Clustering Ensembles of 3D Jet-Stream Core Lines

Abstract

The extraction of a jet-stream core line in a wind field results in many disconnected line segments of arbitrary topology. In an ensemble of wind fields, these structures show high variation, coincide only partly, and almost nowhere agree in all ensemble members. In this paper, we shed light on the use of clustering for visualizing an ensemble of jet-stream core lines. Since classical approaches for clustering 3D line sets fail due to the mentioned properties, we analyze different strategies and compare them to each other: We cluster the 3D scalar fields from which jet-stream core lines are extracted. We cluster on a closest-point representation of each set of core lines. These representations are derived from the extracted line geometry and can be used independently of the line orientation and topology. We cluster on the 3D line set using the Hausdorff distance as similarity metric. In the resulting clusters, we visualize core lines from the most representative ensemble member. We further compute ridges in a single 3D visitation map that is build from the ensemble of core lines, and we extract the most central core line from the ensemble closest-point representation. These ''averages'' are compared to the clustering results, and they are put into relation to ground truth jet-stream core lines at the predicted lead time.