General comments:

The authors present an interesting and novel investigation that attempts to establish a new method for estimating landscape-scale methane emissions from degrading palsa wetlands and for detecting the initial stages of palsa degradation, which represent two urgently pressing issues in ongoing permafrost peatland research. The study combines ultra-high resolution imagery from a UAV with Sentinel-1 and -2 data to extrapolate field-measured methane emissions to the landscape-scale, using relationships with key vegetation types. The authors estimate that for the 50 ha study area, ongoing degradation could increase methane emissions by up to two orders of magnitude, indicating the importance of such observations and measurements for estimating carbon release from palsa sites. The distinction between vegetation types for estimation of CH4 emission rates is much needed for improving our understanding of degradation impacts, as many current studies use more simplistic comparisons between palsas and fens, as the authors note. The introduction, methods and discussion are generally well-written and interesting. I note some points for consideration below.

However, some further work is required on the analysis and presentation of the measured data before this manuscript is ready for publication. Specifically:

1. The method for calculating the headline finding of increasing CH4 emissions in subsiding areas in these sites (i.e. from 116 kg season^-1to 12,960 kg season^-1) is currently not well explained and unclear to me. Further clarification and justification of this extrapolation analysis is vital to ensure a robust interpretation of these results by other readers.

We have added a detailed explanation of how this was done in the data analysis section. The added section reads as follows:

”We also calculated the net emissions of CH₄ across the subsiding parts of the three wetlands (as detected by InSAR). This was done by summing the CH₄ emissions attributed to each subsiding InSAR pixel for all three sites based on the dominant vegetation type (using the UAV vegetation classification) in that pixel. To explore how CH₄ emissions may change as subsidence progresses and results in palsa collapse, waterlogging and vegetation change we multiplied the currently subsiding wetland areas with the seasonal CH₄ emission that was measured in the lower and higher emitting fen vegetation types, respectively to generate two exemplar future emissions scenarios. This “after permafrost loss and palsa collapse” CH₄ emission scenario used the assumption that all of the subsiding area will be covered with a particular fen vegetation type. We want to highlight that this is not a realistic scenario as the fen vegetation that will replace the current vegetation on the raised palsa will be a mixture of vegetation communities suited to permanently waterlogged conditions. Hence the calculations of future CH₄ emissions should be viewed as illustrations of the ranges of emissions that could occur from the sites following full transition to a fen vegetation type based on their current day emissions.”

1. Secondly, several improvements could be made to assist the readability and presentation of figures and figure captions, most particularly Figure 2 where panel D is currently missing any data.

We have corrected the figures as suggested.

1. I recommend a global check of the manuscript for typographical errors, because I encountered several mistakes (for example, incorrect figure numbers referred to in the text).

We have corrected the figure numbers and the symbols that had got corrupted.

Overall, this research is a novel and useful contribution to the field of permafrost peatland research, but currently requires some polish to its presentation to be ready for publication.   

Specific comments:

L29: It is very unclear both here in the abstract, and in the main text, how the results for future CH4 emissions have been reached. As this is the headline finding of the paper, it is important that further clarification is provided for the methods used to calculate these statistics, and some interpretation in the main text for why there is such a large range between these figures.

We have added a description of these calculations in the data analysis section as per our response to the reviewers point made above.

L41–43:  Although the study’s focus is on methane, permafrost thaw can also cause substantial carbon dioxide (CO2) release. It would be useful in the introduction to include some comparison of the relative warming potential or persistence of methane compared to CO2. 

In the introduction we state that the overall carbon balance in permafrost and non-permafrost peatlands is similar based on a synthesis study by Olefeldt et al., 2012. In response to the reviewers comment we have specified the global warming potential of methane.

L44-46: Given the current limited mapping of palsas across much of Siberia (e.g. see Fewster et al., 2022), I don't believe there is enough evidence to state a confident areal extent of palsa peatlands across the total permafrost region. Instead, perhaps it would be better to state geographic regions in which they are most commonly found.

We have changed the text to reflect the reviewer’s point.

L46: Tarnocai et al. (2009) do not discuss palsa extents so this reference is unsuitable for making this point.

We have removed this reference.

