TL;DR
This paper reconstructs global monthly burned area maps from 1901 to 2020 using machine learning, filling a critical data gap prior to the satellite era. These new maps provide a valuable benchmark for understanding historical fire regimes, their impact on the carbon cycle, and for improving Earth system models.
What Problem Does This Paper Address?
Accurate global datasets of burned area are crucial for understanding fire's role in the Earth system, particularly its impact on the carbon cycle through CO2 emissions and subsequent vegetation recovery. While satellite data provides good coverage since the 1980s, historical data before this period is sparse. Existing fire models (DGVMs) can simulate long-term trends but often lack the spatial resolution and accuracy to capture global patterns. This study aims to bridge this gap by creating a consistent, high-resolution (0.5° x 0.5°) global burned area dataset spanning over a century (1901-2020), leveraging machine learning trained on recent satellite observations.
Methods
The authors developed a novel machine learning approach to reconstruct burned area. First, they trained a classification model to distinguish between grid cells experiencing "extreme" fires (defined as exceeding the 90th percentile of burned area within a region) and "regular" fires. Subsequently, they trained separate regression models for each category (extreme and regular fires) to predict the burned area fraction (BAF). These models were trained using a suite of explanatory variables including climate data (temperature, precipitation, wind, vapor pressure deficit, fire weather index), vegetation characteristics (leaf area index), and anthropogenic factors (population density, land use/land cover). Long Short-Term Memory (LSTM) networks were found to perform best for the regression tasks, particularly for capturing temporal dynamics. The models were trained on the FireCCI51 satellite product (2003-2020) and then applied to reconstruct the historical period.
Datasets
- Training Data: FireCCI51 (2003-2020) - a satellite-based monthly global burned area product.
- Explanatory Variables:
- Climate: CRUJRA v2.2 (daily temperature, precipitation, wind speed), calculated VPD, and FWI.
- Vegetation: GIMMS LAI4g (monthly leaf area index), and harmonized LAI data for the pre-satellite era derived from DGVM simulations (TRENDY v11).
- Anthropogenic: HYDE3.2 (population density), ESA CCI land cover maps (1992-2020), and LUH2-GCB2020 (pre-1992 land use data).
- Validation Datasets:
- Charcoal Records: Global Charcoal Database and Reading Palaeofire Database.
- Other Burned Area Datasets: FireCCILT11, Global Annual Burned Area Maps (GABAM), regional datasets from Canada, USA, Brazil, Chile, Europe, Australia (New South Wales, Queensland, South Australia, Western Australia).
Key Insights and Ideas
- Distinguishing Extreme Fires is Crucial: The study highlights that simply modeling all fires together can lead to underestimation, as abundant "regular" fire data can mask the impact of rarer, extreme events. Separating these two categories significantly improves the representation of fire dynamics.
- LSTMs Excel at Temporal Dynamics: Long Short-Term Memory (LSTM) networks proved superior for predicting burned area, likely due to their ability to capture complex temporal dependencies in the input variables over multiple time steps.
- Africa Dominates Global Burned Area Trends: Africa consistently shows the largest burned area globally, and its trends significantly influence the overall global patterns.
- Charcoal Records Offer Broad Validation: While not perfect, charcoal records provide a valuable, independent source for validating fire occurrence predictions, especially in regions with limited historical data.
Key Results
- Reconstructed Burned Area: The study produced global monthly burned area maps at 0.5° x 0.5° resolution for 1901-2020.
- Global Burned Area Trends: The reconstructed data reveals a decrease in global burned area from 1901-1978, followed by an increase from 1978-2008, and then a stronger decline from 2008-2020.
- Model Performance: The trained models demonstrated good performance in reproducing spatial patterns (slopes = 0.70-1.28, R² = 0.69–0.98) and inter-annual variability of satellite observations.
- Accuracy with Charcoal Records: High overall accuracy (>80%) was achieved in boreal North America, southern Europe, South America, Africa, and southeast Australia, though lower accuracy (<50%) was observed in northern Europe and Asia.
- Comparison with Other Datasets: The reconstructed data showed general consistency with various regional and global burned area datasets, particularly for Brazil and Chile.
Limitations and Future Work
- Data Harmonization Uncertainties: The reconstruction relies on harmonizing satellite data with older datasets (e.g., land use, LAI), introducing uncertainties due to differing definitions, resolutions, and potential biases in cross-walking tables.
- Limited Representation of Human Intervention: While population density was included, explicit representation of fire suppression or prescribed burning practices was limited due to data availability, potentially explaining overestimations in some regions like the US and Europe.
- Landscape Fragmentation: The impact of landscape fragmentation on fire regimes was not explicitly considered due to a lack of long-term, high-resolution data.
- Extrapolation Uncertainty: Applying models trained on recent data to a much longer historical period inherently involves extrapolation, assuming that the relationships between drivers and fire remain consistent over time.
- Future Work: The authors suggest that more reliable explanatory data for model training and validation would further improve accuracy. They also produced additional versions of the dataset (GFED5-based and FireCCI51-GDP-based) to explore the impact of different input data and calibration methods.
Summary updated