Taxonomic and functional reorganization of terrestrial mammal communities under anthropogenic influence in a tropical forest

Abstract Protected areas and national legislation are important tools for conserving biodiversity and maintaining ecosystem services. In Brazil, the 2012 Native Vegetation Protection Law requires rural landowners in the Amazon to maintain at least 80% of each property under native vegetation as Legal Reserves (LRs), including within settlement projects. We compared mammal community structure and trait-based responses between human-occupied LRs and a continuous protected forest (CPF) in the central Amazon. Using camera-traps, we assessed species richness, relative abundance, taxonomic composition, functional traits, and ecological attributes in relation to anthropogenic variables. RLQ and fourth-corner analyses revealed similar species richness, but different taxonomic composition and relative abundance patterns between protection types. Herbivores and subsistence-hunted species were associated with the CPF, whereas carnivores, retaliation-hunted species, and mammals with large home ranges were linked to LRs. Forest cover was the only anthropogenic variable significantly associated with relative abundance in the CPF, while mammal communities within LRs exhibited compositional reorganization under multiple anthropogenic pressures. These findings indicate that LRs, despite being embedded in human-altered landscapes, can support diverse mammal assemblages and play a complementary role in biodiversity conservation. Methods (a) Study areas We compared mammal communities in Legal Reserves (LRs) within rural settlement projects to those in a continuous protected forest (CPF) in the central Brazilian Amazon. The LRs were located in the Tarumã-Mirim, Água Branca, and Iporá settlements, whereas the CPF was located in the Alto Cuieiras Reserve. Settlements are anthropogenic landscapes established for agricultural production , characterized by high deforestation rates and dense unofficial roads networks that contribute to forest loss and fragmentation . Under the Brazilian Native Vegetation Protection Law (NVPL), LRs within settlements must maintain 80% native vegetation . The Alto Cuieiras Reserve is a 22,700-ha private reserve located approximately 60 km north of Manaus and managed by the National Institute for Amazonian Research (INPA). It forms part of the Lower Rio Negro Mosaic of protected areas and is bordered by the BR-174 highway, the ZF2 unpaved road, and the Cuieiras River. The regional climate is classified as Af (Equatorial rainforest, fully humid) according to the Köppen system, with annual mean temperature of 26.4 °C and rainfall of approximately 3,000 mm. The dry season extends from June to November and the wet season from December to May. (b) Data collection We established 15 camera-trap stations in each of the three settlements (45 sampling points in LRs), and 40 points in the CPF. Sampling points were located in terra firme forest and spaced at least 1 km apart to ensure spatial independence. Cameras were not placed on roads or trails to reduce detection bias. To ensure seasonal comparability, surveys were conducted during the transition from the rainy to the dry season. Água Branca was sampled from December 2014 to April 2015, whereas Tarumã-Mirim, Iporá, and the Alto Cuieiras Reserve were sampled from March to June 2019. For each camera-trap station, we selected a standardized 31-day sampling window within this seasonal transition. Although some cameras remained active longer because of logistical constraints (Água Branca: 118 ± 1.4 days, CPF: 68 ± 1 days), only the standardized subset was retained for analysis. Sampling coincided with the end of the fruiting season in Central Amazon lowland forests, a period that influences mammal movements, foraging behavior, and detection probabilities. (c) Camera-trap sampling In the LRs, we used Reconyx (Holmen, WI, USA) HyperFire HC600 camera-traps, whereas UltraFire XR6 cameras were used in the CPF. All cameras were motion-activated, unbaited, and operated continuously for 24 h per day. In the CPF, cameras recorded both photos and videos; in LRs, photos only. Cameras were installed on trees 20–30 cm above ground, generally in locations with signs of mammal activity following standard protocols. To reduce false triggers, we cleared vegetation from a 5 × 5 m area in front of each camera. Data were processed using Timelapse2. Independent events were defined as records of the same species separated by at least 30-minute at the same station, whereas records of different species were always considered independent. For gregarious species, all individuals recorded within the same event were counted as independent records. Only independent records were used in the analysis. Species identification followed field guides for Neotropical and Brazilian Amazonian mammals. (d) Mammal functional traits and species-level ecological attribute We evaluated five variables describing mammalian ecological strategies and responses to disturbance: body mass (BM), home range size, population density, hunting level (HL), and trophic level (TL). BM, home range size, HL, and TL were treated as intrinsic functional traits, whereas population density was considered a species-level ecological attribute reflecting interspecific differences in typical relative abundances. BM, home range size, and population density were treated as continuous variables, whereas HL and TL were categorical. We selected these traits and attributes for their direct link to species vulnerability and ecological function. Population density was included to capture interspecific differences in typical abundance and potential ecological influence, using species-level estimates from published databases rather than to estimate local population sizes. Moreover, BM, population density and TL are related to dietary habits and the amount of food that the species or group conspecifics can