This manuscript presents a new database of lightning-ignited wildfires in boreal forests of North America and Eurasia. The new database is produced by selecting lightning candidates (measured by ENTLN) for wildfires detected by MODIS. The selection of lightning candidates is based on a spatio-temporal criterion. To evaluate the method of selecting candidate rays based on MODIS information, the authors compare the results with those obtained using national fire databases in North America, which are more accurate. Lightning-ignited wildfires in boreal forests play a crucial role in climate dynamics. However, significant uncertainties remain regarding these fires, primarily due to the limited availability of detection instruments, particularly in Eurasian regions.

The manuscript is well written, and the results are interesting. The database could be very useful to investigate lightning-induced wildfires in boreal forest. However, there are still significant questions that the authors should address before the manuscript can be published.

Line-by-line comments

- Line 10: "The frequency and severity of fire weather have increased under climate change, particularly in high-latitude boreal

forests.". I think this is not clear. See Xing and Wang (2023, https://doi.org/10.1029/2023JD038946).

- Section 2.3.1:

* I think the authors should provide a estimation of the Detection Efficiency of ENTLN in Russia. Maybe compare with ISS-LIS lightning measurements?

* Do you use flashes or strokes measured by ENTLN?

- Line 194: What are "non-native forests"? Please define.

- Lines 127.128: It seems you have used a very simple classification in which you only mention fire embracer species and post-fire resprouter species. What reference or references did you use to classify tree species into fire-related strategies? You may take a look at the references included within the Table S1 in Moris et al. (2022, https://doi.org/10.1007/s10980-022-01478-w) for some more comprehensive classifications.

- Line 128: What do you mean by “can regenerate independently”, and “that require species members to regenerate”.

- Lines 229-230: Please mention again that this phrase applies for lightning candidates outside the perimeter.

- Line 172:  I think you should include where you downloaded the Canadian and Alaskan data from, like a website address.

-  Table A1: Based on Table A1, I do not think that forest vs non forest are the most appropriate names. Maybe “natural vegetation” or something like that could be more suitable. For example, the forest class included shrublandsand and grasslands. In addition, why is “Permanent Snow and Ice” included in the forest class?

-  Lines 196-197: Why these 3 fires did not have a country?

- Section 2.6:  It is not clear to me that 14-days window is enough for boreal forests in Eurasia, where the holdover times tend to be long compared with other regions. Could you provide some references to support this? If not, could you maybe analyze how the results would change by using another quantiy?

- Table 2. How can the approaches “MaxA” and “Dmin” have a different proportion under the colum “% Total”. If both method use 10 km and 14 days from the ignition point, that should provide the same number of fires matched. However, I can see in Figure 2 that you allowed for a holdover time of -1 day in Dmin. I guess the difference must come from this aspect. Why only for Dmin? And not for MaxA? The start date for a point is the same no matter what selection criterion you apply.

- Figure 2. I think two aspects would be beneficial in this figure. First, to use the same scales for the Y axes. Second, I understand you use difference scales for the Tmin distance distribution given the large differences in values. However, it would help for the comparison if you include a smaller plot within this plot, in which you show the distance distribution between 0 and 10 km in the same wa. This way we can observe the differences with the other two plots from MaxA and Dmin.

- Figure 2. Have you plot the distance distribution using 1km bars for the 3 methods? And if so, do they look similar to other studies of lightning fires? From my personal experience, the distance distribution may be useful sometimes to detect if something is wrong with the lightning-fire matching. For example, when the harmonization of times are done not correctly, the matching suffers from “artificial” time lags (e.g., 1, 2 or more hours due to different time zones) causing that the selection may be different for a part of the fires. Have you double checked that this is not the case in your validation exercise for MaxA and Dmin? Maybe this concern doesn’t apply to you if all times are given in days for the fires. Therefore, everything may be correct and your distance distributions really reflect the distribution from a suitable matching…

- Table B2. You forgot to include the unit of holdover time in days. I don’t think you need to put the distances to centimeters.

- Lines 315-318:  I have missed something here, or something that I don’t really understand. How did you calculate the overall accuracy of Table B3? What is a correctly and an incorrectly matched? For that, don’t you need a “true” match? Or you simply classified as “correctly matched” if the selected strokes was reported inside the perimeter and as “incorrectly matched” if they were outside the perimeter? If so, please, explain it. In my opinion, it’s not clear what you have done here.

