Forest degradation, not loss, drives widespread avian population declines

1 2 In many regions of the world, forest management has reduced old forest and simplified forest 3 structure and composition via reliance on monoculture tree plantations. We hypothesized that 4 such forest degradation has resulted in long-term habitat loss for forest-associated bird species of 5 eastern Canada (130,017 km 2 ) which, in turn, has affected bird population declines. Back-cast 6 species distribution models revealed that despite little change in overall forest cover, breeding 7 habitat loss occurred for 66% of the 54 most common species from 1985-2020. This habitat loss 8 was strongly associated with population declines for 72% of species, as quantified in an 9 independent, long-term dataset. Since 1985, net forest bird abundance has declined in this region 10 by an estimated 33-104 million birds due to habitat loss alone. The effects of forest degradation 11 may therefore be a primary cause of biodiversity decline in managed forest landscapes.

Most conservation policy has focused on reducing deforestation, and this approach remains 13 fundamental to many conservation strategies. Effects of forest loss on global biodiversity are 14 well-known and directly measured 1 and often used as estimates of biodiversity decline 2 . Forest 15 degradation is also expected to be a key driver of biodiversity decline and is a component of 16 broad-scale biodiversity agreements (e.g., Aichi Biodiversity Targets in the Convention on 17 Biological Diversity, REDD+ [Reducing Emissions from Deforestation and Forest 18 Degradation]). However, forest degradation has been much more challenging to measure and 19 there have been few attempts to quantify its effects on species' population trends across entire 20 regions 3,4 . 21 22 From a biodiversity standpoint, forest degradation is defined as the reduction or loss of 23 biological complexity in forest ecosystems 5,6 . Forest management alters forest complexity most 24 commonly in two important ways; first, due to harvesting, managed forests tend to be younger 25 than under a natural disturbance regime 7 , with potential implications for species associated with 26 mature or old-growth forests 8 . Second, because tree plantations tend to yield more wood per area, 27 managers are increasingly converting native forests to plantations 6 . Unlike most natural forests, 28 plantations tend to be comprised of only one or two tree species. Plantation area is expected to 29 rise as they are increasingly considered "natural climate solutions" 9 . Such changes in age-class 30 structure and forest composition may occur without any overall loss in forest cover, and have 31 thus been largely ignored 4 . Nevertheless, quantifying forest degradation is of critical importance 32 to understanding biodiversity responses in regions where timber harvest and regrowth 33 predominates (e.g., Canada, western US, Scandinavia, Russia) 10 . 34 35 The importance of quantifying forest degradation effects is particularly critical considering 36 recent findings by the Intergovernmental Panel on Biodiversity and Ecosystem Services (IPBES) 37 that the planet is facing a biodiversity crisis 11 . Causes of population declines remain poorly 38 understood for many species. For instance, recent work has quantified widespread avian 39 population declines 12 , but researchers are still searching for the mechanisms driving these losses. 40 Hypotheses explaining avian population dynamics include habitat loss on the wintering 41 grounds 13 , direct effects of toxic chemicals 14 , and climate change effects throughout the annual 42 cycle 15,16 . However, the hypothesis that population declines, for birds or other taxa, are driven by 43 forest degradation and resultant breeding habitat loss, remains largely untested. This lack of 44 robust testing is likely for two methodological reasons. First, changes in forest composition and 45 age-class structure are more challenging to detect than forest loss 17 . Second, it is well known that 46 species have different habitat requirements, which often do not correspond to coarse, human-47 defined land-cover categories 18 .

49
Here, we used species distribution models with Landsat TM reflectance bands as predictor 50 variables 19 to quantify habitat amount for each of the most common 54 forest-dependent bird 51 species in the Acadian Forest of eastern Canada (130,017 km 2 ). Since Landsat has been available 52 since 1985, then enabled us to back-cast habitat model predictions to quantify habitat change for 53 each species over 35 years (1985-2020). Under the hypothesis that changes in forest degradation 54 is driving population declines, we predict that we would see (1) little net change in total forest 55 area (due to the rates of forest regeneration matching forest harvest), (2) reductions in breeding 56 habitat across forest-associated species, particularly those associated with mature native forest 57 which is under pressure from timber harvest, and (3) links between habitat loss and long-term 58 bird population trendsas quantified in an independent dataset (the Breeding Bird Survey 20 ). 59 These links between habitat change and populations are not necessarily a given, and several 60 alternative hypotheses are possible. First, initial habitat may have been underutilized, in which 61 case individuals initially occur at low population density and then pack into remaining habitat as 62 it declines over time 21 . Second, over a 35-year period, bird habitat preferences could relax or 63 shift via behavioral plasticity or strong evolutionary selection pressure exerted by habitat loss 22 .

