to the selection of model variables and the regression coefficients of the full sample model, the probability model of forest wildfire occurrence in CYP was established as:

Assessment of model accuracy
The probability of forest wildfires in the CYP was categorized into five levels with equal intervals. This categorization was used to assess the percentages of the burn scars of the model falling into different forest wildfire occurrence probability intervals. Fig. 8 shows a histogram of the burn scars falling into the forest wildfire occurrence probability intervals in the CYP. As shown in Fig. 8, most of the burn scars used for inspection were concentrated between 0.4 and 1, with only a few burn scars occurring in the intervals of 0-0.4, indicating that the model has a high degree of credibility.

Forest wildfire risk grade and spatial analysis in the CYP
The Kriging interpolation method in the ArcGIS software was used to interpolate the probability of forest wildfire occurrence in CYP according to the prediction results of the  logistic model. As shown in Fig. 9, the probability of wildfire was divided into five intervals: (1) extremely low; (2) low; (3) medium; (4) high; (5) extremely high. The generated data for wildfire risk grades were used to estimate the wildfire risk (Table 5) and the area percentage of all levels of fire risk in the four cities of the CYP (Table 6). As shown in Table 5, the areas of fire risk grades 1 to 5 in the CYP were 48,344.50 km 2 , 20,400.00 km 2 , 13,102.75 km 2 , 7,531.25 km 2 , and 8,645.5 km 2 , respectively. Among these areas, the total areas of medium, high, and extremely high fire risk accounted for 29.87% of the entire area.
The results showed a clear risk of forest wildfire in CYP. Areas of extremely high and high wildfire risk were mainly distributed in southwestern Yuxi City and southern Qujing City. Accounting for 26.60% and 42.67% of the total area, respectively, whereas that of Qujing City reached 11.23%. Areas with high fire risk were evident in the southern and central parts of the Chuxiong Yi Autonomous Prefecture. Areas of medium fire risk were mainly distributed in the Chuxiong Yi Autonomous Prefecture, northern Yuxi City, southern Kunming City and central Qujing City. Areas of low and extremely low fire risk were distributed mainly in Kunming, Chuxiong Yi Autonomous Prefecture and the central and northern part of Qujing, 94.23%, 73.23%, and 81.69% of each region, respectively, but at a lower proportion in Yuxi at only 9.32%.

Discussion
The present study is the first to use the spatial prediction model for forest wildfire susceptibility to conduct a systematic forest wildfire risk assessment in the CYP. The CYP is a key forest wildfire risk area in Yunnan Province due to its unique climate, topography, forest types, and human production activities. The diversity and complexity of forest wildfires in the CYP increases the complexity of forest wildfire prevention in central Yunnan. Therefore, the present study considered more comprehensive wildfire risk factors, including climate, topography, vegetation, anthropogenic, and geographic factors.

Model reliability
The present study selected a series of static and dynamic data falling within the categories of vegetation, topography, climatology, human activities, and geography as the independent variables for the logistic forest wildfire risk probability model in the CYP. The assessments of multicollinearity and significance indicted that aspect, vegetation types, and longitude that did not meet the requirements of the model could be 1 3 eliminated. The selected variables and the regression coefficients of the full sample were then used to construct a spatial prediction model for forest wildfire susceptibility for the CYP. When the significance levels of the selected model variables were less than 0.05, the AUC values of the group were high at between 0.882 and 0.890, the AUC value of the full sample reached 0.884, and the prediction accuracy of the sample group and the full data sample model were around 81.6%. Past studies that applied the logistic model similarly obtained relatively high AUC values (Deng et al. 2012;Wang et al. 2017;Wang 2020). The obtained AUC > 0.8 indicated the good predictive ability of the model (Li et al. 2021). In addition, the histogram shown in Fig. 7 indicated that most of the wildfire points used for inspection were concentrated between 0.4 and 1, with only 20.14% of the fire points were within 0.4. This result indicated that the high fire risk areas identified in the present study correspond to areas of frequent forest fires according to data on the burn scar. Therefore, the logistic model constructed in the present study is highly reliable and can be used as an analysis tool for exploring fire risk.

