This study covers all forests of Luxembourg with predominantly temperate coniferous and broadleaf stands, the extent of which are defined by the LIS-L land cover data set (Korzeniowska, 2020). Luxembourg is located in Central Europe (between 49 and 51° latitude and 5 and 7° longitude) and has a forest area of approximately 900 km² representing about ∼35 % of the country's total area. Forests comprise approximately 71 % broadleaf, 23 % coniferous and 5 % mixed forests. Despite being one of the smallest countries in Europe, it is characterised by a diverse landscape, ranging from relatively flat, urban environments in the south (∼250 m a.s.l), which are dominated by broadleaf forests to a more hilly, forested, conifer-dominated landscape in the north, reaching elevations of 560 m a.s.l. (Fig. 1).

We used a high-resolution dataset of conifer and broadleaf canopy mortality covering the entirety of Luxembourg, which was first presented in Schwarz et al. (2024). The dataset covers the drought period between 2018 and 2020 as well as the pre-drought year 2017 and has a ground resolution of 20 cm. It is based on annually collected RGB orthophotos freely available on the open data platform https://data.public.lu (last access: 14 April 2026) of Luxembourg (see Table S1 in the Supplement). Canopy mortality of coniferous and broadleaf trees was mapped using a Convolutional Neural Network, a deep learning algorithm for image segmentation and classification. The very high spatial resolution of 20 cm enabled differentiation between coniferous and broadleaf tree canopies based on fine structural features that cannot be resolved by coarser-resolution sensors such as Sentinel-2 or Landsat. The use of high resolution imagery allowed precise detection of dead canopies, while reducing background noise (e.g. browning of understory vegetation). Because a semantic segmentation model was used instead of an instance segmentation model to derive the mortality data, the resulting canopy mortality polygons include both individual dead trees and groups of dead trees.

To identify the drivers of canopy mortality per forest type for the year 2020 (i.e., conifer, broadleaf) we used a Generalized Additive Model (GAM; for explanation see below) (Hastie and Tibshirani, 1986). The analysis focused on homogeneous forest stands with > 75 % coniferous or > 75 % broadleaf cover, as defined by the Lis-L land cover data set (Korzeniowska, 2020). Forest types of both dead and alive canopies were determined by the land cover data. For 2020, the dataset contains 211 461 polygons of dead coniferous canopy and 84 698 polygons of dead broadleaf canopy. As shown in Fig. 2a, a binary data set of dead and alive canopy data points was generated as follows: First, the centroid of each dead canopy polygon was calculated using the polygons barycenter. Centroids that initially fell outside their respective polygons due to irregular shapes (e.g., L-shaped units) were identified and manually repositioned to a central location within the polygon. Next, the dead canopy polygons were subtracted from the land cover polygons, resulting in areas containing only living forest. An equal number of random points were then distributed across living forests for each forest type, and both living and dead canopy points were combined into a single dataset (Fig. 2a).

In this study, environmental drivers are defined as biotic or abiotic factors that influence whether and where canopy mortality occurs under severe, multi-year drought conditions. Drought represents the overarching environmental setting and primary cause of mortality, while additional factors such as soil properties, topography, forest structure, and biotic interactions (e.g., bark beetle activity) modulate tree vulnerability and shape the spatial patterns of mortality. To assess the influence of these drivers on observed canopy mortality, we considered 14 biotic and abiotic variables available at high spatial resolution. A more detailed overview of the geospatial datasets used in this study is provided in Table S1. All data used for the GAM calculations were resampled to a resolution of 50 cm, which is highly detailed for such a large spatial data set. The availability of canopy mortality data at this fine spatial scale allowed us to explicitly investigate fine-scale processes, such as stand structure, bark beetle occurrence and spread, and local topographic context, that are unlikely to be detectable with coarser (10–30 m) satellite imagery. The use of very high-resolution data enables a direct link between individual canopy or small-patch mortality and corresponding high-resolution drivers, and allows us to assess whether these fine-scale mechanisms aggregate into consistent landscape-scale patterns.

We used Generalized Additive Models (GAMs) to assess the contribution of the aforementioned environmental covariates to broadleaf and conifer canopy mortality in 2020 as a result of three consecutive drought summers. GAMs are non-parametric linear models used to model non-linear relationships as a sum of linear basis functions (Hastie and Tibshirani, 1986; James et al., 2013). As such they are able to represent more complex relationships between drivers and responses that exist in natural environments. Generally, a GAM consists of a response variable and a number of covariates, expressed as smoothed (non-linear) terms and linear functions.

In this study, we fitted the models using the function bam() provided by the mgcv package (Wood, 2011) in R (V4.2.1, R Core Team, 2022) which is more suitable for very large datasets.

