Air pollution is a major environmental and health challenge requiring timely monitoring and prediction. This study develops Efficient air quality classification models deployable on low-power microcontrollers using TinyML. Daily aggregated pollutant and meteorological data from 77 sites in Taiwan were preprocessed and enhanced with feature engineering. Feature selection techniques Random Forest importance, mutual information, and recursive feature elimination identified informative predictors. Logistic Regression, Multilayer Perceptron, and Random Forest models were trained and tested. Results show reduced feature subsets preserved predictive accuracy while lowering inference time and complexity. Logistic Regression and MLP offered favorable accuracy–efficiency trade-offs for TinyML deployment, while Random Forest consumed more memory. Findings highlight the importance of aligning model choice with deployment constraints. This research demonstrates the feasibility of systematic feature selection and light-weight modelling for embedded air quality classification, laying a foundation for real-time microcontroller- based monitoring.