Currently, Reverse Geo-tagging relies on the keywords describing an image and use probabilistic algorithms to guess the localization of the depicted scene. However, such algorithms still perform poorly and show clear limitations Notably, the location estimation only occurs at the landmark level; regions or countries are only processed through their centroid. In this paper, we address this particular issue by exploring a semantic approach, which identifies geographical entities among the keywords to localize the picture (being a landmark or a country). We leverage the Linked Open Data cloud to find possible entities. The benefits of our approach, as opposed to numerical approaches, include an in-depth study of the ``geo-relevance'' of an image.