Simulation of potential habitat overlap between red deer (Cervus elaphus) and roe deer (Capreolus capreolus) in northeastern China

Study area

Our chosen study area was the Tieli Forestry Bureau (TFB) with a total area of 2,058.42 km² (Fig. 1). The TFB is located on the southern slope of the Lesser Xing’an Mountains (47°02′–47°36′N, 127°57′–128°12′E) where forest covers nearly 70% of the landscape. From east to west, the topography changes from mountains to hills, and the elevation ranges from 1,137 m to 224 m a.s.l. The mean summer temperature is 20 °C and the mean winter temperature is −21 °C. Typical broadleaf–conifer mixed forest (Ma et al., 2014) supports more than 100 species of wild animals, such as black bears (Ursus thibetanus), red deer, grey cranes (Grus grus), mandarin ducks (Aix galericulata), and lynx (Felis lynx).

In recent years, logging has decreased and was finally banned in 2003. A state-owned forest tenure reform program was initiated in the TFB in 2006. This program involves the selection of commercial forests for management using 50-year contracts with individuals who are designated as land managers at specific sites. The policy has improved the enthusiasm of local foresters and enriched the management patterns of understory planting (Li et al., 2013), but this human pressure disturbed the natural habitats of deer (Li et al., 2014).

Target species

Characterization of cervid distributions in the TFB is important for three reasons. First, population abundance is very low (0.11/km²) because of long-term excessive logging and hunting, and some species have become extinct in several localities (Qin & Zhang, 2009). Second, human activities (e.g., conversions of forest into cropland) have a substantial effect on reducing abundance and distribution of suitable habitat. Third, they play important ecological roles in forest systems, because they are the major food source for predators and represent a large proportion of ungulate species. Therefore, cervids are ideal target species for a case study on SDMs.

Sampling design and species occurrence data

To collect occurrence records, we used the distance sampling method (Thomas et al., 2010; Waltert et al., 2008) to design “Z” glyph survey routes (1.5–2 km per line, 32 total lines sampled by walking). Along each route, we recorded direct observations, footprints, repose imprints, feces, tracks, and fur, which were clear evidence of deer presence. Depending on observer experience, information about individual deer could be gained based on morphological features such as body size, footprint patterns, and track size. More importantly, a variety of traces can be simultaneously observed in the field, and animal species can be determined based on the specific characteristics of various traces. Field investigations were conducted in the last 3 years (summer 2013, summer and winter 2014, and winter 2015). In total, 196 records were obtained in the field, and their positions were recorded by GPS (90 for red deer and 106 for roe deer, Fig. 2). GPS points were unified in a geographic coordinate system (Beijing_1954) and converted to the format of SDMs as input data within ArcGIS 9.3 (ESRI INC, 2008).

Species ecology and environmental data acquisition

Red deer and roe deer are two closely related species that occupy nearly the same habitat (e.g., broad-leaved forest and bush-grass land) at a local scale. The suitable habitats of these target species partly overlap and the behavioral ecology and habitat requirements are similar (Jiang, 2007). To estimate the environmental niche, we grouped environmental factors into five major categories: terrain, habitat accessibility, land cover, vegetation feature, and interference factors. No bioclimatic predictors were used because these datasets were usually coarsely scaled (Turner et al., 2003) and more suited to regional or continental SDMs (Elith & Leathwick, 2009; Gillespie et al., 2008).

Modeling species distribution

MaxEnt 3.3.3e (http://www.cs.princeton.edu/ schapire/maxent/; Phillips, Anderson & Schapire, 2006) was used to determine a suitable deer habitat. It is currently one of the most commonly used SDM algorithms, and its popularity has increased over the past decade owing to its robust distribution estimates (Elith et al., 2006). We used default values for convergence threshold (10⁻⁵) and maximum iterations (500), randomly splitting the data into 75% training data and 25% testing data, and 10,000 background points for each MaxEnt run. Area under the curve (AUC) was used to measure model performance and mean area under the receiving operating characteristic (ROC) curve. AUC values ranged from 0.5 (for random probability) to 1 (for perfect ability to predict presence; Phillips & Dudik, 2008). All selected variables were unified a resolution of 90 m, and respective calculations were done in ArcGIS 9.3. Finally, we used these 17 predictor variables to generate two model results: one for red deer and another for roe deer.

Identifying overlapping habitat areas

To map the overlapping habitat of red deer and roe deer, the Youden index was applied (Vega Garcia et al., 1995) to determine the cut-off point for each model. The ROC curve was obtained by plotting sensitivity versus specificity for various probability thresholds, and the Youden index was calculated by the following formula: sensitivity + specificity—1. Good performance is characterized by a maximal Youden index (Catry et al., 2009; Vega Garcia et al., 1995). Then, we divided the habitat map into suitable (≥threshold value) and unsuitable (<threshold value) habitat based on the maximal Youden index for each species, and assigned 1 for suitable and 0 for unsuitable habitat.

