the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
Reconstructed global monthly burned area maps from 1901 to 2020
Abstract. Fire is a key Earth System process, driving variability in the global carbon cycle through CO2 emissions into the atmosphere and subsequent CO2 uptake through vegetation recovery after fires. Global spatiotemporally consistent datasets on burned area are available since the beginning of the satellite era in the 1980s but are sparse prior to that date. In this study, we reconstructed global monthly burned area at a resolution of 0.5°×0.5° from 1901 to 2020 using machine learning models trained against satellite observed burned area between 2003 and 2020, with the goal of reconstructing long-term burned area information to constrain historical fire simulations. We first conducted a classification model to separate grid cells with extreme (burned area > the 90th percentile in a given region) and regular fires, and then trained separate regression models for grid cells with extreme or regular fires. Both the classification and regression models were trained against a satellite-based burned area product (FireCCI51) based on explanatory variables related to climate, vegetation, and human activities. The trained models can well reproduce the long-term spatial patterns (slopes = 0.70–1.28 and R2 = 0.75–0.98 spatially), inter-annual variability and seasonality of the satellite-based burned area observations. After applying the trained model to the historical period, the predicted annual global total burned area ranges from 3.46 to 4.58 million km2 yr-1 (M km2 yr-1) over 1901–2020 with regular and extreme fires accounting for 1.36–1.74 and 2.00–3.03 M km2 yr-1 respectively. Our models estimate a global decrease in burned area during 1901–1978 (slope = -0.009 M km2 yr-2), followed by an increase during 1978–2008 (slope = 0.020 M km2 yr-2) and then a stronger decline in 2008–2020 (slope = -0.049 M km2 yr-2). Africa was the continent with largest burned area globally during 1901–2020, and its trends also dominated the global trends. We validated our predictions against charcoal records, and our product exhibits a high overall accuracy in fire occurrence (>80 %) in boreal North America, southern Europe, South America, Africa and southeast Australia, but the overall accuracy is relatively lower in northern Europe and Asia (<50 %). In addition, we compared our burned area data with multiple independent regional burned area maps in Canada, USA, Brazil, Chile and Europe, and found general consistency in the spatial patterns (linear regression slopes ranging 0.84–1.38 spatially) and the inter-annual variability. The global monthly 0.5°×0.5° burned area fraction maps from 1901 to 2020 presented by this study can be freely downloaded from https://doi.org/10.5281/zenodo.14191467 (Guo and Li, 2024).
- Preprint
(3217 KB) - Metadata XML
-
Supplement
(4985 KB) - BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on essd-2024-556', Anonymous Referee #1, 24 Mar 2025
Guo et al. have developed a novel machine learning model that can reconstruct the global monthly burned area at a spatial resolution of 0.5°×0.5° from 1901 to 2020. This model can be used to provide a benchmark for historical simulations of fire modules in Dynamic Global Vegetation Models (DGVMs). This approach employs various machine learning models to distinguish between extreme large fires and regular fires, using climate, vegetation, and human activities as explanatory variables and satellite-based burned area (FireCCI51) as the target variable to build the models. The results of model show high accuracy in some regions when compared with charcoal records. The manuscript is well-written and the results are effectively presented. In general the work is worthy publication in ESSD. I have only minor comments mainly for clarification.
Regarding the methods:
- Machine learning is a data-driven model, and thus the selection of data, especially the choice of explanatory variables, is crucial for model training. In this paper, variables related to climate, vegetation, and human activities were selected, and feature selection was employed to screen these variables. However, lightning, particularly cloud-to-ground lightning with sufficient energy, which is an important ignition source, was not included as an explanatory variable in the training of the machine learning model. Additionally, large-scale climate forcing, such as sea-surface temperature, can dominate extreme fire activity and the seasonality of fire activity. I would like to know the authors’ thoughts on whether these data were excluded due to insufficient temporal coverage or were screened out through feature selection.
- The terrain can affect the spread of fire. However, among all the explanatory variables listed in Table 1 and Figure S8 in the article, there was no information on terrain. I guess this might be due to limitations in the availability of data?
- It is unclear for me whether you do training-validation separately for each GFED region or you just do training-validation for NHAF and then apply the model globally? I.e., did you build a single model by using NHAF followed by its application everywhere? Or each region has its own model?
- Why not consider using different types of LM model for different GFED regions if we can pick a best type of model for each region? Is this because of computation resource limitation?
- I understand that two types of ML models were built: classification model for extreme fire grid cell and regression model to predict BAF. My question is, if a certain grid cell was classified as an extreme fire grid cell, how its BAF was determined? The grid cells with greater than 90th quantile was classified as extreme fire but still, we need to know its specific BAF?
