On the possibilities for black carbon sensor virtualisation

Traditionally monitored air quality parameters often do not give a sufficiently detailed and comprehensive picture of air quality, thus limiting the ability of citizens and authorities for informed decision making. For example, reducing PM2.5 by the same amount of mass concentration in different places may not deliver the same health benefits, thus highlighting the need for additional metrics. Carbonaceous aerosols concentration in air is an example of one such additional air quality metric, since carbonaceous aerosols are often the largest component of fine particulate matter, yet very diverse, composed of light-scattering Organic Carbon (OC), and light-absorbing Black Carbon, (BC). Presence of BC has negative effects for both, human health and our climate (Novakov et al., 2013). Furthermore, inhalation of BC is associated with health problems including respiratory and cardiovascular disease, cancer, and even birth defects (Janssen et al., 2011, Janssen et al. 2012). However, measuring mass concentration of BC, either using online or offline methods, requires specialized equipment, which is typically not a part of routine regulatory monitoring and BC concentration data is thus not readily available throughout air quality sensor networks. This BC data is typically only available at the specialized monitoring sites, so-called monitoring supersites, where much more comprehensive monitoring is done. Recently, a concept of virtual or soft sensors has been introduced (Zaidan, Martha Arbayani et al. 2023) as a way of increasing spatial resolution for certain pollutants, by utilizing IoT (low-cost sensors) and AI algorithms. Sensor virtualization is gaining traction in the metrology of sensor networks (Tabandeh, Shahin et al., 2025). By using similar research vision of virtual sensors, we attempt to expand the concept to reference air quality automatic monitoring networks, and to determine to which extent information about BC mass concentration can be statistically inferred for the entire reference grade network, and not just the supersites in the network, where such information is produced by the physical sensors. This paper further explores the initial models presented in (Davidovic et al., 2025) based on multiple linear regression, by further considering different training and test periods for the model development, more specifically, assessing virtual sensor model performance for scenarios with different biomass burning and fossil fuel contribution to black carbon concentrations. Sensor virtualization is a promising technique that can be used to estimate BC concentration, in the absence of specialized monitoring equipment. It is a usefull sensor network metrology tool under the condition that the sensor network infrastructure is well developed, and that certain network nodes (supersites or sites conducting additional monitoring campaigns) contain additional information about BC concentration. Under the favourable model training conditions, RMSE as low as ~1µg/m³ can be achieved.