65
The effects of habitat loss should be particularly severe when habitat amount is low 2324 . 66 Alternatively, some species may exhibit rapid population declines even at low levels of initial 67 habitat loss 1 . Although these alternative hypotheses have been tested using space-for-time 68 studies 25 to our knowledge no studies have tested whether species' populations exhibit threshold 69 behavior during the process of habitat loss over the long-term. We formally tested this 70 'extinction threshold hypothesis' using long-term habitat change predictions, along with 71 independent data on bird population trends. We predicted that populations in landscapes with the 72 lowest amount of habitat at the beginning of our time series should experience the strongest 73 negative effect of further habitat loss.

78
The Acadian Forest of eastern Canada has shown a pervasive signal of forest change since 1985 79 ( Fig. 1), despite a relatively stable trend in total forest cover ( Fig. 2A). Since 1985, >3 million ha 80 have been clearcut (Fig 1A), with most of this area now occupied by either tree plantations ( Fig.   81 1A, B, D) dominated by single tree species 26 or a mix of early successional tree species (Fig. 1B). 82 This pattern of forest harvest followed by rapid regeneration appears to be common across many 83 forest regions of North America (e.g., central Canada, southeastern US, western US; Fig. 1C) 10 , 84 and can be considered forest degradation in that these practices simplify forest structure, reduce 85 tree species diversity, and truncate old forest age classes 7 .

87
Overall, species distribution models (SDMs) using Landsat reflectance bands as predictors 88 performed well for most species when tested on 50% independent hold-out data ( Fig. S1 (Fig. 3). Most species with strongly declining habitat are associated with mature 99 forests 27 (Fig. 3A, B) which is consistent with forest degradation due to harvesting of mature 100 forest. Indeed, clearcut harvest alone was strongly associated with habitat declines for all mature-101 forest-associated species (Fig. S3). Fifteen species exhibited habitat increases, and most (14/15) 102 of these tend to be associated with young or immature forests ( Fig. 4A and 4B).

104
We tested the hypothesis that habitat loss affected bird population declines using Breeding Bird 105 Survey Data (BBS) 28 for the Maritime Provinces (see Methods). First, we used SDMs to quantify 106 habitat change  in landscapes surrounding BBS routes (N=90; see Methods). We 107 then used Bayesian hierarchical models 20 to test whether SDM-predicted habitat loss or gain in 108 each given year of the time series drove population changes for each species along each route. 109 Importantly, BBS data are entirely independent of our SDMs, so this test also represents a strong 110 validation of our habitat models (18). Bayesian models revealed a strong effect of habitat loss or 111 gain on population abundance (Fig. 4). Abundance changes for all but three species tracked 112 annual habitat change with 80% posterior distributions that did not include zero (vertical line in Given this strong association between habitat and population changes for most species, we 117 estimated the net number of breeding individuals lost due to habitat loss from 1985-2020 using 118 published accounts of territory sizes for each species 27 (see Table S1). Across all species, back-119 cast species distribution models (SDMs) indicate that 28,215,247 ha (282,153 km 2 ) of habitat has 120 been lost, equating to a loss of between 16,779,704 and 52,243,938 breeding pairs (33,559,408 -121 104,487,876 individuals; Supporting Methods, Table S1). One might expect that forest 122 degradation, rather than resulting in broad-scale declines across species, is simply causing 123 species turnover from mature forest bird species to young-forest associates. However, it is 124 important to note that we quantified net bird decline from an unbiased list of the 54 most 125 common forest bird species in eastern Canada. This list included both early and late successional 126 species. Such net bird declines could be due to the fact that (1) even some early seral species are 127 losing habitat (perhaps due to conversion from diverse early successional forest to single-species 128 plantations) and (2) in this region, more species occupy older forests than regenerating forests.

