This paper combined satellite observations of fire behaviour, tower-based CO and CO2 mixing ratios and bottom-up approach to estimate forest fire emissions in France with a focus on improving emissions from smoldering. I found the methods used by the authors in general credible and the paper advances the quantification of fire emissions induced by forest fires. I have a few major comments mainly regarding clarifications of the methods being used and some minor technical comments (detailed below).

Major comments:

• I suggest adding a paragraph giving an overview of the methods, preferably with a flowchart figure, focusing on how different approaches are combined and connected.
• What’s the major purpose of Hysplit model? I don’t really see how it is connected with the selection of tower sites and determination of background measurement…. Is it only used to justify that most of the ICOS sites are free of influences of Mediterranean forest fires and hence their measurement could be considered as background ones? Fig. 3 is nice but also quite unique I guess. Is it a sufficient example to argue that, based on Hysplit simulations, most of the ICOS sites are free of influences of Mediterranean forest fires and hence their measurement could be considered as background ones?
• If I understand well, it seems that the characterization of fire behaviour using satellite data is independent of the bottom-up estimation of fire emissions, in particular, the fire behaviour information has not been used to determine the key parameters in Equation (2) (e.g., SFp) and parameters in lines 330-334. This is somewhat a little disappointing. Following this logic, it then seems that the key strength/advancement of the paper is that the authors compiled a nice Table 2, a range of more credible sources of fuel load, and the used satellite-derived fire information and power-based MCE to *indirectly* verify their bottom-up estimate of emissions? Is this correct? This point has to be made clearer when the authors address my first major comment.
• The authors examined three typical fires, or fires in three typical forests using satellite-based fire behaviour and power-based mixing ratio measurements. These are then used to support their bottom-up approach. Then then the challenge is how we can ensure that the upscaling to the national level using their bottom-up approach is also reliable, given that fires are highly temporal and spatially heterogeneous in terms of fire bebaviour, fraction of flaming versus smoldering, combustion completeness etc. (I believe the authors have tried to address well the spatial heterogeneity in fuel load)?

Minor comments:

Line 139: some introduction on VIIRS data is necessary because it seems an important limitation on what fires have been analyzed.

Line 145: “beyond the fire outbreak ”. What does ‘beyond’ mean here?

Line 146–147: I don’t see how the approach described here (visual examination of RGB spectrum) could be reconciled with BAMTS… So what is exactly the role of BAMTS in burned area detection? And how are these two further linked with random forest classifier and how the classifier is used and for which purpose?

Line 203: “corresponding to a single grid cell. ”. Which model does this grid cell refer to? What is the spatial resolution of Hysplit?

Table 1: Better to report R2 rather than R. The same for the texts.

Line 168: what is this 6-hour data?

Line 199: Is this 600 per hour particle numbers typically used in transportation modeling? How does this influence the results?

Line 201: “By tracking the arrival times of these particles within an influence region surrounding each atmospherictower, we successfully attributed a source to each anomaly”, I don’t understand the latter half. Could you please explain?

Line 298-299: I don’t understand what you mean by ‘baseline’ here.

Table 2: I cannot reconcile/connect Table 2 with lines 330–335. (1) you provide only constant SF values in Table 2. But if SF values do no change among the flaming phase, mixed phase and smoldering phase, then how is this used in Equation (2)? (2) lines 330-335 seems giving proportions of fuels being affected by fire, what is the difference between this and CC in Table 2? Seems that lines 330-335 should be better integrated with Table 2 so that you have only a single source to present the parameters used in emissions calculation. (3) how the information in lines 330-335 is used in Equation (2)? (4) how do you choose CC values between its min and max values in Table 2?

Line 352: TROPOMI data not explained in Methods.

Figure 4: what is the difference between 1-hour and 1-minute? Are they the temporal resolutions of the data ? what is the temporal resolution of measurement over the towers?

