Bisecting stochastic clustering: a new algorithm for PolSAR image unsupervised classification

• Autor(a) principal: Naiallen Carolyne Rodrigues Lima Carvalho
• Data de Publicação: 2020
• Tipo de documento: Tese
• Idioma: eng
• Título da fonte: Biblioteca Digital de Teses e Dissertações do INPE
• Texto Completo: http://urlib.net/sid.inpe.br/mtc-m21c/2020/12.28.13.44

PolSAR (Polarimetric Synthetic Aperture Radar) images can be represented by a set of complex Hermitian positive definite matrices, which have a natural Riemannian metric tensor. PolSAR images are, also, known for following the Wishart distribution, and, by using the information theory contrast function, stochastic distances between Wishart distributions can be derived. This work addresses unsupervised classification strategies, explores the Riemann geometry and studies stochastic distances applied to PolSAR images. The proposed algorithm, named Bisecting Stochastic Clustering (BSC), is a combination between the Stochastic Clustering (SC) algorithm and the hierarchical divisive clustering algorithm. The SC algorithm is technique based on K-means, which uses stochastic distances as similarity metric. The SC algorithm can, usually, be trapped in a local minimum, what led to incorrect clustering results. Therefore, the choice of good initial parameter candidates is essential for the clustering quality. The BSC algorithm is a top-down procedure, it starts with all samples in an unique cluster, that are successively splitted into two new sub-clusters. This algorithm is mainly divided into three steps: the initial parameter determination, the cluster bi-partitioning procedure, and the choice of a suitable cluster to split. In this work, two algorithms for the initial parameter determination are tested: the Expectation-Maximization (EM) algorithm for Wishart Mixture Model and the Riemann Principal Direction Divisive Partitioning (RPDDP). The RPDDP is a new proposed algorithm, whose goal is to perform the bi-partition of a dataset. This algorithm estimates the dataset covariance matrix under the the Riemann geometry, in order to find the principal component, which is used to separate the input data in two sub-clusters. From the RPDDP two estimated sub-clusters, the BSC derives the initial parameters. The BSC second step is performed by the SC algorithm. The BSC builds a dendrogram in order to represent the dataset splitting. Each sub-cluster, or node, links two successor sub-clusters in the dendrogram. When three or more nodes are available in one dendrogram level, the algorithm needs to choose a node to split. The BSC third step uses the information gain as the node choice rule. This work analyses the SC algorithm and two main variants of BSC. The first variant uses the RPDDP as initial parameter determiner, and the second, uses the EM algorithm as initial parameter determiner. The Bhattacharyya (B), Kullback-Leibler (KL) and Hellinger (H) stochastic distances are analysed in this work. In total, nine algorithms are evaluated: SC-B, SC-KL, SC-H, BSC-R-B, BSC-R-KL, BSC-R-H, BSC-EM-B, BSC-EM-KL, BSC-EM-H. The algorithms were analysed in a quantitative and qualitative way. The quantitative analysis consists in the confusion matrix and accuracy estimation, and the qualitative analysis explore the BSC dendrogram and the clusters scattering mechanism by inspecting the Plan H − alpha.