Road networks predict human influence on Amazonian bird communities

Abstract

Road building can lead to significant deleterious impacts on biodiversity, varying from direct road-kill mortality and direct habitat loss associated with road construction, to more subtle indirect impacts from edge effects and fragmentation. However, little work has been done to evaluate the specific effects of road networks and biodiversity loss beyond the more generalized effects of habitat loss. Here, we compared forest bird species richness and composition in the municipalities of Santarém and Belterra in Pará state, eastern Brazilian Amazon, with a road network metric called ‘roadless volume (RV)’ at the scale of small hydrological catchments (averaging 3721 ha). We found a significant positive relationship between RV and both forest bird richness and the average number of unique species (species represented by a single record) recorded at each site. Forest bird community composition was also significantly affected by RV. Moreover, there was no significant correlation between RV and forest cover, suggesting that road networks may impact biodiversity independently of changes in forest cover. However, variance partitioning analysis indicated that RV has partially independent and therefore additive effects, suggesting that RV and forest cover are best used in a complementary manner to investigate changes in biodiversity. Road impacts on avian species richness and composition independent of habitat loss may result from road-dependent habitat disturbance and fragmentation effects that are not captured by total percentage habitat cover, such as selective logging, fire, hunting, traffic disturbance, edge effects and road-induced fragmentation.

1. Introduction

Infrastructure developments, particularly roads, are a ubiquitous feature of human-modified landscapes. While roads and roadsides already account for 1–2% of the land surface in some developed countries [1], rates of road expansion are fastest in developing countries where they are afforded a high priority by governments wanting to encourage growth and reduce poverty by increasing spatial connectivity between population centres and access to natural resources [2–5].

Despite the irrefutable socio-economic benefits that roads bring to humans, their proliferation typically leads to negative impacts on the environment and native biota [5–9]. Plant, insect, herptile, bird and mammal richness and community composition have all been shown to be affected by roads [10–14]. Biodiversity loss may occur directly via road-kill events, disturbance or pollution, or indirectly by stimulating and facilitating loss of habitat, and forming barriers to dispersal and gene flow. Roads also affect biodiversity through reduction in habitat quality, facilitating human access to frontier landscapes, increasing opportunities for selective logging and bushmeat hunting, increased risk of forest fires and the creation of edge effects at road-habitat boundaries [6,15,16].

The impacts of roads on birds have been well studied and roads have been found to negatively affect avian site occupancy [17,18], mortality [19], reproductive success [20], movement patterns [21–23] and even vocal activity [24,25]. Avian dispersal capacity, and hence gap-crossing propensity, is highly species/guild specific; while some species routinely undertake non-stop migrations between hemispheres [26], other species of similar body size are physiologically incapable of flying more than 100 m [27].

Quantifying the link between roads and bird diversity is especially important in tropical deforestation frontiers, where forest-dependent birds are particularly vulnerable to human impacts [28–31]. Here, we focus on Amazonia, the world's largest remaining expanse of tropical forest [32], host to the world's most speciose and increasingly endangered avifauna [33]. Road networks in the Brazilian Amazon grew by almost 17 000 km annually between 2004 and 2007 [34]. This growth is underpinned by major government-led development plans in infrastructure, including the construction of major highways such as the BR-10, BR-163, BR-219 and BR-319 that have largely defined regional development patterns. Individual Brazilian states have their own transport development plans, such as PELT-Pará for the state of Pará. Based on predictions of economic growth and transport logistics, it is estimated that over 27 000 km of roads will be constructed, improved (paved), expanded (more lanes), extended and maintained in Pará by the year 2031 (Pelt Pará. http://www.setran.pa.gov.br/PELT/tranporte/arquivos/evol_futura_transp.pdf). We are unaware of any environmental assessment associated with the PELT-Pará plan, raising concern that biodiversity impacts associated with these infrastructural changes will not be appropriately avoided and mitigated against. Given anticipated Amazonian road network expansion, it is of the utmost importance that links between roads and biodiversity be quantified to facilitate accurate assessment of potential impacts.

