Wildfire risk assessment in Sichuan Province, China: hazard modeling approach considering different combinations of classification criteria and connection values of factor attributes

The MODerate Resolution Imaging Spectroradiometer (MODIS) fire dataset was utilized as the primary source of wildfire samples. Provided by the Fire Information for Resource Management System (FIRMS) of NASA (https://firms.modaps.eosdis.nasa.gov/usfs/map/), this dataset has monitored fire points daily since 2000 via sensors onboard the Terra and Aqua satellites, with a spatial resolution of 1 km and four observations per day, effectively capturing wildfire dynamics (Li et al. 2018). The dataset includes observation coordinates, acquisition time, fire radiative power (FRP) and confidence levels, offering comprehensive support for wildfire spatiotemporal analysis.

The raw MODIS data contains noise and misclassified points. To enhance accuracy, fire points from 2001 to 2020 were filtered by excluding samples with confidence values below 80 and FRP below 10 (Khan et al. 2024). Points labeled as ‘active volcano’, ‘other static land source’ and ‘offshore detection’ were removed to retain only vegetation fires. Further refinement was conducted using MCD12Q1 land use data from 2001 to 2020 (https://lpdaac.usgs.gov/products/mcd12q1v006/) to exclude points in water bodies, buildings and bare land. Duplicate fire points within the same grid were also eliminated, resulting in 4824 high-quality wildfire samples for analysis.

Wildfire hazard factors in this study were selected based on their relevance to both natural and anthropogenic conditions influencing wildfire occurrence and spread. They were categorized into five types: topographic, hydrological, anthropogenic, vegetation and meteorological. Topographic factors, such as elevation and aspect, are considered to influence local microclimate, while slope and curvature influence fire spread patterns (Ghorbanzadeh et al. 2019; Yang et al. 2021; Nikolić et al. 2023). Hydrological indicators, including topographic wetness index (TWI), stream power index (SPI) and distance to rivers are used to characterize surface wetness and accessibility of water sources (Pourtaghi et al. 2015; Zhang et al. 2021). Anthropogenic factors are employed to reflect infrastructure proximity and human activity intensity, including distance to railways and highways, and nighttime lght index (NTLI) (Elia et al. 2020; Wu et al. 2023). Vegetation characteristics, such as fractional vegetation cover (FVC), determine available fuel load and combustion potential (Rihan et al. 2019; Bergonse et al. 2021b). For meteorological factors, precipitation, potential evapotranspiration and relative humidity are used to describe vegetation and environmental dryness, solar radiation determines surface heat conditions and wind speed affects fire spread behavior (Cooke et al. 2012; Bergonse et al. 2021a; de Diego et al. 2021). Information about the factors is shown in Table 1 below.

The selection of wildfire vulnerability and adaptive capacity factors primarily focused on socio-economic conditions and infrastructure levels, assessing both the exposure of human societies to wildfires and their ability to withstand wildfire (Xepapadeas et al. 2024). Vulnerability factors, such as population density, gross domestic product (GDP), building density, road density and railway density, reflect the risks to human life, property and regional infrastructure development. Higher population and building densities indicate increased fire exposure and potential economic loss, while dense infrastructure networks amplify vulnerability to wildfire disruptions. (Oliveira et al. 2017; de Diego et al. 2021; Xepapadeas et al. 2024). Additionally, regions with higher biodiversity values are more ecologically vulnerable to fire disturbances (Woinarski et al. 2023). Adaptive capacity factors, including per capita GDP and the proximity of fire stations and medical institutions, measure the ability to respond to and recover from wildfires. Higher GDP per capita is usually associated with more adequate resources to respond to disasters, and shorter distances to emergency services enhance the effectiveness and timeliness of interventions (Lambrou et al. 2023; Liu et al. 2024a). Infrastructure-related indicators were derived using geographic information systems (GIS) spatial analysis, where building, railroad and highway densities were calculated from vector data using kernel density tools, and proximity-based indicators were obtained using Euclidean distance functions. The information about the relevant factors is shown in Table 2.

To ensure consistency in model development and statistical analysis, all factors were projected onto the UTM Zone 48N coordinate system. The data were resampled to a uniform grid size of 500 × 500 m. Subsequently, the area was divided into 2103 × 1849 grids, resulting in a total of 1,946,748 units.

