Statistical clustering method of spatial point events based on Voronoi diagram
Clustering spatial point events aims to identify hotspots where these events occur frequently in specific spatial regions.Exploring clustering patterns of spatial point events in geospatial data is crucial for disease outbreak warning,crime hotspot analysis,urban facilities planning,and many other fields.Classical methods exist for recognizing spatial point aggregation patterns of different shapes,different densities,and sizes.However,most of these methods lack statistical discrimination of spatial point aggregation patterns,potentially leading to unreliable clustering results.For example,existing clustering algorithms like k-means,DBSCAN,AUTOCLUST may group randomly distributed spatial points into clusters.While methods like spatial scan statistics can statistically infer the significance of spatial point clustering patterns,they are mainly designed for identifying clusters with spherical or elliptical shapes.Limitations arise in recognizing significant spatial point clusters with arbitrary shapes and different densities,especially along streets or rivers.Recently,many attentions have been paid to clustering of spatial point events,discovery of statistically significant point clusters with varied densities and irregularly shape is still a challenge work.In order to identify statistically significant spatial point clusters with different shapes,densities,and sizes,a statistical spatial point clustering method based on the Voronoi diagram is proposed in this paper.The Voronoi diagram is initially constructed based on the original spatial points to measure the aggregation of their distribution.A smaller Voronoi cell area indicates a more clustered distribution of points around a spatial point.Using the Voronoi diagram,this paper transforms the spatial point clustering problem into a task of detecting spatial hotspots the commonly used local Gi* statistic is used to detect statistically significant hotspots(high-density points or seed points)in local areas.Further,these high-density points or seed points are used to hierarchically merge neighboring spatial points,generating candidate clusters.The significance of these candidate clusters is then evaluated using Monte Carlo permutation tests.Experimental results on simulation datasets and real-world crime event data in Portland,USA,show that the proposed method effectively identifes significant spatial point clusters of arbitrary shapes and different densities,avoiding the generation of false clustering patterns.Comparative analysis with existing spatial point clustering algorithms,including DBSCAN and spatial scan statistics,indicates that the proposed method can effectively identify significant spatial point clusters of different shapes and sizes,while being less sensitive to noise points.The performance of the proposed method is shown to be superior to existing representative.
spatial point clusteringsignificant patternsspatial data miningstatistical testingcrime hotspot analysisVoronoi diagram
NETL
NSTL
万方数据
