Clustering of spatial data with implicit polygonal structure based on topological approaches

Clustering is one of the fundamental approaches for data mining, which in the field of geoinformatics and image processing is used to search for knowledge and hidden patterns of spatial information. During automatic vectorization of objects on satellite images due to imperfections of these technologies, missing elements appear on linear and polygonal objects, which prevent full-fledged data analysis and visualization. The paper considers the problem of clustering geometric primitives with implicit polygonal structure with the possibility of eliminating incomplete data in vector models. The proposed method is based on the iterative formation of spatial structures by stretching the original linear objects. Unlike many clustering approaches, elements are grouped into clusters not by the principle of nearest Euclidean distance, but by determining the nearest intersection between segments. This approach allows us correctly dividing adjacent objects into different clusters. For the spatial structures formed at each iteration, their topological features are calculated depending on the stretching coefficient, which makes it possible to detect and filter implicit polygonal structures. The developed method is tested for clustering of linear geometric primitives on vector models of urban infrastructure. The performance of the method is compared with its competitors: k-means, DBSCAN, and agglomerative clustering. The research has shown that using the clustering quality assessment metric in the form of inertia and Jaccard indices, the proposed method has an advantage due to the correct separation of closely located clusters.