Abstract
Air pollution has become a critical global concern, particularly in rapidly developing countries such as Bangladesh and India, where urbanization and industrial growth have severely degraded air quality. Poor air quality poses serious threats to public health, contributing to respiratory diseases, cardiovascular complications, and premature mortality. Beyond human health, air pollution intensifies global warming and disrupts ecosystems, making accurate forecasting essential for effective mitigation. This study aims to develop a robust predictive model for the Air Quality Index (AQI) using advanced machine learning (ML) techniques. Environmental sensor data were analyzed to capture key influencing variables, including geographic location, time of day, and pollutant concentrations such as carbon monoxide (CO), sulfur dioxide (SO2), and particulate matter. Several ML approaches—random forest (RF), deep neural network (DNN), autoregressive integrated moving average (ARIMA), and seasonal ARIMA (SARIMA)—were employed to predict AQI levels and pollutant trends. The results demonstrate that RF outperforms other models, achieving a high predictive accuracy with an overall R2 of 0.88 for India. However, its performance is considerably weaker in Bangladesh, with an R2 of only 0.28, despite showing 0.79 during the training phase. In contrast, DNN, ARIMA, and SARIMA yielded poor results, with low or negative R2 values, emphasizing the complexities of modeling air quality across regions with differing dynamics. These findings underscore the importance of selecting models tailored to regional characteristics. The insights derived from this research can support policymakers and environmental agencies in improving air quality management and designing effective pollution control strategies.
