Forest fragmentation and habitat loss are increasing human–wildlife conflict. The use of fencing to reduce wildlife incursions into agricultural areas, particularly through optimizing fence locations, is gaining interest. However, the spatiotemporal movement patterns of animals relative to hotspot fencing have received scant study. This study evaluated how effectively a single fence reduces seasonal incursions into an agricultural zone by crop-damaging mammals. The seasonal distributions of wild boar (Sus scrofa Linnaeus) and water deer (Hydropotes inermis argyropus) were evaluated, along with the effect of a short fence on their spatial behavior at a junction of a forest fringe and fields during different seasons. Seasonal habitat suitability models were constructed using unmanned aerial vehicle and camera trap data for the environmental predictor variables. Connectivity analysis utilizing habitat suitability models was applied to assess the effects of fencing on wildlife. The seasonal movements of water deer and wild boar were found to be primarily constrained by trails and roads, while fencing was the most important variable for wild boar following fence erection. The fencing effect was evaluated using a connectivity model, which showed increases of 65.5% for wild boar and 100.9% for water deer from the growing season to the harvest season. However, a high permeability area of wild boar disappeared at the fencing location, while that of the water deer was maintained. This suggests that properly designed fencing, based on the behavior of target species, can significantly reduce their incursions. Knowledge of the fencing effects in fragmented areas may help allocate local and regional management resources to solve coexistence challenges and establish wildlife conservation strategies.

1. Study area

The study site was in a forest fringe area of Baekbong mountain in Namyangju City (an area of 458.1 km², located at 37° 30–46ʹ N, 127° 05–22ʹ E), Gyeonggi Province, South Korea (Figure 1). It was selected because it provides accessibility to agricultural fields for wildlife and is a fragmented forest with suitable conditions that enable identification of the effects of fence installation on wildlife movements through monitoring with high-density camera traps. The climate is hot and humid in summer, and cold and arid in winter, and is part of the East Asian monsoon system. The annual mean temperature is 21°C, and the monthly mean temperature fluctuates seasonally, ranging from −4.2°C in January (winter) to 25.0°C in July (summer). The vegetation is primarily composed of a mixed forest of deciduous broadleaf and needleleaf trees, including pine (Pinus densiflora) and nut pine (Pinus koraiensis). Grassland and cropland are the other dominant nonurban land cover types in the area. Lettuce, green onion, pepper, corn, potato, sweet potato, and beans were identified as main crops through manual investigation.

2. Occurrence of data collection and spatial filtering

Camera traps were deployed in 18 locations for 14 months (January 2021–February 2022) and were separated within regular hexagonal cells with 100-m-long sides (Figure 1). Camera trap locations were selected based on traces of wild boar, such as rubbing trees, mud pools, soil rooting, and feces, or along wild animal pathways, to capture multiple events during the period. Cameras were located at least 50 m from each other. Cameras with infrared flashes, types D3 and D9, were used (Hong Kong Ica Industry Co., China). These were tied high on tree trunks beyond the reach of wild boar (1.18 m, SD = 0.20 m) to ensure safety (Supplementary Table S1). Both camera types had an aperture angle of 90° and detected animal movement using a passive infrared sensor (PIR sensor) with a detection range of ≤20 m. Camera PIR sensitivity was set to the normal level from among three choices (low, normal, and high). When triggered, they recorded 30 s of continuous video (resolution: 1920 × 1080 pixels). The cameras were checked and their memory cards replaced every 2 weeks.

During the months representing crop growth, the camera traps recorded 283, 382, 694, and 1209 water deer appearances during the seedling season (March), growing season (June), harvest season (September), and winter season (December), respectively. The appearances by wild boar numbered 176, 419, 86, and 97 for the same months, respectively. Local Moran’s I applied based on a distance weight of 160 m and a power value of 2 in GeoDa (version 1.20), and spatial autocorrelation of the records was removed (Kwon et al. 2016). After spatial filtering, 981 and 146 appearances by water deer and wild boar, respectively, were retained in the final dataset used for habitat suitability modeling (Figure 2). Occurrence points were randomly generated within a circular sector buffer that considered camera trap shooting range, 90° camera angle, and a detectable range of up to 20 m that included occurrence data in high spatial resolution. To decrease the spatial aggregation, one occurrence point was randomly selected per pixel per 1 m.

3. Environmental predictors of habitat modeling

To establish environmental variables, high-resolution UAV images were acquired in June 2021 using a DJI Phantom 4 Pro, V2.0, (DJI, Shenzhen, China) equipped with a built-in ground control point (GCP) unit and a 1-in. 20 MP CMOS sensor (5480 × 3648 pixels) with an 84° field of view and storage of 4K images at 60 fps. The flight mission was planned using DJI Go 4 and the Pix 4D capture applications. Altitude was set at 150 m, and flying speed was set to 5 m/sec, yielding a ground sampling distance of 7.85 cm/pixel. To generate the orthomosaic, a front and side overlap of 80% was set between the images to ensure coverage of every part of the study area. This resulted in 162 collected images. Image correction was conducted using GCP coordinate values, which were manually measured using the Gaia GPS app. Non-natural constructs, such as grave headstones, centers of manhole covers, and corners of road paint that were in low-slope open spaces, were selected as reference locations that could be precisely distinguished. A total of 15 GCPs and 39 checkpoints (CPs) were used for quality assurance. The digital elevation model (DEM) and orthomosaic were produced based on the structure from motion algorithms using the UAV imagery and Pix4Dmapper software (www.pix4d.com). The resulting DEM and orthomosaic were aligned to each other and used to generate environmental variables.