The variables driving forest wildfires
The study found that climate is an important factor leading to forest wildfires, which is consistent with the findings of GUO et al. (2017). Among them, the temperature is positively correlated with the occurrence of forest wildfires, which is consistent with the research results of Tian et al. (2014). The relative humidity shows a negative correlation, which is inconsistent with the results of some studies (Bravo et al. 2010;Zumbrenne et al. 2011;Liu et al. 2018). The reason may be that the lower the relative humidity is, the higher the air temperature is, the higher the surface temperature is, and the lower the moisture content of combustibles increases the possibility of fire. Precipitation is also positively related to the occurrence of forest wildfires, which seems to be contrary to our common sense. However, one explanation is that more precipitation is conducive to the growth of surface plants, and the accumulation of surface fuel load is more, thus increasing the risk of forest fires. Wind speed had no significant effect on the occurrence of forest wildfires. Among topographic factors, altitude and slope are negatively correlated with forest wildfire occurrence, which is consistent with the research results of Syphard et al. (2008), which means that the probability of forest wildfire occurrence is lower in areas with high altitude. The distance from residential areas was identified to have a negative correlation with the occurrence of forest wildfires, with the probability of forest wildfires higher in areas with frequent human activities, which can be attributed to the active role of human populations in starting forest wildfires. This result was consistent with those by Guo et al. (2016a, b). In addition, the results of the present study found that the distance to roads was positively correlated with the occurrence of forest wildfires, indicating a higher probability of wildfires in areas further away from roads. This can be attributed to road networks mainly being concentrated in industrially developed areas in which firefighting resources are concentrated. In contrast, fires spread easily in remote forest areas in which fire protection resources are scattered (Guo et al. 2017). Fractional vegetation cover and vegetation moisture content showed no significant relationship and a negative correlation with the occurrence of forest fires, respectively. In general, the influence of vegetation factors on the occurrence of forest fire was not very significant, inconsistent with the conclusion of Wen (2019). Among all the independent variables, latitude appeared to have the most significant negative correlation with the occurrence of forest wildfires. Indicating that a single fire risk factor has different effects under different spatial and temporal scales, further emphasizing the complexity of processes leading to the induction and spread of forest wildfire. Areas with high heterogeneity of environmental conditions require scientific analysis of the spatial relationship between fire occurrence and influencing factors based on actual conditions and the optimal selection of related factors.

The spatial pattern of wildfire susceptibility
The spatial pattern of wildfire susceptibility in CYP predicted based on the logistic model ( Fig. 9) shows that Kunming, Chuxiong Yi Autonomous Prefecture, and the north-central part of Qujing are low fire risk and extremely low fire risk areas, and the forest fire sensitivity of these areas is low. However, the forest fire sensitivity in the southwest of Yuxi City and the south of Qujing City is relatively high, of which the high-fire-risk areas and extremely high-fire-risk areas in Yuxi City account for 26.60% and 42.67% of the total area respectively. It is found that the spatial distribution pattern of wildfire susceptibility in central Yunnan is basically consistent with the research results of Wang (2020), Zhang (2019a, b), and Feng et al.(2014), but the fire-risk level of Yuxi City has increased compared with the research results of Zhang et al. (1994), The reason for the rise of fire risk in Yuxi may be that the CYP wildfire risk factors and internal links have changed over time, forming a complex wildfire environment. A large number of recent studies have shown that the central and southern regions of CYP, including Yuxi, are indeed highly susceptible to wildfires, which may be due to the best hydrothermal conditions and higher vegetation density in the central and southern regions of CYP, which have accumulated sufficient combustible materials for the occurrence of wildfires. In addition, due to the strong low-pressure trough in Europe and the high pressure in South Asia during the period from February to April, it is not suitable for cold air to move southward. Yuxi City and Qujing City continue to be controlled by the high pressure in South Asia. A long period of high temperatures, sunny weather, and no rain may be one of the reasons for the high incidence of forest wildfires during the fire prevention period (Yang and Li 2013).
The forest wildfire risk rating map generated in this study provides objective and clear guidance for CYP wildfire management from the perspective of spatial variables. In areas with high wildfire sensitivity, regional and local wildfire managers must correctly understand and perceive the variables directly related to wildfire susceptibility, so as to reduce the serious wildfire hazards in this area.

Conclusions
The present study used satellite wildfire data during the wildfire prevention period in the CYP from 2001 to 2020, as well as a series of static and dynamic data incorporating vegetation, topography, climatology, and human activities in the corresponding period to construct a forest fire susceptibility model. The model was applied to classify forest wildfire risk levels in the CYP. The conclusions of the present study are listed below.
(1) When the significance levels of the selected model variables were below 0.05, the AUC value of the model reached 0.884, indicating the high reliability of the model. Therefore, the model can be used to explore wildfire risk.
(2) Climate, vegetation, terrain, human and location factors all affect the occurrence of forest wildfires. The distributions of temperature, vegetation coverage, distance to water bodies, distance to roads, and precipitation were positively correlated with the occurrence of forest wildfires. Elevation, relative humidity, the global vegetation moisture index, wind speed, slope, latitude, and distance to residential areas were negatively correlated with the occurrence of forest wildfires. Each factor has different degrees of influence on the occurrence of forest fires at different Space-time scales. This conclusion further confirmed the complexity of the potential mechanism of forest wildfire induction and spread. (3) The area of high forest wildfire risk in the CYP was relatively low. The areas of extremely low, low, medium, high, and extremely high fire risk accounted for 49.32%, 20.81%, 13.37%, 7.68%, and 8.82%, of the total area, respectively. (4) The CYP shows a clear risk of forest wildfires. The areas of extremely high and high fire risk are mainly distributed in southwest Yuxi City and southern Qujing City. The areas of medium fire risk are mainly located in the Chuxiong Yi Autonomous Prefecture, northern Yuxi City, southern Kunming City, and central Qujing City. Therefore, forest wildfire prevention efforts should be concentrated in southern central Qujing city, Yuxi City, and central Chuxiong Yi Autonomous Prefecture. Forest wildfire prevention efforts should correspond to local conditions and should follow rational emergency resources planning and allocation.