Evaluating the multi-variable influence on O3, NO2, and HCHO using BRTs and RF model

Abstract

This study addresses significant knowledge gaps in understanding the complex interplay between atmospheric chemistry and synoptic conditions. Using emerging machine learning techniques—Boosted Regression Trees (BRTs) and Random Forest (RF) models—we investigate the influence of synoptic conditions on pollutant levels. Several BRTs and RF models are developed to estimate surface concentrations of ozone (O3), nitrogen dioxide (NO2), and formaldehyde (HCHO). By considering a range of algorithmic structures and explanatory variables for each pollutant, the research aims to identify the most skillful predictive approaches and influential factors governing pollutant levels. The design seeks to highlight key determinants of concentration patterns without constraining the investigation to pre-defined model structures or explanatory variable sets. Introducing a novel methodology, Correlation Coefficient Differential Evaluation (C2DE), we quantitatively assess the influence of explanatory variables. C2DE reveals significant contributions from spatial variables (i.e., trajectory clusters at varying altitudes), formaldehyde to nitrogen dioxide ratio (FNR), and meteorological parameters. Specifically, spatial variables contribute approximately 28 % to O3 concentrations, while the FNR accounts for around 5.2–9.8 % of the overall influence. For NO2 and HCHO, spatial variables contribute around 26.5 % and 32.1 %, respectively. Moreover, when considering the combined influence of meteorological parameters, these collectively explain about 45.34 %, 35.31 %, and 45.41 % of the variations in O3, NO2, and HCHO concentrations, respectively. Thus, C2DE provides valuable insights into the relative contributions of these factors, aiding in the comprehensive evaluation of air quality dynamics. This underscores the need for a multifaceted approach to comprehending and effectively addressing air pollution before devising its control strategies.