Six statistical connection methods, probability statistics (PS), FR, IV, certainty factor (CF), WOE and evidential belief function (EBF), were selected in this study to convert the raw attributes of factors into quantitative connection values, which were combined with the LR algorithm to construct wildfire hazard models.

PS is a common method of sample statistics. It calculates the ratio of wildfire sample points in a given interval to the total wildfire samples, reflecting the relative frequency of wildfires within that range (Huang et al. 2024).

where N_i is the number of wildfire samples in the i-th attribute interval, and N is the total number of wildfire samples in the study area.

FR is a statistical probability-based approach that compares the wildfire frequency within the specific attribute interval to its overall frequency across the area (Pourtaghi et al. 2015). The resulting ratio indicates the degree to which that interval influences wildfire hazard.

where N_i and S_i denote the number of wildfire samples and total grid cells in the i-th attribute interval, while N and S represent the total wildfire samples and grid cells in the entire study area. FR_i > 1 indicates the attribute interval that promotes wildfire hazard, and vice versa.

The IV assesses the impact of each attribute interval on wildfire hazard by calculating the ‘information content’ of wildfires (Polat 2021).

where the parameters are defined in the FR method. IV_i > 0 indicates increased wildfire hazard, and vice versa.

The CF evaluates the influence of attribute intervals on wildfires by measuring the deviation of the conditional probability within a specific interval from the prior probability of wildfires in the study area (Hong et al. 2018).

with P_i representing the conditional probability of wildfires in the i-th interval and P_s denoting the overall wildfire occurrence probability. CF_i > 0 indicates higher wildfire hazard, and vice versa.

WOE is based on Bayesian probability theory and quantifies the impact of each attribute interval on wildfires (Hong et al. 2019). It calculates the logarithmic ratio of wildfire occurrence to non-occurrence within specific intervals.

where Wi+ and Wi− represent the positive and negative weights, respectively; P (B/D) is the probability of wildfires in the i-th interval; P(B/D¯) is the probability of wildfires outside that interval; P(B¯/D) and P(B¯/D¯) are the probabilities of no wildfires within and outside that interval, respectively. C_i > 0 indicates a positive influence on wildfire hazard, and vice versa.

EBF, based on Dempster–Shafer theory, evaluates uncertainty and heterogeneous data in spatial data integration (Nami et al. 2018). It quantifies the effect of factor conditions on wildfire hazard using belief (Bel) in the belief parameters.

where W_i represents belief weights. The higher Bel is, the greater wildfire hazard in that interval.

The delineation of factor attribute intervals and the number of classifications directly influence the ability of factors to indicate wildfire hazard (Qin et al. 2024). To ensure accuracy and objectivity, classification should be based on the actual distribution of data rather than subjective or uniform segmentation. Therefore, Jenks-GVF methodology that combines Jenks natural breaks (Jenks) and goodness-of-variance fit (GVF) was used to dynamically determine the optimal attribute classification for each hazard factor. It helps to eliminate uncertainty associated with manual subjective classification schemes.

The Jenks method classifies data by minimizing within-class variance and maximizing between-class variance (Chowdhuri et al. 2021; Rivière et al. 2023). The breakpoints were optimized to minimize the sum of squared deviations (SSD) by iteratively testing the segmentation scheme. Meanwhile, GVF was utilized to assess the classification effect by measuring the ratio of intra-class variance to total variance (Jenks and Caspall 1971):

where TV is the total variance of the dataset, and GVF ∈ [0, 1]. Higher GVF values indicate more effective classification (Fraile et al. 2016). Compared to classification methods with equal intervals or fixed thresholds, the Jenks-GVF approach better captures natural groupings within heterogeneous data distributions. In this study, the optimal number of classes was determined when the increase in GVF between successive classifications fell below 0.01, indicating that further increase in the number of classifications did not significantly improve the quality of classification. This data-adaptive and statistically grounded method ensures that the classified attribute intervals of hazard factors have higher internal consistency and external differentiation, which is particularly important for improving the interpretability of wildfire hazard modeling.

LR is a statistical model used for binary classification tasks, aiming to establish a mathematical relationship between the dependent variable and multiple independent variables, and transforming the linear model into a nonlinear probability output through the Sigmoid function (Sevgen et al. 2019). For a binary dependent variable Y (where Y ∈ [0, 1]), the predictive function is defined as:

where P(Y = 1) represents the probability that the sample belongs to class 1 (wildfire); β₀ is the intercept, β_1, β₂, …, β_n are the regression coefficients for independent variables x_1, x₂, …, x_n, and e is the base of the natural logarithm.

