Review for manuscript entitled: “The European Forest Disturbance Atlas: a forest disturbance monitoring system using the Landsat archive' submitted to the journal Earth System Science Data (Manuscript ID: essd-2024-361).

Overall:

In the manuscript the authors report the creation of an impressive dataset with multiple forest disturbances in Europe using multi-temporal Landsat data. The authors discuss an important topic that should be of interest to the readers of ESSD and establish large data set that includes a way to detect multiple disturbances within a single timeseries. I think the paper is well-written, but the methods description could be improved. Variable importance/selection is not discussed, and the reader is referred to other published work to assess the validity of the evaluation datasets. I commend the authors on producing such large dataset, that is consistently processed, well evaluated (at least I assume), and made openly available.

Major comments:

L125 onwards: the Explanation of the reference data collection is somewhat inadequate. I assume that interpreters used higher resolution data to interpret disturbance occurrence within Landsat pixels. If not, I think it would be very hard (except maybe for clear cuts) to detect and attribute a disturbance. The reader is referred to a couple of other papers, but since this is a critical step, I recommend the authors to discuss this step in more detail.

Methods: The methodology seems pretty solid, but there is no discussion of variable selection, variable importance, or reduction of auto-correlated variables. I agree that random forest is somewhat robust against overfitting (although opinions vary on this), the authors should at a minimum discuss the variable importance and detail a bit more what variables where chosen and why no model reduction was performed.

Note: In reviewing the final maps (which is an impressive dataset!) I found in some parts of Europe (in particular northern Europe/Scandinavia) >50% of the forest was disturbed and I wonder if this is realistic or whether the model is oversensitive/has many commission errors. Not sure how to test this as the accuracy metrics seem pretty balances. I just wonder if this is realistic… See also Sweden, Fig 5.

Minor comments:

L15: Could add “overlapping” to “…accounting for multiple overlapping disturbance events”. To be a bit more clear.

L24: Change to: “…spanning from timber production and carbon storage (Lindner et al., 2010), to water purification and regulation (Orsi et al., 2020), to recreation and spiritual value (Saarikoski et al., 2015).”

L67: “Shortly after, Senf and Seidl (2021a) created the first pan-European characterization of forest disturbance by combining a trajectory-segmentation algorithm (LandTrendr; Kennedy et al., 2010) with a random forest classification approach.”  this might need a bit more info since about you categorized LandTrendr into “(1) trajectory segmentation approaches”. What was the RF classification used for? Maybe also explain why this method cannot detect multiple disturbances per timeseries.

Figure 1: why does the validation stop -- connect an arrow to accuracy assessment?

L189-L191: How was this applied? Are all forested pixels smaller than 6 adjacent pixels in all configurations set to non-forest? Please clarify.

L197-199: Please provide formulas for the different spectral indices. Could be added to Fig. A5, if you needed to keep appendices to a minimum. Note that there are different TC indices for different Landsat sensors and the DI is somewhat less known.

L211: So the minimum forest patch is 6 pixels and within that the minimum disturbed patch is 3 pixels, if I understand this correctly. You might want to restate the minimum forest patch size again here to remind the readers.

L255-258: This needs more information -- “We thus used an approach developed in Senf and Seidl, 2021b and selected a random background sample from all disturbance patches. As harvest is assumed to be the major disturbance agent in Europe (Patacca et al., 2023; Seidl and Senf, 2024). This background sample will represent harvest conditions in contrast to the agent information available in the existing databases.”

• What do you mean be “random background sample”?
• How did you ensure this is harvest, please add more details here?

L267: I do not agree here: there are many bark beetle outbreaks (and other biotic disturbances) throughout Europe prior to 2017. For instance, a large outbreak in the 1990s and early 2000s on the German and Czech border. See Kautz at al. (2017) and this paper:  https://annforsci.biomedcentral.com/articles/10.1007/s13595-013-0279-7 and the many references therein. In my opinion its ok to combine the 2 classes but I would stay away from stating the bark beetle disturbance was minor without much evidence. For example: Fire and Bark beetles are of the same magnitude from 1980 to 2015 (Patacca et al, 2023, Fig. 5).

Kautz, M., Meddens, A.J., Hall, R.J. and Arneth, A., 2017. Biotic disturbances in Northern Hemisphere forests–a synthesis of recent data, uncertainties and implications for forest monitoring and modelling. Global Ecology and Biogeography, 26(5), pp.533-552.

Figure 10. There are very large discrepancies in 2018 across the different data sets, is there a way to improve this? In L449 this is attributed to dry years, but I assume there are many more “dry” years in the time series. Could you relate (correlate?) drought years to this overestimation?    