Before running the water deer and wild boar habitat models, all the layers created using drones were set at a resolution of 1 × 1 m, and 13 distance-related variables in four types of habitats were selected. These consisted of the distances to the following items: oak trees, coniferous trees, and edge of forested area (forest); sweet potato and corn (distance to crops); distances to roads, informal trails, formal trails, cemeteries, building, and fences (human-related); and distances to mud pool and water bodies (hydrography). In addition, three site characteristics were selected: elevation, slope, and roughness (topography). To reduce the effect of potential multicollinearity on the models, Pearson’s correlation coefficient was calculated, and only the variables <0.75 were selected (Dormann et al., 2013). Therefore, from the 16 environmental variables, 11 were selected to predict the distributions of the species. These consisted of oak tree, coniferous tree, and edge of forested area (forest); corn (crop); distances to roads, informal trails, and fences (human-related); DEM, slope, and roughness (topography); and distance to water bodies (hydrography) (for further details, see Supplementary Table S2 and S3).

4. Maxent model setting and model validation

To model seasonal water deer and wild boar distribution, maximum entropy species distribution modeling (Maxent, version 3.4.4) was used. This applies the principle of maximum entropy to relate presence-only data to environmental variables in order to model the likelihood of the presence of species without the need for absence data, which is difficult to collect (Phillips et al. 2009). In this study, presence data of water deer and wild boar were collected by camera traps as a representative value for each hexagon, and environmental layers were derived from high-resolution mapping resulting from the use of UAVs (Figure 2).

The presence data was solely captured through camera traps; therefore, a bias grid file was created by assigning a sampling probability value of 1 within the camera trap shooting range (90° angle and 20 m detection distance) and the value of 0.01 to other areas to restrict the background sampling area (Vollering et al. 2019). The output format was set to cloglog as it provides a better result compared with logistic when bias correction is used (Phillips et al. 2017). The final output provides relative suitability values from 0 (unsuitable) to 1 (suitable). All other parameters remained on default settings. The K-fold cross-validation method was used to fit the model (Fielding and Bell 1997), with partitioning occurrences for training (75% of the total) to calibrate the model and testing (25%) to evaluate model performance. Models with the area under the curve (AUC) >0.75 corresponded to high discriminatory performances, while a value of 0.5 indicated that the model prediction was only as good as a random guess (Fielding and Bell 1997). The final probability model was transformed into a binary map using the maximum of the sum of training sensitivity and specificity, with the value above for presence and under for absence (Manel et al. 2001). Thus, eight models representing the two species and four seasons (seedling, growing, harvest, winter) were obtained.

5. Connectivity modeling

Functional connectivity analysis can be applied to assess the effects of fencing on wildlife because it can represent their movement capacity within human-modified landscapes (Cushman et al. 2009). Open-source Julia software (version 1.6.5) was used to implement the Omniscape algorithm, a circuit theory approach for omnidirectional landscape connectivity modeling used to assess pathways across landscapes. Omniscape calculates current flow over the study area between all (or a regularly spaced subset of) pixels and the center pixel, treating the landscape as resistance to movement in a circular moving window within a user-specified radius (McRae et al. 2016). Species distribution maps, generated using Maxent, were imported into Omniscape to generate the input spatial source and resistance layer (see Section 2.4). The habitat suitability model was employed as the source layer and was transformed to produce the resistance layer by reversing the values in order to allocate high movement resistance to low-quality areas (1-suitability) (Costa et al. 2021).

Two Omniscape outputs were generated, and within the cumulative connectivity values map, higher values represented higher connectivity and normalized flow potential, where cumulative connectivity was divided by current flow without resistance. The normalized flow potential map was reclassified into three categories: low (impeded < 1), moderate (diffused = 1), and high (channeled or intensified > 1). These better represented the most important areas for water deer and wild boar movement and illustrated the fencing impact upon them. To assess the similarity of intra- and interspecies functional connectivity according to the seasons, the spatial overlap in the cumulative connectivity map was assessed with pairwise Schoener’s D metric using ENMTools in R, version 4.1.3 (Warren et al. 2010). D statistic values range from 0 to 1, where 1 corresponds to identical niche overlap or close to the absence of overlap, and 0 corresponds to no niche overlap.

6. Fence erection in the hotspot area 

To observe the effect of fencing in water deer and wild boar hotspot areas, fencing was erected considering the following factors: (1) a location where the probability of presence and connectivity is relatively high before fencing erection; (2) the presence of a mud pool, which is a preferred habitat; and (3) consultation with the land owner. In addition, there were three aspects of fence construction that had to be considered in its design since large mammals pass through fences by three methods: jumping over the fence, lifting up and crawling under the fence, and penetrating weak areas of the fence material. Therefore, the fence was designed and installed considering these three aspects (Figure 3). Firstly, the upper part of the fence included an approximately 60° angle toward the direction from which animals mostly come. Secondly, the lower part of the fence net was tied with poles to prevent animals from bending or twisting the net and widening gaps. Lastly, the net was comprised of steel wire strong enough not to be broken by animals pushing against it. The erected fence was 200 m in length and was installed from July 26 to August 6, 2021. From June 28 to December 31, 2021, six camera traps facing the fence area were evenly deployed along the fence to observe approaching wild boar and water deer. During the winter, when crops are not generally damaged by wildlife, a door installed in the middle of the fence was opened.