Assessing GEDI data fusions to map woodpecker distributions and biodiversity hotspots

The identification of biodiversity hotspots is a key approach used to prioritize areas for conservation (e.g. Jenkins et al 2013). Hotspot analyses might include different taxonomic groups or a subset of threatened or endemic species (Jenkins et al 2010, Roll et al 2017, Villacampa et al 2019), but the overarching goals of these efforts are typically to identify areas for future conservation actions, particularly in light of environmental change (e.g. Bellard et al 2014, Weinzettel et al 2018, Trew and Maclean 2021, Zhao et al 2022). In forested systems, areas of high biodiversity may be attributed in part to heterogeneity in forest structure, as sites with high structural heterogeneity provide multiple niches that support high species richness (MacArthur and MacArthur 1961, Bae et al 2018, Davison et al 2023). Our understanding of how forest structure attributes (amongst other variables) are related to patterns of biodiversity has been greatly expanded by fine resolution data from airborne lidar (e.g. Goetz et al 2007, Clawges et al 2008, Vierling et al 2008, Műller and Brandl 2009, Melin et al 2018, Toivonen et al 2023); however, airborne lidar data can be expensive to obtain, which can influence their availability in different regions of the world (Davies and Asner 2014). In contrast, the availability of a broader spaceborne lidar data, which has near-global coverage, has great potential for applications for understanding species distributions and patterns of biodiversity (Dubayah et al 2020), given the importance of forest structure on animal distributions, demography, and biodiversity patterns (Vierling et al 2008, Goetz et al 2014, Davies and Asner 2014, Hill and Hinsley 2016, McLean et al 2016, Bae et al 2018).

NASA's Global Ecosystem Dynamics Investigation (GEDI) mission used a full-waveform lidar instrument mounted to the International Space Station to make detailed measurements of 3-dimensional forest structure across the globe. This forest structural data has multiple implications for studies of forest ecology, including applications to better understand animal habitat and biodiversity patterns (e.g. Dubayah et al 2020, Hakkenberg et al 2023). GEDI data have been used to understand relationships between forest structure and multiple mammal (e.g. Smith et al 2022, Killion et al 2023, Barry et al 2024) and bird species (Burns et al 2020, Vogeler et al 2023a). These studies used a variety of GEDI data, ranging from simulated GEDI waveform data (Burns et al 2020), to aggregated forest structure traits at 1000 m around camera trap locations (Killion et al 2023), to GEDI data fusion modeling efforts that have produced wall-to-wall forest structural metrics at 30 m resolution (Smith et al 2022, Vogeler et al 2023a, Barry et al 2024). These studies all illustrate the importance of structural elements in helping to advance our understanding of animal-habitat relationships, but generally, these efforts have focused on individual species that might be of management interest, such as forest carnivores (Smith et al 2022, Killion et al 2023), ungulates (Killion et al 2023), small mammals (Smith et al 2022, Barry et al 2024) and three woodpecker species (Vogeler et al 2023a). The focus on woodpeckers is partly because individual species are recognized as keystone species whose provisioning of tree cavities supports a wide variety of cavity users, especially those that do not excavate their own cavities such as bats and cavity-nesting ducks (e.g. Bonar 2000, Aubry and Raley 2002).

Several studies from around the world demonstrate that diverse woodpecker communities are indicators of songbird diversity (Mikusiński et al 2001, Drever et al 2008, van der Hoek et al 2020). Avian hotspots associated with woodpeckers can occur via two non-exclusive mechanisms functioning at different spatial scales. First, individual woodpecker species often excavate cavities with different entrance hole dimensions and cavity volumes (e.g. Stitt et al 2019), that are then used by a variety of secondary cavity user species (Martin et al 2004, Gentry and Vierling 2008, Remm and Lohmus 2011, Tarbill et al 2015, Trzcinski et al 2022, Cadieux et al 2023). Second, because of the variation in species-specific woodpecker habitat requirements, woodpecker diversity hotspots may also indicate diverse forest structure that can support high diversities of other species across a gradient of forest succession. van der Hoek et al (2020) examined relationships amongst cavity excavator species richness (e.g. woodpeckers and other species that excavate cavities) and the species richness of other forest bird guilds at the global scale, using bird species ranges from BirdLife International and NatureServe at 10 km x 10 km grid cell sizes. They found strong positive relationships between cavity excavator richness and species richness of other avian groups, including forest generalists, forest specialists, and other cavity-nesting birds; they suggest that cavity excavators are a group that could serve as indicator species of overall avian richness in most forests around the world (van der Hoek et al 2020).

