Abstract:

【Objective】Spatial modeling and grade evaluation of forest and grass fire dangers are great significance in fire prevention. In this study, a framework of spatial non-stationarity was introduced to carry out fire danger modeling in Yunnan Province. 【Method】 Data from the Moderate Resolution Imaging Spectroradiometer(MODIS), four fire-influencing factors (vegetation status, temperature, topography, and accessibility), and geographically weighted Logistic regression (GWLR) were employed to model and evaluate the forest and grass fire danger. After checking the spatio-temporal distribution of fires during 2001 and 2014, an equal number of unburned and burned points from December to April were sampled according to the designed procedure. The sampled data were divided into seven groups (one group for two years); of these, four were selected to develop the GWLR model (i.e., training and inner validation subset) and three were used to validate the developed model (i.e., independent testing subset), which created thirty five datasets. After excluding potential multicollinearity among the independent variables, the 35 adaptive Gaussian GWLR models and 35 adaptive bi-square GWLR models were developed. Cross validation was utilized to evaluate the performance of the developed GWLR models. Statistical inference was carried out based on the optimal model. 【Result】 We concluded that attention should be paid to overfitting and spatial autocorrelation of residuals in the application of GWLR to fire danger studies as they may hamper fire prevention operations. The “extreme high” danger zone accounted for 12.7%, “very high” for 18.1%, and “high” for 19.0% of the forest and grass fires in Yunnan Province from 2001 to 2014. Together, the “extreme high” and “very high” zones accounted for 80% of these fires and could be labeled as the “key area” in the danger zone. We suggest that the locations of Wenshan, Honghe, Yuxi, Chuxiong, Lijiang, and Dali, which together account for 71.5% of the “key area,” should be recognized as the most fire-prone areas and must be given priority in receiving labor, funds, and goods for forest and grass fire prevention. High temperature was an all-important forest and grass fire-influencing factor in Yunnan Province, followed by altitude, slope, dead vegetation, and fraction of green vegetation. The accessibility of roads and villages also plays a role in controlling the spread of fire. The influence of environmental factors on the presence of fire presented significant spatial variability in this study; however, these factors are unique to the Yunnan Province. 【Conclusion】 According to the significant areas of fire-influencing factors and their relative weights, Yunnan Province can be divided into six fire prevention regions and requires the formulation of different fire prevention policies for each region.

Key words: forest and grass fire danger, geographically weighted Logistic regression (GWLR), spatial autocorrelation of residuals, cross validation, Yunnan Province