Summary
Accelerating human impacts are reshaping Earth’s ecosystems. Populations1, richness2–4 and composition4 of communities at sites around the world are being altered over time in complex and heterogeneous ways5–7. Land-use change is thought to be the greatest driver of this population and biodiversity change in terrestrial ecosystems8–10. However, a major knowledge gap is whether land-use change drivers, such as forest loss and habitat conversion, can indeed explain the high heterogeneity of temporal population and biodiversity trends9,11. Here, we fill this gap by analysing change in 6,667 time series of populations (species’ abundance)12 and biodiversity (species richness and turnover in ecological communities)13 over one and a half centuries of forest cover change and habitat transitions. We revealed an acceleration in both increases and decreases in population size, species richness and turnover after peak forest loss at over 2,000 sites across the globe. We found that temporal lags in population and biodiversity change following forest loss can extend up to half of a century and were longer for species with longer generation times such as large mammals. Together, our results demonstrate that historic and contemporary forest cover change do not universally lead to population declines and biodiversity loss, though population declines were most pronounced during and immediately following peak forest loss. By explicitly quantifying multi-decadal temporal lags in population and biodiversity responses to land-use change, our findings inform projections of how life on Earth will be reshaped across the Anthropocene.
Main text
Earth’s biodiversity is changing3–5,14. At sites across the planet, populations are increasing and decreasing1,15,16, species are lost17 and gained18,19, yet synthesis studies across sites indicate no net change in local scale species richness3,4 despite marked shifts in species composition (turnover)2,4,5. At present, we have only a limited understanding of how global change drivers produce these complex population and biodiversity patterns over time6,20,21. Our current knowledge of the mechanisms explaining the ongoing reorganisation of ecological communities predominantly stems from space-for-time8,22 and modelling23,24 approaches that attribute population and richness declines to land-use change. Yet, space-for-time methods can overestimate the effects of global change drivers compared to long-term monitoring, because they do not account for ecological lags6,25,26 and community self-regulation27. Such temporal discrepancies in the magnitude of effects have been observed when studying the impacts of warming on community change25, and of habitat fragmentation on populations and biodiversity28,29. The integration of multi-century long reconstructions of past land cover30 and high-resolution remote-sensing observations31,32 with recent compilations of over five million population and biodiversity records12,13 provides an unprecedented opportunity to test the in-situ impacts of land-use change. Only now can we quantitatively attribute the heterogeneous patterns of population and biodiversity change observed over time to land-use dynamics, thus improving projections of human impacts on the world’s biota.
Here, we asked how populations (trends in numerical abundance) and biodiversity (trends in species richness and community composition) across vertebrate, invertebrate and plant taxa vary according to the timing and magnitude of forest cover change and habitat conversions (Figure 1, Extended Data Figures 1 and 3b). We assessed whether population and biodiversity change were different after versus before contemporary peak forest loss (the timing of the largest forest loss event across the duration of each time series). Additionally, we tested population and biodiversity change versus overall forest cover gain and loss experienced across the duration of each time series. In a post-hoc analysis, we categorized population time series based on whether they were recorded before, during, or after the period of all-time historic peak forest loss (the timing of the largest forest loss event at the location of each time series between the years 850 and 2015), and then compared population trends among the three categories. Finally, we investigated among-taxa variation in temporal lags of population and biodiversity responses to peak forest (the time period between contemporary peak forest loss and maximum change in populations and communities). We used a hierarchical Bayesian modelling framework for all attribution analyses, with individual time series nested within biomes33 to account for the spatial and temporal structure of the data. We used the Living Planet Database (133,092 records) and the BioTIME database (4,970,128 records), currently the two largest databases of population and community time series, respectively. We calculated population change using state-space models that account for observation error and random fluctuations34. We quantified turnover by partitioning Jaccard’s dissimilarity measure into its nestedness (change due to communities becoming reduced subsets of themselves or new species colonising in addition to the original species) and turnover35 components. We focused on turnover because it quantifies compositional changes due to species replacement and is independent of changes in species richness. Our data synthesis quantitatively tests the attribution of change in populations and ecological communities to land-use change through time across the world’s woody biomes.