Roadless volume (RV) [35] is a relatively new metric of road network density that accounts for the exact spatial pattern of roads within an area of interest. However, we are unaware of any validation of the RV metric demonstrating a correlation with changes in biodiversity, hindering its potential use in assessing the impacts of road network changes. Here, we use data on forest bird species richness collected in a deforestation frontier region in southern Amazonia to (i) determine whether there is a relationship between RV and biodiversity, (ii) to quantify the relative importance of roads versus habitat amount (measured as forest cover) in determining biodiversity patterns, and (iii) to estimate the impacts of historical road network expansion on biodiversity in a region where road networks are expanding rapidly.

2. Material and methods

(a). Study region

Our fieldwork was carried out in the municipalities of Santarém and Belterra, Pará state, Brazil (figure 1) under the auspices of the Sustainable Amazon Network [38]. The study region was divided into 286 natural water catchments (mean area 3721 ha, s.d. = 2747), delineated using a digital elevation model (DEM) and Soil and Water Assessment Tool (SWAT) in ArcGIS v. 9.3

Figure 1.

Study region (a) location within the Brazilian Amazon Legal, (b) municipalities of Santarém and Belterra, with 18 study catchments highlighted in black, (c) distance to nearest road calculated on a 30 m grid, with river location and roads in 2008 shown in black. The 18 study catchments are highlighted in white (numbered as in [36,37]). (d) RV calculated over an equal area 60 m grid. (e) RV and forest cover for the 18 study catchments.

(b). Bird sampling

A subset of 18 catchments (figure 1) was chosen to represent a gradient of forest cover from 10 to 100%. A stratified-random sampling design was employed to help ensure a representative assessment of bird species richness in each land cover (along a gradient of environmental impact ranging from undisturbed primary forest to soya bean fields) within the catchments, resulting in between 6 and 12 transects per catchment. A sample of 300 m transects were distributed randomly across the catchments to increase the likelihood that important internal heterogeneities in land cover were captured. A minimum separation distance rule of 1500 m between transects was employed to maximize independence between sampling points (see the electronic supplementary material and [36] for more details). Bird surveys were conducted by A.C.L., N.G.M., Christian B. Andretti, Bradley J. W. Davis & Edson V. Lopes, between 16 October 2010 and 8 February 2011. Two repetitions of three fixed radius (75 m) 15-minute point counts per transect were conducted, located at 0, 150 and 300 m along the transect (see figure 1 for location of study catchments, for further details on the project design, bird survey methodology and links to digital vouchers see Lees et al. [36]). All bird species recorded by sight and sound were recorded, but this study only considers forest-associated bird species (any species occurring within extensive areas of intact forest) assigned based on the classifications of Henriques et al. [39] and Lees et al. [37]. In order to evaluate sample representativeness, estimates of bird species richness (chao1) were calculated using EstimateS software [40] individually for each catchment and for total samples at catchment scale. Moran's I [41] was used to test the richness data for spatial autocorrelation.

(c). Calculating roadless volume

Two satellite images from Landsat location 227/062, covering the period between 2000 and 2008, were manually digitized to generate road network maps [34] following the methods described by Brandao & Souza [42] and were used to calculate RV in ArcGIS v. 9.3. RV represents the amount of space existing between roads, with the ecological ‘value’ of that space weighted by distance to the nearest road, such that areas that are less disturbed by roads have a higher RV value. RV is simple to calculate making it an attractive option for estimating the extent to which roads pervade landscapes at multiple spatial scales. All visible roads on the satellite images were included in the maps and treated equally. We first generated a ‘distance to nearest road’ raster grid, using the ArcGIS Euclidean Distance tool at 30 m resolution (figure 1, inset (c)), then using the Hawth's tools Zonal Statistics tool [43], we calculated the sum of raster cells for each catchment, i.e. the sum distances to nearest road. These values were divided by catchment area to give a standardized metric of RV for each catchment. RV was log(e) transformed for subsequent analyses.