Woodpecker occurrence and habitat suitability has long been linked to specific aspects of forest structure, and several studies have used airborne lidar to describe distribution patterns of different woodpecker species (Holbrook et al 2015, Vogeler et al 2016, Virkkala et al 2021, Hu and Tong 2022). Vogeler et al (2023a) recently incorporated the forest structure samples provided by GEDI along with continuous remote sensing data sets within modeling frameworks (hereafter GEDI-fusion metrics), and related these metrics of canopy cover, canopy height, and foliage height diversity (fhd) to examine occurrences of three woodpecker species in western North America. Predictive performance and ecological insights from these habitat models indicated that GEDI-fusion metrics of forest structure capture essential habitat elements that are commonly associated with woodpecker distributions (Vogeler et al 2023a). In this study, we sought to build upon that work to explore areas of high woodpecker species co-occurrence (e.g. biodiversity hotspots) using GEDI-fusion structural metrics, bioclimatic conditions, and forest composition.

To characterize biodiversity hotspots, we expanded the analyses from Vogeler et al (2023a) to include the full suite of woodpecker species for whom we had sufficient data within our focal ecosystem. This is the first study to our knowledge to utilize GEDI data to map the distributions of a keystone guild with the aim of characterizing diversity hotspots. Moreover, we assessed how woodpecker richness was partitioned across land ownership to inform conservation efforts regarding biodiversity. While woodpecker richness patterns have been described regionally (e.g. Mikusiński et al 2001) and globally (e.g. van der Hoek et al 2020), these studies utilized bird atlas data or other sources of bird distribution data and were not using remotely sensed data to examine underlying habitat patterns of forest structure associated with woodpecker distributions and hotspots. Additionally, our work provides critical information concerning benefits and potential limitations of GEDI-derived data for multi-species distribution modeling.

2.1. Study area

We focused our modeling efforts on the Marine West Coast Forest (MWCF) level I ecoregion (U.S. EPA 2010), in Oregon and Washington, USA, which comprises three smaller (level III) sub-ecoregions: the Coast Range, Willamette Valley, and Puget Lowland (figure 1; Omernik 1987). This region is generally temperate but inland areas experience larger daily and seasonal temperature and precipitation shifts. The Coast Range is wetter, more topographically rugged, and less densely populated than the Valley and Lowland (e.g. Portland and Seattle, respectively). Tree species composition changes with precipitation, elevational, and latitudinal gradients but is generally dominated by western hemlock (Tsuga heterophylla), Douglas-fir (Pseudotsuga menziesii), and Sitka spruce (Picea sitchensis) but shifts to Oregon white oak (Quercus garryana) woodlands in drier and warmer areas, (e.g. the Valley; Franklin and Dyrness 1973). Major disturbance agents are timber harvest, fire (natural and prescribed), insect mortality or defoliation, and weather-driven events (e.g. drought, wind blow-down). Land ownership in the Coast Range is a mix of private industrial timber and public lands, while the Valley and Lowland are heavily dominated by small woodland lot owners and other private non-industrial lands.

Zoom In Zoom Out Reset image size

2.2. Bird data

We examined and predicted the distribution of 12 woodpecker species in this ecoregion associated with forest structural elements that ranged across canopy cover, tree height, and edge gradients. Of these 12 species, only 7 had sufficient sample sizes for our analysis. These species include: Acorn woodpecker (M. formicivorous; ACWO), red-breasted sapsucker (S. ruber; RBSA), American three-toed woodpecker (Picoides dorsalis; ATTW), the downy woodpecker (Dryobates pubescens; DOWO), hairy woodpecker (Leuconotopicus villosus; HAWO), the northern flicker (Colaptes auratus; NOFL), and the pileated woodpecker (Dryocopus pileatus; PIWO). In addition to representing a range of habitat associations, these species capture a gradient in body sizes and consequently a gradient in cavity size.

We obtained encounter data for each species from the eBird database (eBird 2021). While we were ultimately interested in the probability that a species occupies a site, semi-structured citizen-science data, such as eBird data, better lends itself to examining the probability that a species is encountered by an observer (Strimas-Mackey et al 2020). To better use these encounter rates as an index of occurrence and account for varying detectability, we followed best practices and techniques described in Johnston et al (2019) and Strimas-Mackey et al (2020; e.g. inclusion of effort variables, calibrating models). We included only stationary checklists to better match the spatial scale of observations with the pixel size of GEDI-fusion metrics. As a general workflow, we obtained eBird checklists conducted in the study area between June 1 and July 31 of 2016–2020 (totaling 45 336 checklists), conducted spatiotemporal subsampling (following Strimas-Mackey et al 2020), obtained effective sample size of positive observations for each species, and retained 20% of the checklists for model evaluation (table 1, figure 2).