We predicted greater population and species richness declines with increased forest loss. Forest degradation and land-use conversion reduce habitat and resource availability8,19,36 and are the most common global threats for our studied species37 (Figure 1b, Extended Data Figure 4e). Conversely, we predicted greater increases in population abundance and species richness with larger gains in forest cover. Forest restoration and natural regeneration, two examples of forest cover gain, can lead to positive biodiversity responses38,39. We expected greater turnover of species within ecological communities with greater change in forest cover (both loss and gain), as these extremes of the forest cover change spectrum both create novel environmental conditions prompting local extinctions but also colonisations by new species27,36. Secondly, we predicted that the largest population declines will occur during the periods of all-time peak forest loss across sites as that is the baseline for maximum intensity of forest cover change26. Finally, we predicted that temporal lags will be greater for species with longer generation times, as they typically respond more slowly to environmental change40 and have more limited dispersal41. If we find support for our overall prediction that population and biodiversity loss will be greater with higher forest loss, this would indicate that forest loss is a key driver of rapid and pervasive declines over time. Alternatively, if we find support for heterogeneous and temporally-delayed responses of populations and ecological communities to forest loss, this would imply that the effects of land-use change over time are more complex than previously thought, which has implications for improved prediction of future ecological change and development of biodiversity and conservation policy.
We found that forest loss and habitat transitions did not universally lead to population declines and biodiversity loss (Figures 2-3) and are instead reshaping populations and ecological communities in more complex ways than previously recognised8,15,22,24. Surprisingly, forest loss acted as a catalyst for both positive and negative change and intensified population declines, population increases and species richness losses over time, despite equally long monitoring periods before and after peak forest loss. In 72% of the populations which were in decline before peak forest loss, the declines became more acute after forest loss (slope = −0.04 CI = −0.04 to −0.03, Figure 2a). Similarly, 66% of the increasing populations experienced even more positive population change after peak forest loss (slope = 0.02, CI = 0.02 to 0.03, Figure 2b). In contrast, among time series, we did not find directional relationships among greater forest loss and population and species richness declines (Figure 3, Extended Data Figures 5-7). This disconnect between the magnitude of land-use change and population and biodiversity change could be due to a number of factors including temporal lags in population or community responses26,40 and/or less forest cover change having occurred during the monitoring period relative to historic forest clearing26,42,43 (Extended Data Figures 2-3). In further contrast to our first prediction, larger magnitudes of forest loss often led to greater increases, rather than decreases, in species richness over time, particularly among time series comparisons with shorter durations (Extended Data Figure 6e). Forest loss is a key driver of habitat fragmentation which can lead to rapid colonisation by new species due to increased landscape heterogeneity and larger breadth of ecological niches across sites5,28,36. Our results highlight that the same global change driver, forest loss, affects populations and ecological communities in heterogeneous ways at different sites around the world, and accounting for this heterogeneity is key when scaling from local impacts of human activities to global scale biodiversity patterns and attribution of change10.
Our results revealed that population declines were most pronounced during the period of all time peak forest loss and turnover was highest when primary forests were converted to agricultural and urban areas (Figure 3). Following peak contemporary forest loss, turnover increased by over 10% in 19% of the time series (Figure 2), further testifying to the high rates of compositional change detected across the Anthropocene4,5,14. However, within 22% of the time series, turnover declined by over 10% after forest loss, suggesting that biotic homogenisation might also be occurring following human-induced environmental change44. Taken together, our findings suggest site-specific impacts that were stronger and more common when the population and biodiversity monitoring captured the largest forest loss events and the most dramatic habitat conversion events across time relative to when monitoring is mismatched with forest cover change (Figures 2-3). A greater proportion of the planet is projected to experience an unprecedented amount of land-use change in the coming decades45, highlighting the importance of improved biodiversity monitoring in current and future hotspots of forest loss and habitat conversion.
