Abstract: Traditional land cover classification process is very complicated, timeconsuming and labor-intensive, which requires huge amount of imagery data and involves many people. Recently, crowd-sourcing data have been used for land cover classification with lower costs, but they are still time-consuming due to the process of interpreting data. We examine the potential of textual information in point of interest (POI) as a new reference source. Firstly, POI textual data is analyzed to calculate the word distributions and topic distributions of POI using latent Dirichlet allocation (LDA) topic model. Secondly, support vector machine (SVM) algorithm is applied with topic distributions of POI to build a land cover classification model. Finally, we evaluate the land cover classification result by taking a random sample of remote sensing images. In the experiments, 1.9 million POIs from Weibo, Baidu and Gaode are used to test the proposed method, and result shows that a classification accuracy of over 80% is achieved.