Predicting Ozone Pollution in Urban Areas Using Machine Learning and Quantile Regression Models

Ozone is the most harmful secondary pollutant in terms of negative effects on climate change and human health. Predicting ozone emission levels has therefore gained importance within the field of environmental management. This study, performed in the Andean city of Cuenca, Ecuador, compares the performance of two methodologies currently used for this task and based on machine learning and quantile regression techniques. These techniques were applied using cross-sectional data to predict the ozone concentration per city block during the year 2018. Our results reveal that ozone concentration is significantly influenced by nitrogen dioxide, sedimentary particles, sulfur dioxide, traffic, and spatial features. We use the mean square error, the coefficient of determination, and the quantile loss as evaluation metrics for the performance of the ozone prediction models, employing a cross-validation scheme with a fold. Our work shows that the random forest technique outperforms gradient boosting prediction, neural network, and quantile regression methods.