HU Su-xian ? JIANG Guang-shun ? MA Jian-zhang ZHANG Ming-hai ? LIU Hui

Introduction

Moose (Alces alces cameloides) is typically representative of the fauna of the frigid-temperate zone. In China, it is distributed only in the Greater Khingan Mountains and part of the Lesser Khingan Mountains. China lies on the southernmost edge of Moose’s distribution area in Asia. In recent years, both the area of distribution and the population of moose in China have been declining as a result of disturbance from anthropogenic factors; therefore,the moose has been put on the Chinese second priority list of protected animals (Wang 1998). Historically, moose were ubiquitous throughout the Greater Khingan Mountains and part of the Lesser Khingan Mountains. A survey in the 1970s estimated the total Chinese population of moose to be 18,538 individuals (Ma 1989). In 1987, another census showed a population of about 10,000 individuals, with about 6,000 individuals in the Greater Khingan Mountains and about 3,000 individuals in the Lesser Khingan Mountains, and the average density was 0.0519 individuals/km2. Hence, in 17 years, its population decreased by 46.2% at an annual decline rate of 6.3%. Its area of geographic distribution, which initially was 190,000 km2, shrank toward north and west by nearly 100?200 km, resulting in lost habitat of 55,102 km2accounting for 21.6% of its historical distribution. At the present situation, its population size and distribution may still be in decline (Piao et al. 1995).

The moose is the largest member of the deer family in body size and has no natural predator within the study area (Yu et al. 1993).However, human impacts pervade most ecosystems in northeastern China. The over exploitation of forests for timber production,the expansion of rural and suburban housing, and tourism has resulted in a greater presence of people across areas that were once exclusively habitats for moose. The impacts may result in a direct loss of habitat, but there may also be indirect loss of habitat as a consequence of avoidance behavior by affected wildlife(Mace and Waller 1996; Stevens and Boness 2003). These human disturbances have led to the identification of habitat degradation as a potential threat to the species.

A detailed understanding of habitat selection for moose is required to address the impact of habitat factors on the moose including human disturbances. In addition, landscape structure has important impacts on spatial distribution of a population (Stenseth 1980; Dempster and Pollard 1986; Dunning et al. 1992).Also, consideration of multi-scale variables is necessary for deciding how habitat data should be applied in resource management. It has been argued that abiotic factors of the environment are primary determinants of broad-scale distribution patterns for large herbivores (Bailey et al. 1996; Fortin et al. 2003). The purpose of our study was to better understand the influence of vegetation and human disturbance on the distribution of moose.

Materials and methods

Study area

The study was conducted at Heilongjiang Provincial Shengshan Forestry Farm (49°24′?49°35′N, 126°27′?126°55′E), located in the northwestern slope of Lesser Khingan Mountains, northeastern China, covering 155.6 km2.

The forest farm is located in the lower elevations of the mountains at average altitude of 450 m, with undulating terrain in relative height of 100?200 m, with average slopes of l0°. The weather is characterized by long cold winters and short hot summers. The annual average temperature is -2°C, and extreme temperatures range from -38.8°C to 34.3°C. Average annual precipitation ranges from 500 to 700 mm. The frost-free period is 100?120 days lasting from late April to late September. The vegetation flourishes and provides good cover for wildlife. In addition, a comprehensive drainage system merges to form a large area of wetland within the study site. Other ungulates in this forest region include roe deer (Capreolus pygargus bedfordi) and wild boar (Sus scrofa ussuricus). The only carnivorous species is the lynx (Lynx lynx), but it is extremely rare in our study area. There is only one village in the forest farm. A relatively sparse network of functioning roads and trails is associated with forestry operations.