Zoom In Zoom Out Reset image size

Table 1. Species' sample sizes and detection frequencies for stationary eBird checklists recorded between June 1 and July 31 of 2016–2020 in the Marine West Coast Forest ecoregion. Effective sample size includes detections of the species, whereas training and testing dataset sample sizes are the total number of spatially subsampled checklists in each dataset, respectively.

	Sample size			Detection frequency
Species	Effective	Training dataset	Testing dataset	Before subsampling	After subsampling	Training dataset only
PIWO	1501	7046	1765	0.033	0.041	0.057
NOFL	8522	7011	1753	0.188	0.174	0.186
DOWO	3450	7009	1752	0.076	0.073	0.077
HAWO	1800	7009	1752	0.040	0.045	0.058
RBSA	1328	7009	1752	0.029	0.033	0.041
ATTW	1328	7009	1752	0.029	0.033	0.041
ACWO	630	7009	1752	0.014	0.009	0.005

For each species, we obtained home range sizes from the literature and used this reported range to identify three appropriate buffer distances within which to consider GEDI-fusion metrics (table 2). These buffer distances were in multiples of the 30 m × 30 m pixel size used in the GEDI-fusion products (90 m, 180 m, 360 m, 540 m, 1530 m, 2520 m) so that the resulting 90 m × 90 m raster grid centroids remained consistent for predictive maps. The three buffer sizes selected for each species included one buffer size larger than the maximum reported territory size, one size smaller than the minimum reported territory size, and one buffer size that fell within the reported range in territory sizes.

Table 2. Body mass size, home range data, and buffer distances for woodpecker species found in the Marine West Coast Forest ecoregion in Oregon and Washington, USA.

Species	Mass (in g)a	Home range radius (m)	Home Range Citations	Buffer diameters (m)
PIWO	324.6 (range = 293–355, n = 18)b	1139–1658	Bull 1993, Mellen et al 1992, Bull and Jackson 2020, Aubry et al 1996	540, 1530, 2520
HAWO	173.62 ± 0.636 SE (n = 33)d	58–91	Jackson et al 2020	90, 180, 360
NOFL	160.6 ± 8.99 (n = 1390)c	282	Elchuk and Wiebe 2003	180, 360, 540
ACWO	80.9 ± 4.9 (n = 706)e	124–205	Hooge 1995	180, 360, 540
RBSA	58.3 ± 5.23 (n = 23)f	137	Walters et al 2020	90, 180, 360
ATTW	58.3 ± 3.5 (n = 16)g	410–984	Tremblay et al 2020	540, 1530, 2520
DOWO	27.8 ± 1.3 (n = 40)h	90–200	Ritchison 1999	180, 360, 540

^aMass is reported as mean ± SD (n), as reported in the Birds of the World account for each species unless noted otherwise, for a male in the reported geography that most closely matches the study area (e.g. British Colombia rather than western Pennsylvania is used for Northern Flicker). ^bBull and Jackson (2020). ^cWiebe and Moore (2023). ^dJackson et al (2020). ^eKoenig et al (2020). ^fWalters et al (2020). ^gTremblay et al (2020). ^hJackson and Ouellet (2020).

2.3. Environmental variables

2.4. MODIS

2.5. PRISM

2.6. GEDI-fusion metrics

2.7. Random forest analysis

2.8. Hotspot analysis and land ownership

2.9. Results

Our single-species encounter rate models showed good fit, with AUC ⩾ 0.77 (table 4). Model specificity was quite high (⩾0.80) but sensitivity was relatively low (0.25–0.64), indicating that our models were better at identifying areas where species were unlikely to be encountered than areas where the species was actually encountered.

Table 4. Model evaluation metrics for random forest models of single species' encounter rates, including Mean Square Error (MSE), Sensitivity, Specificity, and Area Under the Curve (AUC).

Species	MSE	Sensitivity	Specificity	AUC
PIWO	0.04	0.38	0.95	0.81
NOFL	0.12	0.64	0.80	0.78
DOWO	0.05	0.38	0.97	0.81
HAWO	0.05	0.38	0.98	0.87
RBSA	0.03	0.24	0.98	0.77
ACWO	0.00	0.55	1.00	0.98
ATTW	0.03	0.25	0.98	0.78

Variable importance rankings differed across species (figure 3; supplemental figures 1 and 2). Process variables were highly ranked for several species. Likewise, bioclimatic variables had the highest Gini Index for three species. In particular, isothermality (biovars_3) was highly ranked for all species. Forest composition variables were highly ranked for only a few species. Overall, bioclimatic variables had the highest mean predictor importance values for all species, and the mean predictor importance of GEDI fusion metrics was similar to those of the bioclimatic variables for several species (supplemental table 1).