We found evidence for up to half-century ecological lags in changes in population abundance, species richness and community composition following forest loss (Figure 4). On average, we documented maximum change in populations and biodiversity six to 13 years after contemporary peak forest loss across taxa. Yet, nearly half of population and biodiversity change (40%) occurred within three years of peak forest loss, demonstrating that rapid shifts in populations and ecological communities occur frequently (Figure 4a). As predicted, the period between peak forest loss and peak change in populations and biodiversity was longer for taxa with longer generation times (e.g., large mammals, Figure 4b, Extended Data Table 1), further confirmed by a post-hoc analysis of lags in population change versus mammal generation time (Extended Data Figure 8a). Population declines and increases occurred on similar time scales (Extended Data Figure 8b-c), potentially explaining why previous temporal analyses of population change have not found evidence for net population declines1,20. Losses in species richness lagged behind richness gains only by approximately half a year (slope = 0.5, CI = 0.1 − 1.05), indicating that potential extinction debts and immigration credits accumulated at roughly the same speed across taxa. The similar pace and temporal delay of richness gains and losses could be the source of the previously observed findings of no net local richness change3,4, yet substantial compositional change2,4 across sites and taxa. Such temporal lags in biodiversity change have also been observed in post-agricultural forests2,46 and fragmented grasslands40, where agricultural activity has ceased decades to centuries ago, yet richness and community composition change continue to the modern-day. Overall, our results indicate that increasing rates of land-use change in the Anthropocene11,45 will alter ecosystems on both short- and long-term timescales.
In summary, our analysis reveals an acceleration of increases and decreases of populations and biodiversity after forest loss and habitat conversion at sites around the planet. Our findings that all is not loss contrast with our hypothesis and challenge the widely-held assumption that land-use change universally leads to population declines and species richness loss8,15,23. Nevertheless, the increased magnitude and likelihood of population declines during and following peak forest loss highlight that human impacts are altering the biodiversity of the planet and emphasize the importance of expanded biodiversity monitoring in current and future hotspots of land-use change. A critical assumption underlying existing projections of biodiversity responses to land-use change8,23 is that space-for-time approaches accurately reflect longer-term population and biodiversity dynamics11. On the contrary, we find that temporal lags in population and biodiversity change following forest loss varied by taxa and generation time and extended up to half of a century. Over the Anthropocene, ecosystems could be responding to a suite of global change drivers, in addition to land-use change, and a key next research step is to test the synergy and discord between the effects of multiple anthropogenic threats on Earth’s biota. Our results highlight the complex biological responses to habitat conversion across sites, taxa and time scales that are leading to the reorganisation of ecological communities. Thus, indicators used to assess biodiversity change regionally and globally10, including progress towards Aichi targets47, must capture the full spectrum and temporal spread of population and biodiversity responses to human impacts across the Anthropocene.
Author contributions
G.N.D., M.A.D. and I.M.S. conceptualised the study. G.N.D. integrated databases and conducted statistical analyses with input from S.B., I.M.S., A.D.B. and M.A.D. G.N.D. created the figures with input from co-authors. S.B. and S.S. wrote the code for the rarefaction of the BioTIME studies. I.M.S. was the primary supervisor and M.A.D. the co-supervisor for G.N.D. M.A.D. and A.M. funded the compilation of the BioTime database. G.N.D. wrote the first draft and all authors contributed to revisions.
Methods
For an illustration of the workflow of our analyses of forest cover and population and biodiversity change through time, see Extended Data Figures 1 and 3b. All data and statistical analyses are described in detail below. We did not predetermine sample size and instead worked with all available temporal population, biodiversity and forest cover change data that met our duration criteria. For analyses of population change, we included time series with five or more survey points. For analyses of biodiversity change, we included time series with five or more data points when analysing the full time series, and time series with two or more data points when matching the duration of time series comparisons to the 16-year duration of the Global Forest Change Database from 2000 to 2016).
Databases
Forest cover change data
To quantify historic and contemporary forest cover change, we extracted historic forest loss from the Land Use Harmonisation (LUH; 850 – 2015, forest loss and habitat transitions at a 0.25° degree resolution)30 and contemporary forest cover change and habitat conversions from the Global Forest Change (GFC, 2000 – 2016, forest loss and gain at a 30 m resolution)31, and MODIS Landcover (2000 – 2013, land-use transitions at 500m resolution)32 datasets.
