Effect of forest loss and fragmentation per se on arboreal and ground mammals of the Lacandon rainforest, Mexico

This study was carried out in the Lacandon rainforest, Chiapas, Mexico (16º19’–16º02’N, 91º06’–90º41’W; Table 1; Fig. 1). This region holds the largest remnant of tropical rainforest in Mexico: the Montes Azules Biosphere Reserve (330,000 ha; Instituto Nacional de Ecología 2000). Due to its high and unique biodiversity, this region is considered to be of high conservation value within the country (Arriaga et al. 2000). Yet, land-use changes since the late 1970’s have resulted in the loss of more than 55% of its original forest cover (Carabias et al. 2015). The remaining forest cover is distributed in different-sized forest patches surrounded by an anthropogenic matrix mainly composed of secondary forests, annual crops, and cattle pastures.

We focused on forest-dependent rodents, arboreal mammals, and ground mammals from the Lacandon rainforest. In particular, we recorded four ground rodent species in this region (Arce-Peña et al. 2019a): Desmarest’s spiny pocket mouse (Heteromys desmarestianus), rice rat (Oryzomys sp.), Mexican deer mouse (Peromyscus mexicanus), and Toltec cotton rat (Sigmodon toltecus). However, given the purposes of the present study, we focused on Desmarest’s spiny pocket mouse, which is a well-known forest-specialist species that forages principally in old-growth forests (Reid 1998; San-José et al. 2014). Regarding medium- and large-sized ground mammals, we recorded 21 species in the region (Arce-Peña et al. 2022), but after excluding nine rare (≤ 5 records) species (e.g. Panthera onca, Puma concolor, Herpailurus yagouarundi, Conepatus semistriatus), and species that frequently use anthropogenic land uses (e.g. Dasypus novemcinctus, Urocyon cinereoargenteus, Nasua narica, Procyon lotor, Odocoileus virginianus, Pecari tajacu), we assessed here four forest-dependent species: lowland paca (Cuniculus paca), margay (Leopardus wiedii), white lipped peccary (Tayassu pecari), and Baird's tapir (Tapirus bairdii). These four species can occasionally occupy disturbed forests, but they are all associated with old-growth forests (Sowls 1997; Huanca-Huarachi et al. 2011; Garmendia et al. 2013; Lira-Torres et al. 2014; de Oliveira et al. 2015). Finally, we also recorded 15 arboreal mammal species in the region (Cudney-Valenzuela et al. 2021, 2022a, 2022b), but after excluding five species with few records, we assessed here ten species that are particularly dependent on forests: Derby’s woolly opossum (Caluromys derbianus), Mexican hairy dwarf porcupine (Condoeu mexicanus), tayra (Eira barbara), margay (Leopardus wiedii), Mexican mouse opossum (Marmosa mexicana), gray four-eyed opossum (Philander opossum), kinkajou (Potos flavus), northern tamandua (Tamandua mexicana), spider monkey (Ateles geoffroyi) and the black howler monkey (Alouatta pigra). Although these species may be able to utilize anthropogenic matrices, they rely heavily on old-growth forests for food and shelter (Emmons 1997; Wallace 2008; González-Zamora et al. 2011; Ortega-Reyes et al. 2014; Animal Diversity Web 2024). Some of these species (e.g. P. opossum, T. mexicana) can be considered scansorial, but we included them in the arboreal group because they have great climbing skills and can spend a considerable amount of its time foraging in the canopy (Emmons 1997).

Mammal surveys were carried out between 2016 and 2019 in the geographic center of old-growth forest patches and a few sites within the continuous forest of the Montes Azules Biosphere Reserve (Fig. 1). Depending on the group, we sampled between 12 and 26 forest sites (Table 1). The minimum distance between sampling sites was ~ 1500 m. In small rodents and arboreal mammals, we standardized the sampling effort within each forest site to avoid potential confounding effects related to the ‘sample-area effect’ (Fahrig 2013). In medium- and large-sized ground mammals, however, the sample-area effect was statistically controlled by including the covariate “patch size” in the models because sampling effort was proportional to patch size (Arce-Peña et al. 2022). All study sites were in lowland forests (100–200 m a.s.l.) with similar soil and weather conditions to avoid the confounding effect of these variables.

