The Indo-Gangetic Plain of India continues to face a serious environmental and public health problem due to air pollution, as particulate matter (PM2.5 and PM10) continuously surpasses permissible limits. In order to predict particulate matter concentrations in Lucknow, India, this study creates a hybrid Convolutional Neural Network–Long Short-Term Memory (CNN–LSTM) model utilising an eight-year dataset (2017–2024) of meteorological and air quality indicators. To guarantee dependability, the data underwent preprocessing using min–max normalisation, wind vector reconstruction, interpolation, and outlier correction. Through the integration of CNN's feature extraction and LSTM's sequential learning, the CNN–LSTM model is able to capture temporal relationships as well as spatial correlations. R2, RMSE, MAE, and MAPE were used to compare performance to standalone models. According to the results, the hybrid method successfully reproduced seasonal variability, including winter peaks and monsoon-driven falls, with the maximum accuracy (R2 = 0.658 for PM2.5; R2 = 0.754 for PM10). The CNN–LSTM outperformed other models in terms of robustness and generalisability, although somewhat underestimating intense episodic surges. Under India's National Clean Air Programme (NCAP), the results highlight the model's potential as a decision-support tool for early warning systems and policy actions. The importance of deep learning hybrids for long-term air quality control in heavily polluted metropolitan areas is demonstrated by this work.