130
We also quantified overall population trends for 54 species of forest birds using data from the 131 BBS (Fig. 5). These estimates give the total magnitude of population changes which include, but 132 are not limited to, habitat loss or gain effects. Thirty-nine of the 54 species examined (72%) are 133 in population decline (defined as having 95% credible intervals that do not bound zero). The  For several species, rates of population decline seemed to outpace rates of habitat decline 193 (compare x-and y-axes in Fig. 3A). For instance, Blackburnian Warbler populations have 194 experienced a ~70% decline over 35 years (4.5%/year; Fig. 4B), but only 33% of habitat has 195 been lost. One explanation for this apparent mismatch is that populations show particularly 196 strong declines at low habitat amounts, which supports the 'extinction threshold' hypothesis 23 . 197 Indeed, our results indicate that the effect of habitat loss is much greater in landscapes with low 198 habitat amounts ( Figure S5). The mechanism for such threshold effects could be due to habitat 199 fragmentation 32 ; once patches shrink below the size of an individual territory, it is likely to have 200 higher rates of local extinction. Disconnected patches are also less likely to be colonized 33 . 201 202 However, the mismatch between population versus habitat declines could signal that additional, More subtle mechanisms for habitat loss due to forest degradation reported in this study would 221 likely have remained undetectable without a species-specific habitat modeling approach 19 . Given 222 that no two species associate with identical habitats 18 , our model enabled us to quantify habitat region have resulted in substantial increases in single-species tree plantations (Fig. 1A) 26 1B) and are unlikely to be succeeded by such species given current truncated harvest 246 rotations. We predict that similar effects of forest change could be prevalent in other temperate 247 forests globally that are heavily managed for timber production (e.g., southeastern USA, Pacific 248 Northwest USA, Chile, Scandinavia). These regions show little net loss of forest cover but high 249 rates of forest reductions and regrowth (e.g., Fig.1C) 10 , which is symptomatic of intensive forest 250 management with potential for forest degradation.

252
Overall, our results point to broad-scale declines in forest birds of the Acadian Forest of eastern 253 Canada. For most species we assessed, these declines are driven by habitat loss that is primarily 254 due to forest degradation rather than forest loss. We expect that similar consequences for 255 biodiversity may hold in other intensively managed forests of the world. This mechanism for 256 bird population declines would have been invisible using coarse, human-defined categories of 257 'habitat' (i.e., forest cover) 18 . in the spectral trajectory. The advantage of this approach is that it capitalizes on (1) within-year 301 changes in reflectance (e.g., differential rates of leaf out across tree species), and (2)  To test whether habitat change, measured using back-cast SDMs, predicted population trends we  Fig. 1A Study area in context of other regions of North America that have similar rapid rates of forest loss (pink), then gain (purple)which is likely a signal of commercial forest harvest followed by rapid regeneration. Panel B shows cumulative clearcut disturbance (pink) across the Maritime provinces of eastern Canada from 1985-2020 along with the area that has been converted to plantations (purple). Panel C shows cumulative area clearcut and planted across the study area. Methods for mapping plantations and disturbance are given in the SI. Panel D shows the area of forest that has been clearcut since 1985 (left bar) for public land and private woodlots for a subset of the study area (New Brunswick; 72,908 km 2 ), and forests that have not been clearcut since that date (right bar). Most forest cut since 1985 has been planted, or pre-commercially thinned to favor conifer species (purple bar) or has regenerated as shade-intolerant hardwood (IH) or balsam fir (BF); pink bar).
In contrast, forest that has not been recently clearcut is comprised of shade-tolerant tree species (green bar). Intolerant hardwood/balsam fir stands in areas not recently harvested likely originated from clearcutting before 1985. Data in D were derived from the NB Forest Inventory (2010) so do not include changes over the past decade. Panel E shows native mixed deciduous/conifer forest (left) in relation to older conifer plantation forest (right) that replaced the original mixed forest. Habitat trends (1985-2020) for the seven bird species exhibiting the greatest population declines according species distribution models (SDMs). All of these species are mature-forest associated (see Fig.  4a, 4b). Over the same time interval, total forest cover did not decline (black line), indicating that habitat loss is a function of forest degradation rather than loss. B-F show predicted habitat distributions for three example species in 1985 and 2020 respectively; 25-33% of habitat (shown in blue, non-habitat is in yellow) has been lost over this period across the entire region. Habitat loss quantified using SDMs strongly predicted population trends for forest bird species.

