Machine learning-based models are effective and practical tools for analyzing complex relationships between air quality parameters and meteorological variables. They are also capable of identifying key factors influencing air quality. Therefore, the present study evaluates the performance of three tree-based models—Random Forest (RF), XGBoost, and CatBoost—alongside a multilayer perceptron (MLP) neural network model, using daily data of seven variables, including five meteorological parameters (wind speed, temperature, pressure, relative humidity, and rainfall) and two air pollutants (O3 and PM2.5), to estimate the daily wintertime concentration of carbon monoxide in Kerman city’s atmosphere.

The models’ performance was assessed using statistical indices including the coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE). Results indicated that the CatBoost model achieved the highest accuracy in estimating CO concentrations, with an R² of 0.778, an RMSE of 0.284 (ppb), and an MAE of 0.209 (ppb) during the test phase. The performance of the XGBoost and RF models was relatively similar, with R² values of 0.747 and 0.728, respectively. The MLP model yielded the lowest accuracy, with an R² of 0.693, an RMSE of 0.308 (ppb), and an MAE of 0.236 (ppb). These results confirm the superior capability of tree-based models in comparison to the neural network-based model for estimating CO concentration.