consume. Home range size reflects the spatial scale of resource acquisition. HL captures the type and intensity of threats, if any, that the species may encounter. We defined HL as a species’ intrinsic vulnerability to hunting. Our classification reflects a literature-based consensus on susceptibility, derived from biological traits like body size and behavior. HL is categorized into three classes: ‘consumed’ (hunted for human consumption as food), ‘retaliation’ (hunted due to perceived threats to livestock or human safety), and ‘not hunted’ (not targeted for either purpose). TL is categorized into four dietary groups: herbivore, insectivore, omnivore, and carnivore. (e) Anthropogenic variables We used four predictor variables: distance to roads, water, and houses - proxies of human accessibility and disturbance - and percentage of forest cover within a 500 m buffer around each camera-trap location (hereafter “forest cover”), representing habitat availability. Although distance to water and forest cover are natural variables, we classified them as anthropogenic proxies in our study context. Human features like ports and dams are commonly associated with water sources across Amazonian landscapes, potentially altering their ecological function by increasing perceived predation risk and triggering behavioral avoidance under the ‘Landscape of Fear’ concept. Thus, distance to water represents a gradient of human exposure and risk, not only resource availability. Similarly, forest cover reflects not only habitat quality but also refuge from the human footprint. All spatial analyses were conducted in QGIS v3.10. We calculated Euclidean distances from each camera-trap station to roads, waterways, and houses. Road shapefiles were obtained from the Instituto do Homem e Meio Ambiente da Amazônia, and waterway shapefiles from the Brazilian Institute of Geography and Statistics. House locations were georeferenced during fieldwork using a handheld GPS (Garmin Ltd., Olathe, USA); when unavailable, we digitized inhabited houses from satellite imagery corresponding to the sampling year. Forest cover was derived from MapBiomas Collection 5.0 using Google Earth Engine), calculated within a 500 m buffer around each camera-trap station to reduce spatial autocorrelation. The study areas represented a gradient of anthropogenic influence, with the CPF farther from roads with slightly higher forest cover than the LRs, whereas LRs were generally located closer to water sources (table 1). This configuration ensured a clear environmental gradient in human accessibility and habitat availability, which was used to test associations with mammal community structure. (f) Data analysis A detailed description of all statistical analyses is available in the Electronic Supplementary Material (electronic supplementary material, Methods S1). All statistical analyses were conducted in R. We calculated relative abundance using two approaches based on independent camera-trap records: [1] species-level relative abundance per sampling point for community composition analyses, and [2] community-level relative abundance per protection type (LRs vs CPF) for overall abundance modeling. We first assessed differences in mammal community structure between LRs and CPF. Differences in species richness and relative abundance were tested using permutational t-tests, while sampling completeness was evaluated using rarefaction and extrapolation curves. Compositional differences were visualized through Non-Metric Multidimensional Scaling (NMDS) and tested using Permutational Multivariate Analysis of Variance (PERMANOVA). Prior to PERMANOVA, homogeneity of multivariate dispersion among groups was assessed using permutation tests. To evaluate the influence of anthropogenic variables, we performed two sets of analyses. First, we used PERMANOVA to test the relationship between species composition and anthropogenic variables within each study area. Second, we modeled species richness and relative abundance as functions of anthropogenic variables using Generalized Linear Mixed Models (GLMMs) for LRs and Generalized Linear Models (GLMs) for the CPF. GLMMs and GLMs were fitted separately because LRs comprise spatially disjunct settlements, whereas the CPF represents a single continuous forest area. Finally, to assess functional trait/attribute filtering, we applied RLQ and fourth-corner analyses. These methods integrated matrices of anthropogenic variables, species relative abundance, and functional traits/attributes, and tested different null models of trait–anthropogenic association using permutation tests with false discovery rate correction. Methods S1: Comprehensive Description of Statistical Analyses All statistical analyses were conducted in the R language and environment for statistical computing. The following suite of analyses was selected to directly address our hypotheses regarding taxonomic and functional differences between protection types and their drivers. We calculated relative abundance using two different approaches, depending on the analytical objective: 1. Species-level approach (for community analyses):We calculated the relative abundance of each mammal species at each sampling point as the number of independent records of that species divided by the total number of independent records of all species at the same point. This standardization was performed using the decostand function (margin = 1) from the vegan package. These values were used in species composition analyses, including non-metric multidimensional scaling (NMDS), permutational multivariate analysis of variance (PERMANOVA), as well as RLQ and fourth-corner analyses. 