Use and categorization of Light Detection and Ranging vegetation metrics in avian diversity and species distribution research

Vegetation structure is a key determinant of animal diversity and species distributions. The introduction of Light Detection and Ranging (LiDAR) has enabled the collection of massive amounts of point cloud data for quantifying habitat structure at fine resolution. Here, we review the current use of LiDAR‐derived vegetation metrics in diversity and distribution research of birds, a key group for understanding animal–habitat relationships.

LiDAR metrics, usually from airborne laser scanning data with low point densities (<10 points/m 2 ) and small footprints (<1 m diameter). Our metric harmonization suggests that 40% of the currently used metric names are redundant. A categorization scheme allowed to group all metric names into 18 out of 24 theoretically possible classes, defined by vegetation part (total vegetation, single trees, canopy, understorey, and other single layers as well as multi-layer) and structural type (cover, height, horizontal variability and vertical variability). Metrics related to canopy cover, canopy height and canopy vertical variability are currently most often used, but not always effective.

| INTRODUC TI ON
A key goal of biodiversity and conservation research is to predict and understand species distributions (Franklin & Miller, 2009) and diversity patterns (Ricklefs & Schluter, 1993). In this context, climate and habitat heterogeneity are among the most widely studied determinants of terrestrial species diversity and distribution (Hawkins et al., 2003;Stein, Gerstner, & Kreft, 2014). While climate is often thought to be the key determinant at large spatial extents (Pearson & Dawson, 2003), vegetation structure often becomes particularly relevant at small spatial extents and/or fine grain sizes (Zellweger, Braunisch, Baltensweiler, & Bollmann, 2013). Vertical and horizontal structure of vegetation has long been recognized as an important factor driving animal diversity and species distributions, especially of birds (Cody, 1985;Dunlavy, 1935;MacArthur & MacArthur, 1961). However, measuring the fine-scale habitat structure and 3-D characteristics of vegetation across large regions is difficult and labour-intensive, and hence limits most animal-habitat studies to the local or landscape scale (i.e., <200 km). Nevertheless, recent developments in remote sensing now provide unprecedented opportunities for measuring vegetation structure across broad spatial extents (Kissling, Seijmonsbergen, Foppen, & Bouten, 2017;Lausch et al., 2016;Skidmore et al., 2015).
Light Detection and Ranging (LiDAR)-an active remote sensing technique-is a technology that provides fine-grained information about the 3-D physical structure of ecosystems (Davies & Asner, 2014;Lefsky, Cohen, Parker, & Harding, 2002;Simonson, Allen, & Coomes, 2014;Vierling, Vierling, Gould, Martinuzzi, & Clawges, 2008). LiDAR sensors, for example installed on airplanes or satellites (Figure 1a), measure the return time of an emitted pulse of laser light and convert this time to a distance. Especially, airborne laser scanning (ALS) data are often used because they cover large areas and country-wide datasets are becoming increasingly available. For ALS, a plane or helicopter serves as the platform from which the LiDAR sensor is operated. In most cases, near-infrared light is used which partially penetrates the vegetation and therefore allows to measure both the canopy and subcanopy vegetation structure. LiDAR data are recorded as full waveform (FWF) or discrete echoes (i.e., returns of one or more energy reflections from the FWF; Figure 1a). The resulting dataset is called a "point cloud" (Figure 1a) and consists of the x,y,z coordinates of the points together with intensity values. LiDAR is widely used in forestry to measure 3-D forest structure (Maltamo, Naesset, & Vauhkonen, 2014), but applications in ecology, biodiversity and conservation planning are also increasing (Davies & Asner, 2014;Simonson et al., 2014). Nevertheless, using and processing the massive amounts of LiDAR data across large spatial extents remains challenging (Bergen et al., 2009;Kissling et al., 2017) and most current LiDAR applications are restricted to local study sites as well as forest habitats (Wulder et al., 2012).
