TL;DR
This paper introduces a novel global fire emissions forecasting system that predicts monthly emissions 1-6 months in advance. By employing an ARIMAX statistical model tailored to over 1,300 distinct fire regions, the system leverages past fire observations and climate data to significantly improve prediction accuracy compared to simpler models. This advancement is crucial for better managing the increasing impacts of wildfires on climate, ecosystems, and human health.
What Problem Does This Paper Address?
Wildfires are a major source of atmospheric gases and aerosols, significantly impacting climate, ecosystem function, and human health. With a warming climate and expanding human populations in fire-prone areas, the frequency and intensity of wildfires are increasing, necessitating more effective management tools. While near-real-time fire weather forecasts and longer-term seasonal outlooks exist, there's a critical gap in operational systems that can provide reliable global fire emission predictions on subseasonal to seasonal (S2S) timescales (1-6 months ahead). Such a system is vital for anticipating fire extremes, understanding their impact on atmospheric CO2, and guiding fire and air quality management.
Methods
The core of this research is a global fire forecasting system built upon the Autoregressive Integrated Moving Average model with eXogenous variables (ARIMAX). This statistical time series model was applied independently to over 1,300 "fire cohesive regions" (FCRs) identified globally.
Key methodological aspects include:
- Region-Specific Modeling: Instead of a single global model, the ARIMAX model was optimized for each FCR, allowing for regional nuances in fire behavior and its drivers. FCRs were defined by progressively aggregating grid cells based on fire emission variability and continuity criteria.
- ARIMAX Components: The ARIMAX model incorporates:
- Autoregression (AR): Captures the dependence of current emissions on past emissions.
- Integration (I): Handles non-stationarity in the time series (e.g., long-term trends) through differencing.
- Moving Average (MA): Accounts for the influence of past forecast errors.
- Exogenous Variables (X): Incorporates external predictors like climate data.
- Predictor Selection: The system uses:
- Ocean Climate Indices (OCIs): 14 indices representing large-scale sea surface temperature patterns (e.g., ENSO-related indices).
- Vapor Pressure Deficit (VPD): A local meteorological parameter indicating atmospheric moisture stress, linked to fuel moisture.
- Lagged Predictors: The model optimizes the time lag between climate predictors and fire emissions to maximize correlation, recognizing that climate influences can manifest with a delay.
- Hyperparameter Optimization: For each FCR and forecast lead time, optimal ARIMAX orders (p, d, q) and predictor lags were determined using Akaike's Information Criterion (AIC) based on historical data.
Datasets
- Global Fire Emissions Database (GFED) Version 4s: Used as the primary target variable for monthly fire emissions (carbon emissions). Data from 1997-2014 were used for model development and training, and 2015-2018 for prediction and evaluation.
- Ocean Climate Indices (OCIs): 14 indices from NOAA, representing sea surface temperature anomalies in various ocean regions.
- MERRA-2 Reanalysis Data: Used to derive monthly Vapor Pressure Deficit (VPD) at a 1°x1° resolution.
The paper does not introduce new benchmark datasets but rather utilizes existing, well-established datasets for its modeling and validation efforts.
Key Insights and Ideas
- Regional Specificity is Crucial: A "one-size-fits-all" global model is insufficient. Tailoring the ARIMAX model to over 1,300 distinct fire regions significantly improves forecast skill by accounting for local fire dynamics and drivers.
- Past Fire Activity Matters: The autoregressive component (using past fire emissions) is a significant predictor, highlighting the "ecological memory" of fire regimes. This suggests that recent fire events and their anomalies contain valuable information for future predictions within the same season.
- Climate and Local Weather Synergies: Combining large-scale climate drivers (OCIs) with local meteorological conditions (VPD) yields better forecasts than using either alone, especially at longer lead times. This reflects the complex interplay of global climate patterns and immediate fire weather conditions.
- Lagged Relationships are Key: The optimal timing (lag) of climate predictors relative to fire emissions is critical and varies by region and predictor. This acknowledges that climate impacts on fuel moisture and availability are not instantaneous.
- ARIMAX as a Flexible Framework: The ARIMAX model proves to be a robust and adaptable tool for this S2S forecasting task, allowing for the integration of temporal dependencies (AR, MA), trend handling (I), and external influences (X).
Key Results
- 52% Variability Explained: The reference OCIVPDAR model (ARIMAX with OCI and VPD predictors) explained 52% of the variability in global fire emissions anomalies for a 1-month forecast lead time during the evaluation period (2015-2018).
- Superior to Persistence/Climatology: This performance significantly exceeded that of a persistence model (assuming current anomaly continues) and a climatology model (using historical monthly averages).
- Improved Accuracy: The OCIVPDAR model achieved the lowest RMSE and highest correlation with observed anomalies compared to other tested ARIMAX variations and simpler models.
- Forecast Skill Declines with Lead Time: As expected, forecast skill generally decreased with increasing lead times (from 1 to 6 months), but the ARIMAX models maintained considerable skill longer than simpler approaches.
- Spatial Patterns Captured: The system successfully resolved detailed spatial patterns of fire emissions anomalies in regions with significant fire activity.
Limitations and Future Work
- Human Drivers: The model implicitly accounts for some human impacts through regionalization and differencing but does not explicitly include human predictors (e.g., land use, policy). This is a significant challenge due to the complexity and rapid changes in human drivers.
- Data Latency and Uncertainty: The latency of OCI data and uncertainties in MERRA-2 VPD estimates can limit operational forecasting. Near-real-time data products may have even higher uncertainties.
- Model Structure Simplicity: The ARIMAX model, while flexible, is a statistical model. It may not fully capture complex physical processes or sudden shifts in fire regimes (e.g., due to policy changes).
- Ocean Teleconnection Optimization: Using only one OCI per region might not be optimal. Future work could explore using machine learning to identify optimal combinations of OCIs or to find more flexible proxies for large-scale climate influence.
- Dynamical Model Integration: The authors suggest exploring the integration of dynamical seasonal forecasts from climate models to improve predictions, especially for OCIs.
- Multivariate Regional Models: Extending the model to consider synergistic variability between nearby regions could further enhance forecasts.
- Higher Resolution Data: Utilizing finer-resolution fire detection data could help resolve within-region variability.
Summary updated