The Lacandona rainforest, Chiapas, Mexico (91˚6’42.8”–90˚41’8.7’’ W; 16˚19’17.1”–16˚2’49.3” N) has a warm (mean annual temperature 24-26°C) and humid climate (mean annual precipitation ranges from 2500 to 3500 mm). The original vegetation is tall evergreen rainforest (Carabias et al. 2015). The Lacantún River separates a large protected forest tract on the western side of the study area, the Montes Azules Biosphere Reserve, from the Marqués de Comillas region on the eastern side, a heavily impacted area with approximately 50% of remaining forest cover (203,999 ha; Arce-Peña et al. 2019), dominated by cattle-ranches, annual crops and oil palm plantations. The study was conducted in 20 forest patches in the Marqués de Comillas region. Patches ranged in size from 5 to 2300 ha and were separated from each other by a distance of at least 2.5 km, measured from their geographical centres (Fig. 1).
Arboreal mammal surveys
Mammal surveys are detailed elsewhere (Cudney-Valenzuela et al. 2020), but a brief overview is given here. At the geographical centre of each patch, and avoiding vegetation gaps, we selected five trees with suitable climbing conditions (branches ≥ 20 cm wide, preferably hard wood species) and whose architecture allowed to install a camera trap facing other main branches. At each tree, we established a single-rope climbing system. Focal trees in the same patch were separated by 30 – 150 m. Of the five focal trees per patch, four reached the canopy (mean ± SD = 21.8 ± 6.2 m, range = 10.2 to 36.6 m) and one the midstory (9.1 ± 4.7 m, 3.4 to 19.6 m). This allowed us to capture a greater vertical range of strata potentially used by arboreal mammals.
We used one camera trap (Bushnell Trophy Cam HD Aggressor Low Glow ©) per patch. Cameras were placed at varying heights depending on the characteristics of the focal tree (camera height of canopy and midstory trees was 15 ± 4.3 m and 2 ± 0.6 m, respectively). Cameras were continuously active from May 2018 to May 2019, and they were serviced once a month (change of batteries, downloading of pictures, replacement of malfunctioning cameras). We rotated the location of the cameras once a month among the five focal trees in each patch, except from October to December when they remained on the same focal tree. Total sampling effort was 7,387 camera trap nights (average per patch = 369 ± 11.6 nights), with 6,233 active camera trap nights (average per patch = 311.7 ± 19.9 nights).
To increase the probability of photo-capture we used baits in the midstory trees (tuna fish, peanut butter with oatmeal and a banana). As revealed by photographs, bait was consumed during the first two nights in all cases. Since we did not provide more bait while the camera was active on that tree and no camera malfunctioned on the period the trees contained bait, all sites had the same bait sampling effort. We processed all photographs with the program Digikam© and extracted photograph metadata with the package ‘camtrampR’ (Niedballa et al. 2016). We considered photo captures as independent events when there was at least a 24 h interval between captures of the same species, since individuals photographed on the same day are likely the same ones (Royle et al. 2009). We identified each mammal species using Reid’s (2009) field guide, and obtained their body masses from Ceballos and Oliva (2006; Supplementary Material, Table S1). Except for the Mexican hairy porcupine (Coendou mexicanus) and squirrels, all other rodents were excluded from the analyses, due to imprecision in identification. For further analyses, we excluded rare species that appeared in less than half of the sites (i.e. Eira barbara, n = 4 patches; Leopardus wiedii, n = 2 patches; Procyon lotor, n = 1 patches) to avoid spurious relationships. We finally included 12 species in the analyses that are described below.
We adopted a site-landscape approach (sensu Brennan et al. 2002), with response variables measured in same-sized sample sites (i.e. 5 focal trees at the centre of each forest patch), and landscape variables measured within 13 concentric circular landscapes (100- to 1300-m radius, at 100 m intervals, measured from the geographical centre of each site; Fig. 1). We used recent and high-resolution Sentinel S2 satellite images (from 2016) to produce land cover maps of each landscape surrounding the focal patches using ENVI 5.0 software. Land covers were classified into six types: (i) old-growth forest cover; (ii) secondary vegetation; (iii) tree crops (e.g. oil palm plantations); (iv) annual crops and cattle pastures; (v) human settlements; and (vi) water bodies (Fig. 1). The area covered by each land cover type was calculated using ArcGIS software with the ‘Patch Analyst’ extension. We then estimated the following landscape variables: (i) the percentage of old-growth forest cover (i.e. area covered by old-growth forest divided by landscape size × 100), (ii) matrix openness (i.e. area covered by treeless land divided by matrix area × 100), (iii) patch density (i.e. number of old-growth forest patches divided by landscape size), and (iv) edge density (i.e. the sum of the length of all old‐growth forest edges within the landscape, divided by landscape size).
We calculated the number of species per forest patch. We also calculated each species’ relative abundance index (O’Brien 2011) by dividing the number of events for a given species by the number of days the camera was active in the patch, and multiplied by 100. This index is widely recommended as a proxy of mammal abundance in studies using camera traps (Mandujano and Pérez-Solano 2019; Benchimol and Peres 2020). We rounded up each species’ relative abundance to the nearest whole number to calculate species-specific abundance per patch, and summed the relative abundance of all species in each patch to calculate total abundance per patch.
To identify the SoE, i.e. the scale at which each landscape variable best predicted each response (total relative abundance, number of species, and relative abundance of each mammal species), we used generalized linear models with a Poisson distribution error. We excluded the smallest scale from the analysis (100-m radius) since it did not show variation in forest patch density (Fig. S1). We first quantified the relationship between each landscape variable and each response at each scale (4 landscape variables × 12 landscape buffers × 14 response variables = 672 models). To identify the SoE, we calculated the percentage of explained deviance by each landscape variable measured at each scale to identify the scale at which each variable best predicted each response variable (Fig. 2). For the analyses at the species level, we only considered this scale as the SoE if it showed a relatively higher empirical support (i.e. it showed a difference in Akaike Information Criterion (ΔAIC) > 2) when compared with the null model (i.e. including only the intercept) (see Table S2). Then, following the protocol proposed by Galán-Acedo et al. (2018), San-José et al. (2019) and Martinez-Ruiz et al. (2020), we used ANOVA to test whether the SoE differed among landscape variables, and a t-test to test whether the number of species had greater SoE than total relative abundance. We finally used a linear regression to assess whether the SoE increased with increasing body mass.