Overall this is an interesting topic and highlights the various limitations in burned area products. I applaud the effort that went into creating the validation dataset, however, as highlighted below, there are some crucial details lacking in the training and validation descriptions that are necessary for a full review. Please see below for specific comments.

Specific/Major Comments

• Section 2.1: I assume the November and December 2018 images were only used for the pre-fire window analysis? How did the authors account for any burns that occurred during those 2 months?
• Section 2.2: What happens if a pixel burns more than once in a year? This might occur in the agricultural regions?
• Line 162: Are the authors not calculating dNBR here? Also, does the moving window continue beyond the first instance of detecting a potential burn? i.e. if the drop occurred on February 1^st, does the moving window continue to Dec 31 to see if it burned again?
• Line 177: Can the authors please provide a map of the locations of the 988 training pixels and their associated land cover types in the supplementary? Is 988 training pixels enough? How was that number decided upon?
• Line 183: dNBR already shows burn severity (here is an article with more information: https://un-spider.org/advisory-support/recommended-practices/recommended-practice-burn-severity/in-detail/normalized-burn-ratio). Did the authors quantify these values over their training pixels or simply rely on the color? I suggest the authors quantify these values to ensure the training pixels are in fact medium-to-high severity especially since the authors are prioritizing mapping high burn severity fires to reduce false positives.
• Section 2.4.1: As with the training samples, what land cover types were associated with the validation samples? Secondly, what size burn scars were these validation pixels associated with? For example, if all validation pixels were associated with very large burn scars then the validation results will be biased because large burns are easy to detect. Also, I assume the training and validation samples did not overlap?
• Line 310-311: Please explain why the authors only chose the cardinal directions?
• Line 389: While doing some reading into power laws and fire size, I came across this paper from the US Forest Service with the following quote: “Newman (2005) specifically excludes fire size distributions, while admitting that they might follow power laws over portions of their ranges. Current opinion is divided among those who would globally assign power laws to fire-size distributions (Minnich 1983; Bak et al. 1990; Malamud et al. 1998, 2005; Turcotte et al. 2002; Ricotta 2003) and those who would attribute them only to portions of distributions or rule them out altogether in favor of alternatives (Cumming 2001; Reed and McKelvey 2002; Clauset et al. 2007; Moritz et al., Chap. 3)” - https://www.fs.fed.us/rm/pubs_other/rmrs_2011_mckenzie_d001.pdf

             Please can the authors go through the literature and ensure their power-law assumption is correct and justify it in the paper.

• Line 399: The current analysis does not support this finding regarding agricultural burning. Based on Figure S3, the small patches are likely associated with the small burn patches surrounding the larger burn scars. Agricultural burning is a very difficult fire type to map and although the current methodology is likely to map more agricultural burning than MODIS (due to the finer resolution) the mapping methods and validation assessment was not adequately designed for agricultural burning. The authors can mention that the S2 mapping is better suited for identifying “small fires”. Furthermore, it was noted on lines 427 – 433 that the approach omitted hard–to–detect fires (e.g. savanna grasslands) which are much easier to detect than agricultural fires therefore that statement is not supported.

Minor Comments  

• Line 23: change to “..which occur on..”
• Line 27: Should the size of the intermediate fires read (100ha – 1000ha) similar to what you have in the main body of the paper?
• Line 88: change “excepting” to “except”
• Introduction: When are the peak burning months? It seems based on the GWIS country profiles (https://gwis.jrc.ec.europa.eu/apps/country.profile/charts) that the peak occurs August – October and since the authors are also referring to agricultural fires then please also include the cropland burning months.
• Line 137: Remove “and finally” after “Fourth. The authors go on to a final step on line 139.
• Figure 3: There is no h panel (line 679)
• Line 237: Should that be referencing Figure 3? Also, there is no panel h
• Line 696: change to “minimum mapping unit”
• Figure 5: Please add to the caption that the light grey represents countries outside of Indonesia. This was confusing at first before I looked at a map.
• Table S3: please add the meanings of Am and Wh to the caption
• Line 370: add a comma between “Figure 6 Figure S2”
• Line 702: Change to “MCD64A1”. There were a few other instances where the A was lowercase (i.e. MCD64a1)
• Line 377: It would be interesting for the authors to create a 3-panel figure showing this scar from each of the 3 products to show the omissions made in MCD64 and the Official dataset.
• Line 443: change to “addressing”