Small rodents were sampled in 12 forest sites (9 forest patches and 3 continuous forest sites; Table 1) following a random order to avoid potential confounding effects related to temporal variations in resource availability. In each site, we placed 120 Sherman traps in a grid of 90 × 110 m located in the center of each forest site, but avoiding tree-fall gaps when needed. The traps were active for 8 consecutive nights (960 trap-nights per site) and baited with a mixture of oats, sunflower seeds, and vanilla. Sampled individuals were marked with gentian violet, so that we could count the total number of individuals per site, excluding recaptures. The grid within the continuous forest sites was located > 1 km apart from the nearest forest edge. See further details in Arce-Peña et al. (2019a).

Medium- and large-sized ground mammals were sampled in 26 forest sites (22 forest patches and 4 continuous forest sites; Table 1) using camera-traps (one camera per forest patch). In particular, we located one camera in each forest site (Cuddeback Capture or Cuddeback Capture-IR Plus; recovery time of 30 s per photograph), in locations favorable for mammal detection (i.e. animal trails, sites with signs of use by mammals). We changed the location of the camera every 30 days, and repeated this procedure for five months (i.e. 30 d × 5 months = 150 camera-trap nights per site) to cover a larger sampling area. To have a more reliable estimation of site occupancy by each mammal species, we complemented the information from camera traps with footprints and direct sightings recorded from slow (1 km/h) walks between 6:00 and 16:00 h, once a month. Yet, the duration of the survey varied with patch size: 3–4 h in small-sized patches (< 10 ha), 5–6 h in medium-sized patches (10–50 ha), 7–8 h in large-sized patches (> 50 ha), and 9–10 h in continuous forest sites (i.e. in an area of ≈100 ha). Combining these three sampling methods, we obtained a more accurate record of the presence of each species within each site.

Finally, arboreal mammals were sampled in 20 forest sites (i.e. 19 forest patches and a continuous forest site; Table 1). In the geographic center of each site, we selected five focal trees, separated from each other by ~ 30 m. We selected four big canopy trees and a midstory tree to capture the vertical range of strata potentially used by arboreal mammals. In each tree, we established a single-rope climbing system, and a camera trap (Bushnell Trophy Cam HD Aggressor Low Glow©), which was rotated among the five focal trees once a month. Cameras were continuously active along a complete year, and were serviced once a month (change of batteries, downloading of pictures). We considered photo-captures as independent events when there was > 24-h interval between captures of the same species, and calculated each species’ relative abundance index by dividing the number of events for a given species by the number of days the camera was active in the forest patch, and then multiplying it by 100. This index is widely used as a proxy of mammal abundance in studies using camera traps (e.g. Benchimol and Peres 2021).

We used high-resolution (10 × 10-m pixels) Sentinel-2 satellite images from 2017 to generate land cover maps of each landscape surrounding the focal forest sites. We used ENVI 5.0 to conduct a supervised classification, which gave an overall classification accuracy of Kappa index C = 0.9. We used ArcGis 10.5 software to estimate the percentage of old-growth forest cover in the landscape (forest cover, hereafter), and the density of old-growth forest patches, which is a simple, direct, and commonly used fragmentation measure (Riva et al. 2024b). Following Jackson and Fahrig (2012, 2015), we calculated forest cover and patch density (fragmentation, hereafter) in multiple concentric buffers from the center of each site. In particular, we considered 13 spatial scales (100–1300-m radius, at 100 m increments). We did not consider larger (> 1300) buffer sizes to avoid spatial overlap among landscapes, and thus prevent potential independence problems among samples (Eigenbrod et al. 2011). Indeed, in most cases the scales of effect of forest loss and fragmentation were 400–600 m (see Results), suggesting that assessing scales > 1300 m is not needed (Jackson and Fahrig 2015).