Land Use Harmonisation Database
To estimate forest cover change across a time period matching the full duration of the biodiversity observations, we derived the change in primary forest cover from the Land Use Harmonisation database (LUH)30 for 96 km2 cells around the location of each population in the LPD database and for the standardised grid cells of the BioTIME database (~ 96 km2 each). LUH includes annual gridded fractions of land-use states for the period from 850 to 2013 at 0.25° x 0.25° resolution. The estimates are based on historical reconstructions using Earth System models, with inputs such as regional and national rates of wood harvest and potential biomass density. The accuracy and precision of LUH increases towards the modern day, when there are more available data to inform the Earth System models. Note that unlike GFC, LUH estimates forest cover as a proportion (bounded between zero and one). For our analyses, we focused on time series from locations that have experienced at least 0.05 (equivalent to 5%) forest loss. To calculate total forest cover change over the period of a given population or biodiversity time series, we subtracted the proportion of forest cover in the first year of biodiversity monitoring from the proportion of forest cover in the last year. The type of forest cover change detected by the LUH database was predominantly forest loss, with forest gain occurring infrequently and at very small magnitudes (<0.001 out of maximum 1), thus we focus our analysis on forest loss.
To estimate the historic baseline of forest cover change, we calculated yearly change in % forest cover in a study cell from one year to the next for each site from 850 to 2015 from the LUH data, and determined the 10-year period when the most forest loss occurred (historic peak forest loss, calculated by adding the yearly proportions of forest loss in each cell over standardised 10-year blocks). Time since historic peak forest loss was a poor predictor of the variation in contemporary population and biodiversity change (Extended Data Figure 10e-f). To determine contemporary peak forest loss for each time series of monitoring data, we calculated yearly changes in forest cover across the duration of each time series and determined the year when the most change had occurred.
Global Forest Change Database
We derived overall forest loss and forest gain across the 2000-2016 period for 96 km2 cells around the location of each population in the LPD database and for the standardised grid cells of the BioTIME database (~ 96 km2 each) from the Global Forest Change (GFC)31 database using the Google Earth Engine48. The GFC database provides high resolution forest cover change data, derived from Landsat satellite observations at a 30-meter spatial resolution. We calculated the total area of forest cover gain and loss separately (measured in km2) for each 96 km2 cell on a yearly time step. We then summed the yearly values for the period that coincided with population and biodiversity monitoring to estimate overall forest cover gain and loss (two separate metrics). For example, for a biodiversity time series spanning 2002 – 2009, our forest cover gain and loss metrics included the total amount of forest cover gained and lost during that same period. For our analyses, we focused on time series from locations that have experienced at least 0.5 km2 of forest gain or loss. GFC does not distinguish between primary forest, secondary forest and plantations, but it does provide a very high-resolution measure of general forest cover. The drivers of the forest loss detected by GFC across our study sites are predominantly forestry, changes in agricultural practices and wildfires49. Note that the GFC database spans from 2000 to 2016, whereas the earliest terrestrial biodiversity record in BioTIME is from 1858.
MODIS Landcover Database
We used the MODIS Landcover Database32 to quantify habitat conversion for locations where we had population and biodiversity monitoring data. The MODIS Database has a resolution of 500 m, and it uses satellite-derived reflectance data to classify land cover around the world. To determine the types of habitat conversion between 2000 and 2013 (the time span of available MODIS data) across all monitoring locations, we calculated the dominant land cover type at the start and end of each population and biodiversity time series and split time series into categories such as “no habitat conversion” and “grassland to woody savannah”. We focused on the eight most frequent types of habitat conversion (Extended Data Figure 7).
By synthesising information from scenario data based on Earth Dynamics Models (LUH) and remote-sensing databases (GFC, MODIS), we were able to determine historic forest loss from the start of the monitoring period to 2015, as well as contemporary forest cover change (gain and loss) and habitat transitions from 2000 to 2016. GFC and MODIS detect forest cover, with no distinction between primary and secondary forests, thus we derived information on transitions from primary to secondary forest from the LUH database. We calculated overall forest cover change because we considered total change in habitat to be more meaningful for long-term population and biodiversity trends as opposed to an annual rate of forest cover change which does not capture cumulative effects. Together, the three databases (GFC, MODIS, LUH) encompass two different elements of land-use change: 1) land cover types and long-term historical reconstructions of past land-use and habitat conversions and 2) high-resolution satellite data from recent years of forest cover change and habitat conversion types. Thus, the combined analysis allows for a comprehensive test of the effects of land-use change on populations and biodiversity around the world.
