the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 04 Dec 2024
A statistical global burned area model for seamless integration into Dynamic Global Vegetation Models
Abstract. Fire-enabled Dynamic Global Vegetation Models (DGVMs) play an essential role in predicting vegetation dynamics and biogeochemical cycles amid climate change. Modeling wildfires has been challenging in process-based biophysics-oriented DGVMs, as human behaviour plays a crucial role. This study aims to reveal a global statistical model for the relationships between biophysical and socioeconomic drivers of wildfire dynamics and monthly burned area (BA) that can be integrated into DGVMs. We developed Generalised Linear Models (GLMs) to capture the relationships between potential predictor variables that are simulated by DGVMs and/or available in future scenarios and the latest global burned area product from GFED5. Predictor variables were chosen to represent aspects of fire weather, vegetation structure and activity, human land use and behaviour and topography. The final model was chosen by minimizing collinearity and by maximizing model performance in terms of reproducing observations. The final model included eight predictor variables encompassing the Fire Weather Index (FWI), a novel Gross Primary Productivity Index (GPPI), Human Development Index (HDI), Population Density (PPN), Percentage Tree Cover (PTC), Percentage Non-Tree Cover (PNTC), Number of Dry Days (NDD), and Topographic Positioning Index (TPI). Given its simplicity, our model demonstrated a remarkable capability, explaining 56.8 % of the burnt area variability, comparable to other state-of-the-art global fire models. FWI, PTC, TPI and PNTC were positively related to BA, while GPPI, HDI, PPN, and NDD were negatively related to wildfire. While the model effectively predicted the spatial distribution of burned areas (NME = 0.72), its standout performance lay in capturing the seasonal variability, especially in regions often characterized by distinct wet and dry seasons, notably southern Africa, Australia and parts of South America (R2 > 0.50). Our model reveals the robust predictive power of fire weather and vegetation dynamics emerging as reliable predictors of seasonal global fire patterns. The presented model should be compatible with most, if not all, DGVMs used to develop future scenarios.
-
Notice on discussion status
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
-
Preprint
(2149 KB)
-
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
- Preprint
(2149 KB) - Metadata XML
- BibTeX
- EndNote
- Final revised paper
Journal article(s) based on this preprint
Interactive discussion
Status: closed
-
RC1: 'Comment on egusphere-2024-3595', Anonymous Referee #1, 27 Jan 2025
- AC1: 'Reply on RC1', Blessing Kavhu, 16 Mar 2025
-
RC2: 'Comment on egusphere-2024-3595', Anonymous Referee #2, 30 Jan 2025
General comment
The manuscript entitled “A statistical global burned area model for seamless integration into Dynamic Global Vegetation Models” by Blessing Kavhu and colleagues develops a Generalized Linear Model (GLM) with 19 predictors. The authors designed and tested 26 models using burned area data from the Global Fire Emissions Database version 5 (GFED5) and combinations of the selected predictors. Model 25 was chosen as the best-performing model, with an explained deviance of 0.568 and a Normalized Mean Error (NME) of 0.718. The authors identified key predictors such as Fire Weather Index (FWI), and Percentage Non-Tree Cover (PNTC), which strongly influence fire occurrence, and Human Development Index (HDI), Gross Primary Productivity Index (GPPI), and Population Density (PPN), which are negatively associated with fire occurrence. While the model demonstrates limited accuracy in predicting global annual burned area variability (Figure 5), it performs well in capturing global seasonal variability (Figure 8). The authors also discussed the comparison between predicted and observed data in terms of spatio-temporal variability at the GFED regional level.
In general, I have concerns regarding the alignment of the manuscript’s title with its methods and objectives. The current title suggests that the authors developed statistical models (GLMs) seamlessly integrated into Dynamic Global Vegetation Models (DGVMs). However, upon reviewing the manuscript, it becomes clear that the GLM was built independently of any DGVM, and the integration is only theoretically explained. According to my understanding, true integration with DGVMs requires a lot and long modification processes, testing within specific DGVM frameworks, reparameterization, new input-output verification, module integration, and validation process. The integration process involves technical adjustments such as modifying input data formats, calibrating modules, and ensuring compatibility with existing model components (e.g., physical, physiological, vegetation, or disturbance, and biogeochemical modules). Without actual implementation and demonstrated results, the claim of "seamless integration" remains unsupported. I suggest revising the title to reflect the study's scope and contributions more accurately. For example, the title could emphasize the development of a GLM, its evaluation of wildfire drivers, and its ability to predict spatio-temporal variability in burned area data.
