A MULTISCALE STUDY OF SPATIOTEMPORAL DATA MINIING-CASE STUDY:PRECIPITATION PATTERN ANALYSIS

Abstract

The scale used in the input data is one of the key issues in conducting cluster analysis in spatiotemporal data mining. Results determined from different scales of the input data could be varied. To reduce the clustering uncertainties by using a fixed time scale, it will be useful to generate a two-dimensional scale-based data which covers a range of scales as input in cluster analysis. A multiscale rotated principal component analysis approach using the continuous wavelet transform is proposed to investigate the non-stationary characteristics of time series for the study of the spatial variability in the cluster analysis of rainfall. Experimental results of precipitation pattern analysis show that the proposed approach is capable of removing the local small features by using one small scale or improving the over-smoothed regions by using one large scale input.