Other Supplementary Materials for this manuscript include the following:
Movie S1 1

2
Breeding Bird Survey Models 3 4 We fit all models in JAGS 1 using the 'rjags' package 2 , in the statistical software R 3 . We used 5 four Markov Chain Monte Carlo (MCMC) chains for each model with random starting values. 6 We optimized MCMC tuning with 1000 iterations and then sampled for 24000 iterations, 7 discarding the first 2000 as a burn-in, and then thinned by discarding every other iteration, 8 leaving a total of 11,000 iterations per chain. We assessed convergence by calculating the 9 Gelman-Rubin diagnostic 4 and examining trace plots of the posterior distributions of every 10 parameter; no parameter diagnostics indicated lack of convergence. 11 12 Regional Trends 13 14 We investigated how bird counts have changed in the study period by modeling each species ' 15 annual count at route i, route-observer combination j, and year t (yijt) as, 16 17 ~ Poisson( ) 18 log( ) = + + first.
year + year where αj are effects of unique route-observer combinations, γt are year effects, η is an effect of an 25 observer being in their first year of conducting surveys (the variable first.year is an indicator (1 26 or 0) of when a route was observed for the first time by a specific observer and a zero otherwise; 27 following Sauer and Link (2011)), βi is the trend for route i (the variable year indicates the year 28 of the survey for each route) and ∈ijt is general dispersion beyond that accounted for by the 29 Poisson variance. The route-specific trends (βi) arise from a distribution with mean µβ and 30 variance σ 2 representing the regional trend for the species, while αj arise from a distribution with 31 mean µα and variance σ 2 , representing the regional average route-observer effect. We used 32 diffuse priors on our parameters as, 33 34 , , ~ (0,10) 35 36 2 , 2 , 2 , 2~ (0.001,0.001) −1 37 38 Population Trend -Habitat Model 39 40 We investigated how bird counts have changed across the study period due to changes in habitat. 41 We followed the same model structures as above, but replace year with the variable habitat as, 42 ~ where 'habitat' is a normalized covariate (centered at zero and scaled to unit variance) of the 50 amount of habitat available for each species in a route and year. The parameter µβ is the average 51 effect of habitat across all routes. Prior distributions are defined similarly as the above models. 52 53 Testing for Habitat Loss Thresholds 54 55 We also investigated how habitat amount in the initial year (1985) mediated the effect of habitat 56 change on bird abundance using the model, 57 58 ~ Poisson( ) 59 log( ) = + + first. year + habitat where 'habitat' and 'init.habitat' are normalized covariates (centered at zero and scaled to unit 67 variance within a species) of the amount of habitat available for each species in a route and year 68 and the initial amount of habitat available to the species in the first year, respectively. The 69 parameter β0 is the effect of habitat change at mean initial habitat, while the parameter ξ is the 70 effect of initial habitat, which modulates the route-level effect of habitat changes (βi). Prior 71 distributions are defined similarly as the above models. All Jags code is available 72 at https://figshare.com/s/72d8da46d2041c984fdb. 73 74 Habitat Associations 75 76 We used Birds of North America (BNA) to establish forest successional associations for each 77 bird species. If BNA accounts indicated that a species was associated with "old" or "mature", or 78 "dead wood" we counted the species as a "mature-forest associate". Alternatively, if the account 79 indicated that the species was associated with "open", "bog", or "shrub" habitat (excluding 80 mature species nesting in shrub understories) we classified the species as a "young-forest 81 associate". We also used an independent bird point count dataset (from 5 to model all available 82 species as a function of both deciduous and coniferous trees >20 cm in diameter. We used 83 logistic regression with a binomial error distribution to model the presence or absence of each 84 species as function of each of these variables. If species showed positive coefficients for either 85 deciduous or coniferous large trees, we deemed these mature-forest associates. If a species' 86 model had a negative coefficient for both of these variables, we deemed it a young-forest 87 associate. However, if confidence intervals overlapped zero, we categorized the species as a 88 forest-age generalist. Only 30/54 species were sufficiently abundant in this dataset to yield model 89 estimates; in all but three cases (Black-and-White Warbler, Swainson's Thrush, Hermit Thrush) 90 these estimates concurred with the BNA accounts. In these cases of disagreement, we used the 91 local quantitative estimates for categorization. Finally, we checked these classifications with 92 existing habitat categories used by the New Brunswick Department of Natural Resources 6 and we 93 found no discrepancies with our final categorization. 94 95 We examined the relationship between habitat change from 1985 -2020 and mature forest 96 association (Fig. 3B) using linear regression with habitat change as the dependent variable. The 97 mature forest association has uncertainty associated with the estimates, so a simple regression 98 using the mean mature forest association estimate would deflate the uncertainty in this 99 relationship. Thus, we conducting this regression in a Bayesian framework, treating the true, 100 unobserved mature forest association as a random variable. We used the mean and standard error 101 of the estimated mature forest association to develop a prior on these values with mean equal to 102 the estimated mean and variance equal to the squared standard error. We fit models in JAGS, 103 running 4 chains each for 20,000 iterations, discarding the first 10,000 iterations as burn-in and 104 assessing convergence using the Gelman-Rubin diagnostic. 105 106 Estimates of Loss/Gain in Bird Numbers 107 108 To estimate the total number of birds lost or gained across the region from 1985-2020, we first 109 calculated net habitat loss or gain for each species. We gleaned territory size of each from the  Fig. S1 Area Under the Receiver Operating Characteristic Curve (AUC)a measure of model prediction success ranging from 0 -1 (perfect predictions) for 54 forest bird species of the Maritime Provinces. AUCs were calculated using 50% of locations held-out as independent test data.