Collection of resource and human disturbance variables

Field work was conducted 24 hours after snowfall from January to March in both 2006 and 2007. A total of 428 10 m × 10 m plots were placed every 200 m to avoid clustering along 28 sampling transects. Twenty eight transects, each more than 2 km in length,were located by randomly selecting the starting point of the first transect and extending that transect 2 km to the south. The remaining transects were established in an approximately parallel array at 2-km intervals extending to the west, traversing the whole study area and covering all habitat types (Fig.1). The GPS location at the center of each 10 m × 10 m plot was recorded. At each plot, we firstly examined the presence of moose and roe deer by identifying fresh snow tracks (left within 24 hours), beds, dung pellet (Chang and Xiao 1988; Zhang and Xiao 1990; Lennart 2002;Jiang et al. 2006), and then recorded or generated some habitat variable data at patch or landscape variables (Table1).

Vegetation types were determined by remote sensing. Digital forest cover and land use maps were generated by visual interpretation of Spot-5 imagery (resolution: 2.5 × 2.5 m, Volume:Network YZP0016533) taken on September 21, 2005. Aerial photographs were used to build more sensitive photo mosaics for classification of vegetation type. Photographs were scanned and mosaics rectified with remote sensing software (ERDAS Inc.ERDAS IMAGINE 8.5 Tour Guides 2001) and then digitized using Arc View GIS Version 3.1 (ESRI Inc.1996). Classification of vegetation types was primarily based on categories established by the Zhanhe Forest Bureau Forestry Management Map (2002).Five forest stands were identified using an unsupervised classification with an iterative self-organizing data analysis technique and a supervised classification based on expert knowledge gained in the field (ERDAS Inc. IMAGINE Subpixel-Classifier 8.5 User Guides 2001). The five cover types were Betula platyphylla stands,mixed deciduous broadleaf forest stands, Larix gmelinii forest stands, mixed coniferous and broad leaf forest stands, and low shrub and swamp.

We defined a vegetation patch as a single or homogeneous collection of pixels representing one cover type. We queried the percentage use of each cover type found across the study area.And then, we generated a second set of variables quantifying the density of patches representing the landscape variation in the availability of each cover type across the study area (Johnson et al.2004). We used a pattern analysis technique, 3-term local quadrat variance, to identify the distributional patterns of each land cover type. First, we randomly placed 10 north-to-south transects across each cover type. We then used an overlapping moving window,consisting of three terms or blocks, to calculate the variance in pixel (i.e. patch) occurrence along each transect. The variance was repeatedly calculated for each transect following incremental increases in block size. When plotted, peaks in variance corresponding with block size indicate patchiness (Dale 2000). We plotted the median variance for the 10 sample transects and identified the major peak in variance as one variable at which the broader-scale analysis of the vegetation differed beyond that of the individual patches (Johnson et al. 2004). We used that scale(i.e. distance) to identify the size of a rectangular moving window that was applied to a binary image of each land cover type to calculate and map the density of pixels (i.e. patches) across the study area. These analyses were conducted at a pixel resolution of 100 × 100 m.

The altitude was derived from a Digital Elevation Model (DEM)for the study area by using ArcView GIS Version 3.1 (ESRI Inc.,1996). The Sun index were calculated by the equation: cos(aspect)× tan (slope) ×100 (Wilson et al. 2001). Slope and aspect were generated using a digital elevation model (DEM). Viewshed wasgenerated from an elevation surface (TIN-Triangulated Irregular Network) with, for each plot, the number of hypothetical observers (one in the centre of each of the other plots) who can see the position. The result is a grid theme with visibility attributes assigned to every cell (plot) (Maichak 2004). A Normalized Difference Vegetation Index (NDVI) was derived from 21 September 2005 SPOT 5 imagery by using remote sensing software (ERDAS Inc., 2001). Greenness, wetness and brightness were produced using a tasseled cap image transformation (Crist and Ciscone 1984) from the 2005 Spot 5 image.

Table1. Variables used to model resource selection by moose, in Shengshan Forest Farm of northwestern slope of Lesser Khingan Mountains,northeastern China (January to March 2006 and January to March 2007).

