Abstract
Volatile organic compounds (VOCs) and odours vary considerably across urban micro-environments because of differences in emission sources, occupancy, and ventilation. Fixed-site monitoring does not always reflect the changing exposures individuals encounter during everyday travel and activity. In addition, the nature and concentration of VOCs can differ considerably between micro-environments. Therefore, real-time identification of the surrounding environment can help interpret the potential risk associated with elevated VOC levels. This study evaluated an automated method for environmental recognition using the PONG4 portable multi-sensor device. A labelled dataset was generated from real-time monitoring across 15 urban micro-environments, including transport, commercial, campus, and community settings. Five machine-learning algorithms were assessed using 5-fold cross-validation. Random Forest and XGBoost achieved the highest classification accuracy, both reaching 81.1%, and outperformed the single decision tree model, which achieved 66.6%. Feature-importance analysis showed that PM4, CO2, and relative humidity contributed most strongly to environmental classification, with importance weights of 0.271, 0.130, and 0.094, respectively. In addition, t-SNE visualisation showed partial clustering across several micro-environment categories. These findings suggest that portable multi-sensor data, combined with machine learning, can support automated environmental identification and may be useful for context-aware VOC and odour exposure assessment.
