A dual spatial clustering method in the presence of heterogeneity and noise

Abstract

The detection of spatial clusters, taking into account both spatial proximity and attribute similarity, plays a vital role in spatial data analysis. Although several dual clustering methods are currently available in the literature, most of them have detected homogeneous spatially adjacent clusters suffering from between‐cluster inhomogeneity and noise, where those spatial points have been described in the attribute domain. This article aims to accommodate both spatial proximity and attribute similarity with the presence of heterogeneity and noise. In this algorithm, Delaunay triangulation with edge‐length constraints, with consideration of arbitrary geometrical shapes, different densities, and spatial noise, is first utilized to construct spatial proximity relationships among points. Then, a clustering strategy employing information entropy is designed to identify clusters having similar attributes. The attribute clustering can adaptively and accurately detect clusters under the consideration of heterogeneity and noise. The efficacy and practicability of the proposed algorithm are illustrated by experiments employing both simulated datasets and real spatial point events.

Incorporation of Spatially Heterogeneous Area Partitioning into Vector-Based Cellular Automata for Simulating Urban Land-Use Changes

“…DSC aims to address the challenges of heterogeneity and noise by incorporating both spatial proximity and attribute similarity [36]. In real-world scenarios, spatially adjacent clusters usually exist in a spatial dataset where the difference of observations in attribute distribution is homogeneous within each cluster but inhomogeneous between clusters.…”

“…However, in the face of uneven distributions in the attribute space, attribute similarity measurements in these algorithms primarily relied on a binary predicate that utilizes Euclidean distance as the fundamental metric; their inherent transitivity could lead to the continuous propagation and accumulation of differences between attribute values during the clustering process. As a result, the clustering results may fail to accurately reflect the transitional nature of geographical features in spatial distributions, eventually leading to over-or under-simulation results (the validation of this point was demonstrated using both simulated and real-world data in [36]). The DSC algorithm primarily addresses the challenge of discovering homogeneous spatially adjacent clusters while dealing with between-cluster inhomogeneity and noise where those spatial points are described in the attribute domain.…”

“…As a result, the clustering results may fail to accurately reflect the transitional nature of geographical features in spatial distributions, eventually leading to over-or under-simulation results [34,35]. One dual spatial clustering algorithm, denoted as DSC, can handle both spatial proximity and attribute similarity in the presence of heterogeneity and noise [36]. The detection of these clusters is valuable for gaining insights into the localized patterns of geographical phenomena, and it has been successfully used for urban element identification and urban spatial structure analysis [37,38].…”

Using Dual Spatial Clustering Models for Urban Fringe Areas Extraction Based on Night-time Light Data: Comparison of NPP/VIIRS, Luojia 1-01, and NASA’s Black Marble

“…DSC aims to address the challenges of heterogeneity and noise by incorporating both spatial proximity and attribute similarity [34]. DSC initiates by employing DT with edge-length constraints, which takes into account arbitrary geometrical shapes, different densities, and spatial noise, to establish spatial proximity relationships among mutation points.…”

“…The dual spatial clustering method can take into account both spatial proximity and attribute similarity features, allowing for a better exploration of the distribution patterns and trends of geographic spatial entities [32]. Among them, there are three representative dual spatial approaches, which include MK-Means [33], DBSC [32], and DSC [34]. The MK-Means algorithm is an extension of the K-means algorithm that incorporates attribute metrics to broaden the focus on the spatial object's attribute distance, aiming to consider both the heterogeneity of spatial positions and the similarity of attributes.…”