Next, the two reclassified habitat maps were overlapped to generate the final potential habitat suitability map. The pixel value of the map varied from 0 (unsuitable for any species) to 2 (overlapping habitat).

1. (00) = 0: unsuitable for two species;

2. (01 or 10) = 1: suitable for one species but unsuitable for the other;

3. (11) = 2: suitable for both species, e.g., the overlapping habitat.

Finally, we calculated the area of major landscape types in the overlapped habitat with the support of ArcGIS 9.3 (ESRI INC, 2008) and R 3.0.1 (http://www.r-project.org/).

Potential suitable habitat analysis

The testing average AUC for 10 replicate runs of the models was 0.936 ± 0.017 for red deer and 0.924 ± 0.018 for roe deer, indicating that the model predictions were better than chance (AUC = 0.5) and the model could be used to predict the species-occurrence pattern. The optimal cut-off values for red deer and roe deer models were 0.195 and 0.170, respectively. The suitable habitat area was 95.70 km² for red deer and 329.87 km² for roe deer, indicating that roe deer had greater habitat selectivity than red deer in the TFB. The final potential habitat suitability map (Fig. 3) revealed that the unsuitable habitat occupied 1,587.51 km², or 77.7% of the TFB. The overlapping habitat represented 89.93 km² of the study area, which included 94% of potential suitable habitat for red deer. The potential habitat modeled for roe deer constituted 329.87 km², of which 27% also contained overlapping habitat. Results demonstrated that the most highly suitable habitat was located in the eastern mountainous region. The southwestern plain area was not suitable because it was close to Tieli City.

Environmental factors

Results revealed that the most important environmental variable for predicting potential habitat distribution for red deer was distance to farmland (Fig. 4A; accounted for 26.8% of the distribution) followed by distance to bush-grass land (14.6%), elevation (13.5%), distance to water source (12.2%), and distance to settlement (9.2%). For roe deer, the most important variables were distance to farmland (22.9%), distance to settlement (21.4%), elevation (11.6%), coverage of shrub-grass (8%), and distance to water source (6.9%; Fig. 4B). Out of the above factors, distance to farmland, elevation, and distance to settlement were important variables for both species.

Response curves revealed the direction of the effect of the four most important factors in each model on the potential distribution of deer (Fig. 5). We found that the probability of occurrence decreased in areas with longer distance to farmland. In areas with longer distances to the water source or settlements, the probability of occurrence first increased and then decreased. Occurrence probability sharply increased with increasing distance from a water source (>1.3 km) or settlements (>5.5 km for red deer and 6 km for roe deer). The occurrence probability sharply decreased at a specific elevation (>320 m). We also found that vegetation feature factors contributed little to habitat suitability for deer.

Proportion of habitat types in suitable habitat classes

Based on raster calculations, comparison of major habitat types in each suitable habitat category showed that farmland was the main habitat type for deer (Fig. 6). It was found that, with a reduced probability of occurrence, farmland covered a fluctuant percentage of areas (21% of overlapping, 12% of suitable, and 21% of unsuitable habitats). This trend was reversed in broad-leaved forests (41% of overlapping, 46% of suitable, and 41% of unsuitable habitats), which is a landscape with low human interference. In general, the overlapping habitat area consisted of a high proportion of broad-leaved forest (41%) and farmland (21%). Broad-leaved forest and mixed forest occupied the main areas of habitat overlap.

Habitat overlap analysis and environmental factors

The MaxEnt models showed that 27% of roe deer distribution overlaps with predicted red deer distribution, indicating that the niches of the two species are similar. About 94% of the red deer distribution was predicted to be suitable for roe deer, indicating that the realized niche of red deer is smaller. The overlap between the two species in farmland and human settlements was significantly lower than that in broad-leaved forests. This is likely because broad-leaved forests offer deer plentiful food and safety from predators. Indeed, Bonnot et al. (2013) noted that a large portion of deer tend to mostly reside in forests.

The important factors in the models for both species were distance to farmland, elevation, and distance to settlement, indicating that the distributions of both species are limited by similar factors. It is possible that deer distribution is largely affected by distance to farmland, because farmland could provide high nutrition food. We found that the distance to road had little influence on the predicted distribution of deer in this study. Our finding is consistent with those of other studies; for example, Piekielek & Hansen (2012) argued that some roads can be natural wildlife corridors and do not pose a significant disturbance. For terrain factors, Li (2005) also found that altitude and aspect are the main contributors to habitat selection of deer in Heilongjiang Province, China. We found that slope had little impact (with the contributions of 0.4% for red deer and 1.5% for roe deer) on deer distribution, which can be explained by the gentle topography in our study area. In summary, deer prefer the following habitat characteristics: low altitude (about 320 m), far away from residential areas, near farmland and a water source, and suitable coverage of shrub–grass (65–70%). Nevertheless, finer-scale analysis of distribution could present different results; however, this needs to be evaluated with further quantitative research.