Firstly, many apologies for the delay in providing my review.  But, this proved a very interesting manuscript which I enjoyed reading in detail.  The authors present analysis of atmospheric CO2 and CO measurements from ground stations influenced by smoke plumes from large fires in France in 2022, demonstrating the presence of significant amounts of smouldering combustion in at least one of the three large fire events that took place that summer.  They subsequently construct a sophisticated bottom-up fire emissions inventory for France which incorporates detailed land cover and carbon content data, and explicitly estimates the emissions contributions from both smouldering and flaming combustion.  Explicit consideration of smouldering combustion appears to result in a better match with the observed combustion efficiencies than the GFAS emissions estimate for CO2 and CO from the same fires. Furthermore they find that due to the inclusion of data on belowground organic soil carbon stocks, the bottom-up emissions estimate predicts a ~2x larger total carbon emissions across France in 2022 than GFAS does, with implications for the country’s carbon budget.

The manuscript highlights the very significant contribution that belowground carbon stocks can make to fire emissions in regions where smouldering combustion has not previously been considered important to include in global emission inventories, as well as how station measurements of atmospheric composition can be used to discern the relative importance of smouldering combustion from individual fires.  It’s highly relevant and should be of significant interest to the readership of Biogeosciences.

I have a few comments and suggestions which I have detailed below.  It’s a slightly long list because I was keen to really read and understand this manuscript in detail!  But hopefully are mostly just clarification.  There are some areas where I feel the methods require more explanation and there is some extra information needed to fully understand the process.  Additionally, from a story-telling perspective the two strands of the manuscript (observations of Modified Combustion Efficiency using the atmospheric tower network, and bottom-up fire emissions estimation) could be linked together more strongly, and it would be nice to see them being compared and used to inform and validate the other more directly.  Finally, I have some concerns that the uncertainties on some of the parameters used in the emissions estimation are being underappreciated, which may mean that the results need to be interpreted with a little more caution until the model can be explicitly validated with other observational data sources.  I regard these all as minor revisions though, since they do not undermine the main objective or results of the manuscript, but perhaps just imply some additional thought on the presentation.  (I’ve also listed some technical comments, which are mostly just small wording/grammatical suggestions).  Subject to the authors addressing these points (or pointing out why I’ve misunderstood!  Which is certainly possible) then I’d very much support the manuscript being accepted for publication.

Minor comments:

L23: “We examined the role of soil smoldering combustion responsible for higher carbon emissions, locally reported by firefighters but not accounted for in global fire emission budgets” – I think this statement is only partially true.  Widely-used fire emission inventories like GFED and GFAS do include a representation of peat and/or organic soil as a distinct land cover/biome type (with spatial distributions of peat and organic soils based on the literature, much like in the present study).  For instance, GFAS has an ‘extratropical forest with organic soil’ land cover class, and a ‘peat’ land cover class.  Consequently, even if they do not necessarily have separate a smouldering versus flaming representation, they can still have different fuel consumption and/or emission factors calibrated for these land cover types, which reflect different carbon pools and contribution of smouldering combustion.  They may well still underestimate the contribution of smouldering combustion, due to it being undetected or due to incorrect assumptions of the land cover/soil types over Europe, but that’s not the same as it being entirely unaccounted for.  N.B. the authors’ discussion of this issue in Sect. 4.5 is very good, and much more nuanced.

L105-106: “We also compared our emissions to the current global models based on standard fire emission factors (GFAS, 2023)” – if you compared only with GFAS, then technically you have only compared against one current global model, not models.  Also ‘based on standard emission factors’ is an interesting point to consider – does the authors’ work here indicate that the GFAS biome-average fuel consumption and/or emission factors are wrong, or that the assumptions of what vegetation/material is burning is wrong?  E.g. the ‘standard emission factors’ could be perfect, but if applied to an incorrect map of vegetation density or soil type, you could still infer the wrong total emissions.

L152-153: “Among these fire polygons, three of them located in the proximity of atmospheric towers were chosen for in-depth analysis, referred to as “main fires”” – it’s rather unclear, is all the subsequent analysis of the atmospheric tower data restricted to the time periods around these three ‘main fires’, or is the analysis carried out for all fire events but with some particular features discussed in more detail for these three fires?  If the analysis is focused on these three fires then they need to be described in a little more detail: e.g. what dates did they start/finish, what area did they burn?

