Abstract
<title>Abstract</title> <p> Air pollution poses a critical health threat in arid urban environments, yet localized forecasting frameworks remain scarce. This study develops a machine learning-based predictive system for hourly concentrations of six criteria pollutants (CO, O <sub>3</sub> , NO, NO <sub>2</sub> , SO <sub>2</sub> , and PM <sub>2.5</sub> ) in Kerman, Iran (2020–2024). A multi-stage preprocessing pipeline integrating temporal synchronization, Extra Trees imputation (average R <sup>2</sup> = 0.843), and feature engineering generated a complete dataset. Temporal analysis revealed distinct diurnal patterns: CO exhibited bimodal rush-hour peaks (~ 1.5 ppm), O <sub>3</sub> displayed a photochemically-driven afternoon maximum (~ 38–39 ppb), and PM <sub>2.5</sub> showed a pronounced winter maximum (31.31 µg/m <sup>3</sup> ). Correlation analysis confirmed the dependence of O3 on solar radiation (r = 0.36) and the dominance of traffic over NO <sub>X</sub> and CO levels. Among seven evaluated models, ensemble tree-based methods outperformed linear models. Extra Trees achieved the highest test R <sup>2</sup> for CO (0.713), NO (0.320), and PM <sub>2.5</sub> (0.726); Hist Gradient Boosting excelled for SO <sub>2</sub> (0.669); and LightGBM performed best for NO <sub>2</sub> (R <sup>2</sup> = 0.622) and O <sub>3</sub> (RMSE = 2.017 ppb). Feature importance identified lagged pollutant concentrations as dominant predictors, confirming strong temporal autocorrelation, while wind direction and photochemical variables (SNSR, temperature) emerged as critical drivers for PM <sub>2.5</sub> and O <sub>3</sub> . The poor generalization for O <sub>3</sub> (negative test R <sup>2</sup> ) highlights the need for incorporating photochemical indicators. This framework offers a scalable, cost-effective approach to operational air-quality forecasting in arid cities, supporting public health advisories and environmental management. </p>