A strong relationship between roads and habitat loss (deforestation) may confound any relationship between roads and species richness, thus we investigated the degree to which RV and percentage forest cover were correlated. Forest cover was derived from Prodes satellite data, from 2001 and 2008 (we used 2001 data as a proxy for 2000, because the 2000 data are not based on as many satellite images as subsequent years and thus are not as robust). Percentage forest cover was log(e) transformed for the Pearson's correlation. The relative importance of RV and habitat cover in relation to species richness was determined using variance partitioning on general linear models with RV and per cent forest cover as explanatory variables.

(d). Comparing roadless volume with bird richness and composition

Linear regression was used to determine the relationship between RV, forest cover and forest species richness. Three models were used: (i) only RV as a predictor of species richness, (ii) only per cent forest cover as a predictor of species richness, and (iii) both RV and per cent forest cover as predictors of species richness. Variance partitioning and ANOVAs were used to compare the models. To examine how RV influences community composition at the catchment scale, the number of log(e) transformed unique species (i.e. species represented by a single record in the sampled data) for each catchment was regressed against RV. We also quantified community composition using detrended correspondence analysis (DCA) scores to represent the composition pattern of bird communities within catchments. We then performed linear regressions of the DCA axis 1 scores to investigate whether the forest bird community changed with changing RV.

We extrapolated the relationship between RV, percentage forest cover and forest species to predict the species richness of forest birds in each of the 286 catchments (including 268 unsampled catchments) in the Santarém region. Catchments were delineated as before, using DEM and SWAT. We used the same ‘distance to nearest road’ surface calculated over a 30 m grid for initial analyses to calculate the RV for all 286 catchments in 2008. A second ‘distance to nearest road’ surface was generated for the earliest digitized road map available (2000) and the RV for each catchment in 2000 was calculated. Based on the relationship established between RV, per cent forest cover and species richness, we estimated species richness in each catchment for both 2000 and 2008, and compared the two to estimate the number of local extinctions over the 8 year period following the road network expansion across the study site. We also estimated the number of local extinctions for the RV and per cent forest cover only models.

To identify catchments that experienced species loss because of the independent effects of RV change, we used a subset of catchments, which experienced low (%) change in forest cover but high (%) change in RVs. Cut-offs of the second quartile (for Forest cover) and third quartile for RV change were used to identify high versus low changes. All detailed analyses were carried out in the statistical platform R v. 2.10.1 [44] unless otherwise stated.

3. Results

A total of 11 028 detections of 384 bird species were recorded during the timed point counts. Of these, 8743 detections of 298 forest-associated bird species were selected for analysis. Chao1 estimates of forest bird species richness were generally higher than observed forest bird richness, with an estimated mean species richness of 152 (s.d. = 39) and an average observed richness of 119 (s.d. = 28) at the catchment scale. Richness estimates suggested sampling had captured an average of 79% of species within each catchment (s.d. = 5.7%). Little spatial correlation was found, with only catchments up to 22 km apart being significantly correlated (Moran's I = 0.35, n = 55, p = 0.013).

RV was positively, but not significantly, correlated with percentage forest cover within the sampled catchments (figure 1, r = 0.45, 95% CI = −0.02–0.76, d.f. = 16, p = 0.063).

Linear regression models showed a significant positive relationship between species richness and RV (figure 2, r² = 0.73, slope = 324, d.f. = 16, p < 0.001), and between species richness and percentage forest cover (r² = 0.57, slope = 19, d.f. = 16, p < 0.001). These two models were not significantly different from each other, although per cent forest cover had a lesser effect on species richness, and both models performed significantly worse than a model including both variables (table 1). The regression model incorporating both variables (RV and percentage forest cover) performed well, explaining 91% of the variation in bird species richness (r² = 0.91, d.f. = 15, p < 0.01). There was, however, no significant interaction between RV and percentage forest cover (F = 53.49, p = 0.84, d.f. = 14). RV alone explained 15% more variance in species richness compared with per cent forest cover alone (figure 1; electronic supplementary material).