Population time series data (Living Planet Database)
We analysed 4,228 population time series, with records distributed around the world. Geographic representation is variable with, for example, an under-representation of tropical regions in the population data (Figure 1). In the LPD, some populations have precise coordinates, whereas the location of others are approximate. Because of the extent over which we are calculating forest cover change (96 km2), we included both types of populations in our analysis. Duration varied across time series (Extended Data Figure 4c-d) and we only included populations with at least five survey points. The overall range of the time series covered the period between the years 1970 and 2014. We calculated population change using state-space models which are particularly appropriate when quantifying change in data with varying collection methodology, as they take into account observation error and process noise50,51. For more details on state-space model calculations, see Humbert et al. 200934 and Daskalova et al. 20181. We scaled the population size data to be between 0 and 1 to analyse within-population relationships and to make sure that we were not conflating within-population relationships and between-population relationships52. State-space models partition the variance in abundance estimates into process error (σ2) and observation or measurement error (τ2) and estimate population trends (μ): where Xt and Xt−1 are the scaled (observed) abundance estimates (between 0 and 1) in the present and past year, with process noise represented by εt ~ gaussian(0, σ2). We included measurement error following: where Yt is the estimate of the true (unobserved) population abundance with measurement error:
We substituted the estimate of population abundance (Yt) into equation 1:
Given Xt−1 = Yt−1 − Ft−1, then:
For each time series, we calculated overall population change (μ) experienced 1) across the periods before and after contemporary peak forest loss, 2) across the full duration of the time series, 3) from 2000 to 2016 (matching the temporal scale of the GFC database), and 4) from 2000 to 2013 (matching the temporal scale of the MODIS database). We standardised the number of years over which we calculated population change before and after peak forest loss on the population-level, meaning that the number of years before and after was the same within populations, but might differ among populations.
Biodiversity time series data (BioTIME Database)
We analysed 2,339 time series from 190 studies from terrestrial biomes across the globe, part of the BioTIME database13 (with the addition of 36 studies that are not yet a part of the public database). Similarly to the LPD, tropical regions and some taxa such as amphibians and reptiles were under-represented. Some of the study locations fall within protected areas (32%). Because those studies only had one time series each, overall only 1% of analysed time series were from inside protected areas. To account for the different spatial extents of the BioTIME database and uneven sampling, studies with multiple locations and extents > 72 km2 were partitioned into 96 km2 grids, and then sample-based rarefaction was applied to standardise sampling within each time series14. Duration varied across time series (Extended Data Figure 4c-d) and the overall range of the time series covered the period between the years 1858 and 2016. For time series with five or more years of monitoring records, we calculated overall richness change and turnover experienced 1) across the periods before and after contemporary peak forest loss, 2) across the full duration of the time series. For time series with two or more years of monitoring records, we calculated overall richness change and turnover experienced 3) from 2000 to 2016 (matching the temporal scale of the GFC database), and 4) from 2000 to 2013 (matching the temporal scale of the MODIS database). The GFC and MODIS databases cover shorter time periods, thus we included biodiversity time series with shorter durations than the five-year cut off point that was used in the rest of our analyses using datasets with longer durations (but note that 76% of biodiversity time series had a duration of three or more years). To estimate richness change, we modelled species richness versus time (year, mean centered) with random slopes and intercepts for each rarefied cell and a Poisson error distribution with a log link. where yeari,j,t is the time in years, β0 and β1 are the global intercept and slope (fixed effects), β0j and β1j are the biome-level departures from β0 and β1 (respectively; biome-level random effects), β0j,i and β1j,i are the (nested) cell-level departures from β0 and β1 (cell-level random effects); yj,i,t is the (rarefied) species richness within the jth biome in the ith cell in year t.