L168: “we leveraged the temporally dated (6-hour intervals) spatial locations of fire hotspots (Fig. 5)” – as I understand it, the fire hotspots come from MODIS and VIIRS.  However, Terra and Aqua overpasses are only 3 hours apart (10:30 am/pm and 1:30 am/pm respectively). S-NPP and NOAA-20 overpasses are also both around the same time as Aqua (1:30 am/pm).  How were the fire hotspot detections converted to regular 6-hourly intervals?  (Also, what happened to figures 2, 3, and 4?  Figures should be numbered following the order they appear in the text; this is the 2^nd figure referenced in the paper).

L182-186: “Data collection for this study spanned from June 15th to September 1st, 2022. In the context of the Atlantic pine forest, the dominant winds were from the northeast, propelling the plume seaward. Notably, a shift in wind direction occurred on July 14th -15th, with the wind veering to the north-northwest. This shift contributed to the highest CO peaks observed at the Biscarrosse station” – again, it’s very unclear whether data is being analysed for multiple fires across the season (June – September), for just for the single large fire that occurred in this region (‘the plume’).  Was a fire event even taking place on July 14th -15^th – the relevance of this date hasn’t been explained.

L191-205: This is probably just me not understanding very well, but: this paragraph describes both back-trajectory and forward-trajectory analysis to determine the sources corresponding to CO mixing ratio anomalies observed at each atmospheric tower.  I’m not clear how these two approaches were combined to ultimately attribute a single source for each anomaly.  What if there were multiple potential sources within the back-trajectory footprint of a tower?  Also, did the authors only consider forest fire hotspots as potential sources?  E.g. What if the wind direction on a particular day was blowing from an urban source which could also result in anomalously high CO - how were these possible contributions ruled out?  Finally, how was the influence of a particular fire temporally co-located with the tower measurements?  I.e. how did the authors decide over what time period did the measured MCE correspond to a particular fire?

L321: “Table 2 provides a comprehensive summary of CC, EF, and SF for each pool” – Table 2 lists both a min and max value of CC from the literature for each pool – but which value was actually used in the emissions calculation?  This doesn’t seem to be described.  For some of the pools (particularly the belowground stocks), the choice of CC within the range reported will make a huge difference to the total C emissions, which doesn’t seem to be acknowledged or included in the uncertainties given for the total emissions in Table 4 and Sect. 3.5.

L329-335: “we delineated three distinctive phases in the propagation of each fire…” – what was the basis for assuming a 50:50 split of AGS between phase 1 and 2, and 25:75 split of BGS between phase 2 and 3?  I notice also that Figure 4 doesn’t seem to show much trend of decreasing MCE with time for any of the fires – even BIS – which would seem to call into question this assumption about the evolution of the combustion phases.  This should be commented upon – it would be nice to see either on the same plots in Figure 4, or else as a separate figure to compare with, what the time evolution of MCE looks like for the authors’ bottom-up estimation of CO2 and CO emissions.  Currently only the time-mean value of MCE is compared with the tower observations, but comparison of the daily time evolution of predicted MCE with the atmospheric tower observations will help to validate the assumption that the fire transitions from fully flaming to fully smouldering.

L338: “biome-specific standard emission factors (in kgDM.MJ-1)” – what makes them ‘standard’?  In fact, GFAS I think is somewhat unique in taking this FRP-based approach – other inventories such as GFED take an approach similar to the authors’, of relating fuel consumption to the burned area and a map of vegetation.  Also, the biome-specific kg[DM]/MJ in GFAS are not emission factors, they are the dry matter combustion rate.  These are used to estimate the amount of DM combusted.  The emission factors are then the mass of CO2 or CO emitted per DM combusted.

