Advanced air quality prediction using multimodal data and dynamic modeling techniques

Accurate air quality forecasting is critical for human health and sustainable atmospheric management. To address this challenge, we propose a novel hybrid deep learning model that combines cutting-edge techniques, including CNNs, BiLSTM, attention mechanisms, GNNs, and Neural ODEs, to enhance prediction accuracy. Our model uses the Air Quality Open Dataset (AQD), combining data from ground sensors, meteorological sources, and satellite imagery to create a diverse dataset. CNNs extract spatial pollutant patterns from satellite images, whereas BiLSTM networks simulate temporal dynamics in pollutant and weather data. The attention mechanism directs the model's focus to the most informative features, improving predictive accuracy. GNNs encode spatial correlations between sensor locations, improving estimates of pollutants like PM2.5, PM10, CO, and ozone. Neural-ODEs capture the continuous temporal evolution of air quality, offering a more realistic representation of pollutant changes compared to discrete-time approaches. Importantly, we use adaptive pooling, a dynamic operation that optimizes spatial feature reduction while preserving critical information, which sets it apart from traditional fixed pooling layers. This adaptive pooling mechanism reduces computational complexity and results in a 22% reduction in training time, as demonstrated by the experimental results in section 4. Our model thus enables real-time environmental monitoring and large-scale forecasting. The experimental results show superior performance (RMSE = 6.21, MAE = 3.89, and R² = 0.988), outperforming existing models. This study highlights the advantages of combining multimodal data sources with advanced dynamic modeling techniques to improve air pollution prediction and inform policymaking.