From the richness over time model, we extracted the posterior means for richness change for each time series (i.e., the cell-level slope estimates), which then became the response variable in the second stage of our analyses where we tested richness change versus forest cover change (see Statistical analyses section).
To determine changes in community composition, we calculated the turnover component of beta diversity (changes due to species replacement rather than changes in species abundances14,35), at the end of each time period outlined above relative to the first year of observation in the same period. Turnover is bound between zero and one, where zero is no change in species composition and one indicates that all of the original species of a community have been replaced with new species.
Statistical analyses
When testing for an attribution signal (i.e., evidence that a predictor variable is a potential driver of population or biodiversity change), we always matched the temporal scales of the forest cover change data and the population and biodiversity data. For example, when testing the effects of forest cover change and land-use transitions as detected by GFC (2000 – 2016) and MODIS (2000 – 2013), we calculated population and biodiversity change for the same time periods. Because of the longer duration of the LUH database, we were also able to extract forest and land cover information for the full duration of the LPD and BioTIME time series. For our analyses of contemporary peak forest loss and overall forest loss (using the LUH database over a time period matching the duration of each time series), we excluded locations which had less than 0. 05 (out of maximum 1) forest cover change. We excluded locations which had no forest cover across the duration of the time series in both the 96 km2 cells and the 500 km2 larger landscape cells from our analyses of population and biodiversity change versus forest cover gain and loss from 2000 to 2016 (using the GFC database). See Extended Data Table 1 for the outputs of all statistical models and their respective sample sizes.
Population and biodiversity change after versus before contemporary peak forest loss
To test if temporal population and biodiversity change differed before and after peak forest loss on the site-level, we split each time series into two periods – before and after peak deforestation – and estimated population change, richness change and turnover for each period separately. Then, to infer if population and biodiversity change differed following peak forest loss, we modelled μ (population change), richness change (cell-level random slopes) and turnover as a function of period (categorical with two levels – before or after forest loss) and time series duration (numeric) as fixed effects, with a biome random effect to account for the spatial clustering of the data. For population and richness change, we modelled the positive and negative components of the distributions of change separately, e.g., one model for populations with positive μ values and one model for populations with negative μ values. This approach allowed us to test if the effects of forest loss differ across the positive and negative dimensions of population and biodiversity change. The models were as follows: where durationj,i,p is the duration of the time series in years of cell i within biome j for period p, and periodj,i,p is an indicator variable for the period (before or after forest loss); β0, β1 and β2 are the global intercept and slope estimates for duration and the categorical period effect, respectively (fixed effects), β0j is the biome-level departures from β0 (biome-level random effects); yj,i,p is the estimate for change in population size or species richness for the ith cell in the jth biome for the pth period.
To model the change in turnover before and after contemporary peak forest loss, we followed the same conceptual framework as outlined above, but we used a zero one inflated beta distribution to account for the properties of turnover (bounded between zero and one, inclusive, where one is a complete change in species composition). The probability density function for the zero one inflated beta distribution is: where α is the probability that a zero or one occurs, γ is the probability that a one occurs (given an observation is a zero or a one), and μ and ϕ are the mean and precision of the beta distribution, respectively. In the parameterisation approach we used53 ϕ is inversely related to the variance. Beta parameterisation is also sometimes expressed through the parameters p and q that can be derived from our framework following ϕ = p + q54. Because only 7% of time series did not experience any change in species composition (y = 0) in the time period after contemporary forest loss, and less than 1% of time series had a completely new set of species (y = 1) occupying the ecological communities, for y = 0 and y = 1, α and γ were modelled assuming a Bernoulli distribution and logit-link function, and models were fit with only an intercept. For 0 < y < 1, we assumed a beta error distribution and a logit-link function: where durationj,i,p is the duration of the time series in years of cell i within biome j for period p, and periodj,i,p is an indicator variable for the period (before or after forest loss); β0, β1 and β2 are the global intercept and slope estimates for duration and the categorical period variable respectively (fixed effects), and β0j are the biome-level departures from β0 (biome-level random intercepts); yj,i,p is the estimate of turnover for the ith cell in the jth biome for the pth period.