L369: “Daily MCE variations (Fig. 4) emphasized a decreasing trend for the BIS fire” – this decreasing trend is quite marginal, and I wouldn’t describe it as being emphatic in any sense.  Over the 5 days shown on Figure 4, the BIS daily median MCE goes up, down, up, up – and ends up back close to its initial value.  So it’s a weak decreasing tendency at best.  I’m also confused because elsewhere the authors state that both the BIS and ROC fires lasted for more than 25 days, and yet in Figure 4 the MCE is plotted only for 5 and 3 days respectively.  It would be nice to see whether over the full 25 day+ duration of the event, there is more of a trend towards lower MCE after the first few days.

L422-423: “we conclude that fires prone to experiencing smoldering combustion, such as BIS and ROC fires, exhibit a prolonged duration of hotspots after ignition” – I would agree that prolonged duration of hotspots after ignition could be an indication that smouldering combustion is happening, however it’s not a pre-existing property of the vegetation or something that can be observed in the initial fire behaviour, so I’m not clear how it could be used to indicate beforehand that a fire is prone to smouldering combustion.

L436-437: “aligning with the median value obtained from the hourly mixing ratios measured at the ROC tower” – this was previously reported as a mean value (L366) – which is correct?  This sentence is also inconsistent with L432-433 where it was suggested that ABS-only MCE diverges from that measured at ROC.  If flaming-only MCE aligns closely with the average reported for ROS, does that mean that overall, smouldering contributed little to the total emissions for this fire, and it was only BIS where smouldering-only emission factors were really necessary to get the correct carbon budget?

L449: “Drawing from our MCE-derived carbon emissions estimates of AGS-BGS combustion” – how was MCE used to derive the carbon emission estimates?  My understanding was that these were derived independently, using the carbon pools estimated by land cover maps, burnt pixels from hotspot detection, and literature values of combustion completeness, emission factors etc.  It would be really cool if the observed MCE could be used to derive the carbon emissions directly, but that doesn’t seem to have been what was done here.  Unless I’ve misunderstood!

L469-470: “the GFAS framework estimated that summer fires were accountable for 3.86 MtCO2 emissions when excluding belowground combustion” – again, I would say that technically GFAS doesn’t exclude belowground combustion, since it has biome-specific fuel consumption and/or emission factors for ‘peat’ and ‘extratropical forest with organic soil’ which in principle should reflect the different carbon stock and the balance of smouldering versus flaming fires.  However, it’s quite likely the course resolution land cover maps GFAS uses don’t categorise these areas of France as belonging to either of those biomes.  So, I would view one of the major take-home messages as being that your emissions estimate is only as good as your fuel cover data.

Sect. 3.5, Figure 6, and Table 4: These all show uncertainty ranges for the various DM and emission values, however as far as I can see no description is given anywhere (in the Methods, in the Results, or in the Figure captions) as to what this uncertainty represents (e.g. is it ±1σ, 95% confidence interval, something else…) or how it was calculated.  (Apologies if it is described somewhere and I’ve missed it).

L542-544: “To maintain a conservative approach, we adopted a ROS of 0.2 cm.h-1 for soil combustion, resulting in a daily consumption of approximately 4.8 cm” – shouldn’t this have been mentioned in the Methods, if this calculation determines how deep the soil layer is burned?  (C.f. my previous comment; it's not made clear what value is used for the combustion completeness, which can dramatically alter the resulting carbon emissions calculation).

L543-545: “we adopted a ROS of 0.2 cm.h-1 for soil combustion, resulting in a daily consumption of approximately 4.8 cm, which roughly corresponds to 40 cm burned over an 8-day period, which corresponds to the average flaming duration of our fires. This 40 cm of consumed peat aligns with the upper bound of our soil combustion parameters” – this seems a little concerning.  The BIS and ROC fires last much longer than 8 days, and yet after 8 days the model assumes that you have already burned close to the upper bound of the maximum peat depth that can be burned.  Most real-world peat fires do not burn nearly as deep as 40cm.  This might imply that the authors’ model is typically assuming too high belowground DM combusted, by predicting unrealistically high depths of burn.

