Multivariate Autoregressive State-Space modelling (MARSS) to explain daily CO variability in Campo Grande-MS (2003–2016)

We investigated daily carbon monoxide (CO) variability in Campo Grande, Brazil, during 2003–2016 using Multivariate Autoregressive State-Space (MARSS) modelling to separate latent state dynamics from direct meteorological influences. Six model configurations were evaluated by allocating meteorological covariates between the state and observation equations. After quality control, 2017–2018 data were excluded because of inconsistencies in humidity records. Model selection identified the configuration including precipitation and dry-spell days in the state equation and temperature, humidity, inverse humidity, wind speed, and wind direction in the observation equation as the best-performing model (AIC = 56,598.5; BIC = 107,608.5). Wind speed and inverse humidity exerted the strongest effects on CO variability, reflecting atmospheric dispersion and dry-season conditions. The model successfully reproduced seasonal and background CO variability (RMSE = 61.9, MAE = 32.6, MAPE = 17.7%), although extreme dry-season peaks were underestimated. Monthly CO concentrations were strongly correlated with satellite-derived Cerrado fire counts (Pearson r = 0.70, p < 0.001; Spearman ρ = 0.52, p < 0.001), indicating that biomass burning is a major driver of episodic CO enhancements. These findings demonstrate the effectiveness of the MARSS framework for separating meteorological and emission effects and support integrating satellite fire products into future tropical air-quality modelling.