Viewshed analysis determines areas visible and not visible on a grid or triangulated irregular network (Wang et al. 1996) from one(e.g. a sample site) to many observation sites. Results are stored as a temporary integer grid where the value of each cell in the grid equals the number of sites from which it can be seen (Ormsby and Alvi 1999). For example, cells with a value of 1, 2 and 3 can be seen from one, two and three observation sites, respectively. Cells with a value of 0 can not be seen from any observation site.Therefore, the higher Viewshed value a site has, the easier the site can be seen and the poorer its usefulness for concealment in the terrain. We calculate the Viewshed index of all survey plots by using a software application within the 3D extension (Ormsby and Alvi 1999) of ArcView GIS 3.1.

We hypothesized that the distribution of roe deer was correlated with the availability of moose and that the movement and behavior of moose was influenced by the presence of roe deer (Li et al. 1992). Therefore, we firstly modeled and generated maps of habitat selection for roe deer. This map, in turn, then provided spatially explicit covariates to evaluate the strength of interspecific interactions between roe deer and moose. Thus we can use occurrence of roe deer to infer interactions with moose (Johnson et al. 2005).

Spatial distances (i.e. the distance to nearest settlement, roads,trails, rivers), slope degree and aspect were measured by the spatial analysis model provided in ArcView 3.1. In addition,forest harvest interval was quantified by the interval between the present and the time of the most recent forest harvest at a given location, and we converted these logged locations into a grid surface by using ArcView 3.1 (ESRI Inc.1996).

Creation of resource selection function models

We used resource selection functions (RSF) to quantify the relationship between the observed distribution of moose and thevariables representing habitat characteristics and human disturbance. A RSF is any mathematical function that is proportional to the probability of use of a resource or habitat (Manly et al. 2002).Typically, a RSF consists of several coefficients (βi) that quantify selection for or avoidance of some environmental features. Coefficient sign and strength are results of differential variation in the distribution of each environmental feature measured at a sample of animal locations and a comparable set of absence sites. Habitat variables were tested for normality using Kolmogorov-Smirnov and Shapiro-Wilk tests. Some variables were highly skewed and needed to be normalized using standard transformation techniques.All variables were subsequently entered into a Pearson’s correlation matrix to identify collinearities between significant variables(i.e. rs≥0.65) (Loyn et al. 2001). Variables that had collinearities were examined and the variable that explained a greater portion of the deviance was retained. All significant variables were then entered into the model development procedure. We used an information-theoretic approach to guide model development and selection (Anderson et al. 2000). We used the Akaike Information Criterion (AICC) difference adjusted for small sample size and Akaike weights (w) to evaluate and choose the most parsimonious model (i.e., the fewest variables to explain the greatest amount of variation). Akaike weights provide a normalized comparative score for all specified models and are interpreted as the approximate probability that each model is the best from the set of proposed models (Anderson et al. 2000).

Following selection of the most parsimonious of the habitat selection models, we evaluated all combinations of terms for the disturbance covariates. Our work was premised on avoidance as the indicator of disturbance. Although we are confident that observed patterns of avoidance are a proxy for a real effect, we recognize the limitations of our data and the possibility of false negatives. Holding modeled disturbance effects constant and applying predetermined coefficients allowed us to easily assess the simple “footprint” impacts associated with current roads, trails,and the settlement. Similar approaches were used to assess the impacts of human actions on the availability of habitats for grizzly bears (Schoen et al. 1994; Dixon 1997; Suring et al. 1998; Jiang et al. 2007). Consequently we required the final model to have a relative low AICCand may select a convex Gaussian disturbance term indicative of avoidance of a disturbance feature (i.e.,nonlinear variables).

We used 95% confidence intervals to assess the strength of effect of each predictor covariate. Selection or avoidance cannot be inferred from covariates whose confidence intervals include 0.We used the Pregibon’s Δβ and leverage statistics and Hosmer and Lemeshow’s Δχ2statistic to identify cases and clusters that had a large influence on the parameters of the model (Hosmer and Lemeshow, 2000). Predictions generated from an RSF are relative measures of habitat selection, not true probabilities of the occurrence of particular behaviors. Therefore, we used the odds ratio as a relative measure of habitat selection (Menard 2001; Keating and Cherry 2004).