L46-47:  These references for peat carbon stocks do not provide estimates for the total carbon stored specifically in palsa mires. Both studies are now > 10 years old and peat carbon estimates been more recently been improved. It seems more appropriate to describe the permafrost peatland carbon store more generally, using updated carbon maps – for example, see Hugelius et al. (2020) (https://doi.org/10.1073/pnas.1916387117).

We have added the Hugelius reference to this section.

L107: What reference data were used to calculate these MAT and MAP ranges (i.e. weather station observations or gridded climatologies)? A reference to this climate data is important. Please also provide specifics on the time period considered.

We used data from the Swedish Meterological and Hydrological Institute. We have added this information to the text.

L130: Figure 1: I would recommend revising the colour scheme used in Figure 1, because the vibrant green and red classes are not colourblind friendly and it is difficult to distinguish the lake shoreline from the terrestrial green classes.

This colour scheme is the standard CORINE colour scheme so we have kept this for consistency with other users.

L135-136: This sentence is slightly unclear to me. Were these measurements taken from intact palsa tops? Was there a reason why all vegetation types not studied in both subsiding and intact areas?

Vegetation change consistently in response to degradation. We have clarified which vegetation types are associated with the intact drier areas and the water logged degraded areas.

L236-237: How variable were the measured methane emissions within each landcover type? Do the results change substantially if an alternative averaging method is used, e.g. the median?

The greatest variation in CH4 fluxes were found in the sites with the highest fluxes. The pattern in the data with regards to the magnitude of fluxes from different vegetation types is the same for the arithmetic means as the predicted means of the mixed models.

L236-237: When considering the wider applications of these methods, were the methane measurements from these land covers similar or dissimilar to existing measurements from nearby palsa mires? Do sites require ground truthing of methane measurements before such methods can be employed? Some further comparisons to previous measurements in the discussion would be interesting and useful.

The flux measurements compare to fluxes measured in similar system. However, our data show there is high variability among vegetation types, which may be linked to specific species. Hence, we anticipate that if the system for upscaling and assessing future methane emission risk we have tested here was to be applied to larger areas in the region or sites further-a-field additional ground methane data as well as vegetation ground truthing to reflect the most relevant communities in that area would most likely be needed. We have added some indications of fluxes measured elsewhere in the discussion for comparison as suggested.

L278-280: It is unclear what temperature is being measured here – is this soil temperature? Please clarify.

It is soil temperature, We have added this detail to the text.

L285: Figure 2: This figure requires revision. Panel D is missing any data bars. Standard error whiskers only extend above each bar, but should also extend below. For panel C I would recommend the removal of the dry lichen sub. class as it currently reads as no methane flux, when actually no measurements were recorded. In the caption, it is important to clarify where the means and SE are shown (e.g. with bars and whiskers).

The missing data must have occurred in the file upload/processing. We will ensure all the panels display correctly in the revised version. It is currently stated I the caption that mean and SE are shown. We have kept the dry lichen sub in the figure to ensure the x-axis are the same format among panels as this facilitates comparison among the different vegetation types and time points. It is clearly stated I the caption that there was no data from the dry lichen sub vegetation type in September.

L295: Figure 3: Recommend offsetting each data point slightly in the horizontal axis to avoid overwriting.

We have made this change.

L399: Figure 4: Additional information is needed in the caption, detailing what values are shown by boxplot centre lines (medians?), whiskers and closed circles.

We have added this information.

L309: Was there an objective cutoff point for determining a suitable accuracy? How do these performance results compare to previous classification studies?

There is no universal threshold accuracy target. One that is widely used is 85% overall accuracy but this is for Anderson Level 1 (very broad classes) and a lower accuracy would be expected with more detailed classes such as used in the paper. A cutoff of 70% was used to determine a sufficient accuracy for the vegetation mapping. We decided on this threshold because it is high enough to ensure good overall accuracy, whilst not so high to account for the difficulty of mapping classes with very similar spectral signatures e.g. non-subsiding vegetation classes and their subsiding counterparts. Furthermore, it is also important to note that ground data are never perfect. Indeed, the level of error can be quite large even when arising from authoritative sources. For example, trained aerial photograph interpreters are known to differ in their labelling of forest classes by up to 30% in some studies (Powell et al., 2004). We are being upfront on the quality, which is rare. This justification has been added to the manuscript.