Recent reviews on the use of LiDAR data for measuring 3-D habitats of animals show that vegetation structure can be quantified in many ways (e.g., Davies & Asner, 2014;Hill, Hinsley, & Broughton, 2014). A potential confusion for ecologists could be that a multitude of LiDAR metrics exists to quantify structural vegetation attributes.
For instance, canopy height can be calculated as the maximum return, the mean height of the first returns or the height of the 95th percentile of returns in a raster cell (Ackers, Davis, Olsen, & Dugger, 2015;Coops et al., 2016;Smart, Swenson, Christensen, & Sexton, 2012). Additionally, there are also many technical, methodological and data quality aspects that may influence how LiDAR-derived vegetation structure is quantified. For instance, in many cases the point cloud is rasterized (i.e., simplified into pixels) or voxelized (i.e., into three-dimensional voxels; Figure 1b). LiDAR-derived vegetation metrics (e.g., mean canopy height) are then calculated to summarize the vegetation structural information from the point cloud within each pixel or voxel. An advantage of such area-based approaches (i.e., rasterization) and voxel-based approaches (i.e., voxelization) is that they are computationally efficient, but they also lead to information loss regarding the 3-D information of the point cloud (Ciuti et al., 2018). An alternative is to use object-based approaches ( Figure 1b) where either the point cloud or a derived rasterized map is segmented into objects such as trees, forest stands, reed beds or hedges based on similarities and differences in neighbourhood information around points or grid cells (Höfle, Hollaus, & Hagenauer, 2012;Koch, Kattenborn, Straub, & Vauhkonen, 2014). This is computationally demanding, but also provides new ways of quantifying 3-D habitat structure compared to area-based or voxel-based approaches. Other specific technical and methodological constraints include point cloud density, footprint size and conditions of the flight campaign such as flight altitude and season (leaf-on vs. leafoff). The season of LiDAR data acquisition might also be influential when mapping 3-D animal habitats (e.g., understorey in deciduous or broad-leaf forests; Hill & Broughton, 2009). Currently, there is no overview available on how LiDAR data characteristics are represented in ecological studies of animal-habitat relationships.
Diversity and distribution studies usually aggregate raw data from opportunistic observations or standardized and structured surveys to represent the presence, presence-absence, population size or richness of species across space and time (Gaston, 1996;Kissling et al., 2018). Such aggregated field observations (e.g., presence-only or atlas data) can then be mapped and used in species distribution models (SDMs; Guisan, Thuiller, & Zimmermann, 2017) (Figure 1c). In most large-scale studies, environmental information related to climate, productivity, topography and land cover is used to quantify predictor variables (Guisan et al., 2017;Hawkins et al., 2003), but information on fine-scale habitat structure (e.g., density, cover and openness of vegetation) and 3-D vegetation distribution (e.g., vertical and horizontal distribution of biomass) is often lacking (Kissling et al., 2017). An exciting opportunity to improve statistical models of species distributions and F I G U R E 1 Sampling and processing of Light Detection and Ranging (LiDAR) data and its use in species distribution modelling. (a) LiDAR data can be obtained from spaceborne or airborne sensors with a certain footprint size, recorded as full waveform or discrete echo and then processed into point clouds with low or high point densities. (b) LiDAR point clouds can be further processed using area-based approaches (i.e., rasterization into grid cells), voxel-based approaches (i.e., voxelization into 3-D voxels) or object-based approaches (e.g., segmentation into objects such as trees). (c) Applications of species distribution models use mapped species observations together with environmental data and LiDAR metrics in a statistical model to predict the spatial distribution of species

