A spectral clustering approach for multivariate geostatistical data

Spectral clustering has recently become one of the most popular modern clustering methods for conventional data. However, applied to geostatistical data, the general spectral clustering method produces clusters that are spatially non-contiguous which is certainly undesirable for many geoscience applications. In this paper, a spectral clustering approach is proposed, allowing to discover spatially contiguous and meaningful clusters in multivariate geostatistical data, in which spatial dependence plays an important role. The proposed spectral clustering approach relies on a similarity measure built from a nonparametric kernel estimator of the multivariate spatial dependence structure of the data, emphasizing the spatial correlation among data locations. It integrates existing methods to find the relevant number of clusters and to assess the contribution of variables in the formation of the clusters. The results from both synthetic and real-world datasets demonstrate that the proposed spectral clustering approach can effectively provide spatially contiguous and meaningful clusters.