Abstract: Recreation preference is one of the important characteristics of tourists' demand. Identifying the characteristics of national parks' recreation preferences and revealing its influencing factors is the key to evaluate the recreation service capacity of national parks. Taking the recreation exhibition area of Shennongjia National Park in Hubei province as the research area, 9538 samples were obtained by crawling geotagged images from tourism websites. By using computer deep learning algorithm, content recognition and image classification were carried out to clarify the tourists' preference for different types of recreational resources in national parks. ArcGIS spatial analysis tool and social network analysis method were used to explore the spatial and network characteristics of recreation preference. The Maximum Entropy Model was used to analyze the correlation between the influencing factors and recreation preferences. The study found that: (1) Tourists' preferences of recreation resource type with "Geological landscape, Biological landscape, Hydrological landscape, Celestial and climatic landscapes, Architecture and Facilities", a total of five categories including 21 basic types. (2) Recreation preference showed significant spatial distribution differences, showing the spatial network characteristics of "multi-core groups", with poor mobility between groups and high inter-group connection strength. (3) The preference for different recreational resources is mainly affected by landscape attraction, transportation convenience, facilities and other factors. Finally, some suggestions were put forward to optimize the utilization of recreation resources in Shennongjia National Park from the aspects of strengthening the function of science popularization education, promoting the linkage advantage of groups, and improving the attraction of landscape.

Key words: national parks, recreation preference, geotagged images, deep learning, space network structure