In addition, calling that as “correctly or incorrectly” match is misleading in my opinion (if you are using inside vs outside the perimeters). I suggest that you use a different term if possible. A selected stroke may have occurred inside the perimeter of the stroke but because of the location accuracy of lightning data the stroke is simply reported outside the perimeter. That’s the whole point of using spatial buffers to find and select igniting lightning, we don’t know the exact ground location of lightning.

- Line 415:  To avoid misunderstandings, it may be useful to clarify that you are calculating LIE for only a part of the lightning fires (e.g. > 200 ha), and not for all fires. Furthermore, as explained in Table C1, you are missing many lightning fires in NA, and so the LIE given in this manuscript are only indicative to compare LIE among regions using the same methodology, not to give insights about the LIE itself.

General comments

- I understand that the original dataset is the GlobFire Fire Perimeters. A biome, fire size and land use class filters are applied. But after that, all polygons are used to find a matched stroke? If there is a matched stroke, then the perimeters is included in the database? If so, the main limitation of this dataset is that not all fires included in it must be lightning-caused fires. Simply, the dataset includes fires for which there is a match and possibly lightning ignition source, but it gives no indication about the fire cause. This must be clearly reflected, but it can be misleading otherwise. Potential users of the dataset must know that some of the fires are possibly human-caused fires.

It is true that Table C1 gives an indication about how many missing lightning fires there are may be in the BoLtFire dataset, and how many might be human fires (assuming that the 483 fires not match with agency data are actually human-caused fires), but as you know, this does not have to reflect the situation in Eurasia.

In conclusion, I think you should explain a bit clear how the BoLtFire dataset was created, and mention clearly in the section on limitations that this dataset does not explore fire cause (i.e., not all the fires has to be caused by lightning), even if the majority are natural.

- I think that I understood correctly you method. You simply applied the minimum holdover time selection criterion (see Moris et al. 2020 and 2023), and if no lightning strokes (or flashes?) are located within the perimeter and 14 days, then you applied the maximum index A selection criterion using the 10 km buffer around the fire perimeter and the distance to the perimeter as the distance used for the index A. I have a few doubts and comments about your method:

A) In these two sentences, “The first candidate lightning found within the perimeter is designated as the candidate lightning and the ignition point. If multiple potential candidate lightning are found, the one closest in time to the ignition date is chosen”, I assume that this is equivalent to the minimum holdover time method. Thus, I don’t really understand the part “The first candidate lightning found within the perimeter”. If only one stroke is found within the perimeter, no selection is needed, so what is the “first candidate”? In addition, the second sentence is enough to explain that you applied the minimum holdover time for all CG strokes reported within the perimeter (allowing for a maximum of 14 days).

B) I think it would be great if you explicitly cite or use a terminology that allows the reader to be aware that your selection criteria were already used in the past extensively. After all, your approach is based on a 2-step process in which you applied the minimum holdover time and, and if necessary, the maximum index A subsequently. The main novelty of your approach is to use perimeters instead of points for the stroke selection. For instance, Pineda et al. (2022, https://doi.org/10.1016/j.agrformet.2022.109111) applied first a 3-day temporal window, and if no lightning were reported within that period, a maximum of 10 days were then allowed.

C) I am curious about the temporal aspect of your approach. You applied the minimum holdover time, and you only had information on data for the fire discovery. If, for instance, a fire has a “StartDate” on July 12th, the first stroke (within the perimeter) before 23:59:59 on that day will be selected as the ignition source? For instance, if there are only 2 strokes on that day inside the perimeter, one at 23:55 and one at 00:12, which one will be selected?

D) In addition, are the time of fires and lightning strokes using different time zones? I guess lightning data are in UTC, but what about the fire dates? The fire start dates used local times or UTC? For example, the difference in local time between Kamchatka and Alaska must be almost one day. If the start dates that appear in the fires are based on local times, this could have an influence in the matching? For instance, for ignition points with local times reported in hours and/or minutes, in countries like the USA, it’s absolutely fundamental to harmonize the times of lightning and fires before the matching due to the different time zones.

