A self-supervised detection method for mixed urban functions based on trajectory temporal image

Urban function detection plays a significant role in urban complex system recognition and smart city construction. The location big data obtained from human activities, which is cohesive with urban functions, provides valuable insights into human mobility patterns. However, as urban functions become highly mixed, existing feature representation structures struggle to explicitly depict the latent human activity features, limiting their applicability for detecting mixed urban functions in a supervised manner. To close the gap, this study analogizes the latent human activity features to the shape, texture, and color semantics of images, with a contrastive learning framework being introduced to extract image-based crowd mobility features for detecting mixed urban functions. Firstly, by translating human activity features into image semantics, a novel feature representation structure termed the Trajectory Temporal Image (TTI) is proposed to explicitly represent human activity features. Secondly, the Vision Transformer (ViT) model is employed to extract image-based semantics in a self-supervised manner. Lastly, based on urban dynamics, a mathematical model is developed to represent mixed urban functions, and the decomposition of mixed urban functions is achieved using the theory of fuzzy sets. A case study is conducted using taxi trajectory data in three cities in China. Experimental results indicate the high discriminability of our proposed method, especially in areas with weak activity intensity, and reveal the relationship between the mixture index and the trip distance. The proposed method is promising to establish a solid scientific foundation for comprehending the urban complex system.