Analysis of the Fire Risk in Wildland-Urban Interface with Random Forest Model

Objective: The wildland-urban interface (WUI) fires are increasingly frequent and cause serious damage to people's livelihood and economy. In this paper, a random forest (RF) model was used to spatially model the relationship between fire risk and its influencing factors in the WUI at the provincial scale, and the advantages of the random forest model in fitting and interpreting fire risk in the WUI were explored, and the influencing factors in the WUI fire were compared with the factors of forest fire risk to provide a basis for further assessment of fire risk in the WUI. Method: Based on the historical fire data of Anhui Province from 2002 to 2011, in this study 9 factors from climate, geographical environment, human activities and social economy were designated as independent variables, and the monthly average fire density was used as the dependent variable. The feature selection method was used to obtain the contribution of different independent variables within the model, statistical characteristics and the average performance of the internal model, to select the independent variables into the final model. The random forest (RF) model was used to explain the fire risk of the WUI and analyze the important factors affecting fire risk in the WUI and Forest. Result: The ranking of influence degree of key independent variables on fire risk in the WUI was Line density of roads, Line density of rails, Monthly average maximum temperature, Normalized Difference Vegetation Index, Population density and Elevation. The ranking on fire risk in the Forest was Monthly average maximum temperature, Normalized Difference Vegetation Index, Line density of roads, Line density of rails, Population density and Elevation. Through the training and calculation, it was found that the performance of random forest (RF) model in the five sub-models' training set was basically consistent with that of the test set. The simple correlation coefficient between fitted value and actual value reached more than 0.90, indicating that the RF model had remarkable ability to explain fire risk in the WUI and Forest. In addition, the RF model was fitted on the overall data set, and the correlation between fitted value and actual value of fire risk in the WUI was 0.997, and the correlation between fitted value and actual value of the forest fire risk was 0.996, indicating that the RF model had extremely strong fitting performance in the field of fire risk. Conclusion: The most important independent variables that affect the WUI fire occurrence are the line density of roads and the line density of rails, while for forest fires, these variables are the monthly average maximum temperature and the normalized difference vegetation index. It is shown that the occurrence of fire in the WUI is closely related to human activities. Random Forest algorithm can show robust and extremely accurate fitting ability for fire risk in the WUI and Forest, which is a very useful tool. © 2019, Editorial Department of Scientia Silvae Sinicae. All right reserved.