The effects of habitat fragmentation on different taxa and ecosystems are subject to intense debate, and disentangling them is of utmost importance to support conservation and management strategies. We evaluated the importance of landscape composition and configuration, and spatial heterogeneity to explain α- and β-diversity of mammals across a gradient of percent woody cover and land use diversity. We expected species richness to be positively related to all predictive variables, with the strongest relationship with landscape composition and configuration, and spatial heterogeneity, respectively. We also expected landscape to influence β-diversity in the same order of importance expected for species richness, with a stronger influence on nestedness due to deterministic loss of species more sensitive to habitat disturbance. We analyzed landscape structure using: i) landscape metrics based on thematic maps and ii) image texture of a vegetation index. We compared a set of univariate explanatory models of species richness using AIC, and evaluated how dissimilarities in landscape composition and configuration and spatial heterogeneity affect β-diversity components using a Multiple Regression on distance Matrix. Contrary to our expectations, landscape configuration was the main driver of species richness, followed by spatial heterogeneity and last by landscape composition. Nestedness was explained, in order of importance, by spatial heterogeneity, landscape configuration, and landscape composition. Although conservation policies tend to focus mainly on habitat amount, we advocate that landscape management must include strategies to preserve and improve habitat quality and complexity in natural patches and the surrounding matrix, enabling landscapes to harbor high species diversity.

Mammal diversity data

We performed four field expeditions in April 2009, August 2009, May and June 2010, and July and August 2010. This effort was carried over 20 landscapes, distant from each other between 20 km to 634 km, yielding 20 independent samples of terrestrial mammal occurrence with body sizes varying from small (>1 kg) to large (Figure 2). On each expedition, we sampled mammals in five landscapes during five consecutive days and four nights using the following complementary methods: i) identification of vestiges, such as tracks (identified according to Angelo et al. 2008), feces, teeth, and others bones (bones were collected and compared to collection material for identification); ii) direct observation; iii) camera trapping; and iv) capture of small mammals with live traps. The sampling goal was not to estimate abundances, but to get a tally of species in each landscape for calculating species richness and composition.

For the first two methods, we performed walks on foot or by car at different periods of day and night, covering the different environments within each landscape. For the third method, we installed between 11 and 16 camera traps (Tigrinus®, Timbó, Santa Catarina State, Brazil) at 30-40 cm above the ground, in tree trunks of forest or Cerrado patches in each landscape. Cameras were placed on transect lines of 110 m in length containing two cameras in each extremity (in the border and in the interior of each forest fragment), operating 24 hours a day, during four consecutive days and nights. Transect lines were distant at least 150 m from each other (in small areas), but usually a minimum distance of 300 m was set. The total sampling effort was of 1,128 traps-night, with the mean effort per landscape being 56 ± 7 traps-night. We captured rodents and marsupials (<1 kg, Cricetidae, Echimyidae and Didelphidae families) using 65 wire (33x12x12 cm) and Sherman live-traps (30x9x7 cm). Traps were installed in forest ground (wire) and understory (Sherman), between 1.5 and 2 meters above the ground, during four consecutive nights, totaling 6,800 trap-night overall and 340 traps-night per landscape. We baited the traps with a mixture of pumpkin, bacon, peanut butter and cod liver oil. In each landscape, we installed the traps along transects between the camera trap sampling points, 10 m apart from each other in the same transect, separated at least 150 m from each other transect line and at least 20 m from the nearest patch edge. Captured animals were identified and subsequently released. When necessary, we collected voucher specimens for identification, which were deposited in the mammalian collection of the Universidade Federal de Santa Maria (UFSM).

