The authors tackle a timely and relevant research topic, that is a driver analysis of tree mortality in Luxembourg across several years. The study is well written, the proposed methodology is cleanly executed, and the related work is reviewed in depth. The application of GAMs in this specific context is also novel.

However, there are two major methodological design concerns:

The authors report the vicinity of existing mortality to be the best predictor of new mortality, which is not a driver, as promised in the paper title.

1.1) Response: We appreciate the reviewer’s thoughtful comment and partially agree with this assessment. We acknowledge that the spread of mortality is not a direct driver of tree mortality per se, but is tightly linked to it, as bark beetle infestations are often triggered or amplified by drought stress. To address this point, we will further elaborate this causal chain in the Introduction and Discussion, and we will frame our findings more cautiously by emphasizing drought as the primary driver, with bark beetle outbreaks acting as a subsequent biotic mechanism affecting the spatial distribution of canopy mortality across the landscape.

We would like to retain the term driver, as it is commonly used in ecological studies to describe both direct and indirect influences on ecological processes. To avoid ambiguity, we will explicitly define its use in our manuscript accordingly: "In this study, drivers are defined as biotic or abiotic factors that directly or indirectly affect the spatial distribution of canopy mortality during a major, multi-year drought. Under this definition, bark beetle infestation represents a biotic driver that mediates drought-induced mortality, which we characterize using four proxy variables defined as follows: “Distance to previously dead canopies in 2017 and 2019”, “Distance to closest (other) dead canopy in 2020”, and “Change in dead canopy area”.

The only covariate with relevant explanatory power was forest height.  In the case of conifer trees,  that bark beetles spread within some radius and do not normally travel cross-country as a random process is expected. Analogous, potentially for broadleaf, similar locations have similar trees in similar settings causing them to be potentially similarly susceptible to environmental stress. This raises the questions: "What makes a site vulnerable to initiating mortality versus experiencing spread?" or "Why does mortality spread in one region but not another?" if clustering is the only effect you observe. The currently chosen co-location variables might also already be somewhat representing the local conditions and overshadowing the actual drivers, e.g., slope, aspect, soil, etc.. It might make sense to attempt to drop mortality spread covariates in an additional analysis.

1.2) Response: Thank you for your comment. You have raised an important question as to why mortality might spread in some locations but not in others. In fact, we have considered the impact of different groups of drivers (“Mortality occurrence and spread”, “Forest structure” and “Topography”) in the statistical results and have performed analyses in which groups of drivers were dropped during the calculation of partial R² values, also including those concerning spreading mortality (See figure 5). The explanatory value of the other drivers did not increase when dropping the group “Mortality occurrence and spread”, hence we do not consider this effect as dominant and overshadowing the other drivers. We will provide more information on this in the Methods and Discussion sections.

Please discuss the impact of different model performances and thus map accuracies of broadleaf (F1=46%-52%) vs conifer (F1=72%-75%) (Schwarz et al., 2024) on the results, especially given that 71% of the studied forest is broadleaf. The implications also depend on precision vs. recall, which I was not able to extract from the previous study.

1.3) Response: Thank you for raising this point. The F1-scores reported in Schwarz et al. (2024) reflect the spatial agreement between the CNN model results and user-delineated testing data, and thus represent a conservative estimate of accuracy. Both classes performed well in terms of overall visual and structural agreement with the reference data.

Nevertheless, we acknowledge that lower F1-scores in broadleaf stands, which cover most of the study area, may have led to under-detection of small or diffuse broadleaf mortality patches (lower recall) and/or higher false positives in heterogeneous canopies (lower precision). Either bias could reduce the potential importance of the tested drivers in broadleaf forests relative to conifers. We will discuss this limitation in the manuscript to clarify that a weaker statistical signal in broadleaf forests may potentially reflect detection uncertainty rather than ecological differences alone.

Minor:

- Line 87: I'd suggest sticking to the wording "landscape-scale". Countries are only a political structure and heavily vary in size.

1.4) Response: We will change the sentence to: „[...], now make it possible to address these data gaps by facilitating the acquisition of high-resolution mortality data at the canopy level across landscapes (Schwarz et al., 2024).“

- Line 125: Please reference Figure 2a earlier. Without the figure, the described binarization process is very difficult to understand.

1.5) Response: We will move the first mention of Figure 2a to line 125.

- Line 125: How exactly is the centroid of the dead canopy polygon computed? In the naive way, i.e., (maxX-minX)/2, (maxY-minY)/2, this point can very easily land outside the actual deadwood polygon, e.g., with L-shaped polygons. This is especially important since this appears to be a semantic segmentation and not an instance segmentation mask.