The MS has a clear structure and a good description of reference data collection. However, my major concern is the design of the mapping framework. The most confusing part is the integration of NBR and RF classification. As stated by the authors, NBR and dNBR were used to “detect the day when a pixel’s vegetation was disturbed by fire”. If this is valid, what was the point of using RF for burned area mapping? On the other hand, RF can be directly used for classification. What was the benefit of combining NBR and RF? Another concern is that the authors put the research in the context of national burned area mapping in Indonesia only, but there is a lack of clarification on the scientific contribution of the work. There is no literature review of the state-of-art on large-scale burned area mapping. There is no scientific objective (what scientific issue was addressed), and no comparisons with similar studies on the topic (expect for the comparison with the Official and the MODIS results).

Other comments follow:

Introduction: The background info about wildfire is too long, especially for the first four paragraphs. The emphasis should be burned area mapping.

L108: Isn't visual interpretation more accurate than machines? Many field data are from visual interpretation, including yours. It is important to point out the issues in their visual interpretation strategy/method/data.

L109: It is surprising that such a national campaign does not have protocols for accuracy assessment.

Section 2.2.: The purpose of getting the pre- and post-fire composites with NBR is confusing. Why did you use RF classification since you already identified the burned pixels with NBR? Or why not directly using RF to extract burned areas?

L163-165: The description of implementing your method is vague. What NBR and dNBR thresholds did you use? How did you know the variation of NBR was caused by wildfire, not other events (e.g., plant disease)? Do you also need to have a vegetation baseline map?

Section 2.3: How did you tune RF? What parameters did you use?

L177: Aren't the training samples too small for national scale mapping?

Results: You compared your overall results with the Official and the MODIS results. However, it is also important to pick sample locations to demonstrate what types of areas had high agreements and what areas caused discrepancies.

The study presents a new burned area product for the year 2019 in Indonesia based on high spatial resolution Sentinel 2 imagery and machine learning classification algorithm. Given the significant impacts that severe burning episodes in Indonesia have on global carbon cycle and population health across the wider region, the product presented in this study represents an important development and will be of interest to the scientific community. The approach and the dataset, nonetheless, have several limitations which I believe should be better articulated in the revised manuscript. In addition, I don’t think that the comparison of fire patch size distributions between different products adds much to the discussion here due to (i) large differences in spatial and temporal resolutions (or both) between the datasets and (ii) lack of definition what does fire patch represents here.

While the validation methodology does seem robust and the authors do demonstrate that total burned area estimates of the study are more accurate when compared to the alternative sources (MCD64A1 and the Official ba product), it has to be articulated that the algorithm of this study was optimised for the specific region and fire season and for a specific commission/omission error ratio. As a result, it is not clear how the burned area estimate for 2019 would change if the algorithm was optimized to fit training data from different years and regions by different users. In addition, extending temporal coverage of the dataset is not that straightforward as this would require substantial further work (somewhat arbitrary and time-intense selection of training data). Please see the bellow specific comments for further detail.

Specific comments

Lines 56-57: Given the uncertainty in burned area estimates (line 56), the Huijnen et al., 2016 estimate of CO2 emissions quoted in line 57 seems too certain. Do Huijnen et al give uncertainty estimate? Also, would be good to give another estimate for the event, given by GFED or Lohberger et al., (2018) or some other study etc. Large uncertainties in emission estimates is yet another reason why we need better burned area products, hence it would be good to point this out here.

