Environ Monit Assess. 2026 Apr 14;198(5):454. doi: 10.1007/s10661-026-15201-8.
ABSTRACT
Effective air-quality management can benefit from grouping monitoring sites according to their joint behavior across pollutants, yet many existing studies either focus on single pollutants or compress information into composite indices. This study addresses that gap by clustering air-quality monitoring stations in Türkiye’s Marmara Region using a multi-layer Self-and Super-Organizing Maps (Supersom) framework that treats pollutants jointly. The dataset comprises daily PM10, SO2, NO2, NO, and O3 measurements from 13 stations between November 2018 and September 2022. To enhance comparability and robustness, daily series are aggregated to weekly values, missing data and outliers are corrected within a structured pre-processing scheme, kernel smoothing is applied, and seven normalization schemes are benchmarked. A five-layer Supersom model is then trained under 120 layer-order permutations, generating more than 10,000 candidate maps via online learning. Model selection proceeds in stages by discarding solutions with empty units, comparing quantization errors within each normalization type and assessing within-cluster variability to identify a reliable configuration. The best solution, obtained under logarithmic normalization, reveals four interpretable clusters: (1) metropolitan stations influenced by traffic and industry, with elevated PM10, NO2, and SO2; (2) rural and peri-urban stations with higher O3 and lower NO, NO2, and SO2; (3) heavy-industry and port-corridor stations with high NO2 and NO and relatively lower O3; and (4) rural and agricultural contexts with higher SO2 and moderate PM10. The results indicate that modeling multiple pollutants simultaneously and systematically examining layer-order effects can yield stable, policy-relevant groupings that support network rationalization and targeted control strategies.
PMID:41979677 | DOI:10.1007/s10661-026-15201-8
