Predicting spatial patterns of wildfire susceptibility in the Huichang County, China: An integrated model to analysis of landscape indicators

Wildfire Probability Mapping: Bivariate vs. Multivariate Statistics

“…To adequately account for all local characteristics of the study area, twelve independent variables that have been frequently used in the wildfire literature were considered: Altitude, aspect, slope degree, topographic wetness index (TWI), annual temperature, and rainfall, wind effect, land use, normalized difference vegetation index (NDVI), and distance to roads, rivers, and populated areas. We refer the interested reader to the corresponding literature [7,11,[14][15][16][17][18][19][20][21]24] for the information on the significance of these variables on wildfire occurrence and their utility for predictive modeling of future fires. To produce a topographic dataset describing altitude, aspect, slope, and TWI, we used a Digital Elevation Model (DEM) at 30-meter resolution.…”

“…Furthermore, we generated the land use and NDVI maps of the study area using Landsat-8 OLI 30 m (http://earthexplorer.usgs.gov). Finally, we categorized each explanatory variable into several classes based on the previous works [14][15][16][17][18][19][20][21]24,25] and local conditions of our study landscape ( Figure 3). roads, rivers, and residential areas.…”

“…The goodness-of-fit and predictive capability of the modeling approaches adopted in this study were evaluated employing the receiver operating characteristic (ROC) curve and its two components, i.e., sensitivity and specificity, that calculated the success rate and prediction rate and their associated area under the curve (AUC). The philosophy and mathematical formulation of this method have been fully presented in the corresponding literature [14][15][16][17][18][19][20][21]. In summary, the sensitivity (i.e., probability of detection) answers the question of what fraction of the observed fire pixels are correctly classified, and its perfect value is 1; specificity (i.e., negative predictive value) answers the question of what fraction of the non-fire pixels are correctly classified, and its perfect value is 1.…”

“…The ROC of the training dataset yields the success rate of the model and measures the goodness-of-fit of the model. The ROC of the validation dataset yields the prediction rate of the model and indicates how well or poorly the model predicts the future events [7,[14][15][16][17][18][19][20][21]. In terms of the AUC value, the values of <0.6 indicate a poor, 0.6-0.7 a moderate, 0.7-0.8 a good, 0.8-0.9 a very good, and >0.9 an excellent model performance [36,37].…”

“…Nami et al [20] used the evidential belief function and successfully mapped the wildfire probability with a predictive accuracy of 81.03%. Most recently, Hong et al [21] compared the weight of evidence and regression logistic and reported predictive accuracies of 85.4 and 79.1%, respectively. These reviews of the literature clearly indicate that there is no single best model that can be effectively used in all fire-prone landscapes.…”