What does a typical landscape in the continental United States look like? Is it a corn field? A deciduous forest? A suburban backyard? A shrubland? One's mental image of typical landscapes may well be quite subjective, perhaps strongly related to the landscapes of one's childhood or current field plot. Until recently, such a subjective image was perhaps the best approach, since an objective understanding of representative real-world landscapes has been limited by at least three factors: the lack of consistently classified land-use and land-cover data; skepticism of the robustness of landscape metric measures; and a tractable strategy for determining a representative set that appropriately spans the great variety presented by the nation's land surface. In this work we present and apply such a strategy to reveal "exemplar landscapes" and in them, the widespread signature of human effort upon the land cover and land use of the continental United States. Beginning with the National Land Cover Data Set classification, we partitioned the continental US into 190,000 landscapes of 6.48 x 6.48km. For each of 9500 randomly located landscapes in this set, we calculated 1200 class-level and landscape-level landscape metrics using the Fragstats software. Using this database of measurable characteristics of land use/land cover, we calculated the pairwise "similarity" between all 45 million pairs of landscapes. Using a recently developed algorithm for selecting the items best representing a given set, we extracted a small set of exemplar landscapes, then employed regression trees to explore factors underlying a random landscape's association with its chosen exemplar. We asked (1) whether the exemplars were visibly "different" from each other and distributed spatially as expected; (2) whether human factors played a role in distinguishing exemplars; and (3) whether the exemplar landscapes chosen by successive executions of the algorithm gave consistent results.
Results/Conclusions

Results were strikingly tied to human activity. In multiple runs, we found that the amount and/or configuration of Row Crops was frequently the most important distinguishing factor among exemplars. Other human land uses, such as Low-Intensity Residential and Fallow Land, were frequently seen as additional distinguishing factors. The "look" of the landscapes selected by independent runs of the algorithm on different random sets were remarkably similar, and appear to indicate the robustness of the representation presented here. These exemplars reveal the ability of landscape metrics to credibly identify differences among landscapes, and these results have multiple potential practical applications, including sampling design and habitat conservation.