1.6) Response: We used the standard QGIS Centroid tool, which uses the barycenter of the polygon and can occasionally place centroid points outside the polygon for irregular shapes. We will explain this in more detail in the revised manuscript. We identified these cases (using inverse selection by location) and manually adjusted the points toward the approximate center of the polygon. We will add the following clarification to the manuscript: “Centroids that were initially generated outside their respective polygons due to irregular shapes (e.g., L-shaped units) were identified and manually repositioned toward a central point inside the polygon.”

- Line 143: Are these distances calculated between the above-defined centroids or between polygon edges?

1.7) Response: The distances were calculated between the centroid of the dead canopy in 2020 and the polygon edge of the corresponding dead canopy in 2017 or 2019 as we did not calculate centroids for previous years. We will add the explanation to the manuscript in line 142: „[...] by calculating the minimum distance to the closest dead canopy polygon edge of 2019 (dist₁₉) to the data points (centroids) in 2020 […]“.

- Line 146: Writing "other dead canopy" and "spread from tree to tree" is misleading as the polygons are not at the tree-level, but represent canopy groups (Figure 2). I understand the use of this metric, but the implication is misleading.

1.8) Response: We will change the wording in both paragraphs accordingly (l. 140 - -147)

- Line 160: How accurate is Lis-L at detecting forest roads? This is a non-trivial task and the displayed polygon in Figure 2 appears fairly coarse, suggesting that (small) forest roads might not be captured. This might also explain the high correlation between both variables.

1.9) Response: Korzeniowska (2020) do not report a quantitative accuracy assessment for the road classification in the Lis-L data set. However, based on visual inspection and comparison with official road network maps, major and rural roads including those traversing forested areas are well represented. In contrast, small forest tracks and agricultural access roads are not captured by the data set. This limitation may indeed contribute to the observed correlation between road proximity and forest edge metrics, as fine-scale forest fragmentation caused by minor roads is not explicitly resolved. We will clarify the level of road representation and its implications for our analysis in the revised manuscript.

Figure 2 does not display forest edges created by roads. Instead, it shows the boundaries between different forest stand types (e.g., coniferous vs. broadleaf). For clarity, we chose to illustrate only a single coniferous stand polygon in this example and omitted adjacent forest polygons.

- Figure 3: Why is the colorbar scaled the way it is?

1.10) Response: The colourbar was square-root-transformed to best represent the distribution of values. A normally-distributed scale would have led to too small values not being visible well in the plot, leading to a relatively uniform color in the broadleaf plot. We will modify the Figure caption accordingly.

- For readability, I'd replace "dist_e" and "dist_r" with "Distance_Edge" and "Distance_Road", and similarly for dist_19, ...

1.11) Response: We will replace the abbreviations accordingly.

- I think it is called "deadtrees.earth" and not "deadtrees.org".

1.12) Response: Great catch! We will change the name. Thank you for pointing it out.

Suggested additional citations:

- Line 30: Most recent study supporting this: "Gazol, A., Pizarro, M., Hammond, W. M., Allen, C. D., & Camarero, J. J. (2025). Droughts preceding tree mortality events have increased in duration and intensity, especially in dry biomes. _Nature Communications_, _16_(1), 5779." - Line 87: - "Cheng, Y., Oehmcke, S., Brandt, M., Rosenthal, L., Das, A., Vrieling, A., ... & Horion, S. (2024). Scattered tree death contributes to substantial forest loss in California. _Nature Communications_, _15_(1), 641."

- "Junttila, S., Blomqvist, M., Laukkanen, V., Heinaro, E., Polvivaara, A., O'Sullivan, H., ... & Peltola, H. (2024). Significant increase in forest canopy mortality in boreal forests in Southeast Finland. _Forest Ecology and Management_, _565_, 122020."

1.13) Response: We appreciate the suggestions for literature. We will consider adding those to the revised manuscript.

The author's work addresses an important topic of drivers of drought-induced tree mortality at the country scale (Luxembourg). The study is well written, and the results are supported by appropriate figures. However, there are several major points regarding the data selection and their corresponding methodologies that need clarification before publication.

Major points to clarify:

The advantage of using high-resolution (sub 1 m) data compared to freely available data, such as Sentinel-2 and Landsat is not clearly stated. Especially, as the authors do a large-scale country-wide study, Sentinel-2 could already have sufficiently high resolution to observe patterns, and the results would be less noisy. Additionally, more frequent acquisitions would be possible.