We populated Equation 1 with coefficients (β1,β2… βi) from regression models.

Here, we assumed that our definition of availability was functionally related to an animal’s response to human disturbance.

The AICCprovides evidence for selection of the most parsimonious model, but does not permit evaluation of discriminatory performance (Pearce and Ferrier 2000). We use Kappa statistic for evaluating classification effectiveness of a model. Kappa records overall agreement between predictions and observations, corrected for agreement expected to occur by chance. The statistic ranges from -1 to +1, where +1 indicates perfect agreement while values of zero or less suggest a performance no better than random(Cohen 1960). Important advantages of Kappa are the use of agreement, instead of association (traditional classification tables).This is particularly important when the prevalence of the species is low (Fielding and Bell 1997), a frequent phenomenon in wildlife RSF modeling because models often are developed for rare,threatened, or endangered species.

We completed the creation of logistic models and Kappa statistic evaluation using R2.4.1, a statistical data analysis program(R Development Core Team 2006), and its SPDEP, DESIGN and GLM packages (Ihaka and Gentleman 1996).

Results

While surveying for moose over two separate years, from January to March, 2006 and January to March, 2007, a total of 428 10 m ×10 m plots were examined within the study area (Fig.1). Fresh snow signs (within 24 hours) of moose use were found in 19 plots.Due to the random layout of the survey in each year, the majority of these plots were visited only once. Only a few plots, by coincidence, were near to plots of the past year. When the distance between plots from each of the two years was less than 200-m, we deleted one of them to avoid clustering.

Fig.1 Distribution of moose across the study area of the Heilongjiang Shengshan Forest farm (over two separate years from January to March, 2006 and January to March, 2007).

Resource selection function models

The top four RSF models were considered for resource and human disturbance variables at multiple scales, respectively, exhibiting significant univariate difference between “use” and “non-use”survey sites.

Table2. Number of model parameters (k) , AICc Δ and wi, and Kappa value for the top four logistic regression models for moose in Shengshan Forest Farm (January to March, 2006 and January to March,2007).

Table3. Coefficients (β) and 95% confidence intervals from resource-selection models for moose in Shengshan Forest Farm.

For the resource selection of moose, the first most parsimonious resource model consisted of nine covariates for vegetation patch density, NDVI, altitude, wetness, occurrence of roe deer, human disturbances and their nonlinear terms for human disturbances(k=11, AICw=0.311); the second most parsimonious resource model incorporated Larix gmelinii patch density and had a slightly lower AICw (k = 12, AIC w =0.298) (Table2). Although the top four logistic models all had predictive capacity (Kappa ＞ 0), the first most parsimonious model has the greatest AIC weight (AIC w =0.311) and the best predictive capacity among them (Kappa =0.77) (Table2). Consequently, we selected the first most parsimonious logistic model as our final predictive equation. Moose distribution was positively related to areas with higher densities of mixed deciduous broadleaf patch and mixed coniferous and broad leaf patch, and a higher NDVI value. It was negatively related to altitude and wetness, positively to occurrence of roe deer, for these covariates had confidence intervals that approached 0 (Table3).

Response of moose to human disturbances

That model contained linear and non-linear terms for distance to human disturbances, which demonstrated a convex relationship with a peak at 6 km from the settlement and odds ratio of moose occurrence of moose remained relative low within 3 km from roads, whereas relative high within 4 km from trails (Fig.2). Our results showed that moose avoided the disturbance from settlement and roads drastically, but selected the areas near the trails.

Fig.2 Plots representing the likelihood of occurrence of moose relative to three types of human disturbance features (a, b and c) found across the study area of the Heilongjiang Shengshan Forest Farm.Animal occurrence was allowed to vary with distance from disturbance features while other covariates were held at their mean values.

