Identifying winter atmospheric patterns in the Central Valley of Chile (O’Higgins Region) using FAUM, K-means, and a hybrid clustering approach

Abstract

This study proposes a reproducible methodological framework for identifying winter atmospheric patterns associated with the accumulation and dispersion of fine particulate matter (PM₂.₅), designed as a decision-support tool for air quality analysis and environmental public policy design. The approach is based on a comparative evaluation of two classical unsupervised classification methods—Fast Autonomous Unsupervised Multidimensional (FAUM) and K-means++—and further proposes a hybrid strategy that combines both techniques to identify winter atmospheric patterns in the Central Valley of the O’Higgins Region, Chile.

The methodology is explicitly formulated as a reproducible pipeline encompassing data integration and normalization, clustering, and pattern evaluation. It is designed to be replicated in other regions with comparable atmospheric, meteorological, and satellite data availability. As a case study, a daily consolidated dataset was constructed for the winter seasons of the 2021–2024 period in the Central Valley of the O’Higgins Region (Chile), integrating PM₂.₅ concentrations, key meteorological variables: temperature, wind dirección, precipitation (Ppt), atmospheric pressure, and planetary boundary layer height (PBLH), and a satellite-based absorbing aerosol index (AII).

For each clustering method applied, the quality of the identified patterns was estimated using internal validation metrics (silhouette coefficient, Calinski–Harabasz index, and Davies–Bouldin index), as well as stability analyses under different initialization conditions. The results show that the proposed FAUM/K-means hybrid approach outperforms the individual methods, exhibiting higher physical coherence, greater stability, and an improved representation of winter atmospheric regimes. The methodology identified two dominant patterns: one characterized by cold, stable conditions and low PBLH, favorable to pollutant accumulation, and another characterized by enhanced ventilation, vertical mixing, and precipitation.

Beyond the specific case study, the results demonstrate that the proposed methodological pipeline is a flexible, transferable tool for characterizing atmospheric patterns in regions with complex topography and limited monitoring coverage. This framework provides an objective basis for strengthening air quality diagnostics and for supporting the design and evaluation of evidence-based environmental public policies.