Insights from five decades of monitoring habitat and breeding populations of American woodcock

1. Habitat loss and degradation are contributing to severe declines in many North American bird species. For American woodcock Scolopax minor (‘woodcock’), loss of preferred young forest habitat matrices are generally attributed as the primary drivers of range-wide population declines in eastern North America, but regional patterns in abundance or habitat availability have not been assessed in Nova Scotia, the northeastern-most portion of the range. 2. Our objectives were to (a) identify regions of similar trends over the past five decades, (b) evaluate spatiotemporal relationships in the effect of habitat availability on abundance across the province and (c) provide recommendations for woodcock management priorities in Nova Scotia to target local population declines. 3.Using50yearsofstandardisedsurveysandhabitatmeasures,weinvestigatedwood-cock population trends and local habitat availability by applying a novel spatially-explicit model with Integrated Nested Laplace Approximation. 4. Province-level declines were primarily driven by losses of breeding woodcock in the north and south of the province, while the central region experienced growth. The proportion of area around a survey route comprised of clear-cut, harvested forest was the most important habitat and ultimately range wide. The temporally and spatially extensive surveys used here provide a remarkably comprehensive avian monitoring dataset. Coupled with modern analytical tools, our study system serves as an example of how long-term avian survey efforts are capable of informing management regionally to achieve broader conservation targets.

and ultimately range wide. The temporally and spatially extensive surveys used here provide a remarkably comprehensive avian monitoring dataset. Coupled with modern analytical tools, our study system serves as an example of how long-term avian survey efforts are capable of informing management regionally to achieve broader conservation targets.

K E Y W O R D S
American woodcock, avian monitoring, clear-cut harvest, forestry, habitat availability, management, population decline

INTRODUCTION
Recent findings report that North America has lost more than one in four birds in the last 50 years, where habitat loss or degradation have been key drivers for many species (NABCI Canada, 2019; NABCI US, 2019; Rosenberg et al., 2019). Long-term monitoring schemes have proven crucial to provide the data required in gaining broad insights into avian trends in relation to ecological traits (Reif, 2013;Rosenberg et al., 2019). Continued efforts applying advanced, spatially explicit, analytical approaches to comprehensive time series of abundance and habitat are urgently needed to identify conservation and management priorities for target groups and species (Rosenberg et al., 2019).
American woodcock Scolopax minor (herein 'woodcock') share the worrisome trends emerging continentally across avian guilds, showing declines range-wide across eastern North America (Seamans & Rau, 2018). However, the relationship between human influence on the landscape and abundance is challenging to disentangle for species like the woodcock with highly specialised, disturbance-dependent niches (Hunter, Buehler, Canterbury, Confer, & Hamel, 2001). Woodcock are forest birds with particular breeding habitat requirements, thriving in local landscapes comprising young, moist woodlands with high stem density interspersed with clearings. Loss and degradation of preferred young forest matrices are generally attributed as the primary drivers of range-wide declines (Masse, Tefft, & Mcwilliams, 2014;McAuley, Keppie, & Whiting, 2013).  (Seamans & Rau, 2018). SGS have been conducted continuously over the past 50 years, with excellent spatial coverage across the entire province. These data provide unique opportunities to assess how regional population trends contribute to the province-wide trend, and how localised habitat alteration and forest management practices may be influencing populations.
In this study, we investigated spatial patterns in woodcock population trends and associated changes in local habitat availability across the province of Nova Scotia. The latter analysis uses a novel application of spatially explicit, Bayesian approximation methods that generate parameter estimates and a latent Gaussian field to address spatial correlation with the Integrated Nested Laplace Approximation (INLA) approach (e.g. Bivand, Gomes-Rubio, & Rue, 2015). Our objectives were to (a) identify regions of similar long-term population trends over the past 50 years, (b) evaluate potential spatial and temporal relationships in the effect of habitat availability on woodcock abundance across the province and (c) based on our findings, provide spatially explicit recommendations for both future research priorities and land management practices in Nova Scotia to target local declines with the goal of ultimately improving the province-wide population trend.

