Landscape context modulates the effect of local canopy cover on forest multidiversity across elevations

Declining forest biodiversity has increased focus on forest conservation and restoration. Many efforts to conserve and restore tree cover focus on the local scale, but their outcomes are frequently modulated by landscape context. While the diversity and composition of communities are strongly driven by local-scale canopy cover, landscape-scale habitat characteristics affect dispersal pathways and determine the species pool available for colonization of local patches. Moreover, local and landscape-scale habitat attributes vary with elevation, but how their effects on biodiversity change with elevation remains poorly understood. We examined how local canopy cover affects forest biodiversity, and how these effects are modulated by the amount of total and disturbed forest available in the surrounding landscape along an elevational gradient. We used remote sensing and multi-taxa biodiversity data covering plants, aculeata, moths, beetles, and birds (a total of 2,319 species) across 150 plots in naturally developing forests in a forest-rich region in the northern European Alps. We calculated multidiversity across all species and for three habitat affinity guilds (forest, mixed, open-habitat) to test for differences based on varying habitat associations. Local canopy cover negatively affected multidiversity, with the weakest effect observed for forest species. An increasing amount of forest in the surrounding landscape amplified this negative effect, while an increasing amount of disturbed forest reduced it. The negative effect of local canopy cover on multidiversity weakened with elevation and became neutral across all guilds close to the tree line. Our findings highlight that disturbances promote forest biodiversity via two fundamental pathways: reducing local canopy cover and creating a more open and diverse landscape context. Moreover, the effects of canopy cover on forest biodiversity are modulated by environmental conditions that change with elevation. Conservation and restoration efforts should consider landscape context more explicitly when planning specific management measures. Our results suggest that canopy openings benefit biodiversity especially in landscapes with high forest cover and in low elevation areas, while conserving and re-establishing tree cover is important in landscapes with low forest cover and close to the upper tree line.cover and high elevation areas.

Methods
Biodiversity data

We collected data from five taxonomic groups across all 150 plots in 2021, except light trapping, which was conducted in 2022. We provide a comprehensive description of the sampling and identification methods in the supplementary material of the publication, but in brief: The vegetation was sampled on 200 m2 quadratic plots, separately for the herb (<1 m height) and shrub layer (>1-5 m height). We used a combination of two pitfall traps, one malaise, and one light trap to collect beetles, moths, and aculeata (bees, wasps, ants). While beetles from pitfall and light traps and moths from light traps were identified to species level by experts, we used DNA metabarcoding to identify beetles, moths, and aculeata from malaise trap samples. We recorded birds with bioacoustics recorders, identified species using BirdNET (Kahl et al. 2021, version 2.4), and applied species-specific thresholds in order to improve the accuracy of identification (Seibold et al. 2024). We then classified all species into three broad habitat-affinity guilds, namely forest, mixed, and open-habitat affinity, based on Schmidt et al. (2011) and Dorow et al. (2020), and calculated multidiversity based on Allan et al. (2014) for an overall dataset and for each affinity guild.

Local and landscape-scale forest characteristics

We assessed local canopy cover as the proportion of returns 5 m above ground within a radius of 12.6 m using high-resolution LiDAR data with an average point density of ~50 points m2, acquired under leaf-on conditions in September 2021 (Mandl et al. 2023).

To quantify the proportion of forest relative to sample area, we used the 2018 High Resolution Layer Tree Cover Density product from the EU Copernicus programme (https://land.copernicus.eu/en/products/high-resolution-layer-tree-cover-density). To quantify the proportion of disturbed forest relative to the area classified as forest by the Copernicus product, we used the 2020 pan-European forest disturbance maps derived from Landsat data (Senf & Seidl 2021). We resampled the disturbance map products (30 m resolution) to match the 10 m resolution of the Copernicus product and selected each pixel that was disturbed since 2006. We quantified both metrics across multiple radii (50 m, 100 m, 200 m, 500 m, 1000 m, 1500 m, 2000 m, 2500 m, and 3000 m) to capture varying responses across the studied taxa inhabiting the complex study landscape.

Statistical analysis

We analysed how local and landscape-scale forest characteristics affect multidiversity along elevation. We fitted Bayesian multilevel models using the brms package (Bürkner 2017) with a beta probability distribution for each habitat affinity guild and radius quantifying landscape-scale forest proportions. We used multidiversity as a response and canopy cover, elevation, proportion of forest relative to sample area, and proportion of disturbed forest relative to forest area as predictor variables. To analyse the interactive effect of local and landscape-scale forest characteristics and to capture the effect of their changes with elevation, we included an interaction term between canopy cover and proportion of forest, between canopy cover and proportion of disturbed forest, as well as between elevation and each of the other predictors. We z-transformed all predictor variables to increase sampling efficiency, and we added a variable that groups plots of the same sample area as a random intercept to account for spatial autocorrelation (Dormann et al., 2007). We then used Pareto smoothed importance sampling leave-one-out cross-validation (PSIS-LOO) in combination with Bayesian stacking to select the model with the best estimated predictive performance across landscape radii (Vehtari et al. 2024b; Yao et al. 2018).