Land use and land cover maps

We generated an 8-km buffer around the camera trap sampling points within each landscape to delimit landscape extent. We chose this extent based on previous studies reporting landscape structure effects on small-, medium- and large-sized mammal assemblage composition within the Atlantic Forest (e.g. Lyra-Jorge et al. 2010, Beca et al. 2017, Melo et al. 2017, Regolin et al. 2017), as well as to avoid spatial overlap (Jackson and Fahrig 2015). We mapped land cover for each landscape using orthorectified images from the RapidEye satellite constellation, with 5m spatial resolution. Images were selected preferably from the dry season, due to lesser cloud cover and greater contrast between land use classes (47 images acquired between January 2011 and August 2013). Image processing was performed over all five spectral bands (blue, green, red, red edge and near infrared) and included: i) atmospheric correction using the ‘Quick Atmospheric Correction – QUAC’ algorithm implemented in the ENVI 5.0 software and ii) unsupervised classification using the ‘Auto Class’ software (github.com/JohnWRRC). Auto Class uses the GRASS function ‘i.segment’ to generate image segments and the K-means Clustering function of the ‘foreign’ R package (R Core Team 2017) to group the segments into classes according to the mean and standard deviation of pixel values. We then converted this unsupervised map into a thematic classification by supervised visual interpretation and manual editing, based on image visualization at 1:2,500 cartographic scale, generating a final map with 11 classes (Figure 2).

Landscape structure metrics

The produced land cover maps in raster format were used as inputs for landscape structure metric calculations. We used the ‘raster’ R package (Hijmans et al. 2017) to load the raster data and define custom functions to calculate the following landscape structure metrics: (i) woody cover — percent woody (forest plus cerrado) cover in the landscape, (ii) patch density — ratio between the number of woody patches and total landscape area, (iii) edge density — ratio between area of woody patch edges and landscape area, and (iv) landscape diversity — Shannon index for mosaic of patches including all cover types. Woody cover and landscape diversity are used as measures of woody habitat composition, whereas edge density and patch density are measures of woody habitat configuration (Villard and Metzger 2014).

Within-habitat spatial heterogeneity

We estimated within-habitat spatial heterogeneity by calculating image texture measures from the normalized difference vegetation index (NDVI). NDVI is a spectral index sensitive to photosynthetically active vegetation, which is related to plant biomass productivity (Justice et al. 1998). We calculated NDVI using the red and near-infrared spectral bands of RapidEye images (5-m spatial resolution) using the ‘spatial.tools’ R package (Greenberg 2018). Image textures are statistical descriptors of the spatial relationship among pixel values within an image region, thus capturing spatial heterogeneity (St-Louis et al. 2009, 2014). When calculated using NDVI, texture therefore represents spatial variability in photosynthetically active vegetation within a given area (Wood et al. 2012). Texture measures calculated from high resolution images have been related with descriptors of vegetation heterogeneity such as leaf-area index and foliage height diversity (Colombo et al. 2003, Wood et al. 2012). Particularly, textures can yield larger explanatory power for species richness than classified images because it captures fine-scale variability within coarse habitat classes in areas of gradual transition between vegetation types (St-Louis et al. 2009, Wood et al. 2013).

We calculated 12 texture measurements from NDVI, using the ‘r.texture’ GRASS GIS function, being seven first order metrics: (i) sum average, (ii) entropy, (iii) difference entropy, (iv) sum entropy, (v) variance, (vi) difference variance, (vii) sum variance; and five second-order metrics based on a pairwise matrix of spatial relationships among pixels (grey-level co-occurrence matrix; Haralick 1979),  (viii) angular second moment, (ix) inverse difference moment, (x) contrast, (xi) correlation, and (xii) information measures of correlation. Each texture was calculated in four directions (0, 45, 90 and 135 degrees) considering a central pixel and its neighbors within the specified window, and then average of texture metrics were calculated to summarize all directions. We derived textures using four different moving window sizes on each pixel (3x3, 5x5, 7x7 and 9x9 pixels of 5m).

Table columns description

Landscape_name: Name of the landscapes. 42 levels. 

x_utm: X value of the landscape centroid in the Universal Transverse Mercator coordinate system (UTM), zone 21S. 42 values. 

y_utm: Y value of the landscape centroid in the Universal Transverse Mercator coordinate system (UTM), zone 21S. 42 values. 

Columns 4 to 51: Occurrence of each species in the landscapes. 2 values: 1 presence and 0 absence. 

Columns 52 to 100: Within-habitat spatial heterogeneity measurements. 

Columns 101 to 105: Landscape metrics.