E) Finally, I am confused about how you named your approach “Temporal Minimum Distance (TMin)”. I thought it was something involving the closest distance, such as the method used by Schultz in the USA (i.e. DMin), but your method mainly applies the minimum holdover time (in combination with the maximum index A if needed). To me, “Minimum time” reflects better what you have done for the selection, although it’s true that it doesn’t add the second potential step on maximum index A.

- It would be nice to see the distribution of how many strokes are considered for the selection in each fire in the TMin approach versus MaxA and Dmin approach. Given that using a perimeter, and especially a large perimeter will ensure that more strokes are gathered before applying the selection criteria. This can be seen in the column “% Total” of TMin, where the number of fires with a selected stroke increases with fire size. I guess the TMin approach uses, generally, more strokes to select the most likely one.

Review Results

1. The study appears to have been undertaken to create a dataset of lightning-caused wildfires across the boreal forest and to characterise lightning-caused wildfires. However, the classification results show no novelty in the characteristics of lightning-induced wildfires that have been identified.

No analysis examples are provided by applying the dataset.

2. It seems to aim to complement data from forest departments in Alaska, Canada and other countries, but this makes no sense at all.　It is well known that there are estimation errors in the area burned and other factors.

3. In addition, there are problems with the data used for the analysis.

In particular, the lightning strike data are from different detection systems in different regions, which means that there are obvious deviations in the data sets obtained.

4. With regard to fire data, hotspot data is only used for the period from 2012 onwards, even though there are 20 years of data available from 2003 onwards. This results in a poor analysis of fire data and no mention of the major fire in the Republic of Sakha in 2021. There are doubts about the assessment of wildfires in Eurasia using the wildfire dataset.

****

2021 Russia wildfires

https://en.wikipedia.org/wiki/2021_Russia_wildfires

Hayasaka, Hiroshi. 2021. "Rare and Extreme Wildland Fire in Sakha in 2021" Atmosphere 12, no. 12: 1572. https://doi.org/10.3390/atmos12121572

Hayasaka, H. Fire Weather Conditions in Boreal and Polar Regions in 2002–2021. Atmosphere 2022, 13, 1117. https://doi.org/10.3390/atmos13071117

In this manuscript, the authors present a dataset of lightning-ignited wildfires that they created using proprietary data on lightning strike times and locations, and wildfire perimeters from GlobFire. The manuscript is generally well-written, but with a few places that could use a little more clarity. To my knowledge this dataset is novel, and could lead to important advances in understanding of boreal wildfire occurrence. I think it is publishable after some minor revisions.

My specific comments are below:

89-92: I think this statement could be modified. MCD64 and FIRECCI are raster-based products that do not delineate events from pixels of burned area, and therefore ignition location would not make sense. FRY does not to my knowledge provide fire events. There are 3 event-based global products, where MCD64 pixels are delineated into events, that I am aware of. These are FIRED (Mahood et al 2022), the global fire atlas (GFA; Andela et al 2020); and GlobFire (Artes 2019). FIRED does have (crudely) estimated ignition locations (the centroid of all pixels that burned on the first day of the event), and I'm not sure about the other two. I do know that none of those products have cause of ignition. Therefore, I think that while it is worth mentioning MCD64 and FireCCI, those are datasets of burned pixels, not wildfire event datasets. 

98: suggestion: reword to "... that do not rely on pre-existing ignition locaitons..." or something along those lines

170-171: Please cite the Canadian and Alaskan datasets

243: the DMin and MaxA methodologies don't seem to be described anywhere. Please provide a short description.

250-255: The sentence "The BoLtFire dataset successfully located a fire if it was within seven days (before or after the start date) and 10 km (Fusco et al., 2019)" seems to be repeated.

289: Wording is a little weird to me, essentially due to the use of the word 'time' twice. Maybe instead say longer holdover times were exponentially less frequent.

Figure 1: specify in the figure or in the caption that the grey outline is the wildfire perimeter and the dotted line is the buffer.

Figure 3: It might be nice to include the candidate lightning ignition points in these figures. The Landsat 8 basemap is a little busy and perhaps unnecessary. It might be a little easier to look at if it's in the natural projection of the polygon.

Figure 5: increase font size and resolution: legends are illegible.

Table 1: Spell out the acronyms either in the caption or in the table.