Visualizing Confidence in Cluster-based Ensemble Weather Forecast Analyses

Abstract

a) (b) Fig. 1: (a) Analysis of the variation in cluster membership over 81 different clusterings of the case "Tropical Cyclone Karl", an ensemble of 51 potential vorticity fields. Circular elements represent ensemble members, colors distinguish clusters (member 45 is enlarged: color of inner circles denotes reference cluster, surrounding pie-charts show how often the member was grouped into another cluster). Dashed outlines highlight cluster representative members. Member 26 is picked, for all members with similar cluster membership variation "variability matrix plots" (squared elements encoding cluster membership of all 81 clusterings) pop up. (b) A "contour probability plot" (CPP, different greens show probabilities for contour line occurrence) shows the variability of an iso-contour within a selected cluster. Overlaid stipple pattern shows the spatial variation of the plot with respect to the 81 clusterings.Abstract-In meteorology, cluster analysis is frequently used to determine representative trends in ensemble weather predictions in a selected spatio-temporal region, e.g., to reduce a set of ensemble members to simplify and improve their analysis. Identified clusters (i.e., groups of similar members), however, can be very sensitive to small changes of the selected region, so that clustering results can be misleading and bias subsequent analyses. In this article, we -a team of visualization scientists and meteorologists-deliver visual analytics solutions to analyze the sensitivity of clustering results with respect to changes of a selected region. We propose an interactive visual interface that enables simultaneous visualization of a) the variation in composition of identified clusters (i.e., their robustness), b) the variability in cluster membership for individual ensemble members, and c) the uncertainty in the spatial locations of identified trends. We demonstrate that our solution shows meteorologists how representative a clustering result is, and with respect to which changes in the selected region it becomes unstable. Furthermore, our solution helps to identify those ensemble members which stably belong to a given cluster and can thus be considered similar. In a real-world application case we show how our approach is used to analyze the clustering behavior of different regions in a forecast of "Tropical Cyclone Karl", guiding the user towards the cluster robustness information required for subsequent ensemble analysis.

Recent research advances on interactive machine learning

“…It uncovers grouping patterns in the data and stands as a fundamental approach for visually aggregating or summarizing the data. Thus, visual cluster analysis has been applied to many data-driven applications such as weather ensemble forecast (Kumpf et al, 2018) and air traffic optimization (Andrienko et al, 2018). Recent work mainly focuses on cluster-based exploratory data analysis (Andrienko et al, 2018;Badam et al, 2017;Heimerl et al, 2016;Sacha et al, 2018;Wu et al, 2017b,c), comparative clustering analysis (Jarema et al, 2015;Kumpf et al, 2018;Kwon et al, 2018;Zhang et al, 2016b), and bi-cluster analysis (Sun et al, 2016;Watanabe et al, 2015;Wu et al, 2017aWu et al, , 2015Zhao et al, 2018).…”

Interactive Clustering

“…The basic assumption made by the authors of the papers in this category is that an interactive and collaborative process combining the strengths of both human and machine would yield better results than a process that is purely automated or purely manual. Several examples of improving the quality of the clustering results using different strategies are given in the works presented in Andrienko and Andrienko [4], Basu et al [15], Boudjeloud-Assala et al [19], Cao et al [24], Castellanos-Garzón et al [26], Choo et al [30], Dobrynin et al [38], Hadlak et al [50], Hoque and Carenini [53], Hu et al [55], Kumpf et al [64], Lai et al [66], Lee et al [67], Lei et al [68], MacInnes et al [72], Packer et al [79], Schreck et al [86], Srivastava et al [94], Turkay et al [99,101], Zhou et al [116].…”

“…Several authors focus on simultaneous clustering using different parameters and giving the user tools to efficiently compare those results to find the best solution [42,64,71,75,105,106]. For example, in Choo et al [31], users can perform dimension reduction, interactively update the pre-processing options for data and clustering, apply different clustering algorithms, and visually compare the effects of these modifications across parallel coordinates, scatter plots, and cluster label views.…”

“…Finally, in some cases, the focus of interaction is the data that is used for clustering. This is most common in geo-spatial data, where the user specifies an area of interest [19,64]; document clustering, where the user can delete non-important data instances [67]; and biomolecular data [99]. Turkay et al [99] proposes the interactive visualization method to reveal and analyze relationships in heterogeneous datasets at various abstraction levels (stratifications), aiming to find "elements of a cancer subtype that differ significantly from other subtypes."…”

“…In some papers, the user receives help to estimate the optimal number of clusters to generate, such as in Kumpf et al [64], where cluster split/merge diagrams and so-called "elbow plots" mark the optimal parameters. In Srivastava et al [94], the user is presented with a first clustering of the data, and can then either reject, accept, or be ignorant of the result, triggering a re-clustering of the data.…”