The persistence of species in fragmented landscapes relies on landscape connectivity and individuals’ ability in dispersing among habitat patches. Accordingly, matrix structure can affect the orientation of dispersing individuals across the landscape. In this study, we measured the impact of matrix structure on the dispersal performance of the white-eared opossum (Didelphis albiventris). We released individuals in three types of matrix: bare field, corn crops and soybean crops, with distances of 30, 50 and 100 m to the nearest habitat patch. To test if the release height would affect the individuals' dispersal performance, we released animals from the ground and from 2 m high platforms. We released and tracked 14 individuals in bare field on the ground; 30 in corn crops, 22 on the ground and 8 on platforms; 17 on soybeans crop, 12 on the ground and 5 on platforms. The type of matrix influenced the perceptual range. Perceptual range was 100 m in bare field, 50 m in cornfield and less than 30 m in soybean field. The platforms only increased the perceptual range of individuals in the cornfield from 50 to 100 m. Visual and olfactory cues would cause this effect. We conclude that matrix structure affects dispersal performance, and that vertical elements of the matrix, such as scattered trees, may increase orientation in crop fields during inter-patch dispersal.

2.1 Study species

As a model organism, we chose the white-eared-opossum, Didelphis albiventris (Lund 1840), which occurs from French Guyana until Argentina, except in the Amazon Basin (Emmons & Feer, 1990). This species is common in the Brazilian savanna and easily found in urban areas within the region. D. albiventris is the largest Brazilian didelphid; it is nocturnal, solitary, habitat generalist, scansorial and omnivorous (Cabrera & Yepes, 1960; Cáceres, 2002; da Fonseca et al., 1982). Studies on patch occupancy, inter-patch movements, or environmental requirements of this species are rare. In the Atlantic Forest D. albiventris was captured only inside forest patches (Umetsu et al., 2008) and, although it is able to maintain populations in small fragments, it was absent in most of them (Asfora & Pontes, 2009). Regarding matrix use, D. albiventris usually is
not found at crops (Gheler-Costa et al., 2012; Scheibler & Christoff, 2007), or is found in very low abundance and usually near to the forest border (Bilenca et al., 2007). Taken together, these studies suggest that non-forest matrices are not suitable habitats for D. albiventris. This avoidance of deforested matrices, coupled with its easiness to capture, makes D. albiventris an excellent model to understand inter-patch movements and dispersal in fragmented landscapes.

2.2 Animal Capture and Translocation

We captured D. albiventris individuals in urban forest fragments, houses and commercial buildings in the municipality of Campo Grande, state of Mato Grosso do Sul, Brazil. The annual average temperature in the study area is 23°C and the annual average precipitation is 1.500 mm (Coleti et al., 2007). Forest fragments in Campo Grande are woodland savanna remnants, with tropical climate (Aw Köppen), dry winter and wet summer. Most captures occurred around and inside the Federal University of Mato Grosso do Sul campus (20°29'59"S, 54°36'49"W). We captured individuals using Tomahawk live traps (40.6 x 12.7 x 12.7cm). We distributed five to forty traps randomly in the urban forest fragments, houses, and building roofs, where the species was reported. We did not use in the experiments, and we released on the capture sites, young individuals and pregnant or lactating females. Other individuals were placed in individual cages and kept at the laboratory from 4 to 72h before translocation and release. The laboratory was a large climate-controlled room. We offered to the animals enough food and water for the duration of their stay. All individuals used in the experiment were weighted, sexed, marked with ear-tags, and had their age estimated by counting the number of molars (Macedo et al., 2006). We observed the individuals, and individuals who appeared unhealthy, or had atypical behaviour were not used in the experiment. From the laboratory, we carried the individuals in opaque fabric bags to the experimental area. These bags precluded individuals from seeing the environment, thus reducing possibility of orientation before release; they also reduced the stress of manipulation during transport. All procedures were approved by Brazilian Institute of Environment (License n°.: 51900) and Committee on Animal Research and Ethics (CEUSUFMS n°.:745/2016).