Discussion

Scale is an often used, but poorly defined and understood term(Csillag et al. 2000). Scalar principles need to be considered in a context specific manner. When representing scalar processes, one could consider the scale of the phenomenon or the scale of measurement, statistical summary, and modeling (Turner et al.1989; Dugan et al. 2002). Our study is the first to reveal resource selection and response of moose at patch and landscape scales in northeastern China. Previous studies of moose winter habitat selection in China only conducted at microhabitat or forest stands scales and did not emphasis the influence of more broad-scale and human disturbance on their ecology (Yu et al. 1991, 1993; Zhang,2001). In this area, our results suggested that moose selected areas with higher densities of mixed deciduous broadleaf patch, mixed coniferous and broad leaf patch, and a higher NDVI value and were disturbed by human beings at landscape scale. This suggests that moose responded to broad-scale elements. This concurs with Maier et al. (2005) who reported that moose, in interior Alaska,tended to occur in areas with large high-density patches of varied habitat and selected heterogeneous habitat. NDVI was our surrogate for vegetation quality and quantity (Cihlar et al. 1991).NDVI is highly correlated with leaf-area index (Cihlar et al. 1991).Annual shoots are the main food for moose during winter (Yu et al.1993) and food availability is a limiting factor for moose distribution during winter (Dussault et al. 2005), which together mean that there would be an increased use by moose of those forest stands that provide more abundant food (Dussault et al. 2005).The places where NDVI in September was higher would provide more abundant food, in the form of annual deciduous shoots, for moose during winter. During this time, in choosing particular relatively small areas of forest, moose make trade-offs among the available habitat resources to achieve the most favorable energy balance (Daussalt et al. 2005). Consequently, land and forest managers should arrange the logging sites to maintain or create the heterogeneous and abundant food habitat to benefit the survival of moose.

In this study, human disturbance has been revealed to be important factors influencing the distribution of moose populations.However, response of moose to different human disturbance significantly differed. Moose selected area near trails, whereas avoided settlement and roads drastically. Presence of trails showed a significant tendency for moose to be attracted to trails.These trails without human disturbance could serve as convenient travel corridors for moose and could also provide large and easily accessible amounts of forage along their edges. As such, it is understandable that trails within the study area would be utilized rather than avoided by moose (Wisdom and Cook 2000).Avoidance of roads and settlement by moose may lead to the indirect loss or fragmentation of the habitat (Jeager and Fahrig 2004; Jaeger et al. 2005). Similarly study by Jiang et al. (2007)indicated that movement of red deer avoided forest roads from 1.6 km, village from 8.2 km and abandoned roads from 2.2 km.During recent decades, with the rapid development of forest economics, more human settlements and denser road network have been built up in northeastern China. In addition, avoidance responses were greatest for major developments associated with the highest level of human activities and our results approach the maximum of a widely reported threshold distance of 500 m to 5 km from a disturbance (Czech 1991; Nellemann and Cameron 1998; Vistnes and Nellemann 2001; Mahoney and Schaefer 2002;Frid 2003). Thus, human disturbance in this area were seriously impacting spatial distribution of moose populations.

Management implications

The study concludes that moose selected heterogeneous habitat with high-density patches of mixed deciduous broadleaf patch and mixed coniferous and broad leaf stands, and the behavior of avoidance for human disturbance (i.e. settlement and roads) may lead to indirect habitat loss. Hence, land and forest managers should maintain such habitat by adopting suiTableforest harvest patterns. They should focus on quality of habitat by restoring and managing the vegetation density and cover. Moreover, they should take measures to mitigate avoidance of moose for village by guiding activities of local people. They can minimize the influence of roads by closing some roads in key habitat, and protect the areas near the trails by controlling human activities disturbances (Millspaugh et al. 1998; Nielsen et al. 2009).

Acknowledgements

We extended our thanks to the staff of the Heilongjiang Shengshan Naure Reserve Management Bureau for their great help,especially associate director Mr. W.G. Zhang.

