BuRNN (v1.0): A Data-Driven Fire Model

Lampe, Seppe; Gudmundsson, Lukas; Kraft, Basil; Hantson, Stijn; Kelley, Douglas; Humphrey, Vincent; Le Saux, Bertrand; Chuvieco, Emilio; Thiery, Wim

doi:https://doi.org/10.5194/egusphere-2025-3550

Preprints

https://doi.org/10.5194/egusphere-2025-3550

Preprints

01 Sep 2025

| 01 Sep 2025

BuRNN (v1.0): A Data-Driven Fire Model

Seppe Lampe, Lukas Gudmundsson, Basil Kraft, Stijn Hantson, Douglas Kelley, Vincent Humphrey, Bertrand Le Saux, Emilio Chuvieco, and Wim Thiery

Abstract. Fires play an important role in the Earth system but remain complex phenomena that are challenging to model numerically. Here, we present the first version of BuRNN, a data-driven model simulating burned area on a global 0.5° × 0.5° grid with a monthly time resolution. We trained Long Short-Term Memory networks to predict satellite-based burned area (GFED5) from a range of climatic, vegetation and socio-economic parameters. We employed a region-based cross-validation strategy to account for the high spatial autocorrelation in our data. BuRNN outperforms the process-based fire models participating in ISIMIP3a on a global scale across a wide range of metrics. Regionally, BuRNN outperforms almost all models across a set of benchmarking metrics in all regions. However, in the African savannah regions and Australia burned area is underestimated, leading to a global underestimation of total area burned. Through eXplainable AI (XAI) we unravel the difference in regional drivers of burned area in our models, showing that the presence/absence of bare ground and C4 grasses along with the fire weather index have the largest effects on our predictions of burned area. Lastly, we used BuRNN to reconstruct global burned area for 1901–2019 and compare the simulations against independent long-term historical fire observation databases in five countries and the EU. Our approach highlights the potential of machine learning to improve burned area simulations and our understanding of past fire behaviour.

Received: 23 Jul 2025 – Discussion started: 01 Sep 2025

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Seppe Lampe, Lukas Gudmundsson, Basil Kraft, Stijn Hantson, Douglas Kelley, Vincent Humphrey, Bertrand Le Saux, Emilio Chuvieco, and Wim Thiery

Status: final response (author comments only)

RC1:
'Comment on egusphere-2025-3550', Anonymous Referee #1, 03 Oct 2025
Review of “BuRNN (v1.0): A data-driven global burned-area model”
The manuscript presents a monthly 0.5° machine‑learning emulator of burned area trained on GFED5 with region‑blocked cross‑validation. It uses a Long Short‑Term Memory (LSTM) architecture to capture temporal dependence on the input features. The trained BuRNN model is then applied to reconstruct global burned area for 1901–2019, and the simulations are compared against independent long‑term fire observation databases from five countries and the EU. Overall, the manuscript is well written, and the trained model shows promising results. Yet several core claims are insufficiently supported. In particular, BuRNN is target‑aligned to GFED5 when compared with process models, which makes the comparison unfair. Model biases and observational uncertainties must be clearly distinguished. The transferability of a model trained on recent conditions to historical periods requires further discussion. The interpretability of the SHAP analysis assumes feature independence, which is strongly violated; this needs further evaluation.
Major comments
Because BuRNN is trained on GFED5 while process‑based models are not, any superiority shown for 1997–2019 primarily reflects alignment to the chosen satellite target rather than absolute accuracy. The manuscript notes this when introducing the FireCCI51 cross‑check (“we repeat this evaluation… using the 2001–2019 FireCCI51 dataset… because our model is specifically trained to predict GFED5…”), yet the framing still concludes that BuRNN outperforms process models (see Table A17). Where BuRNN appears closer to FireCCI51 than the process models, this likely indicates that GFED5 and FireCCI51 are more mutually consistent than FireCCI51 is with the process models, not that BuRNN is more accurate in an absolute sense.

