Rethinking Collinearity in Self‐Organizing Maps: Evidence From Geophysical Data Classification

Abstract This study examines the impact of collinearity on unsupervised machine learning algorithms (UMLAs), specifically Self‐Organizing Maps (SOMs), for detecting lithological boundaries in geophysical data. Using a multi‐scale experimental framework that includes bivariate isotropic clusters, geologically complex Noddy simulations, and real‐world data from Victoria, Australia, we show that SOMs exhibit inherent robustness to collinearity in variable geological settings due to their distance‐based optimization and neighborhood smoothing. Performance evaluation across UMLAs reveals that distance‐based algorithms maintain stability under collinear conditions, while other methodologies such as Agglomerative Clustering show sensitivity or K‐means and Gaussian Mixtures Models (GMMs) show classification performance degradation. Critically, collinear features improve classification when cluster separation is minimal in synthetic models, and derivative transforms enhance boundary detection, despite high predictor correlation. We propose the use of a Geological Complexity Index (GCI) analysis to identify areas prone to collinearity issues by geospatially mapping a novel distinction between geologically meaningful, model‐relevant “good” collinearity in high‐GCI settings and redundancy‐related “bad” collinearity in low‐GCI regions, which are typically characterized by high noise‐to‐signal ratios.