Dynamic savanna burning emission factors based on satellite data using a machine learning approach

Vernooij, Roland; Eames, Tom; Russell-Smith, Jeremy; Yates, Cameron; Beatty, Robin; Evans, Jay; Edwards, Andrew; Ribeiro, Natasha; Wooster, Martin; Strydom, Tercia; Giongo, Marcos Vinicius; Borges, Marco Assis; Menezes Costa, Máximo; Barradas, Ana Carolina Sena; van Wees, Dave; Van der Werf, Guido R.

doi:https://doi.org/10.5194/esd-14-1039-2023

Articles | Volume 14, issue 5

https://doi.org/10.5194/esd-14-1039-2023

© Author(s) 2023. This work is distributed under
the Creative Commons Attribution 4.0 License.

https://doi.org/10.5194/esd-14-1039-2023

© Author(s) 2023. This work is distributed under
the Creative Commons Attribution 4.0 License.

Articles | Volume 14, issue 5

Research article

|

10 Oct 2023

Research article |

| 10 Oct 2023

Dynamic savanna burning emission factors based on satellite data using a machine learning approach

Roland Vernooij, Tom Eames, Jeremy Russell-Smith, Cameron Yates, Robin Beatty, Jay Evans, Andrew Edwards, Natasha Ribeiro, Martin Wooster, Tercia Strydom, Marcos Vinicius Giongo, Marco Assis Borges, Máximo Menezes Costa, Ana Carolina Sena Barradas, Dave van Wees, and Guido R. Van der Werf

Download

Final revised paper (published on 10 Oct 2023)
Supplement to the final revised paper
Preprint (discussion started on 06 Mar 2023)

Interactive discussion

Status: closed

CC1:
'Comment on egusphere-2023-267', Paul Laris, 19 Apr 2023

Great work. I have enjoyed following this project. Drones seem the ideal way to determine how emissions vary from savanna fires given the heterogeneity of landscapes and fires. I have several thoughts and a few critical comments on the findings and methods.
Can you clarify that these fires were all "head" fires as opposed to backfires? And, if so, can you comment on why the following dimensions are adequate? We question wether 10m is wide enough for head fires to fully develop. This width is fine for backfires. Also, if only head fires were examined, can you justify given that many fires are purposefully set as backfires in Africa. Headfires have long been used in research on African fires, yet research finds more backfires are set.
You do not appear to have published local or ground data on weather conditions. While T and H can be collected from regional weather stations, wind speed is critical to determining fire intensity and will influence MCE as well. Do you have wind data, it would seem critical for accurate fire intensity and MCE results.
I wonder about this comment: "
The grasslands with the highest EFs (found in high-rainfall savanna Dambos) were "uncharacteristically green for the time of the season" given that many fires are set to "green" grasses in African savannas, especially the perennials (See Le Page who documented this back in 2010 as well as many West African case studies).
I think Laris et al found very similar results to: "The strongest predictors for the MCE and the CO and CH4 EF were the tree density in the plots, the grass to litter ratio, the combustion completeness and the moisture content of the consumed fuel. It might be useful to compare and to consider the hypothesis that burning of green leaves on shrubs and trees vs. dried leaves on the ground may explain why EF CH4 is not linerally related to MCE. This reasoning may also explain the following finding, "For CO and CH4, the dominant effect is a spatial redistribution with higher CO and CH4 EFs in mesic, high-tree cover savannas and lower EFs in xeric savannas compared to previous estimates. The Higher CH4 EF in mesic may well be a function of leaf burning. This is logical given the findings from Senegal research by Barker finding burning trees emitted smoke with the highest methane EF.
This needs further explanation: "Although CO and CH4 followed the same spatial pattern, we found that MCE affected the CH4 which resulted in lower CH4 to MCE ratios in open (lower tree density?) savannas…. Do you mean higher CH4/MCE in wooded savannas as compared to grass-dominated ones? What is “open”? Clarify. Again, see works in Mali and Senegal which agree with this finding.
Not sure I agree with this logic: "Contrary to previous research which indicated that dryer conditions in the LDS would lead to higher-MCE fires late-LDS conditions (Fig. 3). In part, this may be because our measurement campaigns missed the peak-season fires when the fires may be hotter..." Winds are the critical factor here. When do they peak in areas studied. High winds (especially if fires studied are head fires) result in higher intensity regardless of fuel moisture. Laris also found lower MCE in LDS due to leaf litter in the fuel load and lighter winds with much higher winds in MDS for the region studied. Note that these factors are key reasons why binary (LDS/EDS) is problematic for determining emissions.
Again, see research in Mali and Senegal which support this finding: In accordance with previous studies (e.g. Korontzi et al., 2003b; van Leeuwen and van der Werf, 2011), we found steeper CH4 EF to MCE regression slopes in woodlands compared to grasslands.
Comments
Figure 3. What is “typical savanna” there is no such thing.
Also, use more specific terminology, what is "open"?
This and other data rely on 500 x 500 MODIS is this relevant given efforts to burn patchy EDS fires which operate at a hectare level scale? Can you justify using 500m data for the following? For fire severity proxies we used the differential Normalized Burn Ratio (dNBR) and 5 the differential Normalized Difference Vegetation Index (dNDVI) retrieved before and after the fire. These were based on the MODIS surface spectral reflectance, corrected for atmospheric conditions (MOD09GAV6; Vermote
Respectfully submitted

Citation: https://doi.org/10.5194/egusphere-2023-267-CC1
- AC1: 'Reply on CC1', Roland Vernooij, 20 Jun 2023
  
  We sincerely thank Paul Laris for taking the time and effort to read and comment on our manuscript, and the detailed and constructive comments both on this platform and in earlier conversations, which helped to improve the quality of this paper. Please find below our point-to-point response to the review. The revised text and updated figures are included in the updated manuscript. A separate ‘track-changes’ document is included to highlight the changes we made to the manuscript.
  
