Air Quality Index prediction using satellite and ground based data and Deep Learning models

The cleanliness of the air we breathe is essential for a healthy lifestyle. To assess the extent of air pollution and to create awareness among the public about its health effects, an Index for Air Quality (AQI) is used. The prediction of AQI is essential for the government agencies as well as the society to formulate action plans to reduce the effects of air pollution. Conventional Machine Learning (ML) based prediction models suffer from their limited capacity to handle nonlinear and multi-dimensional data, along with their dependence on manual feature engineering and reduced adaptability to complex data trends. The Deep Learning (DL) approaches outperform ML models by handling large and high-dimensional datasets with their inherent feature extraction capability, leading to more accurate AQI predictions. However, it has been observed that the traditional ground-based datasets put an upper boundary on the performance of DL models due to the deficiency in data features and poor correlation. Hence, this study proposes a novel, multi-platform dataset containing ground-based and satellite-based data to be introduced into the established hybrid DL models to improve their AQI prediction performances. CNN-BiLSTM, CNN-LSTM, CNN-GRU and CNN-AGU models are used in the study. After running the models using 8-year long dataset, their respective performances have been evaluated. The wide spatial resolution of satellite data helps the models identify spatial patterns and interactions that ground-based data alone may not have accounted and thus improved the predictions of all these models significantly. The CNN-AGU model achieved 3.06% higher R²-score compared to ground-data-only models. Hence, this model, along with multi-platform dataset can be used to predict the AQI very accurately.