We fitted one model per forest type, i.e. coniferous and broadleaf forest. Due to the binary form of the response variable, referring to dead or alive canopies, we used a binomial distribution and a “logit” link function. A combined smooth of longitude and latitude was added to model spatially structured variation in the data that cannot be attributed to any of the environmental variables. All continuous covariates were added as smooths and soil type was added as a categorical term. We used the REML (Restricted Maximum Likelihood) method to automatically adjust the smoothing parameter λ, which controls the wigglyness of the smooth and therefore the flexibility of the model. We kept the parameter k= 10, which determines the number of basis functions for the smooth and is reasonably small to achieve interpretable relationships expected for environmental variables.

Prior to modelling, variable independence was checked by calculating Pearson's Correlation Coefficient for linear correlations and Spearman's Correlation Coefficient for monotonic relationships, including non-linear trends for each covariate pair (Supplement Fig. S1). The degree of correlation between the covariates was low to moderate for almost all variables. The highest correlation was found between the two forest edge variables (∼0.8). Due to the large size of the dataset, we consider the statistical power to be large enough to keep both variables in the model and distinguish between different effects of forest fragmentation on canopy mortality. Model assumptions were checked with simulated residuals using the package DHARMa (Hartig, 2022) and revealed no major violations. We report the explained deviance of the model.

To determine the contribution of each variable to the deviance explained per GAM model we calculated partial McFadden's Pseudo-R2 , along the lines of the implementation in the R-package rsq (McFadden, 1974; Zhang, 2022). Additionally, we quantified the contribution of groups of variables by calculating partial McFadden's Pseudo-R2 for topography (elevation, slope, TWI, northness, eastness), forest structure (forest height, distance_edge, distance_road), and indicators of spreading mortality (distance_2017, distance_2019, distance_2020 and Δarea). For each group, a reduced model was fitted excluding all variables in that group, and the resulting McFadden's Pseudo-R2 was compared to the full model using Eq. (1). 1PartialR2=1-1-RF21-RR2⋅n-pRn-pF where RF2 and RR2 represent McFadden's Pseudo-R2 for the full and reduced model, respectively, and n the number of total observations. pR and pF denote the number of parameters in the reduced and full models, respectively. The formula adjusts for the difference in degrees of freedom between the two models.

Canopy mortality in Luxembourg increased from a total of 64 ha in 2017 to 590 ha in 2020. In 2020, about 420 ha of dead canopies were coniferous and 170 ha were broadleaf. As broadleaf forests dominate in Luxembourg, the disproportionate conifer loss indicates a notably higher conifer mortality rate. Hotspots of conifer mortality were observed in the north-east but appeared more scattered across the south (Fig. 3a). In some locations canopy mortality reached almost 90 %, although most areas experienced mortality rates between 0 % and 10 %. In broadleaf forests, the mortality rate was generally much lower, particularly in the north of Luxembourg. Isolated areas of higher mortality occurred in the south and south-east but did not exceed 5 % (Fig. 3b).

Regional gradients in canopy mortality were evident, including bands of elevated mortality in coniferous forests and a north-south gradient in broadleaf forests (Fig. 3). However, these broad-scale patterns alone do not 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 (50 cm). The relative contributions of the 14 environmental variables to the fit of the broadleaf- and conifer-specific GAM models are summarised in Fig. 4.

Thematic groups of environmental drivers differed in their contribution to model performance. Combined, the variables explained a substantial proportion of the fine-scale spatial distribution of canopy mortality in 2020. Overall, the GAM models accounted for 44.7 % of observed mortality in coniferous forests and 25.3 % in broadleaf forests. Variables related to the clustering and spread of mortality were most important in both forest types (Fig. 4), contributing partial R2 values of 29.43 % in conifers and 14.29 % in broadleaf forests (Fig. 5). Distance to dead canopies in the previous years had a similar effect in both types, but conifers were more likely to die in 2020 when dead canopies occurred close by. In broadleaf forests, mortality was strongly associated with a marked increase in dead canopy area between 2019 and 2020 (Δarea). Forest height was an important predictor of mortality in both forest types (Fig. 4). In contrast, spatial coordinates (longitude and latitude), included to capture unexplained spatial patterns, contributed little according to partial R2 values, and topographic variables as well as soil type had minimal influence (Figs. 4, 5).

In order to more closely examine the influence of individual environmental variables on canopy mortality, we used partial effect plots that account for other predictors in the GAMs (Fig. 6). The following subsections explain these effects per group of environmental variables in more detail.