- Although fire is prevalent, I guess there are many zero-BAF grid cells compared with relatively small number of grid cells with BAF>0? Did you encounter the issue of imbalanced sample size? Like there are many pixels without fire but only a small fraction with fire, will this have an impact on the model building?
- Line 172-175: could the authors give more details on how model parameter optimization was made and which parameters have been optimized? This part is interesting. Is it a 5-fold cross validation or a circular process?
- Line 176: After the 80%-20% 5-fold CV, here I think we call it ‘model evaluation’, which is better than ‘model validation’. We cannot really ‘validate’ a model.
- Line 183-184: here is confusing. I get confused by whether the model was validated by 80%-20% as in line 172-175 or by leave-one-year-out?
Other minor comments:
- On lines 161–162 of the manuscript, the authors stated that LSTMs have the best performance among all the machine learning models used in this study, and this result was presented in Fig. 2k and Fig. S7. As far as I understand, Fig. 2 was intended to display the results of LSTMs, but the description in the figure title was unclear. I would like to see a clearer statement and further explanation of what the “absolute and relative difference” mean in Fig. 2b.
Citation: https://doi.org/10.5194/essd-2024-556-RC1 -
RC2: 'Comment on essd-2024-556', Anonymous Referee #2, 09 Apr 2025
This study reconstructed global monthly burned area from 1901 to 2020 at 0.5°×0.5° resolution using machine learning models trained on satellite data (2003–2020). Separate models were developed for extreme and regular fires based on climate, vegetation, and human activity data. The models accurately captured spatial patterns, seasonal trends, and long-term changes in burned area. Results show a global decline in burned area from 1901–1978, an increase from 1978–2008, and a stronger decline from 2008–2020. The reconstruction aligns well with charcoal records, offering a valuable tool for historical fire analysis.The manuscript is well written, and the simulation work—including data preparation, model selection, training, validation, and regional application—is rigorously conducted. It can be shown that the authors put quite a lot effort to build such machine learning and data driven pipeline. The use cases from such data asset construction are wide. The results are clearly presented through both text and figures, and the discussion addresses the burned area distributions, trends, limitations, and uncertainties. However, I have a few specific comments regarding the methods that I believe should be clarified in more detail in the Methods section. Therefore, I recommend a minor revision to further strengthen the manuscript prior to publication.Specific comments:#1 Line 95 Please add more content / citations on how the burned area on cropland is excluded to eliminate agricultural fires in this study.#2 Line 120 Please clarify how the reclassification is done to split them into five land use types (forest, shrub, natural grass, cropland and others). Was it via machine learning classifier?#3 Line 148 Is there reasoning why an initial classification has to be conducted to classify regions into non-BAF, moderate and extreme BAF? Why couldn't the parameterization be applied to all regions without this initial classification. Conducting such classification could introduce extra errors / uncertainties. Please add more content to clarify.#4 Line 152 I am not quite clear why 90th percentile instead of other percentiles is used to determine the extreme BAF area. Please explain and clarify.#5 Line 167 From the content there are only 16 features plus some others in NHAF are used in the model. I am not sure why a recursive feature selection has to be performed. Recursive feature selection is usually applied when there is 1000+ variables used in GBM models. This process is necessary because the larger model package size would directly cause latency issues in the live production env. In this case 16+ variables won't cause such issues plue the model will not be deployed onto live env. Is such feature selection necessary?#6 Figure 2 For the global map of burned area difference please please make it a separate figure as it is hard to see the small dots. And what is the unit? Is it a relevant difference? Please specify in the figure.#7 Figure 2 What does N mean here? Is it the data sample size for validation grids per year? Please specify.#8 Figure 3 It looks like figure (a) (c) (e) vs figure (b) (d) (f) are giving the same messages. Also it is hard to find any insight from figure (a) (c) (e). Consider removing or merging them with (b) (d) (f).#9 Figure 5 It is hard to tell the accuracy difference in (a) (b) as most of them are dark red (are most of them 100% or 70%?). Please consider using a different color scale to make them more distinguishable.#10 Figure 6 Again, (a) (b) (c) convey important messages of spatial distribution of burned areas. but they are too small in terms of figure size. Please make them bigger / clearer.Citation: https://doi.org/
10.5194/essd-2024-556-RC2
Data sets
Reconstructed Global Monthly Burned Area Maps from 1901 to 2020 Zhixuan Guo and Wei Li https://doi.org/10.5281/zenodo.14191467
Viewed
HTML | XML | Total | Supplement | BibTeX | EndNote | |
---|---|---|---|---|---|---|
385 | 75 | 5 | 465 | 20 | 7 | 6 |
- HTML: 385
- PDF: 75
- XML: 5
- Total: 465
- Supplement: 20
- BibTeX: 7
- EndNote: 6
Viewed (geographical distribution)
Country | # | Views | % |
---|
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1