City imagery is essential for enhancing city characteristics and disseminating city identity. As an emerging medium, short videos can intuitively reflect people’s perception of complex urban environment. In this study, we proposed a short video-driven deep learning perception framework to sense city imagery. To quantitatively deconstruct spatial imagery of urban space, deep neural network is used for pixel-level semantic segmentation. K-means clustering and hierarchical clustering analysis are carried out to extract and reveal the spatial imagery characteristics at the landmark level and the city level. Taking the Guangdong-Hong Kong-Macao Great Bay Area (GBA) as the study area, an experiment was carried out with TikTok short videos. The results showed that (1) the spatial imagery of the GBA cities are divided into four categories: Green Waterfront, including Jiangmen, Huizhou, Zhuhai, Zhaoqing, and Zhongshan; Humanistic Capital, including Hong Kong, Guangzhou, and Foshan; Modern Green City, including Shenzhen and Dongguan; Sky City, that is, Macao; (2) the landmark imagery in GBA can be characterized into five groups: Green Water and Blue Sky, Ancient Architecture of Greenery, Modern Architecture, Staggered Roads, and Urban Green Lung. It further investigated spatial distribution of landmark-level spatial imagery. These results prove the feasibility of sensing city imagery with short videos and provide useful insights into city imagery studies. It provides a new approach for understanding and spreading the city imagery over Internet.