In coniferous stands, distance_2017 had no effect on the mortality distribution of conifers, this is not surprising given the small partial R2 (Fig. 6a). By contrast, proximity to dead canopies in 2019 (distance_2019) was a more influential predictor (Fig. 6b), indicating higher 2020 mortality near trees that had already died in 2019. This pattern was consistent with the observed increase in dead canopy area between 2019 and 2020 (Δarea), within a 50 m radius (Fig. 6d). The distance_2020 variable was the most important predictor. This indicates that mortality in conifers showed strong spatial clustering, with dead canopies tending to occur in patches within ∼20 m of each other (Fig. 6c).

In broadleaf stands, self-reinforcing patterns were also evident. Similar to conifers, 2017 mortality showed no influence on 2020 patterns (Fig. 6e), but 2019 mortality was the most important predictor (Fig. 6f). The increase in dead canopy area (Δarea) was also strongly associated with mortality patterns, with local maxima (Fig. 6h). However, unlike conifers, distance_2020 did not explain mortality distribution (Fig. 6g).

We found conifer canopy mortality to increase steeply with forest height > 10 m and remained high in tall stands (Fig. 7a). In contrast, variables related to forest fragmentation, both outer forest edges (Fig. 7b) and inner edges from roads (Fig. 7c) had only minor influence.

In broadleaf stands, the probability of canopy mortality rose steadily with height, levelling off at around 30 m (Fig. 7d). Unlike conifers, broadleaf mortality showed a clearer effect of forest edges, with higher mortality near the forest edge (Fig. 7e), albeit with a low partial R2 (Fig. 4). This effect is absent for forest roads (Fig. 7f).

As shown before, topography played a minor role in explaining canopy mortality patterns for both coniferous and broadleaf stands, with similar responses between forest types (Figs. 4 and 5). Of the tested variables, TWI appeared as the most relevant, indicating higher mortality at low TWI values (< 6) for both conifers (Fig. 8c) and broadleaf forests (Fig. 8h).

In conifers, there was no clear effect of elevation (Fig. 8a) or slope (Fig. 8b). Broadleaf mortality increased slightly at elevations above ∼250 m (Fig. 8f) and on gentle slopes or plains (Fig. 8g), though these patterns are highly uncertain, reflected in the low partial R2 values. Other topographic drivers, such as northness and eastness, showed no discernible trends (Fig. 8e, k).

Our findings indicate that topographic and soil-related environmental drivers explained little of the observed canopy mortality in both coniferous and broadleaf forests. Of the six tested drivers, the Topographic Wetness Index emerged as the most important. In contrast, other environmental drivers such as slope, elevation, northness, eastness, and soil type had little to no effect on the observed mortality patterns. Luxembourg's relatively low elevation range (150–600 m a.s.l.) may have reduced the effectiveness of elevation as driver of spatial mortality patterns. Previous studies that found elevation-related differences in canopy mortality were conducted in regions with significantly higher elevations exceeding 1500 m a.s.l. (Cailleret et al., 2014; Paz-Kagan et al., 2017; Rita et al., 2020; Sturm et al., 2022). In such situations, climatic differences along the elevation gradients may have more pronounced impacts on potential pest activities as well as precipitation patterns, both of which are likely to influence tree mortality. We suggest that in our study, the small elevational gradient together with the strong country-wide drought effects likely overrode any differences in elevation or slope orientation.

Although high-resolution data are generally advantageous for capturing fine-scale variability (Lannuzel et al., 2021; Pellegrino et al., 2024), they may introduce noise to topographic drivers such as slope, northness, and eastness. This small-scale terrain variability, which may be ecologically irrelevant, could obscure broader topographic patterns (Guillaume et al., 2021; Šilhán et al., 2022). This might explain why variables like TWI, which aggregate over broader spatial extents by accounting for upslope contributing area, were less influenced by small-scale terrain features and therefore more effective at capturing relevant hydrological gradients.

Meanwhile, the soil type data, comprising 27 classes, also showed minimal explanatory value. Although soil moisture is a known driver of drought-induced tree mortality (Belmonte et al., 2022; McDowell et al., 2022), the categorical nature and relatively coarse spatial resolution of the data most likely did not capture the specific soil water holding capacities relevant at the examined individual tree level.