Additionally, the study workflow needs to be presented more systematically. I recommend referencing workflows from published manuscripts in this field and ensuring that critical methodological details, such as data sources, temporal coverage of input data, and prediction periods, are clearly outlined. The abstract section is also not structurally strong enough, it should be rearranged. A clear and detailed workflow will greatly aid readers in understanding the study's methodology. Furthermore, the term "prediction" should be adjusted to "historical prediction" to reflect the study’s temporal scope (2002–2018).
Overall, I recommend major revision before this manuscript can be considered for publication in Biogeosciences. Addressing the points mentioned above, along with detailed reviewer's comments, will significantly enhance the manuscript's clarity and alignment with its objectives. Please find detailed comments in the supplement.
- AC2: 'Reply on RC2', Blessing Kavhu, 16 Mar 2025
Interactive discussion
Status: closed
-
RC1: 'Comment on egusphere-2024-3595', Anonymous Referee #1, 27 Jan 2025
egusphere-2024-3595
General comments:
The authors developed the statistical simulation model for fire burnt area which was supposed to be incorporated into DGVM. The proposed GLMs as the functions of multiple environmental factors, would give the potential for a very light simulation tool on global scale wildfire.
However, one of my concerns is that the wildfire would be ignited considerably by lighting and expand their burning severity depending on the dry litter amount as fuel, both of which are not considered this time. The predictability for the long-term future is not secured as long as you do not use an accumulated ecosystem carbon stock, which DGVM has the advantage to simulate, but you use the GPP related index which is not necessarily related to fire severity. Statistical models cannot guarantee accuracy when a regime-shift for fire occurrence has happened mechanistically.
Considering the fewer predictors selected for the general representation of wildfire, the linear regression is shown not to be the best model for wildfire though still easy to use compared to machine learning or process-based models. Correlation coefficients for each term should be rearranged when this model was introduced to DGVM to use their simulated GPP and other ecosystem information which must be more or less different from MODIS or satellite products.
CEAM, BONA, MIDE, and TENA showed a 2-fold difference in interannual variations, which suggests that this model could be parameterized or separately made for vegetation types. A single statistical model cannot estimate the average BA for specific areas.
Minor Comments:
Page 3, Lines 90-91, you have to identify the typical model name that you criticize here. What you propose here is still a simple GLM based module, so I feel this sentence as contradiction
Page 3, Line 94-95, Use of remote sensing data will reduce the advantage of DGVM which enables the longterm vegetation shift simulation in future. also wanna now th direction from start to end
Page 17, Line 330 does this mean that HDI is going down in these years? you would better show the number in the decadal trend of averaged globally
Page 22, Lines 425-427: sounds repeated. you merge these two sentences into one.
Page 23, Lines 463: you also have to mention about the SPITFIRE-based model performance. GlobFIRM is an old version, and we know this is not accurate
Figure 1: you should add more explanation on the shape of frames, rectangles, rounds, diamonds. what is the data and what is the process
Table 2: explain the condition for color
Figure 4: Specify the years for the average
Figure 5: The y-axis should be in 10^6 to reduce the number of digits.
Citation: https://doi.org/10.5194/egusphere-2024-3595-RC1 - AC1: 'Reply on RC1', Blessing Kavhu, 16 Mar 2025
-
RC2: 'Comment on egusphere-2024-3595', Anonymous Referee #2, 30 Jan 2025
General comment
The manuscript entitled “A statistical global burned area model for seamless integration into Dynamic Global Vegetation Models” by Blessing Kavhu and colleagues develops a Generalized Linear Model (GLM) with 19 predictors. The authors designed and tested 26 models using burned area data from the Global Fire Emissions Database version 5 (GFED5) and combinations of the selected predictors. Model 25 was chosen as the best-performing model, with an explained deviance of 0.568 and a Normalized Mean Error (NME) of 0.718. The authors identified key predictors such as Fire Weather Index (FWI), and Percentage Non-Tree Cover (PNTC), which strongly influence fire occurrence, and Human Development Index (HDI), Gross Primary Productivity Index (GPPI), and Population Density (PPN), which are negatively associated with fire occurrence. While the model demonstrates limited accuracy in predicting global annual burned area variability (Figure 5), it performs well in capturing global seasonal variability (Figure 8). The authors also discussed the comparison between predicted and observed data in terms of spatio-temporal variability at the GFED regional level.