2.3 Release Experiment

We released individuals on a crop matrix close to a single isolated woodland savanna forest fragment in the rural area of Campo Grande (20°36'39" S, 54°33'14"O) (Fig. 1). Using this strategy, we were able to standardize soil, relief, forest horizon angle and woodland high, variables that could interfere in our treatments (Zollner & Lima, 1997). The trees within the forest fragment are 15m high on average, the canopy is continuous and the understory is open.The forest is semi deciduous, with loss of some of its canopy during dry season. Is possible to observe some human interference, likely due to selective logging. The target fragment had about five hectares in area, and it was located at least 800m from other fragments, and at least 10km from the places where we captured individuals. We assumed that this distance was enough to avoid animals returning to their original environment (homing behavior) or having any previous knowledge about the release site (Forero-Medina & Vieira, 2009). The matrix was composed of corn (April to August), soybean (December to February) or bare filed (between cultures, typically March, October, and November; Fig. 1). We are aware that the effect of our treatments (matrix type) could be confounded with the time of the year and climate variables. To mitigate the possibility of a climate interference, we released individuals at each treatment at least in two different years. In addition, all release events happened in similar temperatures and in sunny days. The releases on bare field occurred in March 2017 and from September to November 2018. The mean temperature (temp.) and the mean air humidity (humi.) during such releases were 27.8±4.5°C, and 62±16%, respectively. On corn, we released individuals from May to July 2017, and in July 2018 (temp. 27.2±3.4°C, humi. 38±13%). On soybean, we released individuals in 2016 December and 2019 February (temp. 28.9±3.4°C, humi. 65±16%). We released all individuals on the southern face of the fragment. This standardization aimed to reduce environmental variation such as wind direction, vegetation high, row direction, and relief that could affect our results. We released animals at  round 30 m from the forest fragment (mean = 26.95m, varying from 17.05 to 33.95m), 50m (48.89m, from 38.04 to 55.75m) and 100m (100.59m, from 88.48 to 108.36m), at ground height at the 3 types of matrix (Figure S1). We chose those distances, to match previous studies (e.g. Prevedello et al., 2011) and allow better comparisons and future  eta-analyses. Seeding, harvesting and the maintenance of crops were mechanized in the study area. Therefore, well-defined plantation rows and spray lines characterized both crops (Fig. 1). The plantation rows (~0.6 m) were diagonally oriented in relation to the fragment for both crops. However, at 20 m to the fragment, the rows changed the directions and went around the fragment as a halo, which had 20 m width. From walking in the field during the study and observing the crop structure at the ground level, we noticed that a substantial difference in blockage between corn and soybean crops. When individuals were moving perpendicularly to these plantation rows, corn provided less obstructions to the small animals because corn plants do not develop ramifications at small mammals height (~30 cm). Spray lines (~80 cm) were in pairs, spaced 5 m within pair and 20 m among pairs. These spray lines were arranged perpendicularly to plantation rows and resulted from the soil compaction caused by machines tires during spray pulverization. Because our experiments involve animal translocation and could affect the populations from where the animals came from and the population at the release site, we tried to reduce the number of individuals used. Therefore, we only performed platform experiments at distances beyond the D. albiventris' perceptual range, based on our ground experiment. Using this strategy, we were able to test if the platform increases individual perceptual range, orientation, tortuosity, and dispersal success using the minimum number of translocated individuals as possible. Platform experiments were only performed on corn and soybeans matrices, at the 50 and 100m distances to the fragment. We used the Forero-Medina and Vieira (2007) apparatus to release the individuals. To make this apparatus we used two wooden rods. One was  m long, and the other was 0.7m long. We fixed them with spikes in an upside-down "L" shape. We set the longer side of the "L" in the soil. The corner and the shorter "L"  side received a pulley each. We used a bucket as a release box. From this box, a rope was driven through the pulley and we could, using the rope, release the animals from at least 30m distance (Fig 1). The operator location was chosen haphazardly but the he would never stand between the fragment and the bucket. This distance aims to reduce the influence that the presence of the researcher could have on animal behaviour. After the box rose, individuals had 360° view of the surrounding landscape and could decide the way to follow. To simulate the crow's nest effect of scattered trees, independent from other resources that scattered trees may offer, we developed a wood platform. The platforms were 2m height in average, with a tripod shape. The three legs could be removed using a simple system of nuts and bolts. At the top of the legs, there was a wood squared surface where the same upside down “L” structure was present (Fig 1). 2.4 Post-release Animal tracking We tracked individuals using spool-and-line cocoons glued on the fur between the individuals' shoulder blades with Super Bonder® (see Vieira & Loretto, 2004). The cocoons weighted approximately 5g and had no internal reel. The line goes out through the centre of the cocoon that were wrapped with plastic wrapping and adhesive tape. After wrapped the cocoons were glued and the free side was tied at the release apparatus. After released the animal could walk normally. While it walks, the line could exit from the cocoons, recording its movements. Walked distances, hereafter referred to as “step lengths”, and direction changes, hereafter “turning angles”, were measured using measuring tape and compass. We considered only turning angles greater than 10°. The distance between two turning angles is the step length. We tracked the D. albiventris from the release site until the end of the line or until the animal reached the fragment.