Note some minor spelling comments/edits in the attached pdf.

General comments

It is certainly a great paper, describing a huge processing effort to identify long-term changes in forested areas in Europe. The generated dataset will certainly be very useful to monitor changes throughout the European territory and identify the main drivers which should lead to better conservation of European forest patrimony.

I have a few general remarks and a longer list of specific comments that may help the authors to improve the current version of the manuscript, but my general impression is very positive.

First, it is not clear why the authors selected pixel instead of plots as reference data for calibrating and validating the outputs. Single pixels are difficult to characterize and may be affected by geolocation errors, particularly when detecting changes. In this regard it is not clear why the authors start commenting at length about this pixel reference database, but then make several analyses at patch level when they do agent attribution. It seems inconsistency here, which would be convenient to clarify. It would be convenient to clarify the text now included in sections 2.2 and 2.6 and explain why two different approaches were necessary to classify and to assign agents.

Second, it is not clear whether the authors used multitemporal changes in the detection of disturbances or classified single years without any consideration of previous years. There is a comment in lines 214-215 that indicates that double disturbances were removed in post-processing, but it would have been simpler to classify with both current year and previous year indices, so the change would have been included explicitly in the classification, thus removing double detections.

My third comment refers to the confusion that appears in some paragraphs of the paper between validation and uncertainty characterization. They are not the same, the former meaning the level of agreement with the reference data, and the latter the probability of being sure that the classification was properly done. This is clear when authors use the probability of the RF classifier, but they do not mention uncertainties derived from the pre-processing or the compositing periods or even the lack of observed areas. In addition, in lines 404-05 it is indicated that the authors aimed to provide a full characterization of uncertainty, but they talk about validation results.

Finally, the spatial validation is well explained but nothing is very little is said about the temporal reporting accuracy, that is how longer after the actual disturbance your product detects it. Include this on your validation approach with a proper consideration to the impacts of your annual compositing method.

Specific comments

Lines 114-15. Please further clarify the methods used for the annual compositing. The reader should not need to read another paper to understand yours.

Lines 118-19. Gap-filling with a previous year is quite controversial. In my opinion, it would have been more convenient to fill the date with a pre-summer image or just keep the unobserved data as unobserved and more the analysis to the following year.

Line 189: Not clear to me why you used a single forest mask for the whole period. What happens with areas that converted from crop to forest or from forest to urban during your time period? You indicate in line 191 that “All non-forest land use pixels were excluded from following analyses”, non-forested when, anytime in between 1983-2023. How about areas of new urbanization in that period?

Line 198, you include here the use of NDVI and Disturbance index but none of the two are quoted in table 2. Please clarify why?

Line 239-240: Not sure why do you consider that a burned area cannot occur in two consecutive years in a neighbour region. In Europe is less frequent but in other ecosystems is quite common.

Line 241: How patches were created? What is the impact of your compositing period to create those patches?

Figure 5. It is not clear what the colours mean, as three different colour legends are included above.

Figure 7. Include country abbreviations (perhaps in Table 1)

Line 350. Indicate why accuracy is reduced before 2000. I assume it will be a question of image availability, but this should have been considered when computing the uncertainty, and therefore the estimations done with broader confidence intervals.

Line 394. Include a comment of the disturbance return times in the paper, as you indicate is a novelty of your product.

Line 426. You are right that is difficult to compare you map with static land cover. Maybe you can make a comparison with the LC CCI product, which provides annual changes, even though at coarser resolution product than yours.

Figure 11 shows most likely ETM derived strips in the fire area of central Portugal. This was not commented in the paper and it should, as it creates an evident noise in the output. Please include the map location of the three sites.

Line 461. You indicate the potential problems to detect grading changes. Have you explore the use of spectral unmixing approaches?

Line 485. You indicate potential problems with several years with poor image availability. Have you tried to use MODIS 250m resolution reflectances for those years? A few methods that compare Landsat-MODIS to merge reflectances are available in the dedicated literature.

Line 505. I agree on the need of having publicly access validation databases. May be you can refer to the BARD, which is a global (not just European, but with European sites) validation database of burned perimeters (Franquesa et al. 2020).

References

Franquesa, M., Vanderhoof, M.K., Stavrakoudis, D., Gitas, I.Z., Roteta, E., Padilla, M., & Chuvieco, E. (2020). Development of a standard database of reference sites for validating global burned area products. Earth Syst. Sci. Data 12, 3229-3246.