#wildfires
#machine_learning
#seasonal_forecasts
Yu Yan from Peking University and co-authors in 2020.
The objectives of this study are to (1) deepen the understanding of the environmental drivers of African fire, (2) improve the capacity to accurately predict seasonal fire activity in Africa. These contribute to better fire management in these vulnerable ecoregions.
1. Background
The Africa subtropical savannah ecosystems emit ~50% of the global fire-related carbon emissions, and occupy ~70% of the global burned area. The interannual variability in atmospheric CO2 concentration is mediated by African fires. African wildfires also affect human health by emitting atmospheric pollutants.
Previous studies indicate that variations in the extent of burned area in Africa are attributable to vegetation composition and distribution, and air & soil controls on fuel drying. Human activities like cropland expansion and population growth also cause decline in African burned area since 1998. However, these factors cannot predict wildfire activities at the seasonal timescale. Previous seasonal climate forecasts resulted in insignificant correlations between the observed and predicted burned areas across most of Africa.
Since ocean sea-surface temperature (SST) and the land surface have longer memory than the atmosphere, the seasonal predictability of African wildfire may be improved if we focus on SST and terrestrial controls, instead of air controls. But past studies mainly focused on ENSO, ignoring the tropical Atlantic SSTs or the interactions between different modes of SST variability.
2. Method
In order to disentangle the vegetation impacts on fire from the fire impacts on vegetation, and the oceanic and land surfaces influences, the Stepwise Generalied Equilibrium Feedback Assessment (SGEFA) method was used. The SGEFA can disentangle the individual linear influences, but machine learning is necessary for building seasonal forecasting models.
The final results show that SST, leaf area index (LAI), and soil moisture are the most important predictors of African fire, and the machine learning model based on these predictors can effectively predict African fire activity 1 month ahead.
Key findings
I. SGEFA results
(1) Northern Africa's fire is sensitive to tropical Atlantic Ocean SST during the dry season (Nov-Mar), North Atlantic Ocean SST during the boreal winter (fire-active season), and tropical Indian Ocean SST during the wet season (Apr-Sep).
(2) Southern Africa's fire is sensitive to South Atlantic Ocean SSTs during the fire-active season (May-Nov/boreal summer).
(3) Soil moisture exerts important controls on the northern Africa fire during the wet season (Apr-Jan), and slightly smaller control on the southern Africa fire than the South Atlantic SSTs during the fire-active season (May-Nov).
(4) LAI is not an important control in Northern Africa. In Southern Africa, the impact of LAI was akin to that of the North Atlantic, and smaller than soil moisture or the South Atlantic.
(5) If one focuses on specific oceanic modes (ENSO, Atlantic Nino), the response to both of these modes are the highest during the fire-active season (boreal winter in Northern Africa, and boreal summer in Southern Africa). The fire-incuding SST anomalies are generally conducitve to warm and/or dry conditions.
(6) Soil moisture generally supresses fire through low-level cooling and elevated precipitation. LAI enhances fire in Northern Africa and the grasslands in Southern Africa, but supresses fire through surface cooling and reduced wind speed in most of Southern Africa.
II. Seasonal predictability
Fig. 3 shows the predictability of African wildfires by leading time. In both Northern and Southern Africa, the predictability at 1-month lead is significantly enhanced if one uses oceanic + terrestrial predictors, instead of atmospheric + socioeconomic predictors. Using all predictors resulted in only very slight improvements compared to using oceanic + terrestrial predictors.
Also, using season-specific predicting models resulted in much better performances than annual models.
3. Implications of the work
The work will aid the development of fire models, which currently can capture the spatial distribution of observed global fire, but not the seasonal or interannual variations.
The work also has some limitation caused by observational data availability (e.g. lightening, agricultural practice).
The results are good, outperforming past prediction models.
4. Data sources
The data sources are in their Supplemental Table 1. The fire carbon emissions and burned area fractions are from the Global Fire Emissions Database (GFED), which covers 1997-2016 at 0.25 degrees resolution. This is an observation-based analysis.