In general, I have concerns regarding the alignment of the manuscript’s title with its methods and objectives. The current title suggests that the authors developed statistical models (GLMs) seamlessly integrated into Dynamic Global Vegetation Models (DGVMs). However, upon reviewing the manuscript, it becomes clear that the GLM was built independently of any DGVM, and the integration is only theoretically explained. According to my understanding, true integration with DGVMs requires a lot and long modification processes, testing within specific DGVM frameworks, reparameterization, new input-output verification, module integration, and validation process. The integration process involves technical adjustments such as modifying input data formats, calibrating modules, and ensuring compatibility with existing model components (e.g., physical, physiological, vegetation, or disturbance, and biogeochemical modules). Without actual implementation and demonstrated results, the claim of "seamless integration" remains unsupported. I suggest revising the title to reflect the study's scope and contributions more accurately. For example, the title could emphasize the development of a GLM, its evaluation of wildfire drivers, and its ability to predict spatio-temporal variability in burned area data.
Additionally, the study workflow needs to be presented more systematically. I recommend referencing workflows from published manuscripts in this field and ensuring that critical methodological details, such as data sources, temporal coverage of input data, and prediction periods, are clearly outlined. The abstract section is also not structurally strong enough, it should be rearranged. A clear and detailed workflow will greatly aid readers in understanding the study's methodology. Furthermore, the term "prediction" should be adjusted to "historical prediction" to reflect the study’s temporal scope (2002–2018).
Overall, I recommend major revision before this manuscript can be considered for publication in Biogeosciences. Addressing the points mentioned above, along with detailed reviewer's comments, will significantly enhance the manuscript's clarity and alignment with its objectives. Please find detailed comments in the supplement.
- AC2: 'Reply on RC2', Blessing Kavhu, 16 Mar 2025
Peer review completion
Journal article(s) based on this preprint
Data sets
Data for fitting a statistical global burned area model for seamless integration into Dynamic Global Vegetation Models Blessing Kavhu, Matthew Forrest, and Thomas Hickler https://doi.org/10.5281/zenodo.14110150
Model code and software
A statistical global burned area model for seamless integration into Dynamic Global Vegetation Models(Submission release v1) Blessing Kavhu, Matthew Forrest, and Thomas Hickler https://doi.org/10.5281/zenodo.14177016
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 782 | 134 | 30 | 946 | 32 | 53 |
- HTML: 782
- PDF: 134
- XML: 30
- Total: 946
- BibTeX: 32
- EndNote: 53
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Matthew Forrest
Thomas Hickler
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
- Preprint
(2149 KB) - Metadata XML
egusphere-2024-3595
General comments:
The authors developed the statistical simulation model for fire burnt area which was supposed to be incorporated into DGVM. The proposed GLMs as the functions of multiple environmental factors, would give the potential for a very light simulation tool on global scale wildfire.
However, one of my concerns is that the wildfire would be ignited considerably by lighting and expand their burning severity depending on the dry litter amount as fuel, both of which are not considered this time. The predictability for the long-term future is not secured as long as you do not use an accumulated ecosystem carbon stock, which DGVM has the advantage to simulate, but you use the GPP related index which is not necessarily related to fire severity. Statistical models cannot guarantee accuracy when a regime-shift for fire occurrence has happened mechanistically.
Considering the fewer predictors selected for the general representation of wildfire, the linear regression is shown not to be the best model for wildfire though still easy to use compared to machine learning or process-based models. Correlation coefficients for each term should be rearranged when this model was introduced to DGVM to use their simulated GPP and other ecosystem information which must be more or less different from MODIS or satellite products.
CEAM, BONA, MIDE, and TENA showed a 2-fold difference in interannual variations, which suggests that this model could be parameterized or separately made for vegetation types. A single statistical model cannot estimate the average BA for specific areas.
Minor Comments:
Page 3, Lines 90-91, you have to identify the typical model name that you criticize here. What you propose here is still a simple GLM based module, so I feel this sentence as contradiction
Page 3, Line 94-95, Use of remote sensing data will reduce the advantage of DGVM which enables the longterm vegetation shift simulation in future. also wanna now th direction from start to end
Page 17, Line 330 does this mean that HDI is going down in these years? you would better show the number in the decadal trend of averaged globally
Page 22, Lines 425-427: sounds repeated. you merge these two sentences into one.
Page 23, Lines 463: you also have to mention about the SPITFIRE-based model performance. GlobFIRM is an old version, and we know this is not accurate
Figure 1: you should add more explanation on the shape of frames, rectangles, rounds, diamonds. what is the data and what is the process
Table 2: explain the condition for color
Figure 4: Specify the years for the average
Figure 5: The y-axis should be in 10^6 to reduce the number of digits.