Zoom In Zoom Out Reset image size

Although GEDI-fusion metrics were not the most important variables for any species, each species had at least one highly ranked GEDI-fusion metric within the top nine ranked predictors (figures 4, S1 and S3). RBSA had only one highly ranked GEDI-fusion metric: fhd_360; ATTW had two: fhd_180 and fhd_360; and all other species had three to five highly ranked GEDI-fusion metrics. Of these, fhd was the most consistently important GEDI-fusion variable and was important for all species at one or more scales. Three species had fhd highly ranked at all three scales. Cover was highly ranked for DOWO, HAWO, and ACWO, while rh98 was highly ranked for PIWO and ACWO. Partial dependence plots showed species-specific associations with these GEDI-fusion metrics, as with non-GEDI-fusion metrics (figure S3).

Zoom In Zoom Out Reset image size

Individual species encounter rate maps varied spatially (figure 5). PIWO and HAWO had especially high encounter probabilities in the Coast Range (figures 5(a) and (b). In contrast, NOFL and DOWO encounters concentrated in the eastern Willamette Valley and Puget Lowland (figures 5(c) and (g)). RBSA and ATTW encounters were especially rare in coastal areas (figures 5(e) and (f)). ACWO had especially high encounters in Oregon and the northern Coast Range (figure 5(d)). Following these individual species patterns, species richness was notably lower in the northern Coast Range, the Willamette Valley, and coastal Oregon (figure 6).

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Most hotspots are located inland rather than along the coast, with especially dense concentrations of hotspots along the eastern edge of the ecoregion, and in the central and northern portions of the ecoregion (e.g. figure 6(a)). Most hotspots (87.4%) were privately owned. Hotspots had a mean rh98 of 26.9 m (±3.8 SD), mean cover of 0.56 (±0.08 SD), and mean fhd of 2.87 (±0.14 SD). Areas outside of the hotspots had a mean rh98 of 23.7 m (±11.3 SD), mean cover of 0.51 (±0.24 SD), and mean fhd of 2.56 (±0.57 SD; table 5).

Table 5. Environmental characteristics of hotspots and areas outside hotspots.

	Hotspots		Outside hotspots
Variable	Mean	SD	Mean	SD
rh98_2020	26.92	3.77	23.69	11.33
cover_2020	0.56	0.08	0.51	0.24
fhd_2020	2.87	0.14	2.56	0.57
deciduous_combined_540	0.25	0.2	0.08	0.15
evergreen_combined_540	0.67	0.2	0.67	0.39
biovars_1	11.11	0.49	10.66	1
biovars_2	10.21	0.74	9.94	1.1
biovars_3	40.97	1.5	42.43	3.27
biovars_4	542.08	18.69	505.03	59.13
biovars_11	4.89	0.37	4.94	1.18
biovars_12	1252.61	130.93	1921.02	835.69
biovars_14	16.02	6.32	20.75	12.67
biovars_15	61.31	6.69	65.75	7.37
biovars_17	81.93	21.03	109.17	53.62
biovars_18	83.51	21.21	115.83	62.18
biovars_19	502.83	67.68	808.79	359.98

GEDI data have been limited in their applications towards mapping of diversity hotspots to date, and our effort represents one of the first to address these patterns based on animal habitat associations. Multiple studies have used GEDI data to examine patterns of tree diversity (Marselis et al 2022, Crockett et al 2023, Hakkenberg et al 2023, Ren et al 2023, Torresani et al 2023), but there have been a small number of studies that have addressed bird and GEDI relationships (e.g. Burns et al 2020, Vogeler et al 2023a). This study represents the first to our knowledge that focuses upon using GEDI-fusion metrics to describe biodiversity hotspots as represented by co-occurring woodpecker species, and our results show that the inclusion of GEDI-fusion metrics enhanced modeling of biodiversity hotspots that can be useful for future management and conservation efforts.

