首页 > , Vol. , Issue () : -
Geographic objects typically include both physical and human elements. The big data produced by remote sensing and the ubiquitous social media data provide rich sources for feature classification of these two types of objects. The extraction of physical objects based on remote sensing classification and the extraction and classification of social media information based on web text are the current mainstream methods of extracting geographic objects. The former is supported by image processing technology, whereas the latter is achieved with natural language processing technology. With the application of artificial intelligence classification methods, such as machine learning, the characteristics of the classification of the two types of elements are becoming increasingly common. Using the evolution of machine learning methods as a medium, in this study, we analyzed the similarities and differences between the remote sensing classification of single-and multiple-element physical geographic elements and the natural language processing classification of web text elements. Since the 1940s, the development of machine learning methods has experienced five stages: germination, development, bottleneck, recovery, and outbreak. Machine learning and related information classification methods have become the current focus of researchers. We described the principle applied by machine learning methods for geographic element classification. We divided the classification process of geographic elements into data acquisition, data preprocessing, feature construction or model training, and accuracy evaluation. Many similarities exist between the methods of physical-element-oriented remote sensing classification and human-element-oriented text classification in terms of the process and model; however, text and remote sensing classifications also differ because of their differences in data and specific tasks. Using single objects, compound object classification, and microblog social media topic classification extraction as three examples for remote sensing classification, we further determined the process of the completion of different geographic element classification tasks. We built a pixel-based CNN model to classify the water bodies in the Tuul River region of Mongolia. Land cover classification mainly adopts random forest, decision tree, maximum likelihood, support vector machine, and pixel-object-knowledge methods to map the global land cover. Social media classification was based on latent Dirichlet allocation and a random forest algorithm to classify the public sentiment on COVID-19 topics in microblogs. From the discussion, we noted commonalities and differences in the use of machine learning methods for the classification of the two types of geographical elements. The processes of remote sensing and text classification are generally consistent; remote sensing image classification and web-based text classification can learn from each other in many cases. Their differences include various focuses on data processing and the diverse targeting of methods. Text classification focuses on word separation and word vector construction, whereas image classification focuses on obtaining feature information, such as the spectrum, textures, and band indices of target objects. The combination of geographic element classification and artificial intelligence has considerable potential. With the development of big data and the mining and use of multisource heterogeneous data, the multimodal learning of joint text and images can provide new directions and ideas for geographic object research. The integrated fusion of geoscience-domain knowledge and deep learning methods is expected to become a mainstream trend for advancing remote sensing information extraction in the future. The mutual reference among classification methods for remote sensing and social media big data can widen the applications of intelligent classification of physical and human geographic elements.