2.5 Data analysis
2.5.1 Perceptual range

For each tracked individual we calculated both: 1. the direction (angle) that should be used to travel from the release place to the fragment using the shortest path, and 2. The angular mean of turning angles along the individual path, weighted by their respective step lengths. Then, we used these two values to calculate the angular divergence, which represented how oriented an individual moved toward the target fragment. Angular divergence ranged from -180º to 180º, with values around 0º depicting highly oriented individuals (i.e. individuals following theshortest path direction), while higher values depicted divergence for right (positive values) or left (negative values), until a complete opposite path (~|180º|). For each combination - distance, matrix type and release height - we used the Rayleigh test to check if angular divergence were different from 0º. Based on Rayleigh test, we considered as perceptual range the longer release distance at which angular divergence did not differ from 0º (following Zollner & Lima, 1997; Flaherty et al., 2008; Bridgman et al., 2012) 2.5.2 Effects on tortuosity, orientation ability and dispersal success We calculated three metrics to represent the path of individuals: sinuosity, orientation ability and dispersal success. We used the sinuosity index to measure tortuosity. Sinuosity index combines the mean cosine of direction changes with the mean step length, and it has been shown to have superior performance over alternative indexes to capture tortuosity of simulated paths (Benhamou, 2004). Orientation ability was measured as the angular divergence between the angular mean of turning angles, and the direction of the shortest distance to fragment (see details in perceptual range section: sensu Schooley and Wiens 2003; Forero-Medina and Vieira 2009; Prevedello et al. 2011). Dispersal success was measured in a binary way, in which individuals that reached the fragment edge were considered successful (coded as 1’s) and those that did not reach were unsuccessful (coded as 0’s) (Prevedello et al., 2011; Zollner & Lima, 1999, 2005). To statistically analyse these metrics, we built a total of 6 generalized linear models, in two sets of three models. In the first set, we considered only ground releases. We tested the effect of release distance and matrix type on the three metrics: tortuosity, orientation ability, and dispersal success. In the case of dispersal success, we used binomial distribution error because of the binary nature of this response variable. The other two variables had a normal error distribution. For the second set of models, we also tested the effect of the platforms. However, to do so, we had to restrict our analysis to corn and soybean fields at 50 and 100 m distance to the fragment. Once again, we had three models, one for each metric, with the dispersal success model having a binomial error distribution. For all six models, we ran a full parametrized model (including all interactions). Non-significant interaction terms were dropped, and models were ran again. Because the reduced number of individuals released in platforms, we performed a set of complementary analyses grouping releases at 50 and 100m, and compared the platform effect separately for each crop using Wilcoxon test. This set of follow up analyses must be interpreted cautiously All data was analysed using R (R Core Team, 2019). Complete dataset is provided in Table S1.

id: a sequential number of the individual according with the capture order 
sp: Species of the individual that was captured
id_l: ear ring attached at the left ear
id_r: ear ring attached at the right ear
date_cap: the day that the individual was captured
mass: mass of the individual in kilogram 
l_skull: length of the individual skull in centimeter
w_skul: width of the individual skull in centimeter
sex: sex of the individual
age: age class estimated using the molar teeth
date_rel: The day that the individual was released
dist: the distance from the fragment from where the individual was released 
x: the latitude of the place where the individual was released
y: the longitude of the place where the individual was released
matrix: the kind os matrix where the individual was released
time: the time of the day when the individual was released
azi_l: at the release place, using a compass, this is the azimuth to te left limit of the target fragment
azi_r: at the release place, using a compass, this is the azimuth to te right limit of the target fragment
azimuth: using a compass, the direction to where the individual moved
move: the distance moved in the previous azimuth
time_orientation: we measured to some individuals, how long they took to iniciate the movement. These data were not used at this papper.
arrive:if the individual reached the target fragment, the value is 1, if not, the value is 0
row: if the movement was realized at the plantation row, the value is 1, if not, the value is 0
platform: if the release happened from a plataform, the value is one, if not, the value is 0