Understanding and Simulating Wildfire Changes Using Advanced Statical and Process-Oriented Models

This study aims to investigate the spatiotemporal dynamic of global wildfires, their underlying climate-driving mechanisms, and their predictability by utilizing multiple data sources (both process-based model simulations and satellite-based observations) and multiple analytical methods including machine learning techniques (MLTs).

We first explored the global wildfire interannual variability (IAV) and its climate sensitivity across nine biomes from 1997 to 2018, leveraging the state-of-art U.S. Department of Energy’s Energy Exascale Earth System Model (E3SM) land component (ELM-v1) simulations with six sets of climate forcings. Results indicate that 1) ELM simulations could reproduce the IAV of wildfire in terms of magnitudes, distribution, bio-regional contribution, and climate-covariance; 2) tropical savannas, tropical forests, and semi-arid grasslands near deserts contribute to 71.7%–99.7% of global wildfire IAV; 3) precipitation dominants global wildfire IAV with suppression impacts while temperature and radiation positively contribute to wildfire IAV; and 4) major regional discrepancies between simulations and satellite observations lies in semi-arid grasslands, croplands, boreal forests, and wetlands.

We further applied a two-step correcting MLT with multiple classification and regression algorithms in boreal peatland to investigate the MLTs’ predictability and key drivers of peat fires. Results show that 1) the applied oversampling algorithm helps tackle the unbalanced fire occurrence data and improves recall rate with a growth ranging from 26.88% [percent] to 48.62% [percent]; 2) the Random Forest performs the best on simulating the peatland fire occurrence and counts with the Area Under the Receiver Operating Characteristic Curve value ranging from 0.83 to 0.93 across multiple fire datasets; 3) a range of factor-control simulations indicate that temperature, air dryness, and frost day frequency (freeze-thaw cycles) dominant the peat fires, overweighting the impacts by biomass, soil moisture and human activities.

To obtain reliable fire model simulations and amend the physical model failures in reproducing global fire trends, magnitudes, and spatial distributions, six ELM fire simulations with different climate forcings and seven other models from Fire Modeling Intercomparison Project (FireMIP) with different structures were collected to evaluate model performance. Applying unweighted and weighted post-processing algorithms—Multi-Model Ensemble Mean (MME) and Bayesian Moving Averaging (BMA), we analyzed the ensemble calibration efficiency of two approaches. Results show that both ELM and FireMIP models underestimate the magnitude, variance, and trend of burned area in Savannas and Woody Savannas, but the BMA outperforms MME and dramatically raised the modeling correlation coefficient with observations across eco-regions from about 0.4 to over 0.9, indicating the effectiveness of BMA—an optimally combined model ensemble—in uncertainty reduction and reproducibility promotion.