To assess whether the effect of forest loss on each species is linear or nonlinear, we used different statistical models depending on the response variable (Table 1). In particular, we used generalized linear models (GLMs) with a Poisson distribution error to assess the linear effect of forest cover (opposite to forest loss) on rodent abundance and tested for overdispersion in the models. As we found evidence of overdispersion, we ran the models using negative binomial distribution instead of Poisson (Lindén and Mäntyniemi 2011). For continuous responses (i.e. the relative abundance of each arboreal mammal species), we also used GLMs, but with a Gaussian distribution error, and checked for normality of residuals using Shapiro–Wilk test. When models failed to accomplish the normality assumption, we used log-transformations to the response variable. We also tested the nonlinear effect of forest loss on rodents and arboreal mammals with a four-parameter logistic regression model, which uses a sigmoidal curve to fit data and has been widely used for assessing extinction thresholds in tropical forests (Morante-Filho et al. 2015; Benchimol et al. 2017; Saldívar-Burrola et al. 2022; Martínez-Ruiz et al. 2024). We adjusted one model per landscape scale (100–1300 m radius) for a total of 27 models for each response variable (i.e. 13 linear models, 13 nonlinear models, and the null model). Regarding the presence or absence of medium- and large-sized ground mammals, we only tested the null model and a GLM with binomial distribution, which represents a logistic model analogous to the previously described four-parameter logistic regression. In this case, each set included 14 models (13 binomial models, and the null model), all with forest patch size as a covariate to control for variations in sampling effort among patches after verifying that patch size and forest cover were independent predictors (VIF < 1.5 in all cases). Patch size of continuous forest sites was set to 100 ha, as this value represents the total sampled area in each continuous forest site. We ranked each set of models from lowest to highest Akaike Information Criterion value corrected for small samples (AICc, Burnham and Anderson 2002) to identify the model (and spatial scale) that best predicted the response of each species to forest loss. We also calculated the effect sizes for each of the selected (ranked-above-the-null) models using Fisher’s r-to-z transformed correlation coefficients to assess the species-specific and general (merging all species) influence of forest loss.

To assess the effect of fragmentation per se (independent of forest loss; Fahrig 2003, 2017), we followed the statistical protocol proposed by Watling et al. (2020). In particular, we compared a linear GLM with forest cover one, a linear GLM including both forest cover and patch density, and a null (intercept only) model by estimating the AICc scores and evidence ratios based on corrected AICc scores. If fragmentation has a significant independent effect on each species, a model including both terms (forest cover and patch density) should provide a more plausible fit to the species data (i.e. lower AICc) than a model including only forest cover. In all GLMs, forest cover and/or fragmentation were measured at their respective scales of effect, and when including both terms in a model, we verified that they were independent predictors (VIF < 1.5 in all cases). Finally, to assess whether the scale of effect of forest loss and fragmentation were related to body mass (obtained from Animal Diversity Web 2024), we used simple linear regressions (one model per predictor variable). All analyses and figures were performed using R ver. 4.0.4 (R Development Core Team 2022).

We found that in six of 14 species (43%; Desmarest’s spiny pocket mouse, white-lipped peccary, margay, lowland paca, tayra, and Mexican mouse opossum), the null model performed better than linear or nonlinear models, suggesting that landscape forest loss has a weak effect on almost half of the species (Table 2). When the linear or binomial models ranked above the null model (8 of 14 species; 57%; Table 2), most species (7 of 8 species, 88%) were positively related to forest loss (Table 3, Fig. 2), including six arboreal species (gray four-eyed opossum, Derby’s woolly opossum, Mexican hairy dwarf porcupine, kinkajou, northern tamandua, and black howler monkey) and a ground species (Baird's tapir). Only the relative abundance of the Geoffroy’s spider monkey was negatively impacted by forest loss (Fig. 2; Table 3). These contrasting (positive and negative) responses to forest loss resulted in a non-significant overall effect of forest cover (Fig. 3).

When assessing the empirical support of models including forests cover only, and models including both forest cover and fragmentation (i.e. fragmentation per se), we found that forest cover alone was, on average over 2.4 times more likely to provide the most plausible fit to the mammal species data than a model of fragmentation per se (Table S3). Fragmentation per se provided the most plausible fit to the data in four species, but its effect was as likely to be negative (Geoffroy’s spider monkey, margay) as positive (Derby’s woolly opossum, northern tamandua), and in all cases the parameter estimate was non-significant (Table 4; Table S3).

The percentage of forest cover decreased with increasing landscape size (Fig. S1). The scale of effect of forest loss and fragmentation varied from 100-m to 1300-m radius. However, in most cases (64%), the scale of effect of forest loss fell within the 400–600 m range (mean ± SD = 664.2 ± 379.4 m; Table 2). Similarly, in most cases (57.1%) the scale of effect of fragmentation fell within the 400–600 m range (mean ± SD = 628.6 ± 487.4 m; Table 2). The scale of forest cover effect was not associated with body mass (R² = 0.08, F = 1.18, P = 0.31), and the mean values and 95% confidence intervals showed no difference between arboreal (555.5 m, 297.4–813.7 m) and ground species (860 m, 704.1–1015.8 m). Similarly, the scale of fragmentation effect was not related to body mass (R² = 0.02, F = 0.27, P = 0.61), but tended to be relatively smaller in arboreal (411.1 m, 140.8–681.3 m) than ground mammals (1025 m, 602.2–1447.7 m).