Effects of habitat types on habitat suitability

Broad-leaved forest and farmland are the two most important habitat types selected by deer in our study (Fig. 4), a result that is consistent with those of previous studies. Bonnot et al. (2013) indicated that roe deer prefer living in forests during the day and feeding in farmland at night, which may reduce the risk of being hunted. Behdarvand et al. (2014) found that environmental variables related to anthropogenically influenced land, such as irrigated farms, were most important for predicting wolf-attack risk. We suggest that both habitat types and structures affect deer habitat selection.

Advantages and limitations

MaxEnt software has proven to be a robust tool in analyses that include current species data with limited biological information (Gogol-Prokurat, 2011; Pena et al., 2014; Phillips, Anderson & Schapire, 2006). In our study, effective occurrence was recorded in field investigations during 2013–2015. Our models performed well when predicting potential distribution, and our findings were more accurate compared with historical data modeling. Past studies supported this view and showed that current species data could improve the accuracy and credibility of results (Brambilla & Saporetti, 2014; Razgour, Hanmer & Jones, 2011; Zeng et al., 2015). Because the accuracy of historical data can be called into question, it cannot be used with confidence in simulations to represent the current distribution state of target species.

However, we admit that this method has limitations, such as that the data sources were not comprehensive enough across the study area. In our study, the sampling was somewhat concentrated near roads, and high elevation areas may have been underrepresented. Such biases may exacerbate statistical problems when models are based on small sample sizes, although our models had robust predictions. In general, it is hard to avoid non-systematic sampling or sampling biases when constructing SDMs (Stolar & Nielsen, 2015). However, Wisz et al. (2008) suggested that MaxEnt had moderate sample size sensitivity combined with excellent predictive ability. MaxEnt was the best and generally outperformed other methods for evaluating low sample sizes. Hu & Liu (2014) pointed out that, although the availability of presence records is a considerable limitation for the application of ecological niche models for threatened species, MaxEnt is recognized to be accurate and stable across all tested sample-size categories.

In addition, we could not fully examine snow as a factor in our study, because its impact was not obvious and data acquisition was difficult on such a spatial scale. However, there have been reports showing that snow becomes an influencing factor for deer at depths greater than 50 cm or in alpine areas (Dou et al., 2013; Ohashi et al., 2014; Ossi et al., 2015). However, in our field investigations, the average snow depth was 5–20 cm and the maximum value was 40 cm. We found that red deer would still feed in suitable foraging areas by pushing the upper snow aside. Generally, we observed that snow did not hinder the daily activities of the deer. Future surveys would benefit from more data, which would allow for better characterization of the suitable overlapping habitat between red deer and roe deer and a better understanding of the factors that shape their distributions. Other important variables to include in future models could be habitat fragmentation and interspecific interactions.

Management and conservational implications

The Lesser Xing’an Mountain forest has suffered from intense harvesting (including over-hunting of wildlife) since the Japanese invasion in the 1930s. Such a long history of human-driven unsustainable exploitation has resulted in the degradation of wildlife populations and increased sensitivity to disturbance. In addition, many non-forestry land transformations have occurred in the southern TFB areas over the last 30 years, although no quantitative data for landscape change are available at such a spatiotemporal scale. It is obvious that people have been cultivating farmland, building roads and expanding settlements, which have caused the destruction of natural habitats, inevitably transforming the TFB area into a farming–forestry ecosystem. Although recent years have seen few predators and ample food supplies, deer population density has remained very low (Jiang, Zhang & Ma, 2005; Jiang, 2007), a finding we suggest is caused by continued illegal hunting. At the same time, development of agriculture and under-forestry economy (such as understory planting) has led to increasingly isolated large patches, which should be protected. Large herbivores are still very vulnerable to negative anthropogenic effects. For example, results from our roe deer model (Fig. 4B) show that the total contribution of interference factors was 26.8%, which consisted of distance to settlement (21.4%), distance to management area (4.8%), and distance to road (0.6%). In our opinion, the population decline of deer in northeastern China is primarily driven by poaching, which could not be considered in the simulation.

Under conditions of low population density, an understanding of individual species is usually gained by studying habitat distribution. The predicted distribution generated by our study will thus help elucidate the state of cervid habitats. Previous research showed that cervids have similar behavioral characteristics and habitat preferences (Jiang, Zhang & Ma, 2005; Zhang & Zhang, 2010). Jiang (2004) demonstrated that red deer and roe deer feed on the same plant species in winter, whereas red deer and elk (Elaphurus davidianus) consistently overlapped with regard to height of vegetation for feeding. These findings indicate that cervids have similar dietary habits. Our research demonstrates that habitat overlap exists between red deer and roe deer. Overlapping habitats of similar species (e.g., red deer and roe deer) might also be used by rare species (e.g., sika deer and elk; Hu & Jiang, 2010).