Inferring population indices from singing ground surveys
The SGS is a standardised protocol used to estimate annual population indices for woodcock across their range (Seamans & Rau, 2018). The SGS exploits the conspicuous courtship display, whereby males repeatedly vocalise a series of loud 'peent' calls followed by an aerial display. These exhibitions are performed by males throughout spring dusk and dawn periods. Beginning in 1968, SGS route locations were chosen along secondary roads in the center of random 10-minute degree blocks within every state and province of the breeding range, totalling roughly 1,500 survey routes. Each route is 5.4 km long and consists F I G U R E 1 Total woodcock detected on 52 SGS routes surveyed across Nova Scotia, 1968-2019. Total detections depicted by colour (hotter = higher number detected) and circle size of 10 listening stops spaced at 650 m apart to avoid detecting the same individuals at more than one stop. Surveys commence shortly after sunset during the peak spring courtship period at a given latitude.
Observers record the number of individuals heard at each stop, and the total displaying males detected along the route serves as an index of local population abundance. The time of a survey relative to sunset and adverse weather conditions affect both displays and observer detection, and thus surveys are only conducted or deemed acceptable when conditions are within prescribed limits. If surveys on a given route fail to detect woodcock for two consecutive years, the route is not surveyed again for a period of 5 years (Seamans & Rau, 2018). All

Singing ground surveys in Nova Scotia
In Nova Scotia, 52 SGS routes have been active and geo-referenced during the period 1968 to 2019. The mean distance between nearest routes is 21.3 km (± 8 km standard deviation; range 3-43 km apart).
Routes were initiated between 1968 and 2007 and have been surveyed for periods of 12-51 years (mean 43 years across routes). During this time, between 5 and 51 acceptable annual surveys have been recorded for each route, totalling 1,584 surveys conducted by 129 observers.
Observers run their routes for as many consecutive years as possible to reduce observer bias on trend estimates (Seamans & Rau, 2018); in Nova Scotia, the mean number of observers per route from 1968 to 2019 was low at six, ranging from 2 to 13.

Estimating route-level trends in population indices
Initial examination of survey data over time revealed high variation across routes in both absolute abundance and change across the period 1968-2019 ( Figure 1). Plotting indicated that spatial patterns in trends may be significant. To model the data at the route level, annual count (Poisson distribution, log-link) was regressed against survey year. We estimated parameters with INLA via the R-INLA package (http://www. r-inla.org). INLA is an extremely fast Bayesian approximation technique that integrates using a second-order Taylor expansion from the mode (Rue, Martino, & Chopin, 2009). It has become increasingly popular in ecology because of its speed, ease of use and functions for making Bayesian approximations from modelling latent Gaussian models (described below; Cosandey-Godin, Krainski, Worm, & Flemming, 2014;Gutowsky et al., 2020). For each model presented here, survey year was subtracted from the first survey year to give the intercepts a meaningful interpretation, that is, a baseline count estimate.
The count coefficient for each route was plotted by the geographic coordinates to reveal temporal and spatial patterns. Credible intervals that overlapped zero were not considered important. Only routes with 10 or more years were included for the trend analyses (93% of routes, n = 48). All analyses were conducted in the R statistical environment (R Core Team, 2019).

Characterising habitat around SGS routes
A buffer area of 1 km around the length of each SGS route (Supp.   Fig. 1).