Population change before, after and during the period of all-time historic peak forest loss
To determine if population change differed based on whether population time series were recorded before, during, or after the period of all-time historic peak forest loss (the timing of the largest forest loss event at the location of each time series between the years 850 and 2015), we modelled μ (population change) as a function of when monitoring started (categorical with three levels – before, during or after peak forest loss) and time series duration (numeric) as fixed effects, with a biome random effect to account for the spatial clustering of the data. Low sample size precluded a similar analysis for biodiversity change (Extended Data Figure 3). The model was as follows: where durationj,i,m is the duration of the time series in years of cell i within biome j for monitoring start m, and monitoring startj,i,m is an indicator variable denoting when monitioring started; β0, β1 and β2 are the global intercept and slope estimates for duration and the categorical monitoring start variable respectively (fixed effects), β0j is the biome-level departures from β0 (respectively; biome-level random effects); yj,i,m is the estimate for change in population size or species richness for the ith cell in the jth biome for the mth monitoring start.
Habitat conversion and population and biodiversity change
To determine the influence of the type of forest cover change (i.e., land-use transitions) on population and biodiversity change, we compared the distributions of population and biodiversity change across transitions types (from primary forest to secondary forest, from primary forest to non-natural habitat, and from secondary forest to non-natural habitat, to which we refer as habitat conversion). Small sample sizes (on average 10 time series per transition type) precluded statistical analysis, thus we report findings from a visual inspection of distributions of population and biodiversity change across habitat conversion types.
To test the effect of forest cover change on population and biodiversity change among sites, we modelled population and biodiversity change versus overall forest cover change (calculated as forest cover gain and forest cover loss (GFC database, 2000-2016) and forest loss (LUH database, across the duration of the time series). Models of population and richness change versus forest cover change were fitted assuming Gaussian error. where durationj,i is the duration of the time series in years of cell i within biome j, forest changej,i is the forest cover change in cell i within biome j; β0, β1 and β2 are the global intercept and slope estimates for duration and forest cover change respectively (fixed effects), and β0j are the biome-level departures from β0 (biome-level random intercepts); yj,i is the population or richness change metric (a separate model for population declines, population increases, richness losses and richness gains) in the ith cell within the jth biome.
Models of turnover versus forest cover change were fit with a zero one inflated beta distribution to account for the properties of turnover (bounded between zero and one). We used the same probability density function for the zero one inflated beta distribution as in the model for turnover before and after contemporary peak forest loss. For y = 0 and y = 1, α and γ were modelled assuming a Bernoulli distribution and logit-link function, and we fit models with only an intercept. For 0 < y < 1, we assumed a beta error distribution and a logit-link function: where durationj,i is the duration of the time series in years of cell i within biome j, forest changej,i is the forest cover change in cell i within biome j; β0, β1 and β2 are the global intercept and slope estimates for duration and forest cover change respectively (fixed effects), and β0j are the biome-level departures from β0 (biome-level random intercepts); yj,i is turnover in the ith cell within the jth biome.
Lags in population and biodiversity responses to contemporary peak forest loss
To test for temporal lags in population and biodiversity responses to contemporary peak forest loss, we first calculated when population and biodiversity change were greatest following peak forest loss for each time series. Rates of population change were calculated using state-space models and a Kalman filter20,34. Peak richness change and peak turnover were calculated as the maximum value of the absolute differences between consecutive observations of species richness and turnover. We then quantified lag as the number of years between contemporary peak forest loss and peak population/biodiversity change. We modelled lag as a function of taxa, as we expect that species with longer generation times will respond to disturbance more slowly. where taxaj,i is the taxa of the cell i in the biome j time series, β1 is the slope for taxa effect (fixed effect), and β0j are the biome-level random intercepts; yj,i is the temporal lag in the population or biodiversity change metric (a separate model for population change, richness change and turnover) for the ith cell within the jth biome.
We conducted a post-hoc analysis where we tested our temporal lag and generation time hypothesis in a more quantitative manner by modelling lag as a function of generation time in mammals, the taxa for which generation time data were freely available55. where generation timeg is the mammal generation time in years, β0 and β1 are the global intercept and slope (fixed effect); yg is the temporal lag in population change for a species with generation time g.
