Quantifying the drivers and predictability of seasonal changes in African fire

Fires significantly impact African ecosystems, atmospheric composition, and human societies. Approximately 50% of fire-related carbon emissions and 70% of global burned areas originate from African subtropical savannahs. These fires influence atmospheric greenhouse gas concentrations and regional climate, affecting the hydrological cycle through aerosol emissions. Understanding the drivers of African fire evolution is crucial for accurate prediction and improved fire management practices. Fire activity in Africa is highly variable, influenced by weather patterns, vegetation, and human activity. Previous studies have noted a decline in burned area in northern Africa since 1998, attributed to both climatic and socioeconomic factors. While existing knowledge informs predictive models, seasonal prediction remains challenging due to limited predictability of relevant environmental and socioeconomic factors on seasonal timescales. Past studies have focused primarily on the El Niño-Southern Oscillation (ENSO) as an oceanic driver, neglecting other potential influences such as tropical Atlantic SSTs. Terrestrial factors, such as leaf area index (LAI) and soil moisture, also play complex roles, affecting both fuel supply and burning conditions. Challenges in disentangling the effects of fire on vegetation and the intercorrelation of oceanic and land surface anomalies necessitate advanced analytical techniques. This study uses a combined approach of SGEFA, for quantifying the impact of intercorrelated oceanic and terrestrial drivers, and MLTs, for capturing nonlinear effects and building prediction models.

Existing literature highlights the significant impact of African fires on various aspects of the Earth system. Studies have extensively explored the connections between fire activity and climate variability, demonstrating the influence of factors like ENSO and changes in vegetation and human activity. However, previous research has often focused on individual drivers or specific regions, lacking a comprehensive understanding of the interplay between oceanic, terrestrial, and human-induced factors. The use of seasonal climate forecasts for fire prediction has yielded limited success, indicating a need for a more refined approach that accounts for the slower-varying oceanic and terrestrial components. While previous studies have acknowledged the importance of soil moisture and LAI, a systematic exploration of their individual and synergistic contributions to seasonal fire predictability has been lacking. The challenge of disentangling the effects of fire on vegetation and the intercorrelation of oceanic and land surface anomalies has limited progress in understanding the underlying mechanisms.

This study employs a combined approach of Stepwise Generalized Equilibrium Feedback Assessment (SGEFA) and Machine Learning Techniques (MLTs). SGEFA, a lagged covariance statistical method, quantifies the impacts of oceanic and terrestrial drivers on regional climate by exploiting the differing memory timescales of forcing and response variables. It separates the linear contributions of intercorrelated oceanic, vegetation, and soil moisture forcings on seasonal timescales. This study applies SGEFA to assess the effects of sea-surface temperature (SST) from eight oceanic basins, leaf area index (LAI), and soil moisture on fire carbon emissions and burned area fraction across northern and southern Africa. The analysis uses data from 1997 to 2016 from sources such as GFED4s, Hadley Center Sea Ice and Sea-Surface Temperature dataset, and various satellite-based datasets for LAI and soil moisture. To enhance robustness, the stepwise selection method optimizes the Akaike information criterion. To account for nonlinear relationships and complex interactions, MLTs are employed to develop a seasonal fire prediction system. Five different MLTs (Random Forest, Support Vector Machine, Artificial Neural Network, Least Absolute Shrinkage and Selection Operator, and Gradient Boosting Machine) were used to create an ensemble. The 20-year dataset is randomly split into a 15-year training set and a 5-year testing set. Model parameters are optimized using 10-fold cross-validation, focusing on minimizing the root-mean-square error. The performance of the prediction models is evaluated using the correlation coefficient (R²) between observed and predicted fire emissions. The entire process is repeated for 100 random data splits to assess prediction uncertainty.

SGEFA analysis reveals that African fire activity exhibits significant sensitivity to SST, LAI, and soil moisture variations, especially during the dry, fire-active season. Northern African fires are sensitive to tropical Atlantic SSTs (dry season) and tropical Indian Ocean SSTs (wet season), while Southern African fires are more sensitive to Southern Hemisphere ocean SSTs. Soil moisture changes exhibit more robust effects than LAI changes, particularly during the northern African wet season. El Niño events generally enhance fire activity in northern and western Sahel, while the Atlantic Niño mode leads to increased fire activity across most of Africa. Positive LAI anomalies increase available fuel, potentially increasing fire activity, but this effect can be counteracted by associated surface cooling and reduced wind speeds. Including slowly evolving oceanic and terrestrial predictors significantly enhances the predictability of fire anomalies. The MLT-based approach demonstrates skillful prediction of African fire activity one month in advance (R² = 0.60 at 1 month lead time). The inclusion of oceanic and terrestrial predictors outperforms models using atmospheric and socioeconomic predictors alone. Seasonal predictability varies, with higher predictability in northern Africa during the fire-active season, except for August-October. Soil moisture is the most important predictor across seasons, reflecting its substantial impact on fire activity.

This study's findings directly address the research question by identifying key oceanic and terrestrial drivers of seasonal African fire variability and their contribution to predictability. The identification of SST, LAI, and soil moisture as primary drivers provides mechanistic insights into fire dynamics, going beyond simple correlations with atmospheric variables. The successful prediction of fire activity one month in advance demonstrates the practical utility of the combined SGEFA-MLT framework for fire management. The superior performance of models including oceanic and terrestrial predictors underscores the importance of considering slow-varying environmental drivers. The identified seasonal variability in predictability suggests the need for season-specific prediction models. The framework developed in this study offers a valuable tool for improving fire risk assessments and informing fire management strategies in Africa.

This study presents a novel combined SGEFA-MLT analytical framework for quantifying drivers and predicting seasonal changes in African fire. The framework successfully identifies key oceanic and terrestrial drivers (SST, LAI, soil moisture) and achieves skillful one-month-ahead fire predictions. Future research could explore the role of finer-scale factors, improve data quality, especially for lightning and socioeconomic factors, and apply the framework to other fire-prone regions. This research provides a crucial basis for building a global fire early warning system.

The study's conclusions are based on a limited number of MLTs and relatively short datasets, potentially introducing uncertainty in the fire predictability estimates. The limited availability of high-resolution data on lightning and socioeconomic factors might underestimate their true impact. The focus on area-averaged fire activity in broad ecoregions might oversimplify the complex drivers and predictability at finer spatial scales. Improved data availability and resolution would enhance the accuracy and reliability of the findings and allow for more nuanced analyses.