<p>Grasping where and what kind of sightseeing spots are distributed is important when making a tourism plan for an unfamiliar area by yourself. Many of the existing tourism websites introduce popular spots in a ranking format for each administrative division. However, tourism resources are often distributed over administrative divisions, so it is not always useful. Therefore, in this study, we propose a method of extracting areas where features that are not in the surrounding area gather, using geotagged photos posted on Flickr, which has many users of foreign tourists. Specifically, we calculate location information and class probabilities of subjects from geotagged photos. We construct a minimum spanning tree by connecting POIs so as to satisfy the positional cohesiveness from their location information. For each edge of the minimum spanning tree, we calculate how much the feature distribution differs between the sub-areas at both ends, and divide the minimum spanning tree into semantically cohesive sub-areas by cutting the edges that differ significantly. Finally, we extract area-specific photographs rarely taken in the surrounding regions and utilize them as annotations for each divided area. In evaluation experiments using photographs taken in four regions in Japan, we quantitatively demonstrate that the proposed method is superior to existing methods in terms of positional cohesiveness and semantical cohesiveness of POIs in the divided areas. Furthermore, we quantitatively evaluate that each divided area has unique feature distribution compared with that of the entire target area.</p>