Anderson DR, Burnham KP, Thompson WL. 2000. Null hypothesis testing:problems, prevalence, and an alternative. Journal of Wildlife Management,64: 912?923.

Bailey DW, Gross JE, Laca EA, Rittenhouse LRR, Coughenour MB, Swift DM,Sims PL.1996. Mechanisms that result in large herbivore grazing distribution patterns. Journal of Range Management, 49: 386?400.

Chang H, Xiao QZ. 1988. The selection of red deer habitat in Dailing region during the winter. Acta Theriologica Sinica, 8: 81–88. (In Chinese with English summary)

Cihlar J, Laurent LST, Dyer JA. 1991. Relation between the normalized difference vegetation index and ecological variables. Remote Sensing of the Environment, 35: 279–298.

Cohen JA. 1960. A coefficient of agreement for nominal scales. Educational and Psychological Measurement, 20: 37–46.

Crist EP, Ciscone RC. 1984. Application of the tasseled cap concept to simulated thematic mapper data. Photogrammetric Engineering and Remote Sensing, 50: 343–352.

Csillag F, Fortin MJ, Dungan JL, 2000. On the limits and extensions of definition of scale. Bulletin of the Ecological Society of America, 81: 230–232.

Czech B. 1991. Elk behavior in response to human disturbance at Mt. St.Helens National Volcanic Monument. Applied Animal Behaviour Science,29: 269–277.

Dale MRT. 2000. Lacunarity analysis of spatial pattern: a comparison. Landscape Ecology, 15: 467–478.

Dempster JP, Pollard E. 1986. Spatial heterogeneity, stochasticity, and the detection of density dependence in animal populations. Oikos, 46: 413–416.

Dixon BG. 1997. Cumulative effects modeling for grizzly bears in the greater Yellowstone Ecosystem. MSc Thesis, Montana State University, Bozeman,Montana, USA: 15–35.

Dugan J, Perry JN, Dale MRT, Legendre P, Citron-Pousty S, Fortin MJ,Jakomulska A, Miriti M, Rosenberg MS. 2002. A balanced view of scale in spatial statistical analysis. Ecography, 25: 626–640.

Dunning JB, Danielson BJ, Pulliam HR. 1992. Ecological processes that affect populations in complex landscapes. Oikos, 65: 169–175.

Dussault C, Ouellet JP, Courtois R, Huot J, Breton L, Jolicoeur H. 2005.Linking moose habitat selection to limiting factors. Ecography, 28: 619–628.ERDAS Inc. 2001. ERDAS IMAGINE 8.5 Tour Guides.

ESRI (Environmental System Research Institute). 1996. Using ArcView GIS.ESRI, Inc., Redlands, Califonia, USA.

Fielding AH, Bell JF. 1997. A review of methods for the assessment of prediction errors in conservation presence/ absence models. Environmental Conservation, 24: 38–49.

Fortin D, Fryxell J M, O’Brodovich L, Frandsen D. 2003. Foraging ecology of bison at the landscape and plant community levels: The applicability of energy maximization principles. Oecologia, 134: 219–227.

Frid A. 2003. Dall’s sheep responses to overflights by helicopter and fixed-wing aircraft. Biological Conservation, 110: 387–399.

Hosmer DW, Lemeshow S. 2000. Applied logistic regression. John Wiley and Sons, New York, USA.

Ihaka R, Gentleman R. 1996. R: A language for data analysis and graphics.Journal of Computational and Graphical Statistics, 5: 299–314.

Jaeger JAG, Bowman J, Brennan J, Fahrig L, Bert D, Bouchard J, Charbonneau N, Frank K, Ruber B, Tluk von Toschanowitz K. 2005. Predicting when animal populations are at risk from roads: an interactive model of road avoidance behavior. Ecological Modelling, 185: 329–348.

