基于低通保边滤波和尺度约束大津法的网格空间多尺度聚类方法

Grid-Based Spatial Multi-scale Clustering by Synthesizing Edge-Preserving Low-Pass Filtering and Scale-Aware Otsu Method

现有网格空间多尺度聚类方法未能将尺度因子作为模型参数实现尺度驱动的阈值提取,导致算法调参困难,难以全面挖掘空间数据的分布模式。海量空间点数据蕴含的信息更丰富,层次结构也更复杂,对聚类算法的参数自动化和计算效率提出了更高的要求。针对上述问题,该文从数据尺度和观察尺度提出了一种适用于海量数据的多尺度聚类挖掘方法:分别通过网格多分辨率和低通保边滤波器的尺度拓展机制实现了数据尺度和观察尺度上的尺度变换;将观察尺度层级作为参数引入大津法中,实现了观察尺度的密度阈值自动提取。实验结果表明:相比于传统低通滤波,该滤波方法具有良好的去噪保边效果;多尺度密度阈值提取算法能够有效地捕捉数据集中丰富的多层次信息,且计算复杂度低,可用于快速挖掘各类海量空间点数据中的多层次空间结构。

The existing grid-based multi-scale clustering methods fail to take scale factors as clustering model parameters to achieve the scale-driven threshold extraction,which impedes clustering parameter adjustment and optimization,and hinders the fully exploration of the spatial distribution patterns of the study dataset.Improvement of parameter automatization and calculation efficiency of clustering algorithms needs to do to process the emerging big spatiotemporal point data which contain more abundant information and have more complex hierarchical structure,comparing with traditional spatial dataset.To address above issues,this paper designs a multi-scale clustering algorithm that suitable for big spatial point dataset based on data scale and view scale.This method achieves data scale transformation and view scale transformation through the scale expansion mechanism of the grid multi-resolution and the edge-preserving low-pass filtering respectively.By introducing the view scale as a parameter into the Otsu method,the automatic extraction of multiple view scale density thresholds can be achieved.The experimental results show that,the proposed filtering method has good denoising and edge-preserving effect comparing with traditional low-pass filtering methods.Meanwhile,the adaptive multi-scale density threshold extracting algorithm can capture multi-level spatial structure information of the dataset more effectively with low computational complexity.The proposed clustering method can be utilized for rapid mining and visual analysis of multi-level spatial structures of various big spatial point dataset.