Our models for all species included combinations of process variables, bioclimatic variables, and GEDI-fusion metrics. Process variables associated with ebird surveys were important in modeling encounter rates, but our primary focus is on the ecological variables from our study. Bioclimatic variables were among the most important ecological variables for mapping species distributions, and isothermality was a highly ranked predictor for all study species. Isothermality is a representation of temperature variability across daily and annual time steps and captures a longitudinal gradient from coastal to inland areas, which may be important for distribution patterns at broad extents. Goetz et al (2014) found that bioclimatic variables may be amongst the strongest predictors of bird species richness for some avian guilds when examining relationships between bird species richness and forest structural metrics, bioclimatic variables, and vegetation characteristics. Goetz et al (2014) found that the relative importance of bioclimatic variables differed across avian guilds, which was broadly consistent with the range of variable importance variation we saw in our analysis.

In addition, GEDI-fusion metrics were consistently highly ranked predictors in individual species' encounter rate models. fhd was the most consistently important GEDI-fusion metric, and was important for all species at one or more scales. Many bird species forage in structurally diverse forests. For instance, NOFL are noted to forage on the ground in forest gaps and are often associated with forest edges (Wiebe and Moore 2023). Rh98 was highly ranked for both PIWO and ACWO, and these two species have been noted to require larger trees for nesting (e.g. Bull and Jackson 2020, Koenig et al 2020). Cover was highly ranked for DOWO, HAWO, and ACWO but the influence of cover differed among these species. Modeled encounter rates increased at higher cover values for ACWO whereas modeled encounter rates for DOWO and HAWO decreased across the range of similar cover values. These different trends for the same variable likely reflect species-specific spatial patterns that might be associated with forest structural elements used for foraging or indicative of predation risk (e.g. Chalfoun and Martin 2007).

Different forest characteristics were important to different species at different spatial scales, possibly for different and/or competing reasons (e.g. Chalfoun and Martin 2007). For example, fhd was important at the 90 m by 90 m grain size for species such as HAWO and RBSA. At the other end of the spectrum, the wider-ranging PIWO were associated with fhd at scales ranging from 540 m to 2520 m grain sizes. Our results demonstrate the value of moderate resolution of GEDI-fusion metrics for versatility in summarizing forest structural characteristics at a variety of species-appropriate spatial scales, particularly finer resolutions that have historically been difficult to measure across large geographic areas but are especially important to assess for species with smaller home range sizes.

The land ownership of these hotspots is of particular interest in the framework of wildlife management; while wildlife do not respond directly to land ownership, ownership informs habitat management practices (Ahlering et al 2019), which in turn influences the forest structure. For example, in the MWCF ecoregion, old-growth forest is more likely to occur under federal ownership, while clear-cutting is currently more prevalent on lands under private ownership (Phalan et al 2019). There was a greater than expected proportion of private ownership across the hotspots that we identified, and previous studies (e.g. McComb et al 2007) have noted that habitat availability for species like pileated woodpeckers are likely to be influenced by private land ownership patterns in this region. Our results similarly suggest that forest management practices on private lands would be an important consideration for habitat management for multiple woodpecker species.

We have demonstrated that GEDI data have provided structural information that are ecologically relevant for woodpecker species and pertinent to management planning. While bioclimatic variables were important in our analyses, forest structural characteristics included in our analysis, such as height or cover, could provide targets for managers during silvicultural prescriptions. As a global product, GEDI can provide structural information for woodpecker species that are likely to be useful in mapping both individual species distributions (e.g. Vogeler et al 2023a) as well as expanding those analyses to examine patterns of cavity excavator hotspots. Such analyses will evolve as eBird data become more available for rare species in under-surveyed areas, and alternate sources of bird data (BirdLife International and NatureServe) will also provide useful information for such broad scale investigations (e.g. van der Hoek et al 2020). Several authors have noted the potential of woodpeckers to serve as indicator species in different regions due to their conspicuousness, their sensitivity to forest structures, and their role in supporting biodiversity (Mikusiński et al 2001, van der Hoek et al 2020, Menon and Shahabuddin 2021). van der Hoek et al (2020) further suggests that woodpeckers (and cavity excavators) can be a focal group with which to monitor and assess areas of high biodiversity across all forest types around the globe (excluding Australasia). We suggest that the global nature of GEDI data can likely assist in identifying biodiversity hotspots in a variety of forest types around the globe by providing forest structural data which might enhance our understanding of woodpecker distributions and associated diversity hotspots.

This work was supported as a funded GEDI Competed Science Team project (NASA grant 80NSSC21K0192). We would like to thank the GEDI Science Team for advice on guidelines for GEDI data filtering and use during project development. We would also like to acknowledge Sophie Gilbert for discussions early in project development. We also would like to thank the reviewers for their constructive comments.

The datasets presented in this study can be found in online repositories cited below; associated code is available at https://github.com/VogelerLab/GEDI_fusion_hotspots_Elliott_2024 (Elliott et al 2024).