Jaeger JAG, Fahrig L. 2004. Effects of road fencing on population persistence.Conservation Biology, 18: 1651–1657.

Jiang GS, Ma JZ, Zhang MH, Stott P. 2009. Multiple spatial scale resource selection function models in relation to human disturbance for moose in northeastern China. Ecological Research, 24: 423–440.

Jiang GS, Ma JZ, Zhang MH, Stott P. 2010. Multi-scale foraging habitat use and interactions by sympatric cervids in northeastern China. Journal of Wildlife Management, 74: 678–689.

Jiang GS, Ma JZ, Zhang MH. 2006. Spatial distribution of ungulate responses to habitat factors in Wandashan Forest Region, northeastern China. Journal of Wildlife Management, 70: 1470–1476.

Jiang GS, Ma JZ, Zhang MH. 2007. Effects of human disturbance on movement, foraging and bed selection in red deer Cervus elaphus xanthopygus from the Wandashan Mountains, northeastern China. Acta Theriologica, 52:435–446.

Johnson CJ, Boyce MS, Case RL, Cluff HD, Gau RJ, Gunn A, Mulders R. 2005.Cumulative effects of human developments on Arctic wildlife. Wildlife Monographs, 160: 1–36.

Johnson CJ, Boyce MS, Mulders R, Gunn A, Cluff HD, Case RL. 2004.Quantifying patch distribution at multiple spatial scales: application to wildlife-habitat models. Landscape Ecology, 19: 869–882.

Keating KA, Cherry S. 2004. Use and interpretation of logistic regression in habitat selection studies. Journal of Wildlife Management, 68: 774–789.

Lennart H. 2002. Mammal movements and foraging at remnant woodland inside coniferous forest landscape. Forest Ecology and Management, 160:109–114.

Li YZ, Xiao QZ, Chen HP. 1992. Interspecies relation among moose, red deer and roe deer during winter in Shengshan Forest Farm, Heilongjiang Province.Acta Theriologica Sinica, 2: 110–116. (In Chinese with English summary)

Loyn RH, McNabb EG, Volodina L, Willig R. 2001. Modelling landscape distributions of large forest owls as applied to managing forests in north-east Victoria, Australia. Biological Conservation, 97: 361–376.

Ma YQ. 1989. Wildlife in Greater Khingan Mountains. Harbin, China: Northeast Forestry University Press, 79–93.(In Chinese)

Mace RD, Waller JS, Manley TL, Lyon LJ, Zuuring H. 1996. Relationships among grizzly bears, roads and habitat in the Swan Mountains, Montana.Journal of Applied Ecology, 33: 1395–1404.

Mahoney SP, Schaefer JA. 2002. Hydroelectric development and the disruption of migration in caribou. Biological Conservation, 107: 147–153.

Maichak EJ. 2004. Applicability of viewshed analysis to wildlife population estimation. The American Midland Naturalist, 152: 277–285.

Maier JAK, Ver Hoef JM, McGuire AD, Bowyer RT, Saperstein L, Maier HA.2005. Distribution and density of moose in relation to landscape characteristics: effects of scale1. Canadian Journal of Forest Research, 35:2233–2243.

Manly BFJ, McDonald LL, Thomas DL, McDonald TL, Erickson WP (eds).2002. Resource selection by animals. Second edition. Dordrecht: Kluwer Academic Publishers, pp.1–221.

Menard S. 2001. Applied logistic regression analysis. Sage University Paper series on Quantitative Applications in the Social Sciences, series no.07–106.Thousand Oaks, California, USA.

Millspaugh JJ, Raedeke KJ, Brundige GC, Willmott CC. 1998. Summer bed sites of Elk (Cervus elaphus) in the Black Hills, South Dakota: considerations for thermal cover management. The American Midland Naturalist, 139:133–140.

Nellemann C, Cameron RD. 1998. Cumulative impacts of an evolving oil-field complex on the distribution of calving caribou. Canadian Journal of Zoology,76: 1425–1430.