(c)
Probability of occurrence

Low point density
High point density species richness is therefore the increasing availability of airborne and spaceborne LiDAR data which allows one to derive a diversity of metrics that quantify many aspects of the 3-D structure of vegetation. This is particularly relevant because individual species show strong preferences for particular habitat structures (Cody, 1985) and because species richness tends to increase with larger available niche space in structurally heterogeneous and complex habitats (Stein et al., 2014).
Here, we provide a review on the use of LiDAR data in avian species distribution and diversity research. We focus on studies of species distributions (based on occupancy, presence-only, presence-absence and abundance data) and species richness (based on range maps, inventories or atlas data) because these are widely studied aspects in spatial ecology and biogeography (Guisan et al., 2017;Kissling et al., 2018;MacArthur & MacArthur, 1961;Stein et al., 2014). We further focus on birds because they strongly depend on 3-D habitat structure (Cody, 1985;MacArthur & MacArthur, 1961) and because they are among the most widely studied organisms in this particular research field (Davies & Asner, 2014 (Davies & Asner, 2014;Maltamo et al., 2014), predominantly at local (1-10 km) or landscape (10-200 km) scales (Simonson et al., 2014), and with area-based rather than object-based or voxel-based approaches (Blaschke et al., 2014;Kissling et al., 2017;Maltamo et al., 2014). We further hypothesize that metrics related to canopy height and cover are most widely used (Davies & Asner, 2014) and that metrics describing vertical heterogeneity and distribution of vegetation are most effective (MacArthur & MacArthur, 1961;Stein et al., 2014). In comparison with a recent review on this topic (Davies & Asner, 2014), our study provides more detailed insights and technological background of the employed LiDAR-derived vegetation metrics and how they are currently used in avian diversity and distribution research.

| Article selection
Our review is based on a total of 50 articles which focused on analysing bird species distributions and bird species richness in relation to LiDAR-derived vegetation metrics (see Supporting Information Appendix S1). These articles were found with keyword searches in the Web of Science on 11 September 2017 (see workflow in Supporting Information Appendix S2). Included keywords were "avian" or "bird" for the bird component, "LiDAR," "ALS" or "airborne laser scanning" for the LiDAR component and "vegetation structure," "landscape structure" or "habitat struc- We further improved the article selection by applying several selection criteria (Supporting Information Appendix S2). Articles with a focus outside the field of ecology (e.g., LiDAR applications for urban planning) were excluded (n = 29). From the remaining 99 articles, we further excluded articles that were not related to birds (n = 24), review papers (n = 13), articles that used LiDAR but not for deriving specific vegetation metrics (n = 8), studies that did not use direct field observations of birds (n = 3), and studies that focused on breeding success (n = 2) or beta diversity (n = 1). Finally, we added two papers that were suggested by a reviewer and which matched our inclusion criteria. The final list included 50 articles (Supporting Information Appendix S1). is calculated with waveform data as the highest peak above a certain threshold, whereas it is calculated as the mean height of first returns with discrete data). Third, we recorded the LiDAR point density because it influences the calculation of LiDAR-derived vegetation metrics. With low point densities (e.g., 1 point/m 2 ), it is impossible to calculate vegetation complexity at a metre resolution whereas high point densities (e.g., >20 point/m 2 ) allow to analyse multi-layer LiDAR metrics at very fine (e.g., 5 × 5m) resolution. Fourth, we extracted footprint size which describes the diameter of the laser pulse on the ground (Figure 1a). At small footprint sizes (e.g., <1 m diameter), data points reflect measurements of single branches or parts of trees whereas at large footprint sizes (e.g., 10-25 m diameter) they represent multiple trees or even parts of vegetation stands. Fifth, we extracted in which season LiDAR data were collected (leaf-on or leafoff) because this can have consequences for how many points are received from lower vegetation strata (e.g., in broadleaved forest).