L549-551: “For a comparative perspective, 550 Mickler et al. (2017) using fine resolution LIDAR data revealed that peatland wildfires could exhibit an average burn depth of 42 cm” – this was a case study of a single fire – it cannot be interpreted to mean that 42cm is typical of most peat fires.  As the authors have already noted, peat fires can burn anything from 1 cm to 50cm+, but in field studies most are < 20cm (see e.g. Walker at al (dataset, 2020) https://doi.org/10.3334/ORNLDAAC/1744 or Blackford et al. (2024) https://doi.org/10.5194/gmd-17-3063-2024)

L551: “average belowground carbon emissions estimated at 544.43 t C ha−1” – the depth of burn in Mickler et al. of 42cm is almost exactly the same as the depth the authors report here of 40 cm.  So why is the carbon emitted per ha so much higher in the Mickler et al. study?

L571: “This enables our bottom-up approach to be confronted and evaluated against atmospheric MCEs” – indeed – this is why it would be really good if the MCE predicted from the authors’ bottom-up approach could also be plotted side-by-side with the MCE observed at the tower stations, e.g. in Figure 4, or with an extra figure straight after Fig. 4 that has the same plots but for the bottom-up estimate MCE.

L604: “These low MCE values, which are challenging to account for based on biomass or SOM combustion alone” – how low, and can the authors demonstrate that these values require a contribution from lignite combustion to account for? (N.B. not sure MCE values for lignite are ever described in the text, making it hard to judge).

L692-695: “Finally, we advocate for the widespread use of our updated fire emissions processing chain for France, which could potentially be extended to other European temperate forests” – to try and motivate this further, along with the argument in Sect. 4.4 that these fires were a larger contribution to France’s carbon emissions than previously anticipated: is there any validation that can be done to confirm the accuracy of your CO2 and CO total emission amounts?  This feels like one weakness of the bottom-up approach here; it’s based on a lot of different input parameters from the literature (e.g. land cover type, carbon density, combustion completeness etc.) all of which have large uncertainties in them (as mentioned previously, the uncertainty range in the BGS combustion completeness alone could potentially explain the difference with the GFAS emission budget).  The manuscript has shown that the relative CO2/CO ratio is consistent with the MCE observed at the tower stations, but are there any measurements (from the tower observations or otherwise) that can be used to additionally constrain the total emission amount, to validate that your approach gives the correct total magnitude whereas GFAS is definitely too low?  Otherwise, either could be right for the total magnitude (even if GFAS probably gets the wrong MCE).

Technical comments:

L15-16: “coarse remotely-sensed burned area data… to provide near real-time information” – burnt area-based products, such as GFED or FireCCI, are not normally near real-time.  Near real-time products, such as GFAS, use fire radiative power or active fire counts to achieve this.

L32-33: “the generic GFAS global estimates” – what do the authors mean by this?  The authors are comparing emissions from France, so presumably the value from GFAS can’t be global.  What makes it ‘generic’?

L49: “limited information is currently accessible” – do the authors mean accessible, or available?  I.e. does the information exist but is not easy to get access to, or does the data simply not exist?

L71: “firefighters consistently raised concerns about lingering soil fires…” – please add citation if possible

L73: “wash away the burning soil material” – perhaps these fires were unusual, but are smouldering fires typically extinguished by the burning material being washed away through runoff?  Normally I’d assume it would be extinguished just through fully saturating the soil – at least for peaty soils, and especially where the fire has burnt down into the soil and/or is smouldering underground (one reason why it requires such large volumes of water to extinguish them, as the peat can hold a lot of moisture before combustion is inhibited, and can continue to smoulder below the surface layer meaning that the entire soil column needs to be saturated).  Rainfall may erode the already burnt fine ash, of course, but a priori it seems like you’d need really a lot of soil erosion for this mechanism to extinguish an active smouldering fire that’s occurring any significant depth into the soil.

L87-89 and L90-92: “Various studies of smoke chemical analysis… have determined MCE indices ranging from 0.6 to 0.8 during smoldering combustion” and “Hu and Rein (2022) recently compiled a review on smoldering combustion emission factors, with MCE indices varying from 0.93 for flaming in forests to 0.85 for peatland smoldering combustion” – these statements seem inconsistent.  How can the literature-average value of 0.85 in Hu and Rein, be outside of the range reported by ‘various studies’?