Lines 160: Please explain what “Every two days” means here.

Line 161: Was data from the central day of the window included in prior or after median values (or neither)?

Line 163: This relates to the previous two points regarding temporal precision. Here and elsewhere the authors use “The day of the year”. How day of burn was determined if temporal resolution of Sentinel 2 is ~5 days as stated earlier? This suggests considerable uncertainty in day of burn estimate?

Lines 173-176: It is not clear here what was the total number of features used for classification? Please state in this paragraph.

Lines 177-178: Was the training sample fully independent from the validation sample? This is important to state clearly as it underpins the validity of the study’s findings.

Lines 260-265: What was done with classification of the sites which had either direct fire evidence (flame or smoke) but not indirect evidence (reddening) observed and vice versa? Where these samples (if any) discarded from the analysis?

Line 278: The final validation sample number N=1298 is the same as given in line 269 (all reviewed sites) and line 227 where it is termed as “initial sample”. Please clarify this.

Line 306: This statement needs a reference and an explanation. MODIS burned area pixel size is ~21ha. While sub-pixel burning can be detected, the actual minimum burned scar size will depend on environment/vegetation where burning is occurring. Is this estimate of 6.25ha is specific to Indonesia/tropical regions? In addition, I am puzzled by how MCD64A1 fire size histogram shown in Fig. S2 was computed; i.e. how counts for bins for fires < 21ha were derived given that MODIS pixel size is ~21ha?

Lines 366-367: The sentence is too “wordy” and complex. “greater detection of the realm of fire activity characterized by small-scale…” could be replaced with “greater detection of small fires” to the same effect.

Line 371: “lesser estimation” – perhaps change to underestimation?

Lines 380-389: The paragraph is too wordy. The first two sentences say nearly everything that needs to be said. Sentinel 2 sensor can indeed detect smaller fires enabling the detection of small scale agricultural burning. Perhaps cut shorter or even merge into previous paragraphs.

Lines 407-416: This paragraph is very speculative and not well supported. Not sure I agree with such interpretation of fire size frequency distributions. Any differences in distribution shape may arise from huge differences in sensor spatial (MODIS) and temporal (Official map) resolutions and also from the clustering (patch agglomeration) method. For example, the algorithm of the official product may have merged diagonally adjacent pixels as well and that would result in shift towards larger sizes. In addition, the provided references do not show that power-law approximates fire event sizes in Indonesia. While I’m not aware of fire size studies in Indonesia, log normal fire size distributions are common in some ecosystems (see Lehsten et al., 2014). As a result, it is perhaps better to avoid saying that fire sizes should follow power-law relationship and that this itself is a desirable property.

Lines 423-433: While advantages of the approach are discussed across several paragraphs in the Discussion, this is the only paragraph considering the limitations. Please add discussion on implications of changes in training dataset (due to different selection criteria, addition of data from different year etc.) on burned area estimates for 2019 and beyond in future application of the algorithm.

Lines 434-435: Not clear who are those “commentators” and “us” in “our ability” the authors refer to in the sentence. Please be more specific.

Lines 448-449: I do not understand why “large discrepancy for peatland burning” between the datasets would make the dataset of this study a “gold-standard”? Please explain this bold statement. Also, please consider replacing “gold-standard” with something less flashy as only time will tell how the dataset fares among users.

References

Lehsten, V., de Groot, W.J., Flannigan, M., George, C., Harmand, P. and Balzter, H., 2014. Wildfires in boreal ecoregions: Evaluating the power law assumption and intra‐annual and interannual variations. Journal of Geophysical Research: Biogeosciences, 119(1), pp.14-23.

Lohberger, S., Stängel, M., Atwood, E.C. and Siegert, F., 2018. Spatial evaluation of Indonesia's 2015 fire‐affected area and estimated carbon emissions using Sentinel‐1. Global change biology, 24(2), pp.644-654.