The modeling dataset was constructed using connection values of the factors as inputs and the sample attributes as outputs (1 for wildfire; 0 for non-wildfire). The dataset was randomly split into a training set (70%) and a testing set (30%), maintaining a balanced ratio of positive and negative samples: 3376:3377 for training and 1448:1447 for testing. During training, the 10-fold cross-validation method was used to ensure the robustness and generalization ability of the model (Bera et al. 2022).

The receiver operating characteristic (ROC) curve visualizes the model’s classification ability at different decision thresholds by plotting the true positive rate against the false positive rate. The area under the curve (AUC) value ranges from [0, 1]. The closer the AUC value is to 1, the better the model’s overall classification performance (Dehnavi et al. 2015). Beyond the ROC curve, additional metrics such as precision, recall, F1-score and accuracy further evaluate the classification performance from multiple perspectives, offering a more comprehensive assessment of the predictive performance of model (Hang et al. 2024).

The Jenks-GVF method was used to determine the optimal number of classes and attribute intervals for each factor, using both wildfire samples and the whole area as classification criteria. Using elevation as an example (Fig. 3), the GVF curves (first column) showed a rapid increase followed by a plateau as the number of classes increased, indicating diminishing marginal gains in explanatory power. The optimal number of classes was six based on wildfire samples and five based on the whole area. The second and third columns present the area proportion, wildfire sample percentage and connection values (PS, FR, IV, CF, WOE, EBF) across classification intervals. It can be seen that the classification criteria directly influence the number of factor attribute intervals and the distribution of connection values, which in turn affects their ability to distinguish wildfire hazard levels.

Fig. 3.

Elevation classification and connection value calculation results based on different classification bases. Subfigures (a–c) present the GVF curves, subdivision areas, and wildfire sample coverage under optimal classification using wildfire samples as criteria. Subfigures (d–f) show the same metrics using the whole area as classification criteria (GVF: goodness-of-variance fit; PS: probability statistics; FR: frequncy ratio; IV: information value; CF: certainty factor; WOE: weights of evidence; EBF: evidential belief function).

Wildfire hazard values were predicted and visualized for the whole area based on LR models constructed with different classification criteria and connection values. All visual maps of wildfire hazard were classified into five levels using consistent equal-interval classification schemes: very low ((0, 0.20]), low ((0.20, 0.40]), moderate ((0.40, 0.60]), high ((0.60, 0.80]) and very high ((0.80, 1.00]) (Fig. 4). This ensures that each level corresponds to a fixed numerical range, facilitating clear comparison between hazard maps derived from different models and reducing classification ambiguity. Visually, the overall wildfire hazard has similar spatial distribution across the models, with high hazard areas concentrated in the southwestern and some western mountainous regions, and low hazard areas in the plateau and plains of northwestern and eastern Sichuan. Hazard levels exhibited clear gradients from extremely high to extremely low hazard zones. However, the hazard zoning results of PS, FR and EBF models constructed based on different classification criteria varied significantly. In contrast, the IV, CF and WOE models exhibited smaller differences and maintained consistent clustering with higher stability.

Fig. 4.

Wildfire hazard maps generated using different connection methods coupled with the logistic regression (LR) algorithm. Subfigures (a–f) are based on wildfire samples as classification criteria of factor attributes, and subfigures (g–l) are based on the whole area (PS: probability statistics; FR: frequncy ratio; IV: information value; CF: certainty factor; WOE: weights of evidence; EBF: evidential belief function).

As shown in Fig. 5, the models based on wildfire samples classification had higher proportions of very high and very low hazard areas, better capturing wildfire-prone regions while reducing misclassification of low hazard areas. IV, CF and WOE models consistently identified larger very high hazard areas under both classification criteria, showing greater stability. PS, FR and EBF models displayed more variability, with greater fluctuations in hazard area distribution.

Fig. 5.

Hazard zoning results (PS: probability statistics; FR: frequncy ratio; IV: information value; CF: certainty factor; WOE: weights of evidence; EBF: evidential belief function; LR: logistic regression).