Fig. S2
Habitat change (1985-2020) for 54 species of forest birds according to back-cast species distribution models. Transitions from green, through yellow, to red across cells indicate annual habitat loss. Sixty-six percent of species show net habitat loss over the full time period, and 93% lost habitat over the past 10 years.

Fig. S3.
Relationship between area clearcut occurring in each species' habitat in each year across the study area and habitat loss for each of 23 mature-forest associated species. Each dot represents a year (1985-2020). Clearcut area is the sum of t-1 and t (the year we quantified habitat loss because harvest can occur in winter before, and fall following the breeding season). Blue lines are regression lines and gray bands are 95% confidence intervals. Solid lines indicate the isometric (1:1) relationship between clearcuts and habitat loss. As expected, clearcutting within habitat is strongly associated with habitat loss, which indicates that ingrowth of new habitat has not compensated for by habitat loss (which would have obscured this relationship). Also, amount  clearcut is always greater than habitat loss, indicating that habitat decline is unlikely due to changes in Landsat reflectance bands caused by climatic factors. Species codes are provided in Table S1.

Fig. S4.
Habitat trends within 100 m of BBS routes (red lines) versus the entire Maritimes region (green lines) for 54 species of forest birds. Habitat trends along BBS routes tend to reflect changes in the region except for a few species (e.g., Black-throated Blue Warbler).

Fig. S5
Ten-year population trend estimates for 54 species of forest-associated birds across the Maritime Provinces of Canada. Number below species names indicate the Bayesian posterior probability that the species is declining at a rate >30% over 10 years (9 species fall into this category, 4 of which are mature-forest associated). Note that the effect of habitat loss on bird populations is most severe when habitat amount is low, supporting the habitat threshold hypothesis.

Fig. S7
Habitat distribution and change maps for two examples of mature-forest-associated species within and outside three national parks in eastern Canada (Fundy, Kouchibouguac, Kejimkujik National Parks) and the core area of the study region. Note that habitat loss (red) is common in landscapes surrounding parks, but largely absent within, indicating that the habitat loss we quantified is due to timber harvest, not climate-induced changes in Landsat reflectance, or natural disturbance. White areas indicate non-habitat.   Movie S1.
Animation showing species distribution model using Landsat TM bands for blackburnian warbler (Setophaga fusca) in each year of our study  for the Maritime Provinces of Canada. Habitat for this mature-forest associated species (delineated in blue) declined 33% over the period observed. This habitat change was driven primarily by clearcutting without sufficient habitat regeneration (due to tree species composition changes and age-class truncation), and was a strong predictor of population declines in this species.