L330: Does the misclassification of several raised palsa areas in the study area indicate that there may also be an overestimate of predicted emissions in extrapolated results?

The fact that the soft classification overestimated the fen vegetation area resulted in an over estimation compared to the UAV data see Table 3.

L375: Table 3: The caption seems to indicate that four sets of results should be in the table (i.e. the “one model using the UAV data” appears to be missing). The mean and SE should be labelled in the table.

Yes, one column did not display correctly in the document. We have altered the formatting, so this displays correctly now. We have specified which column shows the mean and se.

L383-385: If such a strong bias-correction is required to correct the results for the trained study site, are these methods really suitable for estimating classifications outside of the study domain, where bias-correction would be more difficult without ground-truthed data? Perhaps this evaluation is worth greater discussion.

The approach presented in this paper, is based on a probability sample of the region mapped. The sample is thus representative of that study area but the referee is correct that it may not be representative of other regions but we don’t go to others in this paper. The methods are transferable, but there is a need for ground data and with that ground data we show the impact of the bias-correction method that is simple but extremely effective. We add this point to the manuscript.

L388-389: It is suprising that the hard classification still returned a highly significant correlation to the UAV estimates when the R2 value is much lower than the other relationships. Does this indicate the low numbers of point estimates (four per model) are unsuitable for significance testing?

We are grateful for the reviewer to pick up on this. We had made a typo and the hard classification regression model was indeed not significant reflecting the low classification accuracy of the hard classification model. We have updated the text and figure to reflect this.

L396: Here it would be useful to indicate a range and average for the measured subsidence in mm.

We have added this detail.

L413-414: Figure 7 does not have any lettered panels. Should this relate to Figure 9?

Yes it should be figure 9. We have corrected this.

L428-430: This section is very vague, and appears to represent a headline finding for the study. Please clarify in greater detail the vegetation types used to estimate the future emissions, and how these calculations were applied. I did not follow the calculation of these estimates.

We have added a detailed description of how these emissions were calculated and the assumptions that they were based on in the data analysis section.

L433: Figure 9: The lettered panel labels are very hard to read. North arrows were only present in one figure panel – please add to all panels.

We have made these changes

L439-441: How comparable are these estimates? If possible, it would be helpful to report a range of estimated values from these studies.

We have added some quantitative info for comparison.

L449-453: Similarly, it would be useful to report the ranges to show readers how close these estimates are to the upper end.

We have added a couple of values from the literature to illustrate.

L477: Is this referring to the bias-corrected hard classification? It would be helpful to use consistent terms throughout.

Yes, we have checked the text to ensure consistent use of this term.

L482-484: How easily available are high-resolution remote sensing data for permafrost peatland regions (e.g. central and eastern Siberia are currently very poorly studied)? Some discussion of the availability of these data would contextualise how easy these methods would be to upscale.

We have added the following sentence to emphaze this:

“Spatially distributed high-resolution studies and data collection in remote areas such as the Canadian Arctic and Siberia are in the context of permafrost degradation and CH₄ feedbacks of high value and necessary to calibrate regional to Arctic estimates due to high spatial variability of tundra terrain”

 Technical corrections:

L56: “which have been”.

We have made this change

L106: add degree symbol.

We have made this change

 L115: Remove additional bracket.

We have made this change

L146: consider removing “point”?

We have made this change

L280: degree symbol missing.

We have added this

L314: Should this read “shadows are not included”?

Yes, we have corrected this

L316: missing symbol before “Sites”

We have added in the sum symbol

L342: change “methane” to CH4?

We have made this change

L344: change “has” to “have”

We have made this change

L372: change “were” to “where”

We have made this change

L447: change “were” to “where”

Were is correct in this instance

L477: “Hence improved understanding of the..”

