Clustering Spectral Line Cubes: A Multiview Prototype Approach

Abstract

We present a new clustering paradigm to segment radio astronomy cubes based on unsupervised machine learning provided by a recent modification of self-organizing maps (SOMs) known as self-organizing uniform manifold approximation and projection (SOUMAP). SOUMAP simultaneously reduces sample size via neural vector quantization and learns an expressive UMAP-based low-dimensional embedding of data upon which further analysis can be based. Several similarity graphs learned by SOUMAP are combined into a single multiview similarity graph upon which graph-based clustering of the learned SOUMAP prototypes is performed. The entire unsupervised learning pipeline, which we call multiview prototype embedding and clustering (MPEC), employs data-driven parameter selection, including the automated selection of the number of clusters C, which lessens the burden of parameter tuning and minimizes sources of user-driven bias. Additionally, we propose a post hoc cluster-wise significance test based on ARIMA time series modeling to identify pertinent clusters for further analysis and to flag noise-dominated clusters. We present results of MPEC-based clustering on NH3 emission from additional star-forming regions observed in the Green Bank Ammonia Survey (B18, L1688, NGC1333, OrionA). Our results show MPEC clustering is sensitive enough to extract very low signal-to-noise clusters, while being self-aware enough to discard any noise clustered in this process. We show that the advantage of MPEC is marked when compared to traditional methods (spectral clustering), and opens up new avenues for robustly analyzing astronomical spectral data.