Machine learning-based observation-constrained projections reveal elevated global socioeconomic risks from wildfire

Wildfires pose significant threats to ecosystems, the global carbon budget, climate, and human well-being. Recent catastrophic fire events have resulted in substantial social disruption and economic losses, underscoring the need for accurate wildfire projections. Anthropogenic activities, including climate change, land-use changes, and fire ignition patterns, have significantly altered wildfire behavior and associated risks. Earth system models (ESMs) offer a potential approach to project future wildfire changes and socioeconomic impacts, yet they face limitations in capturing complex human-vegetation-fire-climate feedbacks, leading to uncertainties in their outputs. Previous studies using fire weather as a proxy or applying simpler bias-correction methods have limitations due to the multifaceted nature of wildfire dynamics. This research proposes a machine learning-based framework to address these challenges by incorporating observational data to constrain ESM-simulated wildfire projections and subsequently assess the socioeconomic risks more reliably.

Existing literature highlights the significant impacts of wildfires, the limitations of current ESMs in accurately capturing fire dynamics, and the challenges in projecting future wildfire regimes and their socioeconomic impacts. While ESMs offer a potential solution, their uncertainties related to the human-vegetation-fire-climate feedback complicate reliable long-term predictions. Previous approaches using fire weather as an emulator or simpler bias-correction techniques have proven inadequate due to incomplete consideration of factors such as terrain, fuel conditions, and ignition sources. The emergent constraint (EC) approach has shown promise in reducing uncertainties in other Earth system variables, but its application to spatially complex wildfire projections is limited due to insufficient model diversity and the nonlinear nature of fire-related processes. This study builds upon these prior efforts by leveraging the strengths of machine learning to handle these complexities.

This study develops a machine learning-based framework to generate observation-constrained projections of global fire carbon emissions and socioeconomic risks. Thirteen CMIP6 ESMs, providing coupled carbon-ecosystem-climate simulations, are utilized. The framework leverages machine learning techniques (random forest, support vector machine, and gradient boosting machine) to establish emergent relationships between projected fire carbon emissions and historical, observed climate, terrestrial ecosystem, and socioeconomic states. The models are trained using historical and future simulations from the SSP5-85 scenario, capturing the complete spatial patterns simulated by the ESMs. Observed historical data from multiple sources, including fire carbon emissions, leaf area index, soil moisture, temperature, precipitation, wind, relative humidity, flash rate, orography, land use, and population, are then fed into the trained models to generate observation-constrained projections. Socioeconomic risks are quantified using the default and observation-constrained ensemble projections of fire carbon emissions in conjunction with SSP5-85 projections of population, GDP, and agricultural area. The framework is validated using historical data from 2007-2016, comparing the observation-constrained simulations with observed wildfire activity. The robustness and efficiency of the methodology are evaluated using different machine learning algorithms and exploring the sensitivity of the framework to the spatial resolution of input data. The importance of drivers is further explored using feature importance scores from the machine learning models, separating the contributions of historical distributions and projected trends to wildfire patterns and trends, respectively. The methodology is detailed in the supplementary information.

The machine learning-based observational constraint substantially improved the agreement between simulated and observed wildfire activities during the validation period (2007-2016), reducing biases in various regions and improving spatial correlation. The observation-constrained ensemble projected a lower increase in global fire carbon emissions (4.1% per decade) compared to the default ensemble (6.0% per decade) for the twenty-first century. However, the observation-constrained ensemble projected a greater increase in global wildfire exposure across population, GDP, and agricultural area (5.5%, 40.6%, and 2.5% per decade, respectively) compared to the default ensemble (3.2%, 12.6%, and 1.8% per decade, respectively). The elevated socioeconomic risks are primarily attributed to increased wildfire activity and rapid socioeconomic development in western and central Africa. Specifically, the observation-constrained ensemble projects increased wildfire activity over the Amazonian and Congo Basins, contrasting with the default simulations. The importance scores from the machine learning models indicate the significant role of fuel abundance (LAI, temperature, precipitation) and fuel moisture (soil moisture, relative humidity, precipitation, temperature) in driving future wildfire trends. The projected fire regimes and their associated socioeconomic risks are shown to be highly dependent on the socioeconomic pathway used, with a lower-emission scenario (SSP2-45) exhibiting different patterns. Figures 1-4 in the paper provide detailed visualizations of the findings, comparing the original and observation-constrained results for historical and future wildfire patterns and socioeconomic exposure.

The findings highlight the importance of incorporating observational data into ESMs to reduce uncertainties in wildfire projections and improve the assessment of socioeconomic risks. The machine learning-based approach effectively captures the nonlinear and complex interactions between wildfire activity and its driving factors. The observation-constrained ensemble reveals a more nuanced picture of future wildfire risks, indicating higher risks in specific regions, such as western and central Africa, than previously anticipated. The enhanced accuracy of historical wildfire simulations obtained through observational constraints is crucial for improving future projections. These results underline the need for targeted mitigation and adaptation strategies in vulnerable regions, particularly those experiencing rapid socioeconomic development. Future research should focus on improving ESMs' representation of wildfire processes, particularly the integration of sub-grid topographic factors and more detailed socioeconomic interactions.

This study demonstrates the effectiveness of a machine learning-based framework for generating observation-constrained projections of global wildfire carbon emissions and associated socioeconomic risks. The approach improves the accuracy of historical simulations and reveals a higher level of future risk than previously projected, particularly in western and central Africa. The findings emphasize the critical need for wildfire preparedness and adaptation strategies in rapidly developing regions. Further research should focus on refining ESMs to better represent complex wildfire dynamics and integrate higher-resolution data.

Several limitations exist in this study. The uncertainty in observational data, particularly for lightning and socioeconomic variables, could influence the observational constraint. The use of relatively coarse-resolution satellite measurements of burned area may underestimate the total burned area and associated carbon emissions, particularly in Africa. The incomplete independence of the analyzed ESMs (some share the same land component) limits the multimodel spread and increases the dependency of results on shared modeling components. The approach does not explicitly account for complex feedbacks from socioeconomic development to wildfire regimes beyond the parameterized socioeconomic drivers in the ESMs, or for potential tipping points in fire regime evolution. The spatial resolution of the input ESM data also affects the performance of the machine learning models.