A novel model-based clustering approach for massive datasets of spatially registered time series. With application to sea surface temperature remote sensing data
Francesco Finazzi, Marian Scott · Aisberg (University of Bergamo) · 2015
Massive datasets of spatially registered time series are common when dealing with environmental phenomena observed through remote sensing. Datasets of this kind often include millions of time series and they are usually characterized by missing data. A possible way to extract useful information from the data is to cluster the time series with respect to their temporal pattern. When the clustering result is mapped over the geographic space, it often exhibits a spatial pattern and this helps to better understand the environmental phenomenon under study. In this work we propose a model-based clustering approach which is suitable for a large number of time series with missing data. A modified version of the EM algorithm is proposed to jointly estimate the model parameters, the cluster membership and the number of clusters. The approach is applied to a sea surface temperature (SST) dataset with more than 4.8 million time series covering 10 years.