Towards Better Land Cover Classification Using Geo-tagged Photographs

A land cover map that represents the land surface of the earth is based primarily on analysis of remotely sensed images. However, the rate of concordance of existing land cover maps is not high. This lack of concordance results from a difference in classification methods and observation conditions of remotely sensed images. Also, conducting field surveys around the world is unrealistic. Therefore, we use ground level photographs from photo-sharing sites instead of field surveys. We propose a method to classify areas into land cover types using image features, geo-tags, titles and tags. Additionally, we create the land cover map using classified photographs. We evaluate the method using ground truth created manually. Results show that the accuracy of the proposed method is about 70 percent in New York.