Authors:
Maryamsadat Hosseini
and
Samsung Lim
Affiliation:
School of Civil and Environmental Engineering, University of New South Wales, High Street, Sydney, Australia
Keyword(s):
Bushfire, Susceptibility Map, Gene Expression Programming, Machine Learning, Kangaroo Island.
Abstract:
Kangaroo Island, South Australia is one of the bushfire-prone areas. A catastrophic bushfire known as the black summer hit Kangaroo Island in 2019/2020. We chose Kangaroo Island as a case study to generate bushfire susceptibility maps using five different methods, namely gene expression programming (GEP), random forest (RF), support vector machine (SVM), frequency ratio (FR) and logistic regression (LR). To generate bushfire susceptibility maps, we used eight contributing factors including: digital elevation model, slope, aspect, normalized difference vegetation index, distance to roads, distance to streams, precipitation, and land cover. The proposed methods were evaluated by area under the curves (AUCs) of receiver operating characteristic. RF performed best with an AUC of 0.93, followed by SVM and GEP with AUCs equal to 0.89 and 0.88, respectively, but LR and FR performed least among the five methods with AUCs 0.85 and 0.84, respectively. The generated bushfire susceptibility maps
show that western and central areas of Kangaroo Island are highly vulnerable to bushfire.
(More)