A framework for predicting impact of wildfires validated based on 20-years of historical data in Australia

Abstract:

Wildfire is a critical ecological and disturbance process in Australian terrestrial ecosystems. Australia has experienced increasingly destructive wildfires over the past two decades. Fire risk is expected to increase significantly due to the projected increases in temperature in future climate. Therefore, understanding and predicting fire occurrence and characteristics (burn severity and extent) is critical to evaluate current and future impact of wildfires on the ecosystems of Australia. In this work we present a novel framework that is used to predict fire severity and the level and duration of vegetation recovery. Specifically, using the pre-fire season drought conditions and vegetation type and the fire season forecasted meteorological conditions as input, we build a predictive model coupling the Quantile Weighted Distance (QWD) and Gradient Boosting Trees (GBTs) to predict the difference Normalized Burning Ratio (dNBR), an indicator of fire severity derived from Landsat imageries. Based on a leave-one-year-out cross-validation experiment, the severity classification error by QWD is shown to be lower than 10%, and the final burn severity estimation agrees well with the observation, with determination coefficient (R²), mean absolute percentage error (MAPE) and centered root mean square error (CRMSE) being 0.70, 64%, and 88. The model can also capture the annual variability and magnitude of burn severity. Proxied by the Normalized Difference Vegetation Index (NDVI) time series, we predict vegetation recovery using the hierarchical Bayesian (HB) model, using topographic, soil and climatological variables. The recovery model shows the ability to capture the seasonal variations as well as the recovery trajectory of vegetation. The predictive framework of burn severity and post-fire recovery is tested using data from the past 20 years (2000-2019) in New South Wales. The predictive models could serve as a tool of evaluating vegetation resilience to natural disturbance at global scale and in future climate conditions.