Evaluating the influence of habitat availability across Nova Scotia
The relative importance of habitat variables was first examined using boosted regression trees (BRT; Friedman, 2001) to model woodcock abundance (route level means for each inventory cycle time period) to the set of nine explanatory habitat variables for each of the three periods. BRT differs from traditional regression in that many models (i.e. trees) are fit and combined to optimise predictive performance. Additionally, BRT is able to cope with complex interactions, non-linearity and outliers (Elith, Leathwick, & Hastie, 2008). Classification trees, as performed here, fit regions with the most probable class. Recursive binary splits are used to grow trees until a stopping criterion is reached.
BRTs were fitted with the R packages dismo (Hijmans, Philips, Leathwick, & Elith, 2017) and gbm (Ridgeway, 2017), which prune trees through cross-validation (Hastie, Tibshirani, & Friedman, 2009). We used 10-fold internal cross validation on a dataset of response and predictor variables associated with each route (Jarnevich, Stohlgren, Kumar, Morisette, & Holcombe, 2015). Two learning rates, two bag fractions and five tree complexities were tested for a total of 20 models with 1,000 trees each. The model with the lowest deviance was pruned to remove predictors based on k-fold cross validation and a procedure similar to backwards model selection that identifies the order in which low contributing predictors should be removed based on the mean change in deviance and standard error (Elith et al., 2008;Hijmans, Tibshirani, & Friedman, 2017). While BRT fits woodcock abundance for each habitat variable, anticipated spatial correlation was expected in the response variable. Therefore, BRT was used only to identify the top predictor variables for a generalised linear mixed model (GLMM) capable of incorporating spatially correlated random effects.
R-INLA contains functions to construct Gaussian random fields (GRFs) that allow for parameter estimation in relation to complex spatial structures (Beguin, Martino, Rue, & Cumming, 2012;Bivand et al., 2015;Lindgren, Rue, & Lindström, 2011). GRFs are estimated using Matérn correlation solved by a stochastic partial differential equation on an irregular grid, that is, mesh (Bivand et al., 2015). The mesh is a series of non-overlapping triangles (edges and vertices) created using F I G U R E 2 Slope (y-axis) and intercept (x-axis) ± 95% credible interval from annual count (Poisson distribution, log-link) regressed against survey year for each SGS route in Nova Scotia. The x-axis was transformed to show change in annual abundance relative to the expected number of woodcock in the first year of the survey whereas the y-axis remains on the natural log scale to indicate the relative rate of expected change (i.e. approximating the annual per cent change). The dashed red horizontal line shows where credible intervals overlap zero functions of the R-INLA package. Our mesh was generated from a GIS shapefile plus an extended domain to avoid boundary effects (i.e. an increase in variance near the edge of the mesh) that can arise from the stochastic partial differential equation approach of GRF estimation (Blangiardo & Cameletti, 2015). Our model was a normally distributed GLMM with route-level mean abundance for each inventory cycle time period as the response, the top habitat variables found by the BRT (proportion of clear-cut area and urban/developed area within a route buffer), time period, all interactions, and the locations of sampling routes as the spatially correlated random effects. Two-way and three-way interactions were estimated to test for complex relationships among the variables identified by the BRT.

Route-level trends in abundance
Our trend analysis showed that there were 32 routes with a 95% probability of either an increasing (n = 17) or decreasing (n = 15) trend in woodcock abundance since the first surveys were completed

Influence of habitat availability on abundance
The BRT with the lowest deviance had a tree complexity of four, learn-  (Table 1). Neither clear-cut (%) nor urban area (%) environments alone explained woodcock abundance; however at the average value for per cent urban area (mean: 3.7% ± 4.0 SD, range: 1.5-28.5%) woodcock abundance increased considerably in the third time period (Figure 6). During the third time period, woodcock numbers increased from two to ten birds on average as the amount of clear-cut area increased from zero to 28% (Figure 6).
While a decrease occurred in woodcock abundance as the per cent of urban environment increased, the trend was not different from zero (Table 1). Approximately 40% of the variation in woodcock abundance was explained in the spatial random field (Figure 5b). The GRF indicated a latent process showing the strongest effects on abundance in the northcentral and northeastern regions of the province ( Figure 5-b).

DISCUSSION
We evaluated spatiotemporal patterns in woodcock population trends and associated changes in local habitat availability across the province of Nova Scotia, Canada, over five decades. To our knowledge, these analyses of population indices and habitat availability are unique in their spatial and temporal coverage for ground-nesting landbirds.
We found clear spatial patterns in woodcock population trends. The province-wide decline reported by Seamans and Rau (2018)