The ROC curves and AUC values for each model on the test set are shown in Fig. 6. Except for PS, models classified based on sample points had higher AUC values than those classified based on the whole area. It can be seen that classification based on sample points better captures wildfire distribution characteristics, enhancing overall classification performance, while models classified based on the whole region show limitations in distinguishing high and low hazard areas. Among connection methods, models using IV performed best under both classification criteria, achieving the highest AUC values, followed by WOE, CF, EBF, FR, and PS. The Point-IV-LR model, classified based on wildfire sample points, had the highest AUC value (0.9254), demonstrating the best predictive performance.

Fig. 6.

Receiver operating characteristic curves and area under the curve values for hazard models (PS: probability statistics; FR: frequncy ratio; IV: information value; CF: certainty factor; WOE: weights of evidence; EBF: evidential belief function; LR: logistic regression).

As shown in Fig. 7, all metrics for the models based on wildfire sample classification outperformed for those based on whole area classification, except for PS. Moreover, Point-IV-LR, Point-CF-LR and Point-WOE-LR outperformed the other models. The Point-IV-LR model achieved the highest recall, F1-score and accuracy, demonstrating superior balance in classification precision and wildfire sample identification. The Point-EBF-LR model had the highest precision, effectively identifying high hazard areas while minimizing misclassification of low hazard areas.

Fig. 7.

Receiver operating characteristic curves and area under the curve values for hazard models (PS: probability statistics; FR: frequncy ratio; IV: information value; CF: certainty factor; WOE: weights of evidence; EBF: evidential belief function; LR: logistic regression).

Overall, models classified based on wildfire sample points, particularly when using IV, CF and WOE, demonstrated stability and superior performance. The Point-IV-LR model emerged as the most effective for wildfire hazard assessment, providing a reliable foundation for subsequent risk assessment.

As shown in Fig. 8, elevation, TWI, distance to railways, solar radiation and potential evapotranspiration consistently ranked high in factor importance across most models, highlighting their critical roles in wildfire hazard. Elevation is thought to influence the local microclimate and vegetation zone, which determines fuel availability and flammability. TWI represents the spatial distribution of soil moisture influenced by terrain, which affects how readily vegetation can dry and ignite. Distance to railroads is commonly used to reflect the intensity of human activities affecting wildfires. Solar radiation is considered to affect vegetation dryness, thereby altering ignition potential and combustion behavior. Potential evapotranspiration indicates atmospheric moisture demand and strongly regulates fuel dryness conditions. Notably, solar radiation and potential evapotranspiration are consistently assigned high importance across most models, suggesting that surface dryness and changes in fuel loads have significant impact on the occurrence and spread of wildfires.

Fig. 8.

Factor significance of the models. Where Dis_riv: Distance to rivers; Dis_rail: Distance to railways; Hum_act: Human activities; Sol_rad: Solar radiation; Pot_eva: Potential evapotranspiration; Rel_hum: Relative humidity; Others: the remaining nine factors; FVC: fractional vegetation cover; SPI: stream power index; TWI: topographic wetness index.

In addition, the models using IV, CF and WOE exhibited high consistency compared to PS, FR and EBF, with the top seven factors being nearly identical with minimal differences in rankings. In models classified based on wildfire samples, the most important factors were elevation, slope, TWI, distance to railways, human activities, solar radiation and potential evapotranspiration. For whole area classification, the top factors were elevation, aspect, TWI, distance to railways, FVC, solar radiation and potential evapotranspiration.

Attribute intervals and corresponding IV values of factors were used to reveal the effects on wildfire hazard based on the optimal Point-IV-LR model (Figs 9, 10).

Fig. 9.

Information values (IV) for different attribute intervals of each factor: (a–d) topographic factors (elevation, slope, aspect, curvature); (e–g) hydrological factors (topographic wetness index (TWI), stream power index (SPI), distance to rivers); (h) anthropogenic factor (distance to railways).

Fig. 10.

Information values (IV) for different attribute intervals of each factor: (a–b) anthropogenic factor (distance to highways, human activities); (c) vegetation factor (fractional vegetation cover (FVC)); (d–h) meteorological factors (precipitation, solar radiation, potential evapotranspiration, relative humidity, wind speed).

For topographic factors, IV values for elevation were positive in the range of (120, 1150 m], where elevation may create the moderate temperature microclimate and vegetation type favorable for wildfires. Positive IV values for the slope factor in the range of (4.882°, –66.863°], suggesting that steeper terrain facilitated wildfire spread. Positive IV values for east, southeast, south, southwest and west aspects reflected increased solar exposure on these slopes, leading to drier surface conditions. Positive IV values were observed when the absolute value of curvature was greater than 0.043, possibly because undulating terrain can accumulate dry fuels and support flame propagation.

