Hierarchical unsupervised nonparametric classification of polarimetric SAR time series data

Clustering (and classification) among other approaches of land-cover type discrimination for Polarimetric SAR (Pol-SAR) data often explicitly or implicitly assume a lot about the shape of the clusters (or the classes, in the case of classification). For example, this is an issue for Pol-SAR classification methods [1,2,3] that initialize clusters in decomposition parameter feature spaces [4], subsequently refining the clusters by Wishart moving-means iterations in coherency matrix (T3) space. Indeed, using the means as cluster (or class) representatives can be successful, provided that clusters in the data are compact, well separated, and convex. However, highly nonlinear features and unusually shaped clusters are often obtained when dealing with PolSAR data. To address this issue we present a data-driven hierarchical clustering technique. This we demonstrate for forest-type discrimination purposes with a multi-temporal Radarsat-2 sandwich.