| Data extraction
Finally, we recorded how LiDAR data were processed before calculating the metrics, that is whether studies used area-based, voxel-or object-based approaches ( Figure 1b). As a special case of rasterizing, we also recorded whether studies calculated metrics for raster cells (pixels) or for areas around a focal observation point because many bird studies use point counts rather than area-based sampling methods (such as line transects or territory mapping; Bibby, Burgess, Hill, & Mustoe, 2000).
For the extraction of metric information, we recorded all metric names as mentioned in the original articles (e.g., canopy height heterogeneity, mean fractal dimension index, patch diversity, foliage height diversity based on Simpson index). We also distinguished whether metrics were derived from discrete LiDAR or from full waveform LiDAR. We further extracted the effect each metric had, that is the relationship (e.g., model parameter estimate, correlation coefficient, predictor variable importance) between the metric and the species distribution or richness variable.

| Harmonization of metric names
All metrics used in the selected articles were compared with each other in terms of metric names, structural elements of the vegetation that they describe (e.g., canopy, understorey or total vegetation), calculation method (e.g., mathematical formula) and the unit in which the metric is measured (e.g., metres, %, an index value). We then harmonized the metric names by grouping them if they had a similar meaning, using those names which were most widely used or which

| Categorization of metric names
To facilitate a conceptual and ecological comparison and interpretation of metric names, we developed a categorization scheme based on which part of the vegetation was described ( Figure 2a) and which structural type was quantified (Figure 2b).

| Relationship of metrics with species distribution and species richness
We used the extracted information on model parameter estimates, regression coefficients and predictor variable importance to quantify the effect of LiDAR-derived vegetation metrics on species distributions and species richness. Because the methods in the selected articles were very heterogeneous and did not consistently report the basic data needed for a meta-analysis (i.e., enabling calculation and analysis of effect sizes; Gerstner et al., 2017), we simplified the information to whether the LiDAR metric showed an "effect" or "no effect" on the bird response variable. A positive, negative or unknown direction of the bird-habitat relationship was classified as an "effect." A positive and negative direction reflected the sign of parameter estimates from the species distribution and species richness models, whereas an unknown direction reflected that information on the direction of the bird-habitat relationship was missing in a specific article. We distinguished studies of species distribution from those of species richness to test whether "effect" versus "no effect" depended on the bird response variable. Our quantification provides a coarse estimate of the effectiveness of currently used LiDAR-derived vegetation metrics. analyses. The majority of studies used bird distributional information from point counts (n = 22), whereas atlas data (n = 5), territory mapping (n = 6), opportunistic observations (n = 5), bird tracking (n = 4), nest locations (n = 4), transects (n = 4) and range maps (n = 1) were less often used. In those articles focusing on species distributions, a total of 62 different bird species were studied with a total of 76 SDMs (i.e., some species were studied in multiple articles, Supporting

