Characterizing Topographic Influences of Bushfire Severity Using Machine Learning Models: A Case Study in a Hilly Terrain of Victoria, Australia

Abstract

Topography plays a significant role in determining bushfire severity over a hilly landscape. However, complex interrelationships between topographic variables and bushfire severity are difficult to quantify using traditional statistical methods. More recently, different Machine Learning (ML) models are becoming popular in characterising complex relationships between different environmental variables. Yet, few studies have specifically evaluated the suitability of ML models in predictive bushfire severity analysis. Hence, the aim of this research is twofold. First, to determine suitable ML models by assessing their performances in bushfire severity predictions using remote sensing data analytics, and second to identify and investigate topographic variables influencing bushfire severity. The results showed that Random Forest (RF) and Gradient Boosting (GB) models had their distinct advantages in predictive modelling of bushfire severity. RF model showed higher precision (86% to 100%) than GB (59% to 72%) while predicting low, moderate, and high severity classes. Whereas GB model demonstrated better recall, i.e., completeness of positive predictions (56% to 75%) than RF (49% to 61%) for those classes. Closer investigations on topographic characteristics showed a varying relationship of severity patterns across different morphological landform classes. Landforms having lower slope curvatures or with unchanging slopes were more prone to severe burning than those landforms with higher slope curvatures. Our results provide insights into how topography influences potential bushfire severity risks and recommends purpose-specific choice of ML models.

An Integrated Grassland Fire-Danger-Assessment System for a Mountainous National Park Using Geospatial Modelling Techniques

“…The efficiency of RF is influenced by two parameters: the number of trees in the forest (ntree) and the number of random variables per split node (mtry) [138,139]. Sharma, et al [26] thoroughly explained the RF algorithm for the RF classification and regression algorithms.…”

“…This pattern might be due to the effect of climate warming. Although the TRI had collinearity with other topographic drivers [26], its superiority to elevation underscores its incorporation in fire-danger modelling.…”

“…In addition, remote sensing (RS) is a popular method in fire modelling. RS is used to classify land cover which is associated with fire-fuel characteristics [25], provide elevation data [26], assess vegetation moisture content [7,[27][28][29], assess soil moisture content [10,[30][31][32], and for active fire detection and burnt-area estimation [33][34][35]. With the development of Google Earth Engine (GEE) and Cloud Engine (CE) [36], platforms that reduce the requirement for expensive information technology infrastructure, the acquiring and processing of RS datasets is rapidly improving the prediction of fire-danger models [37].…”