Influence of habitat on abundance
The proportion of clear-cut within a buffer area around a given   & Pendleton, 1996). It has been found that clear-cutting harvests, in strips or in patches, effectively creates woodcock breeding habitat for singing grounds and night-time roosting areas, and eventually for nesting and feeding cover (Kelley, Williamson, & Cooper, 2008;Williamson, 2010). Land management to support woodcock habitat, particularly for providing early-successional forest, also provides conservation benefits to an array of non-target bird species, making the woodcock an effective umbrella species for early-successional forest birds (Masse, Tefft, & McWilliams, 2015).
We expected that increases in urban area would have a negative effect on woodcock abundance, like urban development has had on many other landbird species (Marzluff, 2001). Few studies have considered the availability of urban area as a potential explanatory factor for woodcock abundance, as much of the work on breeding habitat preference has been carried out in areas that are actively managed or protected. Nelson and Andersen (2013) found that the amount of  Urbanisation is often equated to the creation and expansion of major urban centres with high population density or industrial activity, typically combined with the loss of old fields and early successional habitats from development (Marzluff, 2001). Based on examination of urban/developed polygons within SGS route buffers in Nova Scotia, urban areas identified in this study are more typically single or small groups of homes along secondary roads in regions considered more rural, where large yards and/or fields are associated with each urban property. Nova Scotia is one of Canada's most rural provinces, where 43% of the population live in communities with populations of less than 1,000 and < 400 people per km 2 (Gibson, Fitzgibbons, & Nunez, 2015).
Consequently, we interpret that urbanisation and urban areas along SGS routes in Nova Scotia actually represent two different processes.
While the proportion of urban area around a given SGS route was iden-  males and a maximum of 15). Latent Gaussian models are increasingly being used in ecology (Zuur, Ieno, & Saveliev, 2017;Beguin et al., 2012), in part because of their ability to incorporate spatial correlation and reveal potentially unmeasured processes (Pavanato, Mayer, Wedekin, Engel, & Kinas, 2018;Redding et al., 2019;Selwood, Clarke, McGeoch, & Nally, 2017). For example, a similar modelling approach identified a latent process affected fisheries biomass estimates in one particularly productive region of Ontario (Gutowsky et al., 2019). It is plausible that unique circumstances with unmeasured and important habitat variables exist in localised regions of Nova Scotia (e.g. harvest pressure or habitat variables not identified as most important by BRT). Thus, our analysis identifies the influence of measured and unmeasured habitat variables that warrant further investigation.

Regions of growth and decline
Ultimately, our approach has highlighted important regional patterns, but a more fine-scale annual analysis could identify local drivers of population trends. In particular, further efforts should be made to investigate the importance of habitat creation by clear-cutting practices over time in Nova Scotia. This may be possible because forestry data are available at a finer timescale than many other habitat variables and are more closely monitored than changes in other habitat factors.
Future work could look at assessing clear-cut frequency and patch size at a yearly time step where data are available, for comparison of year- to-year and temporal lag effects on abundance. Careful monitoring of woodcock population size and productivity in areas where active management is undertaken is highly recommended.  (Taylor et al., 2020). Today, it is impractical on many landscapes to allow most natural disturbance agents to act unimpeded, and therefore commercial timber harvests and other proactive habitat management at regular intervals is necessary to ensure the availability of required habitat (Dessecker & McAuley, 2001). Nova Scotia forest management is currently entering a new paradigm of 'ecological forestry' gaining popularity across Canada, where emphasis is placed on practises that emulate natural disturbance patterns (Taylor et al., 2020). Replication of natural disturbance needs to be carried out with careful consideration of the various juxtaposed habitat cover requirements of local species-of-concern, for example with woodcock where clearing size and proximity to other habitat type patches is critical. Our findings strongly support the recent recommendation to expand YFI work into Canada (Weber & Cooper, 2019), and the results of our study can provide guidance for where these efforts would be best spent and how cooperation with provincial forest managers could be vital.

CONCLUSIONS
Using a latent Gaussian model, we captured dynamic, regional relationships between habitat and breeding population change for a declining species in eastern North America. Without this approach, it would be challenging to reveal abundance trends across a broad landscape.
Other long-term and spatially extensive landbird surveys would benefit from a similar analytical strategy to guide conservation and management priorities. In 2008, it was estimated that Nova Scotia had lost 22% of singing males (9,049 individuals) since 1970, likely as a result of a decrease in small-diameter size class forests (Kelley et al., 2008). It had been suggested at that time that active management of roughly 730 km 2 of forestland for small-diameter size class is required to increase woodcock populations to historical levels. With continued population declines, this area is likely now higher. However, our spatially explicit analysis has shown that efforts to improve habitat availability should be targeted in the north and south of the province where populations continue to decline. Fortunately, 14.4% of working forest in Nova Scotia is owned and maintained by the forest products industry with another 30.8% as Crown land, which could greatly aid in achieving these habitat goals towards sustaining regional woodcock populations (Lahey, 2018). Realising a positive population status for woodcock within Nova Scotia and beyond to their continent-wide range will require large-scale and long-term planning through more extensive provincial, national, and international partnerships focused on providing habitat for the suite of wildlife species dependent on young forests.

AUTHORS' CONTRIBUTIONS
S.E.G. designed the study and wrote the manuscript, L.F.G.G. conducted the analyses, M.L.M. and G.R.M. facilitated inception and collaboration of the study. All inputted to the manuscript and gave final approval for publication.

PEER REVIEW
The peer review history for this article is available at https://publons.