| Overview of articles
Information Appendix S4). The majority of SDM applications fo- Most studies were located in North America (n = 26) and Europe (n = 19) (Figure 3a). The continents Asia (n = 2), South America (n = 1) and the Pacific region (n = 1) were less represented, and no studies from Africa and Oceania were detected with the keywords of our systematic literature search (Figure 3a).
Most studies (n = 41) focused on a single habitat type whereas studies using multiple habitat types (n = 9) predominantly studied a combination of forest with one or more other habitat types (8 out of 9 studies). Hence, almost all studies (n = 48) included forest as a habitat type whereas grassland (n = 9), shrubland (n = 6), agriculture (n = 3) and desert (n = 2) were less represented ( Figure 3b). Some habitat types (e.g., savannas, wetlands) were not represented. These results confirmed the expectation that most studies have been conducted in forest habitats of the Northern Hemisphere temperate zone. Information on spatial extent and spatial resolution was available from 43 (86%) and 50 (100%) studies, respectively. Spatial extent ranged from local (0.1 km 2 ) to regional (330,000 km 2 ), with one study (using spaceborne LiDAR) even being conducted at a global extent (Roll, Geffen, & Yom-Tov, 2015). For studies presenting data on spatial extent, the median area size covered was 53 km 2 (Figure 3c). Thus, most studies (n = 26, 60%) covered areas <100 km 2 , supporting the expectation that studies are predominantly conducted at local or landscape scales. Spatial resolution also varied widely from 0.25 to 9,600 m ( Figure 3). Half of the studies (n = 25, 50%) calculated LiDAR metrics with a radius around focal bird observation points (e.g., opportunistic observations, point counts, territory mapping, bird tracking or nest location data), whereas the other half of the studies (n = 25, 50%) used raster cells (e.g., for atlas data, gridded range maps, bird tracking, nest locations, transect counts, point counts, territory mapping or opportunistic observations). At least 8 studies (16%) first calculated LiDAR metrics at a fine spatial resolution (e.g., 0.5, 1, 3 or 20 m grid cell size) and then aggregated (i.e., averaged) the metrics at a coarser resolution (e.g., 50, 80, 100 or even 1,000 m) when linking them to the bird data.
A total of 48 studies used LiDAR data only from ALS, one study used SLS data at a global scale (Roll et al., 2015), and one both ALS and SLS data at a landscape scale (Vierling, Vierling, Adam, & Hudak, 2013  Recorded footprint sizes were generally small, that is ≤80 cm (all data from ALS studies, n = 19). The number of studies using leaf-on LiDAR data (n = 31) was larger than those using leaf-off (n = 19), with a few studies (n = 5) having data available from both seasons. The area-based approach dominated, with all 50 studies applying this way of processing the LiDAR data ( Figure 3e). However, one study additionally applied a voxel-based approach (Sasaki, Imanishi, Fukui, & Morimoto, 2016) and one other study additionally used an objectbased approach to delineate single trees (Swatantran et al., 2012).

| LiDAR metrics
After harmonizing the 128 identified metric names from the 50 se- For most metric classes, the LiDAR-derived vegetation metrics showed more often an "effect" than "no effect" (Figure 4). For instance, metrics within the most widely used canopy height metric class often showed a relationship with species distributions or species richness (n = 54, 58%). Canopy vertical variability (the second most widely used metric class) was less effective than other metric classes and showed less often a relationship (n = 28, 44%). Overall, canopy metrics were most widely used, and those related to canopy height and canopy cover also tend to effectively explain species distributions and species richness (Figure 4). It is important to note that the effectiveness and direction of the relationship (i.e., positive or negative effect) of a particular metric will depend on the ecology of the specific species.

