Graph-based block-level urban change detection using Sentinel-2 time series

Remote sensing technology has been frequently used to obtain information on changes in urban land cover because of its vast spatial coverage and timeliness of observation. Block-level change detection with high temporal resolution image data provides fine detail of urban changes, is suitable for urban management, and has gradually received widespread attention. High-dimensional features are required to express the heterogeneous structure of the blocks. High-dimensional high-frequency time series, namely, multivariate time series, are formed by arranging high-dimensional features chronologically. Classic change detection methods treat multivariate time series as univariate time series one by one. Few studies have analyzed the change in a multivariate time series by considering all variables as an entirety. Therefore, a graph-based segmentation for multivariate time series algorithm (MTS-GS) is proposed in this paper. Specifically, 1) we construct a similarity matrix to explore the changing patterns of multivariate time series for seasonal change, trend change, abrupt change, and noise disturbance; 2) a multivariate time series graph is defined based on the changing patterns; and 3) the corresponding graph segmentation algorithm is proposed in the paper to detect the abrupt and trend changes under noise and seasonal disturbances. Sentinel-2 images of the rapidly developing third-tier city of Luoyang, Henan province, China, are adopted to validate the algorithm. The F1-score in the spatial domain is 84.1%; the producer's and the user's accuracy in the temporal dimension are 81.8% and 80.1%, respectively. Seven change types are defined and extracted, showing the development pattern and the efficiency of land use in the city. Furthermore, the proposed MTS-GS can be used for pixel-level change detection and performs well under various time intervals and cloud covers.