Figure 2.

Regression models of RV (first column) and percentage forest cover (second column) against (a) species richness, (b) the number of unique species present at any given catchment and (c) DCA axis 1 scores representing bird community composition at the catchment scale across 18 catchments.

Table 1.

Variance partitioning of the effects of RV and percentage forest cover (%FC) on bird species richness. All models shown are significant with p < 0.001. Shown in brackets is the individual contribution (IC) to variance explained of RV and %FC.

model predictors	r2 (IC)	d.f.	sig. diff. compared to RV + %FC?	sig. diff compared to RV?	AIC	average species loss (s.d.)
RV + %FC	0.91	15	n.a.	yes	133.37	83 (34)
RV	0.72 (0.34)	16	yes	n.a.	152.81	68 (29)
%FC	0.57 (0.19)	16	yes	no	160.46	10 (14)

RV also exerted a significant effect on species composition, reflected in analyses of the overall community composition (DCA score) and the number of unique species per catchment. There was a significant positive relationship between RV and DCA scores (r² = 0.29, slope = −8.4, d.f. = 16, p < 0.05, figure 2). We also found that the number of unique species present in any given catchment increased with increasing RV (r² = 0.32, slope = 5.8, d.f. = 16, p < 0.05, figure 2). Species that do not occur in catchments with log(e) RVs less than approximately 2 ranged from specialized understorey insectivores such as the ant-following bare-eyed antbird Rhegmatorhina gymnops, and the obligate mixed flock-following tawny-crowned greenlet Hylophilus ochraceiceps to large hunting-sensitive species such as razor-billed currasow Pauxi tuberosum. Similar relationships were found between percentage forest cover and DCA (r² = 0.70, slope = −0.8, d.f. = 16, p < 0.001) and between percentage forest cover and unique species (r² = 0.28, slope = 0.4, d.f. = 16, p < 0.05, figure 2).

Between 2000 and 2008, there were 2773 km of new roads constructed in the study area, primarily offshoots of the central BR163 highway, concentrated in the less-developed south of the study region (figures 1 and 3a). However, the majority of forest loss occurred in the north of the study area (figure 3b). Areas of predicted high bird species richness and high RV corresponded to the remaining areas of largely undisturbed primary forest—the Floresta Nacional do Tapajós forest reserve along the western border and areas of unprotected primary forest in the southeast (figure 3c). These areas were predicted to have retained high avian species richness throughout the 8 year period. The model with both RV and forest cover predicted an average forest-associated bird species richness of 177 (s.d. = 31) per catchment in 2000, reduced by an average of 83 (s.d. = 34) species per catchment by the year 2008. Spatial patterns of local extinction matched those of road expansion. The model considering forest cover alone produced significantly lower estimates of species loss compared with RV alone or models containing both predictors (ANOVA F = 592, d.f. = 2,855, p < 0.001).

Figure 3.

(a) Road network, (b) forest cover and (c) predicted species richness in 2000 and 2008 based on RV and forest cover from corresponding years.

Both percentage change in forest (0–80%) and percentage change in RV (0–25%) varied across all 286 catchments (figure 4). Eighty-four catchments experienced forest cover change within the first quartile of values (less than 5.6%) and RV changes higher than the third quartile (more than 10.7%), with an average forest change of 2% (s.d. = 1.5), average RV change of 13% (s.d. = 2.4) and average predicted species loss of 74 (s.d. = 16.5) within this subset of regions. These reductions in species richness represent species loss predominantly determined by change in RV. Interestingly, these catchments form two ‘fronts' (indicated by arrows, figure 4c) bordering extensive areas of forest.

Figure 4.

