Detecting Pollution Anomalies in Multivariate Air Quality Datasets with Unsupervised Machine Learning

Since air pollution affects both health and the environment, detecting unusual events is very important. The study tests the effectiveness of unsupervised machine learning methods—Isolation Forest, DBSCAN and Autoencoders—on air quality samples. The models were analyzed using multivariate data obtained from the UCI Air Quality Repository, OpenAQ and the U.S. EPA to see how well they can spot unusual levels of air pollution, mainly focusing at concentrations of carbon monoxide (CO). Considerable outcomes were achieved by the Isolation Forest and Autoencoders, with each one discovering 67 significant anomalies in tune with seasonal pollution trends. These methods’ toughness towards a lot of noisy and numerous data and ability to see strange patterns from cycles, made them a good choice for ecological monitoring. On the other hand, DBSCAN only detected one anomaly which means it worked poorly in this case because it requires specific settings and there is not much data over time. A review of the correlation heatmap pointed out that pollutant variables are closely related, proving that multivariate methods work better than univariate analysis. This research can guide selection of the best models for real-time monitoring of air quality. Since the study found significant results, they can be applied to public health, future urban development and making policies about environmental protection.