#wildfires
#emergent_constraints
#machine_learning
Yan Yu from Peking University and co-authors in 2022.
Background
(1) It is necessary to create constrained future projections of fire emissions
Current earth system models only include incomplete and highly parameterized driving processes of fire, and cannot accurately characterize the human-vegetation-fire-climate feedback.
(2) The conventional emergent constraint approach does not satisfy the need of wildfire constraint, because
- Only 13 of the currently available CMIP6 models provide fire carbon emissions in both historical and future simulations, and they lack diversity in structures and parameters
- The linear relationship estimated by traditional emergent constraint cannot deal with the complex, nonlinear relationship between meteorological, ecological, and socioeconomic factors and fire
- The traditional emergent constraint may be suitable for large-scale averaged quantities, but not for the detailed spatial distributions of wildfire
- The emergent constraint approach, however, is still more suitable than bias-correction or model weighting.
Method
(1) The authors developed a machine learning based emergent constraint method to project future global fire emissions and socioeconomic risks
(2) The machine learning models were trained on the Earth system model-simulated relationships between the historical climate, ecosystem, and socioeconomic variables, and future fire carbon emissions under SSP585. The observed variables were then substituted in. The predictors included fuel abundance (leaf area index, temperature, precipitation), fuel moisture (soil moisture, relative humidity, precipitation, temperature), fire spread conditions (wind, orography), ignition sources (flash rate, land use, population).
Results
(1) The machine learning constraint reduces the biases in fire carbon emission bias. Before and after the emergent constraint, the grid-level biases always ranged from -0.32 to 0.16 kg m-2 yr-1, but the grids that have large errors were much, much more before the constraint than after.
(2) The constrained results showed weaker trends in fire carbon emissions compared to the un-constrained results, but the socioeconomic exposure became greater, in terms of population, gross domestic production, and agricultural area.