(a) Per cent RV change, (b) per cent forest cover change and (c) predicted number of species lost between 2000 and 2008, identifying in bold those catchments where there has been low forest cover change but high RV change. As roads are the initial stage in deforestation opening up new areas of land, the presence of road development (a) in areas of low deforestation (b) are areas that are potentially ‘newly’ accessible and likely to undergo deforestation in the future. Such areas may be considered frontiers of development. The 84 catchments that fit these criteria (high RV change and low forest cover change) are clustered around nominally undisturbed forest. Based on where the road network is and where the undisturbed and unprotected forest is, we can assume that deforestation and any future road development will move into the forested areas indicated by arrows, which indicate potential development fronts.

4. Discussion

We found a clear positive relationship between RV and forest-associated bird species richness and composition, demonstrating that road networks can impact biodiversity partly independently, and hence additively to any effects of habitat loss. We estimate that catchments along the deforestation frontier in Pará state have lost an average of 83 species of forest-associated birds between 2000 and 2008. We found that the number of unique species present in a given catchment increased with increasing RV, possibly because species with very low abundances are highly sensitive to road presence and so are unlikely to be detected in catchments with high road densities. Roads are one of the key spatial predictors of deforestation patterns within tropical regions [45–49] facilitating access to frontier areas. While we recognize that our comparison between road networks and forest cover is limited by our small sample size of 18 micro-catchments, it is also possible that the partial independence we observed between RV and forest cover may reflect a time lag between road network development and deforestation. Reductions in RV may therefore act as an ‘early warning metric’ of species loss in forest landscapes, as road construction commonly represents the first sign of anthropogenic disturbance in forest systems detected by means of remote sensing [50]. A great deal of attention has been paid to the effects of deforestation on species richness, which is described well by the metric per cent forest cover [51]. However, habitat degradation is often overlooked, despite being more extensive [52,53]. This is likely related to the fact that deforestation can be monitored remotely, while habitat degradation is much more difficult to monitor [52,54].

By assessing road and forest loss impacts, we have demonstrated that road networks can impact biodiversity both independently and additively to any effects of habitat loss. Road networks may affect bird species richness and community composition in three main ways that are additional to any impacts from associated habitat loss. All three of these processes are linked to habitat degradation, suggesting that RV may be used as a proxy for habitat degradation impacts on biodiversity. Firstly, roads fragment habitat and create barriers to movement between patches. Even very narrow roads can disrupt the movement of many forest-associated bird species, particularly terrestrial insectivorous passerines [21,22,55,56]. The conclusions of these observational studies of bird movements are further supported by experimental and genetic studies of understory bird species which indicate that landscape connectivity may be more important for bird dispersal than distance alone [27,57,58]. By disrupting patterns of movement, roads influence richness and composition by creating unstable meta-populations, isolating sub-populations of species which (if small) may render them vulnerable to stochastically mediated local extinctions [59].

Secondly, road construction precipitates environmental changes [6,7,60], such as the proliferation of habitat edges resulting in altered microclimate, light and foliage levels [61–63]. This may result in a turnover in species composition favouring edge and/or gap species, despite there being only very modest changes in the total amount of forest. Road construction is thus likely to be a significant driver of the loss of sensitive forest-interior species and areas with high RV may even impact populations of more generalist forest species, resulting in local increase in edge and gap species. Laurance [55] and Laurance et al. [21] found that edge and gap-specialist bird species were more prevalent close to road edges than deeper inside the forest, whereas the abundance of specialist insectivorous, terrestrial and solitary species declined closer to roads.

Finally, the presence of roads may be a good proxy for more cryptic forms of disturbance and habitat degradation that depend on, or are exacerbated by, improved human access to forest areas, such as selective timber extraction, fire and hunting [50] which are all known to have significant effects on forest bird communities [31,64–69]. Timber extraction results both in a loss of tree biomass—important for species such as long-tailed woodcreeper Deconychura longicauda, and the associated canopy perforation leads to changes in understory microclimatic conditions and a surge in understorey vegetative growth that renders habitat unsuitable for species such as musician wren Cyphorhinus arada. Both these species were restricted in our landscape to areas with low RV. Hunting almost invariably accompanies timber extraction [70], which may not only be directed at species valued for their bushmeat, e.g. the aforementioned razor-billed currasow, but may also afflict raptorial species which may become targets because of human–wildlife conflicts [71]. We found that raptors such as ornate hawk-eagle Spizaetus ornatus and great black hawk Urubitinga urubitinga were associated with areas with low RV in our study region. Understorey wildfires also open up the canopy and reduce forest biomass [67], but are generally more severe than logging [72] and alter bird community composition for many years after the fire event [73]. Finally, roads cause direct mortality which has been known to affect bird community demography [19].

