Sampling bias for accessible locations skews estimates of community abundance

Landscape patterns of human accessibility influence the spatial distributions of samples across a wide array of studies.  Most ecological surveys are conducted close to roads and trails.  This apparent sampling bias begs the question, are inferences made to whole landscapes from such data correct?  Using a landscape model of human accessibility, I answer this question for a dataset of 1104 vegetation surveys spanning 40-years in Great Smoky Mountains National Park.  All but 16 of the 1104 vegetation plots occurred in the most accessible half of the Park.  I compared the distribution of vegetation communities captured by the surveys to the actual distribution of vegetation communities in the park as interpreted from aerial photos.  Of the communities that were over-sampled relative to their abundance in the park, most were positively correlated with accessibility.  Samples within a community tended to be concentrated close to roads, leaving large interior patches without samples.  Results similar to these are likely for many existing ecological surveys, implying that they may be poor predictors of the distribution and abundance of communities at a landscape scale.