Prior specification
For all models except the model of turnover versus overall forest cover change (which was a zero one inflated model), we used weakly regularising normally-distributed priors for the global intercept and slope:
For the turnover models that had a zero one inflated beta distribution, we used the following priors: where zoi is the probability of being a zero or a one and coi is the conditional probability of being a one (given an observation is a zero or a one).
Group-level parameters (the rarefied cell random effect in the species richness over time model, i, and the biome random effect in all models, j) were all assumed to be gaussian(0, σ), and priors on the σ were the same for all models:
All models were fitted in a Bayesian framework using the brms package v2.1.053 in R v3.5.156. Models were run for 6000 iterations, with a warm up of 2000 iterations. Convergence was assessed visually by examining trace plots and using Rhat values (the ratio of the effective sample size to the overall number of iterations, with values close to one indicating convergence).
Sensitivity analyses
Our analyses were not sensitive to our calculation of turnover in the final year of the time series relative to the first year, and previous examinations of the BioTIME database have found that calculating turnover relative to the second year of observation produced similar results4. We also quantified population change using the BioTIME database (following the same state-space modelling framework as with the LPD) and found similar lack of directional patterns in the relationships between population change and overall forest loss (Extended Data Figure 5f). We found no distinct geographic or taxonomic patterning in the relationships between population change, biodiversity change and forest cover change (Extended Data Figure 9). Furthermore, the relationships between population decreases and increases and forest loss were not influenced by whether species were tightly associated with forests or not (Extended Data Figure 5g-i). Similar post-hoc analysis was not possible for the biodiversity time series because habitat preference data were not available for many of the species included in the BioTIME database. The cell size over which we calculated forest cover change (from 10 km2 to 500 km2) did not influence overall findings, as detected forest cover change scaled proportionately with cell size across locations (Extended Data Figure 10a-b). Landscape context (forest cover in a 500 km2 cell around sites) also did not influence the relationship between forest cover change and population and biodiversity change (Extended Data Figure 10c-d). We did not find directional patterns between population and biodiversity change and time since the largest forest loss event (Extended Data Figure 10f-h). Our findings were not influenced by the type of forest cover (primary vs secondary), as loss of secondary forest cover scaled proportionately to primary forest loss (Extended Data Figure 10e).
Data and code availability
Code for the rarefaction of the BioTIME Database is available from https://doi.org/10.5281/zenodo.1475218. Code for statistical analyses is available from http://doi.org/10.5281/zenodo.1490144. Population and biodiversity data are freely available in the Living Planet and BioTIME Databases12,13. The Living Planet Database can be accessed on http://www.livingplanetindex.org/data_portal. The BioTIME Database can be accessed on Zenodo (https://doi.org/10.5281/zenodo.1211105) or through the BioTIME website (http://biotime.st-andrews.ac.uk/). Land-use change data are publicly available in the Land Use Harmonization Database30, the Forest Cover Change Database31, and the MODIS Landcover Database32.
Acknowledgements
We thank the WWF and ZSL for compiling the Living Planet Database, the BioTime team for compiling the BioTime database (which was supported by ERC AdG BioTIME 250189 and ERC PoC BioCHANGE 727440), the creators of the Land Use Harmonization Database, The Hansen Lab for producing the Forest Cover Change Database and NASA for producing the MODIS Landcover Database. We thank the Forest & Nature Lab at Ghent University for a stimulating discussion on historic and contemporary land-use change and choosing appropriate baselines for comparison of biodiversity change through time. We are grateful to Albert Phillimore and Kyle Dexter for providing advice during the conceptualization of the study, and to Laura Antão and Mark Vellend for providing feedback on the draft manuscript. We thank the German Centre for Integrative Biodiversity Research (iDiv) Halle-Jena-Leipzig and the sChange working group for supporting the initial data synthesis work that has led to this study. G.N.D. was funded by a Carnegie-Caledonian PhD Scholarship and supported by a NERC doctoral training partnership grant (NE/L002558/1).