2. Community-level approach (for total mammal abundance and modeling):We computed the proportion of independent records of all species at each sampling point relative to the total number of independent records across all points within each protection type category (i.e., LRs and CPF), then multiplied by 100. Differences in relative abundance and species richness between protection type categories were tested using a permutational t-test (9,999 permutations), implemented with the perm.test function from the broman package. Additionally, we used interpolation and extrapolation rarefaction curves–with 95% confidence intervals and 10,000 bootstrap resampling iterations–to compare observed and estimated mammal richness and assess the completeness of our inventories. These curves were generated using the iNEXT package. Species composition differences between LRs and CPF were visualized using NMDS ordinations based on Bray-Curtis dissimilarity, implemented via the metaMDS function from the vegan package. To evaluate the influence of anthropogenic variables on species composition, we conducted PERMANOVA tests using the Bray–Curtis dissimilarity index with 9,999 permutations (adonis2 function, vegan package (2)). We performed two sets of analyses: (i) a global test comparing species composition between protection type categories (LRs vs. CPF) without incorporating anthropogenic variables; and (ii) separate tests assessing the relationship between species composition and anthropogenic variables within each individual study area: the three settlements (Iporá, Tarumã-Mirim, Água Branca) and the CPF. Prior to PERMANOVA, variables were log-transformed (log + 1), and multicollinearity was checked using Spearman correlations. “Forest cover”, originally expressed as a percentage, was first converted to a proportion (decimal format) to allow correct application of the logarithmic transformation, which requires positive real numbers. This transformation also facilitated proper scaling and interpretation of the data. A high correlation (r = 0.76) was found between “distance to roads” and “distance to house” in both protection type categories; therefore, the latter variable was excluded from further analyses. The final three anthropogenic variables retained were: “distance to water”, “distance to roads”, and “forest cover”. To verify the assumptions of PERMANOVA, we tested for homogeneity of multivariate dispersion among groups using the betadisper function in the vegan package, followed by permutation tests (9,999 permutations; permutest function). Significant differences in dispersion were detected (F = 5.98, p = 0.0013), indicating heterogeneity in within-group variability. Therefore, PERMANOVA results should be interpreted with some caution, as observed differences may reflect both compositional shifts and variation in dispersion. All PERMANOVA analyses used the relative abundance values derived from the species-level approach (Approach [1]). To examine the influence of anthropogenic variables on species richness and community-level relative abundance (Approach [2]), we applied generalized linear mixed models (GLMMs) for LRs, as sampling points were distributed across three disjunct settlements, and generalized linear models (GLMs) for the CPF. GLMMs were implemented using the glmmTMB package. The relative abundance data exhibited overdispersion, so it was modeled using a negative binomial distribution (with linear parameterization). Species richness modeled using a Gaussian distribution. In both models, settlement was included as a random effect to account for spatial structure. For the CPF, GLMs were performed using the glm** function from the stats package, assuming Gamma distributions with a log-link function for both response variables. Only statistically significant models are reported in the electronic supplementary material, Figure S2. Additionally, to visualize the variation in species detections across protection types and their relationship with anthropogenic variables, we ordered sampling points in ascending order based on the number of independent records. To assess the effects of anthropogenic variables on functional trait/attribute filtering in both protection type categories, we applied the RLQ and Fourth-Corner approaches. These multivariate methods integrate three data matrices: an environmental matrix (R table, representing anthropogenic variables), a species abundance matrix (L table, based on relative abundance from Approach [1]), and a functional trait matrix (Q table, containing traits/attribute of the recorded species). All analyses were performed using functions from the ade4 package. We conducted a correspondence analysis (CA) on the L matrix using the dudi.coa** function, a principal component analysis (PCA) on the R matrix using dudi.pca, and a Hill–Smith PCA on the Q matrix–containing both continuous and categorical traits/attribute–using the dudi.hillsmith function (9). All continuous variables in the R and Q matrices were log-transformed (log + 1) to improve normality and comparability. The multivariate relationships among the R, L, and Q matrices were explored using the rlq** function. To assess the significance of pairwise associations between specific traits/attribute and anthropogenic variables, we applied the Fourth-Corner method. This analysis combined two permutation models: model 2, which tests the null hypothesis that species distributions (relative abundances) are independent of anthropogenic variables, and model 4, which tests the null hypothesis that species distributions are independent of their functional traits/attribute. Statistical significance was evaluated using the randtest** function with 49,999 permutations. Additionally, to examine the strength of associations between traits/attribute and environmental gradients represented by the RLQ axes, we used the fourthcorner.rlq function, also with 49,999 permutations. This analysis evaluates: (i) the relationship between individual functional traits and the first two environmental RLQ axes, and (ii) the relationship between individual anthropogenic variables and the first two trait RLQ axes. To control for Type I error in multiple comparisons, p-values were adjusted using the False Discovery Rate (FDR) correction method.