2.1) Response: The orthophotos used in our study are freely available from geoportail.lu. Our choice of very high-resolution data is based on two considerations:

First, as described in detail in Schwarz et al. (2024), the mortality data set was generated using a Convolutional Neural Network trained on 20 cm/px orthophotos to automatically map canopy mortality across Luxembourg. This resolution was necessary to distinguish coniferous and broadleaf tree canopies based on fine structural features, which cannot be resolved by Sentinel-2 or Landsat. These fine structural features visible in the orthoimages allow a much more precise detection of dead canopies, reducing background scatter e. g. from browning of under-story vegetation.

Second, because the mortality data is available at such a fine scale, we took the opportunity to analyze canopy mortality using environmental drivers at comparable spatial detail (50 cm/px). Our objective was to investigate fine-scale processes, such as stand structure, bark-beetle infestation and spread, and local topographic context, that are likely not detectable with 10–30 m satellite imagery. Coarser satellite pixels mix live and dead canopy and therefore obscure early, localized, or small-patch mortality. Using very high-resolution data allows us to link individual canopy or small patch mortality directly to high-resolution drivers and to evaluate whether such fine-scale mechanisms aggregate into consistent landscape-scale patterns.

We will restate this reasoning more explicitly in the revised manuscript, adding to the Introduction and Methods section.

I do not understand why the bark beetle distribution and patches of already dead trees are used as an environmental driver for drought-induced mortality. I agree that bark beetle infestations can happen more often after trees are already damaged by a drought, but the beetle itself is not a driver for drought-induced mortality.

2.2) Response: We thank the reviewer for this important comment. We fully agree that bark beetle infestations are not the direct cause of drought-induced mortality; rather, they are often triggered or amplified by drought-stressed trees. However, in ecological studies it is common to consider both direct and indirect factors influencing mortality when defining “drivers.” In this context, the presence of bark beetles or nearby dead trees acts as a mediating biotic mechanism that can substantially influence the probability and spatial pattern of additional canopy mortality following drought (see also Response 1.1 to Reviewer 1).

To make this explicit, we will revise the manuscript to clarify the causal chain: drought acts as the primary driver, weakening trees, while bark beetle activity constitutes a secondary, biotic driver that mediates mortality under drought stress. We will also clearly define “drivers” in the manuscript as follows:

"In this study, drivers are defined as biotic or abiotic factors that directly or indirectly influence the probability of canopy mortality following drought. Under this definition, bark beetle infestation represents a biotic driver that mediates drought-induced mortality, which we characterize using four proxy variables defined as follows: “Distance to previously dead canopies in 2017 and 2019”, “Distance to closest (other) dead canopy in 2020”, and “Change in dead canopy area”.

I assume that meteorological factors such as temperature and precipitation also play a major role in the drought-induced mortality, but no climatic variables are discussed in this study. Is there any temperature and precipitation data, or other climatic data available to examine local patterns that could also correlate with the drought-induced mortality?

2.3) Response: We thank the reviewer for raising this point. Indeed, climatic factors such as temperature and precipitation are key determinants of drought-induced mortality. In this study, however, our focus was on fine-scale, spatial drivers of canopy mortality, and the data we analysed are high-resolution in space but limited in temporal frequency. Climatic drivers, in contrast, operate over broader spatial and higher temporal scales, which makes it challenging to link them directly to individual-tree or small-patch mortality in the current data set. For this reason, we treat drought as a background condition and assume that the entire study period experienced sufficient water stress to trigger mortality events.

We are planning a follow-up study that will explicitly integrate climatic variables, such as temperature, precipitation, and drought indices, to examine their contribution to local patterns of tree mortality. We will clarify this rationale in the discussion to make the role of the 2018-2020 drought as overall trigger of canopy mortality clear in the current analysis.

Please give more information about the used data, such as the data source, sensor type, resolution, date/year of acquisition. This could be done, e.g., as a table at the beginning of the Methods section.

2.4) Response: Thank you for this suggestion. We will add a table at the beginning of the Methods section (or in the Supplementary Material if preferred) summarizing all data sets used in this study, including data source, sensor type, spatial resolution, and year(s) of acquisition.

Minor:

35 Can you give examples here of what you understand by abiotic/biotic factors?

2.5) Response: In the context of our study, abiotic factors primarily refer to environmental and topographic variables (e.g., slope, aspect, soil texture), whereas biotic factors include forest structure and bark beetle spread. We will replace the word “factors” with “drivers” to be more consistent and add brief examples in the introduction to clarify these categories without going into study-specific detail: “In addition many abiotic (e.g., topography, soil texture) and biotic drivers (e.g., forest structure, insect outbreaks) can moderate individual tree responses during drought, challenging accurate projections of future forest dynamics.”