| D ISCUSS I ON
Our review provides a detailed overview and quantification of the current use of LiDAR-derived vegetation metrics in avian species distribution and species richness research. The categorization into 24 metric classes-defined by vegetation part (total vegetation, single trees, canopy, understorey, and other single layers as well as multi-layer) and structural type (cover, height, horizontal variability and vertical variability)-provides a general conceptual framework for ecological comparisons of LiDAR metrics. While LiDAR canopy metrics of forests are currently most widely used (Figure 4), especially at local and landscape scales (Figure 3), there is great potential for future studies to extend this to other vegetation layers (e.g., understorey, multi-layer), other habitat types (e.g., non-forest ecosystems), other taxa than birds (e.g., invertebrates, other vertebrates) and broader spatial extents (e.g., >100 km 2 ).
The focus of our review was on bird species distribution and diversity studies which linked LiDAR metrics to the occurrence, abundance and species richness of birds as measured from opportunistic observations, point counts, transects, territory mapping, bird tracking, atlas data, range maps or nest location data. These kind of distribution data are widely investigated when using LiDAR in studies with focus on animal ecology (Davies & Asner, 2014). For other aspects of bird ecology (e.g., breeding success, acoustic diversity, beta diversity), some available studies suggest potential similarities to our findings. For instance, LiDAR-retrieved canopy height allows to predict the breeding success (measured as body mass of nestlings in occupied nestboxes) of Great Tits and Blue Tits in broadleaved woodlands in the UK (Hill et al., 2014;Hinsley, Hill, Gaveau, & Bellamy, 2002). Acoustic diversity (quantified with a diversity index based on acoustic frequency bands derived from soundscape recordings) in a Neotropical rain forest in Costa Rica is well explained by canopy height and multi-layer vertical variability (Pekin, Jung, Villanueva-Rivera, Pijanowski, & Ahumada, 2012). Beta diversity (i.e., turnover and nestedness) of birds across Switzerland can be partly explained by variation in LiDAR-derived canopy height (Zellweger, Roth, Bugmann, & Bollmann, 2017). The current use of LiDAR metrics seems to be largely consistent across different bird data sources ( Figure 4), probably because canopy metrics can be reliably retrieved and also act as a potential surrogate for 3-D habitat structure below the canopy, at least in mature and old woodlands and forests. The success and effectiveness of LiDAR metrics to explain bird ecology may depend not only on the quality of the LiDAR data, but also on the quality of the ecological data, including sampling biases or inaccuracies in geo-referencing (Beck, Böller, Erhardt, & Schwanghart, 2014). Moreover, the effect of LiDAR metrics on bird distributions might also depend on whether below-canopy vegetation metrics reflect microclimatic conditions, for example in montane forest environments (Frey, Hadley, & Betts, 2016). Finally, the spatial resolution at which ecological data are linked to the LiDAR data (e.g., length of radii or grid cell size around bird observation points) could also influence the predictive performance of LiDAR metrics. These aspects should be investigated in more detail in future studies.
The revealed geographical bias in current bird-LiDAR studies towards temperate forests of North America and Europe is consistent with an earlier review (Davies & Asner, 2014) and partly caused by F I G U R E 4 Dominant use and effectiveness of classes of Light Detection and Ranging (LiDAR)-derived vegetation metrics as currently employed in studies of (a) avian species distributions, and (b) avian species richness. The metric classes are arranged by vegetation part (total vegetation, single tree, single layers such as canopy, understorey or other layer, and multi-layer) and within that by structural type (cover, height, horizontal variability and vertical variability). Compare Table  1 and Supporting Information Appendix S3 for metrics within metric classes. Frequency indicates how often LiDARderived vegetation metrics within a class showed a relationship ("effect," dark grey) or no relationship ("no effect," light grey). See Supporting Information Appendix S1 for the 50 reviewed articles, and the method section for how information on relationships was extracted from the articles (b) Species richness  (Maltamo et al., 2014). In a seminal paper (based on field measurements rather than LiDAR), MacArthur and MacArthur (1961) already showed that the vertical distribution of biomass in North American forests can be a good predictor of bird species richness. More recent studies using LiDAR have, for instance, demonstrated that forest stand height and density of large conifers can explain the habitat suitability of the spotted owl in North America (Ackers et al., 2015), that heterogeneity in forest canopy height and vertical structure predicts the occupancy