We have made this change

This paper investigated the effects of subsidence and permafrost degradation on methane emission over northern Sweden under Arctic warming. Appropriate in-situ measurements and drone & satellite remote sensing data records were used in this investigation. The results show finer spatial resolution vegetation map derived from Sentinel-2 data using a support vector machine classification method. They also found that permafrost degradation and subsidence contributed to methane emission in various ways. They concluded that a fusion of EO data provided an ability to estimate regional methane emissions. The paper covers a topic that is suitable to readers of Biogeosciences and should be of particular interest to those interested in climate change impacts on greenhouse gas emissions at higher latitudes. However, the manuscript has concluded with lack of detail in data (e.g., sentinel-1, -2), methods (e.g., methane emission modeling), and without new findings and strong conclusion (e.g., not clear relationship between permafrost degradation and methane emission). The paper should be suitable for publication following the recommend major revision:

- Line 184-186, pg7: how did authors determine three seasons (spring, summer and autumn) using NDVI and/or air temperature? Was NDVI derived from in-situ camera or satellite? Is air temperature measured at 2-m height? Authors should specify it.

We added more detail to indicate how we determined the length of the season. We Did not go into technical details of the methods, as they are outlined in the indicated references.

- Section 2.4: Authors should include brief description on the UAV image processing (e.g., radiometric calibration for reflectance conversion, observation date and time (e.g, noon?), how many images?), even though specific software and a reference are provided.

We have revised the paragraph and added additional information as requested by the reviewer.

- Sentinel-2: Optical data are generally contaminated by frequent cloud, aerosol and low solar radiation and angle at higher latitude? Were atmospherically corrected reflectance data used to classify vegetation map? Did you use cloud-free images? What dates did you use? If you combined two or more images for the study domain, did you use the images observed at the same date and time?

Cloud-free, atmospherically corrected Sentinel-2 reflectance data from 27/07/2019 was used to produce vegetation maps for the sites of interest. Only one image was needed to cover the wider study domain. This has now been added to the text in Section 2.5.

-Line209, pg7: Authors should include Sentinel-2 specification, e.g., spatial resolution, temporal resolution, atmospheric correction scheme, wavelength for each band, overpass times and so on. What are wavebands 2., 3, 4, …?

We have now included information on atmospheric correction scheme in the text. A table summarising wavelength and spatial resolution for each Sentinel-2 waveband used in the study has also been included in Section 2.5. We have not included temporal resolution information because this study has only included one acquisition, rather than collecting a time series of data, therefore we did not feel it was relevant.

-Line241, pg8: in a similar way, authors should include the sensor specification for Sentinel-1 satellite.

Sensor specification has now been included in the text in Section 2.6.

- Vegetation classification accuracy assessment from sentinel-2: Accurate vegetation classification map is critical to estimate methane emission in this investigation. However, the 71% accuracy (line 302, pg12) is not high enough to be used as ground-truth data for the Sentinel-2 classification. Authors should justify it.

This relates to a point made by the other referee. The same response applies here but we also reiterate that ground data are never perfect and a core recommendation in best practice advice is to use ground data that are anticipated to be more accurate than the classification to be assessed. It is also important to not imply the ground data are perfect hence why the paper explicitly states the accuracy (many studies do not).

- Authors have not included how to upscale methane emission from in-situ. Authors should provide the details on modeling methane emission in method section.

We have added this to the data analysis section

-There is not clear new findings compared to previous studies. Generally higher water content, the higher methane emission.

The novelty in this work is the higher resolution vegetation classification using a combination of high resolution UAV data, Sentinel- 1 and 2 to map methane emissions and also develop systems (i.e. the InSAR subsidence data) for predicting future areas at risk of becoming high methane emitters.

- They concluded a fusion of EO data types provides the ability to estimate methane emission. Authors have not included how to fuse EO data records for regional methane emission. Authors should revise conclusions in a better way.

We replaced the word ‘fusion’ by the word ‘combination’. We did not mean data fusion as semantically used within the field of remote sensing. 

Additional edits are noted below:

Line 2-3, pg1: This investigation focuses on Northern Sweden. Sub-Arctic peatlands is not clear to describe the study domain. Authors should include study area into the title, e.g., over Northern Sweden.

We have added this to the title.

Line 47, pg2: which data? Is it observation or model simulation? Authors should specify it. Is the 100 GtC stored in entire Arctic region or study area? Authors should clarify it.

We have specified it is for the Arctic region. The Tarnocai paper uses peat cores from across the Arctic for its estimates and modelled permafrost data and remotely sensed vegetation data.

Line65-66, pg3: It seems to me that there are too many references (8 papers). Authors might make a few separate statements, by grouping based on subsidence detection algorithm, sensor types.

We have streamlined the referencing in this section. Now the references are directly relevant.