47 How do topographic variables differ from abiotic factors, or are they the same?

2.6) Response: As stated in the above response topographic variables would be considered abiotic factors, which we will replace with “drivers”. We think that adding the before mentioned sentence will clarify this sufficiently.

75 What do you understand as "high-resolution" canopy data? You mention Landsat data, but could you also use 10 m Sentinel-2 data for your analysis?

2.7) Response: In our study we operate on the level of 50cm/px as stated in line 93. The original resolution the dead canopy data was acquired at was 20cm/px. This was necessary for the CNN model to differentiate between conifer and broadleaf canopy structures. We will state the resolution in the revised manuscript more prominently now.

In principle, 10m Sentinel-2 data could be used to analyse canopy mortality at broader spatial scales. However, such an approach would require additional calibration and validation, and would be associated with higher uncertainty due to mixed pixels and the limited ability to detect small or early-stage mortality patches (see Schiefer et al. 2025). We will discuss this topic in the revised manuscript.

87 If you are interested in understanding the mortality drivers at the landscape scale, why is it necessary to look at individual trees?

2.8) Response: Thank you for the question. Our approach explicitly links fine-scale processes to landscape-scale patterns. By analyzing canopy mortality drivers at this fine-scale, but broad spatial extent, we sought to identify localized mechanisms, such as stand structure, or fine scale topographic differences. Exploring these signals across the landscape then allows us to test whether the drivers identified at fine scales remain consistent when applied to a heterogeneous, country-wide setting. To our knowledge this is the first study concerning tree mortality following such an approach. In our view, this ability to upscale process-based insights is a key strength of the study. We will make this reasoning more clear in the revised manuscript.

95 From your hypotheses, the main aim of the paper is not clear to me. Can you link it more to the title of the paper?

2.9) Response: Thank you for pointing this out. The aim of the study was to identify and quantify the contribution of topography, forest structure and bark beetle spread in driving drought-induced canopy mortality across Luxembourg separately for conifer and broadleaf forests. We will revise the hypotheses as follows so that each reflects this aim more clearly.

“We hypothesized that drought-induced canopy mortality spread/patterns across Luxembourg is related to three main drivers:

(1) Topography, with higher mortality expected on steeper and south-facing slopes;

(2) Forest structure, with taller stands and stands closer to forest edges exhibiting higher vulnerability; and

(3) Biotic interactions, where conifer mortality, but not broadleaf mortality, shows patterns consistent with bark beetle infestation and spread.”

108: How much of the whole country area is forest (in %). E.g., add it after the 900 km².

2.10) Response: Forests cover ~35% of Luxembourg's land area. We will add this to the revised manuscript as you suggested.

139: 2.3.1 Why is the mortality occurrence and spread under "environmental data" and not under 2.2 where you describe the mortality data? I also do not understand why the distribution of the bark beetle patches is an environmental factor influencing the drought-induced mortality of other trees.

2.11) Response: Thank you for the comment. In our analysis, mortality occurrence and spread act as a driver because they influence mortality patterns during drought. For this reason, these variables were originally listed under “environmental data.” To avoid confusion, we will rename this section to “Mortality Drivers” in the revised manuscript.

140: Is this the distance between the centroids (i.e. patchA_2017 to patchA_2019?)? If your assumption of "spreading" is true, the distance between the centroids between the years would not necessarily change that much (for the same patches).

2.12) Response: You are correct that for patches already present in 2019, the distance between centroids may not change substantially from year to year. To account for the expansion or spread of mortality beyond existing patches, we also included the Δarea variable, which measures the change in size of dead-canopy patches between years. This allows us to represent both the influence of existing mortality and the growth of new mortality areas in our analysis. We will make this clearer in the revised manuscript.

142: data point = centroid position?

2.13) Response: Data points refer to the centroids of dead-canopy polygons as well as the randomly distributed points representing alive trees, as illustrated in Figure 2. Please see also Response “1.6” to reviewer 1 on data points outside of centroids.

174: Why did you use northness/eastness instead of only one parameter “aspect”? Would this not be easier to interpret?

2.14) Response: Thank you for the comment. We used northness to assess the effect of slope orientation along the north-south axis, as south-facing slopes are often more exposed to drought stress. Eastness was included because stakeholders reported higher mortality on east-facing slopes, suggesting a potential influence along the east-west axis as well.

While aspect is intuitive, it is a circular variable (0–360°), which can complicate its inclusion in regression-based models and make statistical interpretation less straightforward. Aspect was transformed into northness and eastness to avoid circularity and enable its inclusion as linear predictors in the models. We will add a short sentence explaining this to the relevant Methods section.