Comparing Figure 4 (regional/global annual BA time series) with Figure A2 (BuRNN vs FireCCiLT11), BuRNN aligns more closely with FireCCiLT11 than with GFED5 during the overlap period, indicating that a meaningful portion of the discrepancy reflects observational product differences rather than model bias. The same issue arises in national/EU validations. Because EFFIS explicitly excludes cropland fires, BuRNN–EFFIS agreement or disagreement is interpretable only after restricting BuRNN/GFED5 to the non‑cropland domain; otherwise systematic differences over cropland regions are expected by construction.

Methods state that SHAP values are computed “relative to the average prediction over a dataset,” using GradientExplainer with integration from the average input to each sample. The manuscript does not specify whether “the dataset” and the average input are defined globally or per region, yet regional SHAP panels are compared across regions. If a global background is used, a regional bar reflects deviation from a global reference rather than the region’s climatology; if the background is recomputed per region, cross‑region bar magnitudes are not directly comparable.

FWI, tasmax, SPEI, ET/GPP and related variables are strongly dependent by construction. Under such collinearity, SHAP can redistribute attribution among correlated predictors, so statements like “FWI is most important everywhere” are not stable. Although the manuscript cautions that SHAP is not causal and is difficult to interpret with interdependent inputs, the subsequent discussion reads mechanistically from the bars. This is inconsistent with the stated caveat.

The reconstruction extends BuRNN to 1901–2019 and is compared with FireCCiLT11 for 1982–2018. Applying a mapping trained on 1997–2019 to 1901–1996 introduces non‑stationarity risk in regions with large changes in ignition, suppression or land‑use practices. For instance, should we expect human play a same current role in the 1910s? In addition, vegetation predictors (GPP, LAI, cVeg) are ISIMIP DGVM ensemble means, which predicted with a process-based fire model. The manuscript should discuss how biases in the vegetation inputs and the omission of fire-vegetation feedbacks in BuRNN affect its predictions.

Minor comments
In the introducing. While the satellite products suffer uncertainties due to cloud cover, spatial resolution, rapid regrowth and obscuration by unburned vegetation, etc, the process-based model or machine model calibrated or trained on satellite-based product inherits those uncertainties.

Since many datasets across different period are compared, the figure captions can be enhanced with better notion of time period, dataset, and units. For instance, confusion can be arised when looking at Fig. 3 which shows the burned area fraction (%) vs Fig. 4 showing burned area in Mha

Line 132, Burned area is right‑skewed and zero‑inflated. How are zero burned area handled under the log transform? Was any zero‑handling or sampling strategy, or a distribution‑/tail‑aware loss is applied. A plot of the empirical distribution (PDF/CDF) by region would help evaluate if any especial treatment is needed.

Line 261, Line 261, Underestimation of interannual variability in several regions and suggests VPD/ET/PET as potential drivers. It would be worth to include them to retrain the model.
Citation: https://doi.org/10.5194/egusphere-2025-3550-RC1
CEC1: 'Comment on egusphere-2025-3550 - No compliance with the policy of the journal', Juan Antonio Añel, 11 Oct 2025

Dear authors,
Unfortunately, after checking your manuscript, it has come to our attention that it does not comply with our "Code and Data Policy".

https://www.geoscientific-model-development.net/policies/code_and_data_policy.html
In your "Code and data availability" section you do not include repositories for the ISIMIP and SPEI data, but cite webpages to get access to them. We can not accept this. You must publish all the data necessary to train your model and their outputs in a suitable repository according to our policy (as you have done with others). Therefore, the current situation with your manuscript is irregular. Please, publish your data in one of the appropriate repositories and reply to this comment with the relevant information (link and a permanent identifier for it (e.g. DOI)) as soon as possible, as we can not accept manuscripts in Discussions that do not comply with our policy.
Also, you must include a modified 'Code and Data Availability' section in a potentially reviewed manuscript, containing the information of the new repositories.
I must note that if you do not fix this problem, we cannot accept your manuscript for publication in our journal.
Juan A. Añel

