Abstract: Land subsidence is a globally recognized important disaster, with different disaster areas experiencing various stages of development. Accurate understanding of the development of land subsidence is crucial for effective prevention and control. Aiming at the different characteristics of land subsidence at each development stage, a clustering method (improved adaptive density peak clustering algorithm based on K-nearest neighbors, IADPC-KNN) was proposed. Combining with granular computing theory, the development patterns of land subsidence and their mapping laws were summarized. Firstly, the dynamic time warping (DTW) method was used as the distance metric between data. IADPC-KNN and other five clustering algorithms were tested on seven public datasets. The results show that IADPC-KNN has higher accuracy and better robustness. Secondly, the land subsidence monitoring data from 14 affected regions around the world were collected, and four types of land subsidence patterns and their mapping relationships were obtained through data processing, sequence extraction, cluster analysis, granular model construction, and rule generalization. Finally, monitoring data from 2017 to 2019 for a specific site were used for validation. The results show that the land subsidence pattern for the site after 2018 has a probability of 0.359 2 of belonging to Mode 4, which is good agreement with the actual subsidence development. The research results provide a theoretical reference for predicting and preventing land subsidence disasters.

Key words: dynamic time warping, density peak clustering, granular model, cluster, land subsidence