239: I do not agree that no large-scale patterns are visible in Figure 3. In the coniferous forest, I see 2 "bands" of high mortality, and in the broadleaf forest, a North-South gradient.

2.15) You are, of course, correct. We will revise the wording in the manuscript to acknowledge the presence of regional gradients in canopy mortality and clarify our motivation for the subsequent fine-scale analysis. The revised text will read:

“Although Figure 3 reveals regional gradients in canopy mortality, such as bands of higher mortality in coniferous forests and a north–south gradient in broadleaf forests, these patterns alone are not sufficiently to fully explain mortality dynamics at the local scale. We therefore conducted a more detailed analysis of site-specific characteristics contributing to canopy mortality at a much finer resolution (50cm).”

251: Where can I find these %-numbers in Figure 4?

2.16) Response: Thank you for the question. The percentage values correspond to the bars shown in the bar plot and are indicated on the x-axis. We will revise the figure caption to make this clearer and guide the reader more explicitly to where these values can be found.

266: …this is not surprising given...

2.17) Response: Thank you, we will fix this spelling mistake.

417 What about the large-scale patterns? Are they related to meteorological drivers?

2.18) Response: Thank you for this important question. Large-scale spatial patterns in canopy mortality are likely influenced by regional meteorological conditions, particularly the severity and timing of drought events. In this study, however, we focus on explaining within-country spatial variation in mortality using fine-scale drivers derived from very high-resolution data. Given Luxembourg’s relatively small geographic extent and the limited spatial variability of meteorological variables at this scale, climate factors primarily act in the background rather than as differentiating drivers of local mortality patterns.

Nevertheless, we acknowledge that the observed large-scale patterns may partly reflect underlying gradients in drought intensity or other climatic conditions. We will explicitly discuss this limitation in the revised manuscript and note that a follow-up analysis integrating satellite-based mortality time series with meteorological data is planned to more directly assess climate-mortality relationships across scales.

423 The authors do not discuss the influence of climate variables on the tree mortality.

2.19) Response: We thank the reviewer for this important comment. We fully agree that climatic variables such as temperature, precipitation, and drought indices play a fundamental role in drought-induced tree mortality. In this study, however, our focus is on fine-scale spatial drivers of canopy mortality derived from very high-resolution data, while the temporal resolution of the mortality observations is limited to annual data. Most climatic data sets, in contrast, have high temporal but relatively coarse spatial resolution, making it difficult to robustly link them to individual-tree or small-patch spatial heterogeneity of canopy mortality patterns analysed here. For this reason, we are treating drought as a background condition that is assumed to affect the entire study area during the analysed period, and focus on how local topography, forest structure, and bark beetle infestation modulate mortality under these conditions. We will clarify this rationale more clearly in the introduction and discussion.

We are currently planning a follow-up study in which the mortality data will be upscaled to a satellite-based time series, allowing for a more explicit integration of climatic variables and a detailed analysis of climate-driven mortality relationships across space and time.

Figures:

Figure 1: I think it would be easier to interpret if the pixel size is in km and the color in % forest area. Is it possible to color the non-forest areas in a different color than white? I find it hard to distinguish it from the color of the low forest area.

2.20) Response: Thank you for the suggestion. We will consider giving the pixel size in km and will use gray pixels for the non-forested areas.

In L123, you write that you only use the “pure” forest sites >75% but some pixels are colored in both maps. Is this map still representing the mixed forest sites as well?

2.21) Response: Thank you for the clarification request. The maps display forest cover aggregated to a 100 ha grid. Although our analysis uses only “pure” forest stands (>75%) from the LIS-L dataset, aggregating these stands to larger grid cells means that individual grid tiles can contain both coniferous and broadleaf forest types. As a result, the maps may show mixed forest composition even though only pure stands were included in the underlying analysis.

Figure 2: Please use another color combination. The red polygons over the forest image are not colorblind-safe.

2.22) Response: Thank you for this suggestion. We agree that the current color scheme is not optimal for accessibility. We will revise Figure 2 using a colorblind-safe color palette.

Figure 6: What does a negative deltaArea indicate? Would this indicate forest recovery, or can it also be that areas are no longer forest?

2.23) Response: A negative Δarea indicates a reduction in the mapped dead-canopy area between consecutive years. This can reflect vegetation recovery or canopy closure following disturbance, but it may also result from management activities such as clear-cutting. We will clarify these possible interpretations in the discussion section of the revised manuscript.

Figure 7: Add the diste and distr in the caption (L295).

2.24) Response: We will add dist_e and dist_r in the caption to improve readability.

Figure 8: It should be “elevation”.

2.25) Response: We will correct this spelling mistake.