Geosci. Model Dev. Executive Editor

Citation: https://doi.org/10.5194/egusphere-2025-3550-CEC1
AC1:
'Reply on CEC1', Seppe Lampe, 13 Oct 2025

Dear Dr. Juan A. Añel,
Thank you for your comment. The ISIMIP and SPEI data is already available on Zenodo and is referenced in the following sentence:
The 1901-2019 burned area simulation of BuRNN is available on Zenodo along with all pre-processed data to train BuRNN (https://zenodo.org/records/16918071; Lampe, 2025b).
We believe the confusion might come from this sentence a few lines down:
The original ISIMIP data is available through the ISIMIP data repository (https://data.isimip.org/), the authentic SPEI data from SPEIbase can be downloaded from https://spei.csic.es/database.html.
We provide a pre-processed version of the ISIMIP and SPEI data in our Zenodo repository (as requested by the topic editor) and additionally provide readers the links to the original sources of the data. We apologize for any confusing wording on our behalf in the Code and Data Availability section.
We will rephrase this to avoid any confusion possible upon receiving the final reviewer's comments. Additionally, we have now added the independent evaluation data and the FireMIP simulations to the repository.
We hope you consider this an adequate course of action.
Kind regards,
Seppe Lampe (on behalf of all co-authors)

Citation: https://doi.org/10.5194/egusphere-2025-3550-AC1
- CEC2: 'Reply on AC1', Juan Antonio Añel, 13 Oct 2025
  
  Dear authors,
  Thank you for your clarifications. My comment was about the original data, not the pre-processed data. It is true that citing in the Code and Data Availability section information that is not relevant and does not serve the purpose of such section, such as the ISIMIP website, is inappropriate, but I insist, ideally you should store the original data, not only the pre-processed data. It would be good if you could do it.
  Juan A. Añel
  Geosci. Model Dev. Executive Editor
  
  Citation: https://doi.org/10.5194/egusphere-2025-3550-CEC2
RC2:
'Comment on egusphere-2025-3550', Anonymous Referee #2, 16 Oct 2025
General Comment
Lampe et al. developed a data-driven fire model (BuRNN) based on Long Short-Term Memory to estimate global gridded burned area from 1901 to 2019. BuRNN was trained with satellite-based burned area (GFED5) multiple climates, land cover, vegetation dynamics, and socioeconomic datasets during 2001 to 2020. The trained BuRNN was then used to reconstruct the burned area from 1901 to 2019. Indeed, wildfire remains to be highly uncertain in process-based models. Data-driven methods have the potential to improve the estimation of wildfire, which has been shown in previous studies. This study contributes to our understanding of wildfire mechanisms by providing long-term global burned area estimation and factor importance analysis. However, I have some concerns about the reliability of the data-driven model. Please find the details of the major and specific comments in the following. I recommend a major revision for the current manuscript version.
Major Comments
BuRRN consistently underestimate the burned area across different regions and at global scales. This underestimation seems to be systematic bias, however, I don’t understand why such error cannot be reduced in BuRRN, which is built on LSTM. I suppose machine learn approaches (including LSTM) is effective in reducing systematic bias. Please take a detailed analysis on this error. I suggest showing the validation at monthly scales. One possible reason is that BuRRN underestimate the maximum monthly burned area, which it may capture lower values well.

There is a lack of validation of temporal extrapolation. The authors did the out-of-sample validation by splitting the global domain to 11 subregions. However, they didn’t split the dataset in time to validate the model performance in temporal extrapolation. As shown in Figure 4, BuRNN cannot capture the annual trend presented in the benchmark dataset (i.e., GFED5). One of the objectives of this study is to reconstruct global burned area from 1901 to 2019. The failure of capturing the sensitivity to the environment changes during training period may suggest the reconstruction in the past is highly uncertain.