of the ground-living hazel grouse in Central Europe (Bae, Reineking, Ewald, & Mueller, 2014;Zellweger, Morsdorf, Purves, Braunisch, & Bollmann, 2014), and that canopy height, total vegetation cover and below-canopy vegetation heterogeneity are important determinants of bird species richness in European and North American forests (Coops et al., 2016;Goetz, Steinberg, Dubayah, & Blair, 2007;Zellweger et al., 2016). Socioeconomic factors most likely explain the strong geographical bias towards North America and Europe (Kissling et al., 2018;Meyer, Kreft, Guralnick, & Jetz, 2015). Such biases towards wealthy countries and temperate forest ecosystems are widespread in ecological studies (Martin, Blossey, & Ellis, 2012) and can limit the scalability and applicability of ecological theory. We therefore recommend to conduct more animal-habitat studies with LiDAR-derived vegetation metrics in tropical forests (Davies, Ancrenaz, Oram, & Asner, 2017;Singh, Tokola, Hou, & Notarnicola, 2017;Wallis et al., 2016), savannas (Loarie, Tambling, & Asner, 2013), wetlands and aquatic reedbeds (Corti Meneses, Baier, Geist, & Schneider, 2017;Zlinszky, Mücke, Lehner, Briese, & Pfeifer, 2012), riparian habitats (Seavy, Viers, Wood, Eavy, & Iers, 2009), or in low-stature terrestrial habitats such as grasslands, tundra, shrublands or agricultural areas (e.g., Svendsen, Sell, Bøcher, & Svenning, 2015;Boelman et al., 2016).
Most bird-habitat studies using LiDAR-derived vegetation metrics have been conducted at small spatial extents (<100 km 2 ) using area-based approaches (e.g., rasterization with grid cells), and with ALS data that have low point densities (<10 points/m 2 ) and small footprint sizes (<1 m diameter). This mostly reflects the current availability of LiDAR data and the challenges related to an effective processing and data management of the massive amounts of point cloud data (Kissling et al., 2017;van Oosterom et al., 2015;Pfeifer, Mandlburger, Otepka, & Karel, 2014;Vo, Laefer, & Bertolotto, 2016).
While there is a rapid growth in the availability and quality of national ALS data (e.g., an increasing open access to country-wide ALS data in Europe), only few studies have so far made use of such largeextent LiDAR datasets in bird-habitat studies. Examples include bird diversity studies at sites across entire Switzerland (Zellweger et al., 2016(Zellweger et al., , 2017 or the Canadian province of Alberta (Coops et al., 2016).
To date, LiDAR bird-habitat studies with a global extent have been rare because the only near-global SLS data currently available are from the NASA full waveform Geoscience Laser Altimeter System (GLAS). This laser scanner has a footprint of ca. 70 m in diameter (the size of a small stand of trees) and has been used to estimate forest canopy height at 1 km 2 resolution (Simard, Pinto, Fisher, & Baccini, 2011). These forest canopy height data have subsequently been used to analyse global bird species richness at 96 × 96 km resolution (Roll et al., 2015). A promising avenue for future research is the new Global Ecosystem Dynamics Investigation (GEDI) LiDAR installed in 2018 by NASA on the International Space Station (ISS; Stavros et al., 2017). This will further advance our ability to monitor 3-D ecosystem structure by measuring canopy heights and foliar vertical profiles with footprints of 25 m globally (between 51.6° N and 51.6° S).
Most reviewed studies used high-resolution ALS data (i.e., footprint sizes <1 m diameter), but point densities were comparably low (<10 points/m 2 ) and the flight season (leaf-on vs. leaf-off) varied among studies. Both point density and flight characteristics can affect the calculation of LiDAR-derived vegetation metrics, for example for the quantification of forest understories (Hill & Broughton, 2009) or in single-tree detection  All studies used an area-based approach for processing the LiDAR data, either by rasterizing point cloud information into grid cells or by calculating LiDAR-derived vegetation metrics with a radius around focal bird observation points. The rasterization of LiDAR point clouds is computationally efficient in terms of data handling, but leads to information loss regarding the fine-scale structure of the 3-D point cloud because grid cells merge the information of various objects. Using an object-based approach would instead have the advantage to classify groups of points (e.g., based on direct neighbourhood information from the point cloud around each focal point) into objects such as structurally similar reedbeds, hedges within agricultural landscapes, or single trees and patches of dominant tree species within a forest (Kissling et al., 2017). While object-based approaches are steadily increasing in geography (Blaschke et al., 2014) and examples of object-based vegetation analysis with LiDAR data exist for urban vegetation mapping (Höfle et al., 2012) and singletree detection within forests (Koch et al., 2014), the bird-habitat studies reviewed here rarely used an object-based approach. One exception was a study on migratory songbirds in a US deciduous forest that employed a canopy height model and an object-based approach to obtain crown radii and height of single trees which then allowed to calculate stem densities per hectare (Swatantran et al., 2012). We envisage that the further development of an object-based methodology will provide exciting new ways to quantify the 3-D vegetation structure and habitat use of animals.