L163-164: “a spatial filtration process to exclude all thermal anomalies… corresponding to non-forest fires” – how did the authors spatially filter to exclude non-forest fires?  Presumably using a high-resolution map of land cover type, in which case the source should be cited here.  Was the map updated and current for 2022?  Or, if complied before 2022, could forested areas have changed over time?

L100: “variations in atmospheric MCE” – technically the MCE is a property of the combustion, not of the atmosphere.  Its value can be estimated from atmospheric measurements, but the atmosphere itself doesn’t have a combustion efficiency (well, not at normal temperatures/pressures anyway)!

L159-160: “we harnessed VIIRS data from the SNPP and NOAA sources” – The authors name the other satellite missions (Aqua, Terra, S-NPP) and so for consistency should name the second VIIRS satellite as well rather than just saying NOAA, who operate many satellites.  The satellite with VIIRS onboard is NOAA-20.

L172: “Fig. 5” – this figure should have been numbered as Fig. 2

L218-219: “The meteorological data encompassed parameters such as…” – ‘such as’ is insufficient; please give a complete list of the variables that were used as predictors in the random forest model.

L234: “we sought to estimate the pools” – carbon pools?

L291-292: “The bulk density of brown coal generally hovers around 700kgDM.m-3” – ‘generally hovers around’ to me implies something that fluctuates over time (e.g. stock prices, exchange rates, etc.), which I’m not sure is the case for bulk density of coal (except on geological timescales, I suppose).

L297-298: “we quantified CO2 and CO emissions” – ‘estimated’ might be a better term, to make it clear that (as I understand it) the CO2 and CO emissions are not directly measured or constrained, but rather inferred bottom-up based on estimated carbon pools and some assumptions of combustion completeness, smouldering fraction etc.

L318: “the entire fire (F)” – ‘F’ has already been used previously to mean “flaming (F)” (L298), making the E_Fx notation ambiguous.

L363-364: “The BIS site shows mostly low minimum values” – what counts as ‘low’?

L367: “ROC exhibited minimum values that reached 0.82” – Figure 4 seems to show hourly MCE values that are < 0.8, which doesn’t seem to match up with the text here.

L367: “far beyond” -> ‘far lower than’?

L431-433: “The resulting MCEs ranged from 0.955 to 0.961 for all the fires, with no significant distinctions between them. While these values closely mirrored the MCEs observed at the OHP tower, they notably diverged from the MCEs captured at the ROC and BIS stations” – the mean MCE for ROC was 0.94, so it’s not really diverging very far from this station either, surely?

L455-456: “Fires mainly altered forest areas in the Atlantic pine region (76.5%) and other forest (75.6%) regions” – I’m not clear what these percentages represent.  It seems to say that 77% of fires have been in Atlantic pine regions, while 76% of fires have been in other forest regions?  Which doesn’t make sense – maybe a typo?

L488: “soil vegetation” -> ‘soil and vegetation’?

L548-549: “172 (± 74) tC.ha-1 emitted, which is slightly higher than the value of 96tC/ha” – ‘slightly higher’ is kind of an understatement here.  172 is a lot higher than 96.

L570: “the existing generic fire emissions assessments” – I’m still not clear what makes them ‘generic’

L591: “excludes these processes” – again, not strictly true, since GFAS has a ‘peat’ land classification with a different fuel consumption rate and CO2/CO emission factors, and a ‘Extratropical forest with organic soil’ land classification with a higher fuel consumption rate than regular extratropical forest.  (Though I certainly agree that it’s only included in a very crude way, and the authors’ method is far more sophisticated!  And of course lignite is not included, and the contribution from lignite here is a very important conclusion to highlight).

L647: “might be actually true” -> ‘might also be true’?

L662: “lower” -> ‘low’

Figure 2: Axis labels on the contour bars are much too tiny to be readable!