资源选择函数拟合藏狐洞穴生境利用特征的有效性分析

The validity of using a resource selection functions model to predict den habitat of the Tibetan fox(Vulpes ferrilata)

资源选择函数(resourceselectionfunctions,RSFs)在分析野生动物栖息地特征以及预测有效生境等方面得到了广泛的运用,但是由于RSFs的理论基础的局限,使得该模型一直以来在研究低密度野生动物种群时的有效性存在很大的争议。藏狐(Vulpesferrilata)是一种低密度物种,我们通过对2001–2003年获得的133个藏狐洞穴样方和随机选取的133个环境样方拟合资源选择函数模型,并将模型结果和主成分分析(PCA)结果进行对比。结果显示RSFs在水源距离、鼠兔洞穴数量、坡向、坡度、坡位和植被类型等6个生境变量中,只对坡向、坡位和植被类型3个变量敏感且总预测率为75.2%,复相关系数为0.485(NagelkerkeR=0.235),同时3个变量的偏相关系数水平也很低。偏差分析(Akaike’sinformationcriterion,AIC)值为309.172,说明模型的预测偏差较大,判别效果不佳,不能有效提炼藏狐洞穴生境的特征因素。而PCA结果显示诸变量的重要性由高到低依次为:鼠兔洞穴数量、水源距离、坡度、坡位、植被类型和坡向,坡向的重要性最弱。我们还着重讨论了RSFs的理论基础和该模型在藏狐洞穴生境中失拟的原因,同时强调为了能对野生动物的生境特征进行比较全面的分析,应该综合多种方法。

Resource selection functions （RSFs） are widely used approach of mathematical modelling for the analysis of presence-absence data to deduce wildlife-habitat relationships. However, the generation of RSFs may be hampered by a wild animal＇s population size and density. Consequently, the prediction accuracy of RSFs for low density species is disputed. We examined the validity of RSFs in a study of the Tibetan fox （Vulpes ferrilata）, a low density species, in Shiqu County, Western Sichuan Province, China. Our RSFs model was constructed based on V. ferrilata habitat data collected from 2001 to 2003. Six environmental variables were considered with reference to V. ferrilata den habitat： water distance, vegetation type, pika （Ochotona sp.） den quantity, den location on slope, gradient, and aspect. In order to examine the validity of our RSFs model, we used a principal components analysis （PCA） to re-analyze our data. The total accuracy of our RSFs was determined to be 75.2%. The AIC value was 309.172. A ROC curve revealed, when the sensitivity was 0.857, the specificity of our RSF model was 0.353. The Nagelkerke R2 was 0.485. Only three variables were judged as important by our RSFs： aspect, position on the slope and vegetation type, of which aspect was the most informative variable of the three. However, partial correlation coefficients of the three variables were very low. These results revealed that the RSFs model to predict den habitat for V.ferrilata did not fit well. Considering the result of PCA, the importance ranks were （in order of decreasing importance）： pika den quantity, water distance, gradient, position on the slope, vegetation type, and aspect. We offer explanations as to why, for this species, a RSFs model may not be generated accurately and why the results of the two analyses were so different. We suggest other analytic methods that may be combined to yield a more comprehensive result for wild animal species with low population sizes and densities.