As a result of the partly independent and additive effects that roads have on birds, RV appears to be an important metric to include when estimating forest bird species richness in human-modified forest landscapes, because it is a more accurate surrogate of habitat degradation events than forest cover alone. Furthermore, models using forest cover alone predicted significantly fewer species losses than the combined model or model with RV alone suggesting that ignoring roads underestimates species loss. However, this is based on total forest cover; it is possible that more stringent metrics of forest condition, such as per cent of primary forest with no detectable disturbance from logging or fire, would yield different results. While RV proved to be a better predictor of species richness, per cent forest cover was a much better predictor of species composition, reiterating the point that assessing changes in species richness and composition is best achieved including both variables.

We found that areas subject to pronounced change in RV but little change in forest cover were concentrated at the edge of development ‘fronts', adjoining undisturbed areas of forest. As such our results indicate that RV may act as an ‘early warning metric’ of species loss, picking up regions that have not yet been deforested, but that are subject to other disturbance events. Further, the presence of roads may indicate that these areas have been subject to habitat degradation despite retaining high levels of forest cover and are likely to have experienced species losses that would not have been detected by assessing forest cover alone. This is very important, as it is easier to establish the presence of a road than it is to determine the level of habitat degradation of the forest from a satellite image, which requires a complex series of analysis to be undertaken successfully [74]. We treat all roads the same regardless of paving or traffic activity, both of which are known to influence the ecological effects of roads [75,76] and future research should explore any differences in road ‘taxonomy’ (e.g. paved versus unpaved, official versus unofficial) on species richness and composition.

Road construction is likely to be a significant driver of the loss of sensitive forest-interior species and areas with high RV may even impact populations of more generalist forest species, resulting in local increase in edge and gap species. However, there may also be some survey bias leading to this pattern, because catchments with low RV had a weak, albeit non-significant, tendency to have more forest cover.

The Brazilian road network is forecast to continue expanding in the foreseeable future [77], requiring as little as 60 years for local road networks to reach maximum observed densities from the first road being laid [34,78]. Considering increasing investment in development plans from regional (e.g. PELT Pará) to continental scales (e.g. Initiative for the Integration of the Regional Infrastructure of South America [77]), it is important that we can forecast potential biodiversity changes resulting from current and future road network improvements. Our observed relationships provide a good basis for large-scale scenario modelling. Such forecasting ability would help road planners to investigate scenarios of ‘ideal’ future road networks that minimize biodiversity impacts. We here demonstrate that forest-associated bird species responses to RV are useful to forecast such changes and may be used as a disturbance proxy to examine the potential impacts of competing road layouts on local biodiversity and improve the design of future road networks.

Funding statement

We would also like to thank Microsoft Research, the Grantham Institute for Climate Change and European Research Council (project number 281986) for funding this work. Additionally, we are grateful to the Rede Amazonia Sustentavel project and the following for financial support; Instituto Nacional de Ciência e Tecnologia—Biodiversidade e Uso da Terra na Amazônia (CNPq 574008/2008–0), EmpresaBrasileira de PesquisaAgropecuária—Embrapa (SEG: 02.08.06.005.00), the UK government Darwin Initiative (17–023), The Nature Conservancy and Natural Environment Research Council (NERC) (NE/F01614X/1 and NE/G000816/1). T.A.G. thanks the Swedish Research Council Formas (grant no. 2013–1571). We also thank the farmers and workers unions of Santarém and Belterra, all collaborating private landowners and the Large-scale Biosphere-Atmosphere Experiment in Amazonia (LBA) for their support.