Quantifying the drivers and predictability of seasonal changes in African fire

SGEFA-based quantification of environmental controls

The multivariate statistical approach, SGEFA, assesses the response of a rapidly changing atmospheric or ecological variable, such as fire carbon emission and burned area fraction (Supplementary Fig. 3), to a set of slowly changing environmental forcings, such as SST, soil moisture, or LAI^{17,18,24,25,26}. At timescale, τ (currently assigned to 1 month), which exceeds the persistence time of the target fire variable, a fire variable at time t, F(t), can be approximately decomposed into both the internal noise, N(t), and the response to a set of slowly evolving variables, O(t), in the forcing matrix, O, such that

R represents the response vector, which quantifies the instantaneous influence of slowly evolving oceanic and terrestrial forcings on a terrestrial flux. Multiplying the transposed forcing matrix at an earlier time, O^T(t − τ), on both sides of Eq. (1) and application of the covariance yield the following equation in covariance, C:

Because oceanic and terrestrial variability cannot be forced by, or drive, subsequent ecological internal noise, C_NO(τ) is theoretically equal to zero, allowing the feedback response vector to be estimated as

Because the present study focuses on the seasonal timescale, the seasonal cycle and long-term linear trend are removed from the forcing and response variables prior to applying SGEFA. The present analysis is conducted for the period from 1997 to 2016, corresponding to the coverage of the Global Fire Emissions Database (GFED)^4,38. The latest dataset, GFED4s, combines satellite information on burned area and small fire fraction with observations of vegetation productivity and meteorology to estimate gridded monthly burned area and fire emission. To increase the effective sample size, seasonal feedbacks are examined by aggregating data from three consecutive months at a time. To obtain more reliable estimates of the feedback response vector, relatively unimportant forcings are dropped from the forcing matrix to reduce the number of simultaneously considered forcings and thus minimize sampling error. This is performed via the backward-selection stepwise method³⁹ that optimizes the Akaike information criterion⁴⁰, an index that quantifies the quality of the statistical model by estimating the goodness of fit and penalizing based on the number of predictors.

In this study, the forcing matrix initially contains 16 oceanic forcings and two terrestrial forcings. The oceanic forcings consist of the principal component time series from the leading two empirical orthogonal functions (EOFs) of SSTs from eight basins, namely, the tropical Pacific (20 °S–20 °N, 120 °E–60 °W), tropical Atlantic (20 °S–20 °N, 70 °W–20 °E), tropical Indian (20 °S–20 °N, 35 °E–105 °E), North Pacific (20 °N–60 °N, 120 °E–100 °W), North Atlantic (20 °N–60 °N, 90 °W–10 °W), South Pacific (60 °S–20 °S, 150 °E–70 °W), South Atlantic (60 °S–20 °S, 70 °W–20 °E), and South Indian (60 °S–20 °S, 20 °E–120 °E) Oceans (as displayed in Yu et al.¹⁷, Supplementary Fig. 7). The terrestrial forcings for each target region include the time series of the area-averaged local LAI and surface-layer (0–10 cm) soil moisture. The SGEFA forcing matrix is computed using SSTs from the Hadley Center Sea Ice and Sea-Surface Temperature dataset, LAI from three satellite-based datasets, and surface-layer soil moisture from two observation- and reanalysis-based datasets, as outlined in Supplementary Table 1. The stepwise selection and subsequent feedback response vector estimation are performed for each fixed 3-month season (January–March, FMA, … December–February). The Monte Carlo bootstrap method, with 1000 random iterations of the time series of fire activity, is applied to assess the statistical significance of the SGEFA feedback triggered by a specific SST, soil moisture, or LAI forcing, checking for 90% confidence level (p < 0.1).

MLT-based prediction system of African fire

MLTs have been used for identifying empirical regulators of fire activity. For example, the random forest (RF) method has been applied to diagnose the emergent relationships between global burned area and environmental and anthropogenic factors in both observations and dynamic global vegetation models¹². Although MLTs lack the capability of quantifying environmental controls on fire variability²³, they provide powerful tools for building prediction systems and assessing the predictability as they account for nonlinear and interactive roles among predictors²⁷, which is particularly essential for fire prediction^12,22,23.

For this study, the MLT-based prediction system uses antecedent atmospheric and socioeconomic factors, as identified in previous studies^12,41,42, as well as the SGEFA-employed ocean and land surface variables. All predictors are listed in Supplementary Table 1 with their data sources, temporal coverages, and spatial resolutions. The oceanic and terrestrial predictors consist of the 16 oceanic and two terrestrial forcings used in the present SGEFA analysis. The atmospheric predictors include the occurrence of lightning and low-level atmospheric temperature, moisture, and wind speed. The socioeconomic predictors include population density and land use and land cover change. Based on data availability, the environmental predictors range from one to three months in advance of the prediction time, while the socioeconomic predictors consider only the more recent statistics typically reported at the end of the antecedent year.

In order to minimize the prediction uncertainty associated with the machine learning algorithms selected, this study examines five MLTs, including RF, support vector machine, artificial neural network, least absolute shrinkage and selection operator, and gradient boosting machine. These five algorithms differ substantially in their function. The combination of these algorithms is thus believed to better capture the complex interrelation between the forcings and response variable than any single algorithm. The 20-year data are randomly split into a 15-year training dataset and a 5-year testing dataset. The prediction model is fitted for each MLT using the training dataset, with parameters optimized for the minimum root-mean-square error via 10-fold cross-validation. The performance of the prediction model for all MLTs is evaluated using the correlation coefficient (R²) between the observed and predicted time series of fire emission or burned area fraction, while reporting the highest R² among all the currently used MLTs as the predictability of fire activity. The whole model fitting and evaluation procedure is repeated for 100 random iterations of data splitting, constituting a 100-member prediction model ensemble and checking for the uncertainty of the predictability associated with interannual variability of fire activity.

The assessed models include season-specific models, in which the MLTs are built and applied by season, and the all-season models, in which data from all seasons are used in the training and testing of the MLTs. For the all-season model, we perform an additional test of the robustness of MLT-based prediction. In the additional test, we randomly split the 20-year data (240 months) into a 180-month training dataset and 60-month testing dataset, regardless of year and season, and perform the same model fitting and validation analysis. The resulting performance of all-season models fitted from the 180-month training dataset with unbalanced sampling from each season is generally worse than the performance of models fitted from the 180-month training dataset with balanced sampling from each season. This additional test further confirms that environmental controls on African fire activity are seasonally dependent.

Source: Ecology - nature.com