Our harmonization of LiDAR-derived vegetation metrics revealed that many different metric names exist for the same metric calculation method. For instance, foliage height diversity, understorey height diversity and Shannon diversity index are three names that have been used for the same metric (Bae et al., 2014;Vierling et al., 2013;Zellweger et al., 2016). Even more confusing is that similar names are applied to metrics that differ in the way they are calculated. The most obvious example is canopy height which is calculated in myriad ways (see Supporting Information Appendix S3), including the maximum return in a grid cell, the mean of the first returns or the height of the 95th percentile of returns (Ackers et al., 2015;Coops et al., 2016;Smart et al., 2012). Most of these metrics might be strongly correlated, but their sensitivity to outliers and point cloud density might differ (e.g., maximum vs. 95th percentile canopy height). The 77 unique metric names were grouped into 24 theoretically possible metric classes defined by vegetation part and structural type. This categorization might help in future studies for the selection, prioritization and ecological interpretation of LiDAR metrics. However, choosing the right metrics is not always easy and might depend on a priori ecological knowledge of animal habitat use.
In cases where such knowledge is lacking, it is also possible to condense the variability in LiDAR point clouds by generating principal components as predictors and using them in exploratory analyses with animal species distribution data to generate hypotheses about animal habitat and space use (Ciuti et al., 2018).
Our categorization of LiDAR metrics allowed a comparison of the current use and relative effectiveness of these metrics. This analysis supported our initial hypothesis that canopy metrics (related to canopy height, canopy cover and canopy vertical variability) are most widely used (Davies & Asner, 2014), although we show that they may not always be a good predictor of individual species distributions or species richness (Figure 4). One reason for their widespread use is that canopy metrics (e.g., canopy height) can be well retrieved even with low point densities (Ruiz, Hermosilla, Mauro, & Godino, 2014) or from large footprint SLS data with global coverage (Lefsky, 2010;Simard et al., 2011) and that they may act as a surrogate for 3-D habitat structure below the canopy. LiDAR metrics from below-canopy layers are more challenging to derive and require LiDAR point clouds with high point densities. For instance, single-tree segmentation is often carried out by detecting the local height maxima from rasterbased canopy height models which might easily miss trees below the dominant canopy (Koch et al., 2014). Single-tree delineation is also computationally demanding and requires high point densities (Höfle et al., 2012), and is often only applied to particular forest stands (Koch et al., 2014) or in urban environments where individual trees are spatially well separated (Koma, Koenig, & Höfle, 2016). Understorey and multi-layer LiDAR metrics are also less widely used because most ALS data have low point densities and are limited in capturing information below the canopy. Nevertheless, below-canopy metrics in forest have high potential for understanding and predicting the distribution and habitat use of birds (Cody, 1985;Dunlavy, 1935;MacArthur & MacArthur, 1961) and potentially of many other taxa, and hence should be of increasing focus in the future.

| CON CLUS ION
Our review shows that LiDAR metrics play an important role for un- methodological and data quality characteristics (e.g., flight heights, season, laser return intensities, sensor angle, full waveform vs. discrete points) differ among LiDAR datasets, which may subsequently influence the calculation of LiDAR metrics. Moreover, the development of voxel-based and object-based approaches will allow novel ecological applications (e.g., via better quantification of subcanopy metrics and multi-layer vertical variability), but require high point density information and efficient software for data handling and processing. We further encourage animal-habitat studies with LiDAR in non-forest habitats, including savannas, wetlands, agricultural landscapes and other open habitats. Together, these advancements will allow unprecedented insights into species-habitat relationships, not only for birds but also for other taxa.

DATA ACCE SS I B I LIT Y
All data extracted from the reviewed literature together with a detailed metadata description are available from the Dryad Digital Repository: https://doi.org/10.5061/dryad.tm28hb6 (Bakx, Koma, Seijmonsbergen, & Kissling, 2019).