Bushfire Severity Prediction Optimization: Feature Combination Study with XGBoost Regression and Satellite-Derived Data

Bushfire severity prediction plays a vital role in mitigating the catastrophic effects of bushfires on ecosystems, human life, and property. Accurate prediction models enable timely interventions and informed decision-making, which are crucial for effective fire prevention and response strategies. This study aims to develop an efficient bushfire severity prediction model by leveraging satellite-derived data, specifically from Landsat 8, which includes spectral indices, topographical and climatic features. The primary objective of this study is to assess the impact of different feature combinations on model performance using the XGBoost regression model with dNBR as the target variable. Eighteen tests were conducted, ranging from single-feature indices like NDVI to more complex multi-feature sets. The findings revealed that a 21-feature combination achieved the highest accuracy of 90.09%, with a MSE of 0.0008, a computational time of 1201.86 seconds, and a memory usage of 106.28 MB. However, adding more features did not always result in proportional improvements and often led to increased computational demands. In contrast, tests with fewer features demonstrated that high precision could be achieved with reduced resource consumption, offering a favourable trade-off between performance and efficiency. The results underscore the importance of strategic feature selection and optimization to enhance model accuracy while minimizing computational over-head. This study provides valuable insights for creating effective, high-performance fire severity prediction models, especially in resource-constrained settings.