Nielsen SE, Cranston J and Stenhouse GB. 2009. Identification of priority areas for grizzly bear conservation and recovery in Alberta, Canada. Journal of Conservation Planning, 5: 38–60.

Norusis MJ. 1999. SPSS for Windows Regression Models. 9.0. Englewood Cliffs: Prentice Hall, p.135.

Ormsby T, Alvi J. 1999. Extension ArcView GIS. Redlands, Calif., USA: ESRI Press.

Pearce J, Ferrier S. 2000. Evaluating the predictive performance of habitat models developed using logistic regression. Ecological Modelling, 133:225–245.

Peng CJ, Lee KL, Ingersoll GM. 2002. An introductionto logistic regression analysis and reporting. The Journal of Educational Research, 96: 3–14.

Piao RZ, Guang GS, Zhang MH. 1995. Population size and distribution of moose in China. Acta Theriologica Sinica, 15: 11–16. (In Chinese with English summary)

R Development Core Team. 2006. R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria.ISBN 3-900051-07-0, URL http://www.R-project.org.

Schoen JW, Flynn RW, Suring LH, Trrus K, Beier LR. 1994. Habitat-capability model for brown bear in Southeast Alaska. Bears: Their Biology and Management, 9: 327–337. (In: proceedings of symposia of theInternational Conference for Bear Research and Management),.

Stenseth NC. 1980. Spatial heterogeneity and population stability: some evolutionary consequences. Oikos, 35: 165–184.

Stevens MA, Boness DJ. 2003. Influences of habitat features and human disturbance on use of breeding sites by a declining population of southern fur seals (Arctocephalus australis). Journal of Zoology, 260: 145–152.

Suring LH, Barber KR, Schwartz CC, Bailey TN, Shuster WC, Tetreau MD.1998. Analysis of cumulative effects on brown bears on Kenai Peninsula,southcentral Alaska. Ursus, 10: 107–117.

Turner MG, Dale VH, Gardner RH. 1989. Predicting across scales: theory development and testing. Landscape Ecology, 3: 245–252.

Vistnes I, Nellemann C. 2001. Avoidance of cabins, roads, and power lines by reindeer during calving. Journal of Wildlife Management, 65: 915–925.

Wang J, Robinson GJ, Wiitie K. 1996. A fast solution to local viewshed computation using grid based digital elevation models. Photogrammetric Engineering and Remote Sensing, 62: 1157–1164.

Wang S. 1998. China red data book of endangered animals: Mammalia.Beijing, Hongkong, New York: Science Press, pp.246–249. (In both Chinese and English)

Wilson BA, Aberton JG, Reichl T. 2001. Effects of fragmented habitat and fire on the distribution and ecology of the swamp antechinus (Antechinus minimus maritimus) in the eastern Otways′ V ictoria． Wildlife Research, 28:527–536.

Wisdom MJ, Cook JG. 2000. North American elk. In: S. Demarais, and P.R.Krausman (eds), Ecology and management of large mammals in North America. Prentice-Hall Incorpotated, Upper Saddle River, pp.694–735

Yu XC, Xiao QZ, Lu YM. 1993. Selection and utilization ratio of winter diet,and seasonal changes in feeding and bedding habitat selection by moose in northeastern China. In: N. Ohtaishi and H.L. Sheng (eds), Deer of China:biology and management. Amsterdam: Elsevier Science Publishers,pp.172–180.

Yu XC, Xiao QZ. 1991. The food composition and seasonal change of the moose in Heihe Forest Area. Acta Theriologica Sinica, 11: 258–265. (In Chinese with English summary)

Zhang MH, Xiao QZ. 1990. A study on feeding and bedding habitat selection by red deer in winter. Acta Theriologica Sinica, 10: 175–183. (In Chinese with English summary)

Zhang MH. 2001. Wintering habitat selection by moose cold temperate zone,The Great Khingan Mountains. Acta Theriologica Sinica, 21: 310–313. (In Chinese)