Specific Comments
Line 29: Please specify how the burn area (i.e. 3.5 – 4.5 million km^2) was estimated. As in later sentence, the authors provide the burn area from satellite observation, it is helpful to know how the previous estimation was made.
Figure A1: Please specify which color represent training.
Line 118: Why not randomly grouping the regions into 11 folds?
Line 141: Is 3 years the sequence length used in LSTM? Please elaborate how the sequence length is determined.
Line 155: Could the authors provide some explanation on how the SHAP handle the correlation between input variables?
Line 214: Are those process-based models calibrated? If not, it is not surprise that data-driven method outperforms the process-based model.
Line 219: “between the FireMIP models” is repeated.
Line 223: The evaluation for the 14 fire regions is not consistent with the global monthly evaluation. According to Figure 3, the BuRNN bias does not seems to be significantly. However, Figure 4 shows BURNN significantly underestimated annual burned area at both regional and global scales. Please explain this inconsistency. In addition, BURNN systematically underestimate the burned area, which I don’t understand why the LSTM cannot resume this error.
Figure 4: I suggest the authors to report the slope for the annual burn area for both GFED5 and BuRNN to validate if BuRNN’s can simulate the annual trend. This will demonstrate if BuRRN can capture the sensitivity of wild fire to the changing climate in the past. For example, the global plot in Figure 4 shows GFED5 exhibit a decreasing trend, while BuRNN exhibit no trend.
Line 227: I cannot agree with “excellently modelled”. Those regions are better than other regions, but they are still underestimated in BuRRN.
Line 228: “BuRNN captures the pattern of the interannual variability well, but consistently underestimates the amplitude and total burned area,” This statement is repetitive, which is reported earlier in this paragraph.
Line 245: “in MIDE, BuRNN…”
Line 248: Could the authors try to explain why BuRNN has this behavior: “e higher the average monthly burned area in a region, the better/easier predictions are for that region”?
Line 262: Why not including those variables in BuRNN if they are considered as important? There are existing global datasets of evapotranspiration can be used.
Figure 5: The current text is not very clear to see, please consider increasing the font size. Is the feature importance plotted as Volin plot? I suggest changing to box plot, which gives the 25^th, 50^th, and 75^th percentile. The current plot is dominated by the outliers, which does not help to interpret the data.
Line 269: Why C4 grass Africa is considered as a feature for EURO, BOAS, SHSA, and GLOBAL? The absence of African C4 grass for explaining the lower burned area does not make sense to me. Shouldn’t we focus on existing land covers in each region for the feature importance analysis?
Line 295: Please report the annual correlation.
Line 298 – Line 302: If EFFIS does not include cropland fire, I don’t see a point to benchmark BuRRN with it.
Citation: https://doi.org/10.5194/egusphere-2025-3550-RC2

Seppe Lampe, Lukas Gudmundsson, Basil Kraft, Stijn Hantson, Douglas Kelley, Vincent Humphrey, Bertrand Le Saux, Emilio Chuvieco, and Wim Thiery

Viewed

Total article views: 1,635 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
1,512	102	21	1,635	12	16

HTML: 1,512
PDF: 102
XML: 21
Total: 1,635
BibTeX: 12
EndNote: 16

Views and downloads (calculated since 01 Sep 2025)

Month	HTML	PDF	XML	Total
Sep 2025	1,338	41	5	1,384
Oct 2025	170	58	15	243
Nov 2025	4	3	1	8

Cumulative views and downloads (calculated since 01 Sep 2025)

Month	HTML	PDF	XML	Total
Sep 2025	1,338	41	5	1,384
Oct 2025	170	58	15	243
Nov 2025	4	3	1	8

Viewed (geographical distribution)

Total article views: 1,570 (including HTML, PDF, and XML) Thereof 1,570 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 04 Nov 2025

Short summary

We introduce BuRNN, a model which estimates monthly burned area based on satellite observations and climate, vegetation, and socio-economic data using machine learning. BuRNN outperforms existing process-based fire models. However, the model tends to underestimate burned area in parts of Africa and Australia. We identify the extent of bare ground, the presence of grasses, and fire weather conditions (long periods of warm and dry weather) as key regional drivers of fire activity in BuRNN.


Total:	0
HTML:	0
PDF:	0
XML:	0