Balanced Random Forest Model is More Suitable for Wildfire Risk Assessment

Zili Wang,Binbin He,Xiaoying Lai

DOI: https://doi.org/10.1109/igarss46834.2022.9883573

2022-01-01

Abstract:With global warming, the number and the burning area of wildfires will continue to increase. Wildfire risk assessment is to assess the occurrence probability and spread risk of wildfire by combining fire risk factors, like weather, fuel and terrain. In the past, linear regression models and general artificial intelligence models were used to assess wildfire risk. But the wildfire datasets are typical non-equilibrium, using these methods often results in a bias towards non-fire class, leading to lower point in fire class recognition. Algorithms for imbalanced datasets, such as balanced random forest, can improve the identification accuracy of the minority class, but are rarely used in wildfire risk assessment. In this study, we discussed the performance of random forest, extreme gradient boosting and balanced random forest models in wildfire datasets of the Great Xing'an Mountains. The results showed that: balanced random forest model greatly improves recall and f-score, and its wildfire risk distribution map shows that high fire risk areas are more accurate representation of the location of wildfire. Therefore, this study believes that balanced random forest model is more suitable for wildfire risk assessment.