  Citation: https://doi.org/10.5194/egusphere-2023-267-AC1
RC1:
'Referee report on egusphere-2023-267', Robert Yokelson, 20 Apr 2023

Please see supplement.

Citation: https://doi.org/10.5194/egusphere-2023-267-RC1
- AC2: 'Reply on RC1', Roland Vernooij, 20 Jun 2023
  
  We sincerely appreciate the considerable time and effort spent in assessing our manuscript, and the detailed and constructive comments which helped to improve the quality of this paper. Please find below our point-to-point response to the review. The revised text and updated figures are included in the updated manuscript. A separate ‘track-changes’ document is included to highlight the changes to the manuscript. Additional explanatory figures which we refer to in the answers are added to the bottom of this document.
  
  Citation: https://doi.org/10.5194/egusphere-2023-267-AC2
RC2:
'Comment on egusphere-2023-267', Anonymous Referee #2, 27 Apr 2023

This study collected a large dataset of savanna emission factors (EFs), including over 4500 EF bag measurements of CO2, CO, CH4 and N20 during 129 individual fires from 2017 to 2022. Based on this in-situ observations, the authors identified the drivers of EF variability and implemented this variability into global models through dynamic EFs. The optimized machine learning reduced the error in EF estimates by 60-85% compared to static biome averages. They also found seasonal drying resulted in a decrease of the EFs with the fire season progressing, with a stronger trend in open savannas than woodlands. Overall, this is an important study to understand the variability and mechanisms for biome-specific carbon emissions, particularly at the spatial scales. The generated global EF products can be used to better estimate fire-induced greenhouse gas emissions. However, I do have some concerns on the methodology parts, which may need to be addressed before publication.
1). Biomass burning EFs are highly dynamic both at the spatial and temporal scales for a given fire. For example, EFs may differ a lot as the fire spreads across different vegetation covers and terrain/moisture gradients at the local scale. How well did the collected EF bag measurements represent the total or averaged EFs for each selected fire? Is there a consistent spatial-temporal framework to integrate the concrete EF measurements to reflect the total EFs for all involved fires? Such processing details need to be provided for better understanding the uncertainty of “in-situ” measurement itself.
2). How did the fire induced EFs match the possible drivers at the spatial scale? Are they overlayed by the actual size of each fire, or just at the grid size of 0.25 degree? The latter may introduce large uncertainties.
3). The authors tested a series of machine learning methods and concluded that random forest performed best. Such a part may need data support. Past experiences suggest that the gradient boost MLs such as lightGBM and Xgboost tend to be better than random forest.
4). To predict BB EFs, the authors included a series of factors for each group driver (seen in Table 2). However, it seems that some of them are highly correlated, e.g., NDVI VS LAI VS FPAR, VPD VS evaporative stress index VS Relative humidity. The rationale for including these redundant factors may need to be clarified. In addition, given the potential uncertainty in remote sensing and reanalysis data, it may not be wise to include all predictors without doing a feature selection. One way to include a specific driver or not is to compare its effect with a randomly generated variable. If its effect is equal or worse than the random variable, it may not be included in the final training.
5). Satellite data over tropical regions usually suffers from the contamination of clouds. When deriving the global EFs, how the authors gap-filled relevant remote sensing data is not clear.

Citation: https://doi.org/10.5194/egusphere-2023-267-RC2
- AC3: 'Reply on RC2', Roland Vernooij, 20 Jun 2023
  
  We sincerely thank the reviewer for the time and effort in assessing our manuscript, and the constructive comments which helped to improve the quality of this paper. Please find below our point-to-point response to the review. The revised text and updated figures are included in the updated manuscript. A separate ‘track-changes’ document is included to highlight the changes to the manuscript. Tables and figures referred to in the answers are added at the bottom of this document.
  
  Citation: https://doi.org/10.5194/egusphere-2023-267-AC3

Peer review completion

AR: Author's response | RR: Referee report | ED: Editor decision | EF: Editorial file upload

ED: Reconsider after major revisions (03 Jul 2023) by Anping Chen

AR by Roland Vernooij on behalf of the Authors (08 Jul 2023) Author's response Author's tracked changes Manuscript

ED: Referee Nomination & Report Request started (13 Jul 2023) by Anping Chen

RR by Robert Yokelson (26 Jul 2023)

ED: Publish subject to minor revisions (review by editor) (02 Aug 2023) by Anping Chen

AR by Roland Vernooij on behalf of the Authors (16 Aug 2023) Author's response Author's tracked changes Manuscript

ED: Publish as is (19 Aug 2023) by Anping Chen

AR by Roland Vernooij on behalf of the Authors (22 Aug 2023) Manuscript

Download

Article (12132 KB)
Full-text XML

Short summary

Savannas account for over half of global landscape fire emissions. Although environmental and fuel conditions affect the ratio of species the fire emits, these dynamics have not been implemented in global models. We measured CO₂, CO, CH₄, and N₂O emission factors (EFs), fuel parameters, and fire severity proxies during 129 individual fires. We identified EF patterns and trained models to estimate EFs of these species based on satellite observations, reducing the estimation error by 60–85 %.

Dynamic savanna burning emission factors based on satellite data using a machine learning approach

Download

Interactive discussion

Peer review completion

Suggestions for revision or reasons for rejection