Line69, pg3: Publically should be better instead of freely.

We have made this change

Line75-77, pg3: is it underestimated all the time? Authors should include some references to support this statement.

We have added a reference that discusses the concept of hotspots in biogeochemistry.

Line77-79, pg3: Authors should include relevant papers to support this statement.

We have added a reference to support this statement.

Line83-87, pg3: Authors should include relevant papers to support this statement.

We have added a reference to support this statement.

Line112-115, pg4: how did you identify vegetation species? By visual inspection?

Yes, this was done by a team member with extensive experience of working with Arctic wetland plants.

Line194-196, pg7: This statement is not clear to me. Did you use DEM measured from Parrot Sequoia multispectral camera?

Yes, that is correct. We have clarified this I the text.

Line197, pg7: Authors should provide the resampling method.

The resampling method (nearest neighbour interpolation) has now been included.

Line201-205, pg7: Did authors use UAV reflectance data to apply support vector machine classification (ArcGIS)? Which bands did you use? Authors should provide more details on the classification method.

The four bands described in the previous paragraph were used for SVM classification. This has now been clarified in the description.

Line208, pg7: Are there any specific reasons why authors used a different classification algorithm (neural network classification) and tools (QGIS) from the UAV based-classification?

Due to the coarser Sentinel-2 spatial resolution (20m) compared to UAV spatial resolution (0.5m), we wanted to use a method that would enable fuzzy classification using proportions of each UAV land cover class within each Sentinel-2 pixel. Neural networks in R were more easily implemented for such purposes, with the benefit of faster processing speeds. We have now added this detail to Section 2.5, as well as corrected the tool used to build the neural network (manuscript previously said QGIS).

Line241, pg8: Authors should include the frequency for C-band.

Wavelength and frequency of the Sentinel-1 C-band instrument has been added.

Line245, pg8: What is APSIS or IBSAS?

ASPIS stands for Advanced Pixel System using Intermittent Baseline Subset. This is set out in the introduction. We have now written out what the ISBAS (Intermittent Small Baseline Subset) abbreviation stands for in the text.

Line271, pg9: how did authors identify signs of permafrost degradation/subsidence? Was it done from fieldwork by visual inspection? Authors should clarify it.

Subsiding areas was identified visually based on changes in topography, subsidence, and waterlogging. We have clarified this I the experimental design section.

Line278-279, pg9: In contrast to temperature, soil moisture does not show large variation from June to September, except for Dry lichen sub and moist moss sub. Dry areas get drier as temperature increases. What is initial thaw period (e.g., snow melting, soil thawing)? Did you define spring thaw timing in the previous section?

The raised parts of the sites tend to get snow free in April/May but snow can remain in hollows into June. We have added this detail to the site description.

Line280, pg9: Figure 3b shows lowest temperature occurs in September, not in June.

We have corrected this mistake.

Line287, pg10: What does season mean? Did you define the season?

The definition of season is in section 2.3 CH4 emissions measurements

Line398, pg18: Fig 6 should be fig 7.

We have corrected this.

Line400, pg18: Fig 7 should be fig 8.

We have corrected this

Line443, pg22: did you present the degradation stage and level of flooding in the results? If not, authors should revise it.

We have clarified this sentence to make it clear what we refer to when we talk about degradation.

Line454-455, pg22: Are these species in wetland vegetation shown in Table 3? Or were these species found in the study domain? Authors should clarify it.

We have specified that these species form part of the vegetation communities in the study area.

Line464-466, pg22: Are there any references on species composition impacts on methane emissions?

 We have added references to support this

Figure 4: Authors should describe what error bars and black dots represent.

We have added this explanation in the figure caption

Figure 9: what background map did you use? Google earth map? Authors should describe it in the caption.

We have clarified in the figure caption that the backgrounds are orthophotos over the sites.

Table 3: What are three different models? What is a model using UAV data? Authors should describe in main text.

The models are described in section 2.5 Sentinel-2 remotely sensed data for vegetation mapping. Unfortunately, a section of the table with the UAV had got cut off in the PDF, we have amended this now. We have improved the description of how this data was derived in the table heading and refers to Table 3 in the section of the text were we describe the results of the methane scaling using the different Selntinel-2 models compared to the UAV data methane estimates.