Overall, topographic effects may have also been obscured by biotic drivers such as bark beetle infestation in conifers and by the overall severe drought conditions. The lack of influence in broadleaf forests, where bark beetles are not a disturbance agent, supports the notion that unaccounted variables, like fine-scaled soil properties and corresponding or soil water holding capacities played a more important role. At the same time, we explicitly tested whether spatial clustering effects, represented by the group “Mortality clustering and spread”, overshadowed the explanatory power of the other groups of drivers. To address this, we calculated partial R2 values after excluding this group of drivers. The explanatory value of topographic and soil-related drivers did not increase when mortality clustering and spread were omitted, indicating that their limited contribution was not simply masked by spatial autocorrelation or local spread processes. Thus, even when accounting for potential clustering effects, topographic drivers remained weak predictors of canopy mortality patterns. However, given the severity and spatial extent of the drought conditions in recent years, it seems plausible that topography played only a minor role across all forest types. This poses a concern for forests in future climate change conditions, as the severity of these hotter drought summers might severely affect even areas thought to be more resilient due to their topographic situations causing even greater forest loss than anticipated (Knutzen et al., 2025).

Our findings largely refute hypothesis (1): topographic and soil-related drivers provided only limited value for explaining the spatial patterns of drought-induced canopy mortality, likely due to low regional variability, data limitations, and overriding effects of severe drought and biotic disturbances.

We found that canopy mortality was related to forest structure and increased in taller stands in both coniferous and broadleaf forests. Taller stands experienced significantly higher mortality rates, consistent with findings from other studies linking forest height to increased vulnerability during drought events (Anderegg et al., 2016; Arend et al., 2021; Stovall et al., 2019; Trugman et al., 2018). Physiologically, tall trees can be more susceptible to hydraulic failure under drought conditions due to longer water transport pathways and higher evaporative demand (Anderegg et al., 2016; Arend et al., 2021; Bennett et al., 2015; Trugman et al., 2018). While trees might not die immediately, the cumulative effects of consecutive drought summers can weaken them over time (McDowell et al., 2022) as for instance reported for beech trees in Switzerland following the 2018 drought (Frei et al., 2022). Additionally, in conifers, increased mortality among taller trees is very likely linked to bark beetles preference for mature hosts with sufficiently thick bark as they are more suitable for the development of bark beetle larvae (Pirtskhalava-Karpova et al., 2024; Raffa et al., 2016).

By contrast, forest edges had negligible effect on observed canopy mortality, suggesting that the severity of the climatic drought conditions extended deep into the forests. While previous research indicates that edge effects can amplify drought stress (Buras et al., 2018; Sturm et al., 2022), our findings show only a minor influence in broadleaf forests and no significant effect in conifers. These findings suggest that the severity of the drought conditions of 2018–2020 might have led to unfavourable conditions even in the forest interior. Similarly, forest fragmentation due to roads had no discernible impact on mortality patterns, consistent with the weak role of edge effects in this study. It is important to note that the Lis-L dataset (Korzeniowska, 2020) does not represent small forest tracks or agricultural access roads, meaning fine-scale forest fragmentation could not be resolved. This may partially explain correlations between road and forest edge proximity.

In summary, our results partially support hypothesis (2): forest height was a strong driver of spatial canopy mortality patterns, whereas forest edge effects were minimal, likely due to the pervasive severity of drought across forest interiors.

Our results indicate that bark beetles were a key driver of canopy mortality in conifers, as evidenced by the significant roles of mortality clustering and spread. Both distance_2019 and distance_2020 were strong indicators of spatial mortality clustering, consistent with the clustered spread characteristic of bark beetle infestations (Zhao et al., 2025). The Δarea variable, which captured changes in the dead canopy area surrounding our data points, was also important and positive Δarea further supports the interpretation of spreading bark beetle activity as the infestation grows from single trees to large patches of dead trees (Potterf et al., 2019; Stereńczak et al., 2019). Negative Δarea may reflect vegetation recovery or canopy closure. Forest height was the only other strong driver, in line with bark beetle's preference for larger trees that have a sufficiently thick bark for tunnelling (Raffa et al., 2016). Taken together, this clearly reflects bark beetle-driven mortality dynamics in conifers.

It is important to note that the model identifies the proximity to existing mortality as a strong predictor of new mortality. While proximity is not a primary mechanistic cause, it reflects bark beetle-mediated propagation of drought-weakened trees. Bark beetle impact appears strongly modulated by drought, consistent with previous research (Hlásny et al., 2021). Our data show limited impact of spreading mortality prior to 2019, suggesting that high bark beetle activity only started following the onset of severe drought conditions. This aligns with evidence that drought weakens tree defence mechanisms (Rouault et al., 2006), while simultaneously accelerating beetle development, enabling multiple reproductive cycles per year (Fleischer et al., 2016; Lindman et al., 2023; Robbins et al., 2022; Webb et al., 2025). At the physiological level, we suggest that trees exposed to drought suffered from both hydraulic failure due to low water availability and rising vapour pressure deficit (Adams et al., 2017), and carbon starvation as stomata closed to conserve water (Ziegler et al., 2024). Reduced water availability likely also impaired resin flow (Kolb et al., 2019; Malone et al., 2024), weakening tree defence mechanisms. Carbon starvation during drought might also have decreased the allocation of nonstructural carbohydrates (NSC) to defensive metabolites in tree resin (Huang et al., 2020). These interacting stresses created ideal conditions for bark beetle outbreaks and explain the mortality patterns observed in our study.

