A robust hierarchical clustering for georeferenced data

The detection of spatially contiguous clusters is a relevant task in geostatistics since near located observations might have similar features than distant ones. Spatially compact groups can also improve clustering results interpretation according to the different detected subregions. In this paper, we propose a robust metric approach to neutralize the effect of possible outliers, i.e. an exponential transformation of a dissimilarity measure between each pair of locations based on non-parametric kernel estimator of the direct and cross variograms (Fouedjio, 2016) and on a different bandwidth identification, suitable for agglomerative hierarchical clustering techniques applied to data indexed by geographical coordinates. Simulation results are very promising showing very good performances of our proposed metric with respect to the baseline ones. Finally, the new clustering approach is applied to two real-word data sets, both giving locations and top soil heavy metal concentrations.