For hydrological factors, the influence of TWI on wildfire hazard generally declined, while SPI tended to increase. Positive IV values were observed for TWI in (5.549, 8.737] and SPI in (3.976, 18.935], suggesting that higher flow intensity and lower topographic moisture favored wildfire occurrence and spread. Distance to rivers exhibited an inverted U-shaped trend with positive IV values in (1118, 7778 m], likely due to limited water availability and specific vegetation distributions.

For anthropogenic factors, wildfire samples were densely distributed within [0, 5700 m] from railways with positive IV values. Although the IV value was negative at [0, 707 m] distance from the highway, this range contained the highest number of wildfire samples. In (707, 5220 m], the IV value turned positive, suggesting increased wildfire hazard due to human activity. Similarly, the NTLI had negative IV values at [0, 2.655], which covered nearly 90% of wildfire samples, while positive IV values in (2.655, 44.425] indicating increased wildfire hazards from increased human activity. The IV for FVC was positive at (0.149, 0.827], indicating the areas typically have sufficient combustible material for higher wildfire hazard.

For meteorological factors, precipitation exhibited an increasing then decreasing trend, with positive IV values at (0.113, 0.207 m]. It may be due to the fact that moderate precipitation promotes vegetation growth without significantly increasing ambient humidity, which promotes wildfire hazard. Solar radiation showed an overall positive correlation with wildfire hazard. It may facilitate wildfire occurrence and spread by enhancing surface heating and fuel drying. In contrast, potential evapotranspiration and relative humidity were negatively correlated with wildfire hazard, indicating that high surface evaporation and low air humidity increase wildfire hazard. Positive IV values for wind speeds in the range of [0.094, 0.571 m/s] suggest that low wind speeds can promote wildfire spread by allowing heat to build up locally and trapping embers in forest combustibles.

The AHP, EWM and AHP-EWM combination weights for each factor are shown in Table 3. Among the vulnerability factors, GDP had the highest weight, reflecting the significant economic losses wildfires cause at the regional level. Biodiversity value also received a high weight, highlighting the severe damage in ecologically sensitive areas. Railroad density and population density were equally important for assessing human exposure risks and infrastructure vulnerability, while highway density and building density had relatively lower impacts. For adaptive capacity factors, distance to firefighting institutions had the highest weight, emphasizing the critical role of firefighting resource accessibility in emergency response. The weights of distance to medical facilities and GDP per capita indicate their importance in emergency response and disaster recovery.

Wildfire vulnerability maps were generated using the TOPSIS model combined with AHP-EWM weights (Fig. 11). The regional vulnerability generally followed a ‘high in the southeast, low in the northwest’ pattern, displaying clear spatial heterogeneity. The very high vulnerability areas accounted for only 0.3% of the total area, mainly in Chengdu, a densely populated and economically developed region. High vulnerability areas covered 9.83%, mainly in the ecologically sensitive southwestern Sichuan and urbanized central and southeastern regions. Moderate vulnerability areas made up 22.137%, scattered mainly across central and eastern Sichuan’s hilly and plain areas. Very low and low vulnerability areas comprised 67.734% of the total area, primarily distributed across the sparsely populated northwestern plateau with limited socio-economic development and lower ecosystem valuation.

Fig. 11.

Vulnerability maps: (a) continuous values; (b) Jenks-based zoning levels.

Adaptive capacity exhibited a spatial pattern of ‘strong in the east, weak in the west’ (Fig. 12). The very high and high adaptive capacity areas were concentrated in the urbanized regions of central and southern Sichuan, where higher per capita GDP and proximity to medical and firefighting institutions enhanced disaster response and recovery. Moderate adaptive capacity areas, accounting for 33.321%, were mainly in the moderately developed regions of central and southern Sichuan. Low and very low adaptive capacity areas accounted for 26.981%, mainly concentrated in the remote mountainous regions of western and northwestern Sichuan, where low economic levels and insufficient firefighting and medical infrastructures hindered emergency response and resource mobilization in the event of wildfires.

Fig. 12.

Adaptive capacity maps: (a) continuous values; (b) Jenks-based zoning levels.