Another layer of complexity lies in the interaction between bark beetle infestations and topography. Previous studies have shown that bark beetle infestations are often more severe on south- and east-facing slopes, at lower elevations, and on steeper terrain (Akkuzu et al., 2009; Blomqvist et al., 2018; Jakoby et al., 2019). Blomqvist et al. (2018) also reported more severe bark beetle infestations on shallow soils with low water storage capabilities. While such topographic patterns were not dominant in our mortality models, this does not imply these relationships were absent. Rather, they may have been masked by the overriding biotic effect of infestation.

Canopy mortality in broadleaf trees also showed an apparent influence of two variables associated with the spread of bark beetle infestation. These were distance_2019 and Δarea, while distance_2020 had little explanatory power. This suggests that although the loss of large-crowned individuals strongly increased the mortality area, mortality did not propagate in a spatially contagious manner (Fig. 9). The apparent importance of proximity to dead canopies in 2019 (distance_2019) should therefore be interpreted cautiously, as it largely reflects the persistence of already-dead trees rather than true neighbourhood-driven mortality processes. In other words, these variables capture changes in mortality extent rather than contagious spread. Taken together, these patterns suggest that in conifers, canopy mortality is shaped by contagious bark beetle outbreaks, whereas in broadleaves it reflects the loss of structurally dominant individuals under stress.

Our findings strongly support hypothesis (3): conifer canopy mortality was closely linked to the spatial and temporal dynamics of bark beetle spread. At the same time, the contrasting patterns in broadleaf forests emphasise that identical drivers can reflect very different underlying processes depending on forest type.

A primary source of uncertainty of this study arises from differences in model performance between forest types in the mortality dataset. Schwarz et al. (2024) reported lower F1-scores for broadleaf forests (46 %–52 %) compared to coniferous forests (72 %–75 %), indicating lower agreement between the CNN-based mortality detection and user-delineated reference data in broadleaf stands. These F1-scores represent a conservative estimate of accuracy and reflect overall spatial agreement of mapped pixels. They are therefore sensitive to small position mismatches, fuzzy boundaries or differences in patch shape, which can lower the F1-score even though the general location of the mortality patch was captured correctly (Maxwell and Warner, 2020). Lower performance in broadleaf stands may have resulted in under-detection of small or diffuse mortality patches (lower recall) and/or higher false positives in heterogeneous canopies (lower precision). Given that broadleaf forests account for approximately 71 % of the study area, this disparity in detection accuracy may have influenced the relative strength of detected relationships between drivers and mortality patterns. However, we stress that as demonstrated in Schwarz et al. (2024) the overall quality of the classification was very good with a reliable detection of the vast majority of dead trees which we consider to be adequate as data base for the study conducted.

Climatic drivers such as temperature and precipitation are well-established drivers of drought-induced tree mortality. In the present study, we focused on fine-scale spatial drivers of canopy mortality, using very high-resolution mortality and environmental datasets (≤ 50 cm). While these data allow detailed analysis of individual tree and small-patch mortality patterns, they are limited in temporal resolution because observations were available only at annual time steps. In contrast, measurements and estimations of climatic drivers typically operate at broader spatial scales and higher temporal frequencies, making it challenging to directly link local climate variability to fine-scale mortality patterns in the current dataset. Consequently, although the observed large-scale spatial patterns in canopy mortality may partly reflect underlying gradients in drought intensity or other meteorological conditions, these effects are not apparent using the available fine-scale, but temporally limited dataset. While moderate-resolution satellite data (e.g. Sentinel-2 at 10 m or Landsat at 30 m) could be used to analyse canopy mortality at finer temporal scales, such approaches are 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), and would require additional calibration and validation. As a result, these data are less suitable for capturing site-specific mortality responses at the scale addressed in this study. Accordingly, drought was treated as an overarching environmental context and the study period (2018–2020) was assumed to represent sufficiently widespread and severe water stress to trigger mortality processes across Luxembourg.

Despite these methodological and scale-related limitations, the consistent spatial patterns observed across forest types and driver groups indicate that the main conclusions regarding the relative importance of forest structure and bark beetle dynamics are reliable.