Determining the key meteorological factors affecting atmospheric CO2 and CH4 using machine learning algorithms at a suburban site in China

Abstract

Atmospheric CO₂ and CH₄ were measured from March 2023 to February 2024 at a suburban site near the northern foot of the Qinling Mountains, China, aiming to determine the key meteorological factors that influencing the seasonal CO₂ and CH₄ dynamics using machine learning (ML) algorithms. Yearly average atmospheric CO₂ and CH₄ were 446.1 ± 10.0 ppm and 2118.9 ± 50.5 ppb, respectively. Anthropogenic emissions dominated atmospheric CO₂ and CH₄ changes in winter, and the excess CO₂ (ΔCO₂) and CH₄ (ΔCH₄) above background levels during the heating period were mainly from combustion emissions. We selected seven ML algorithms to determine the variable importance of meteorological factors, among which eXtreme Gradient Boosting (XGBoost) and Random Forest (RF) demonstrating the best performance and ranking consistency of the factors. Humidity and temperature of the atmosphere and soil had the higher effect (XGBoost: 80.4 %; RF: 78.1 %) on CO₂ in spring, while humidity was crucial in winter, with variable importance of 69.3 % for XGBoost and RF. In summer and autumn, photosynthetic photon flux density and wind speed (WS) (totaling 35.2 % ∼ 50.7 %) dominated CO₂ dynamics. For CH₄, atmospheric and soil humidity (totaling around 40.0 %) were key factors in spring, whereas atmospheric humidity was important in winter. WS had the largest effect in summer (XGBoost: 26.3 %; RF: 33.3 %) and autumn (XGBoost: 19.8 %; RF: 28.8 %). Meteorological processes like cold front passage significantly reduced CO₂ and CH₄ concentrations during haze events. XGBoost and RF have emerged as powerful tools for determining the key meteorological factor that favour seasonal GHGs evolution.