AIoT-Powered Virtual Sensors for Particulate Matter Estimation Using Spatiotemporal Feature Learning and Attention Mechanisms

Accurate fine particulate matter (PM2.5) estimation is essential for air quality monitoring, but high-cost sensors have limited spatial coverage. To address this, we introduced the concept of virtual sensors to estimate PM2.5 levels in unmonitored areas by leveraging data from nearby physical sensors. To support this study, we developed a real-time PM monitoring system integrating long range (LoRa)-based low-cost sensors with a oneM2M-compliant Internet of Things (IoT) platform, enabling data collection. To enhance the accuracy of PM measurements, we applied a mechanistic model that adjusts PM values using environmental factors like humidity and temperature. Using this calibrated data, we propose the geo-temporal transgraph attention network (GTA-Net), a deep learning model that estimates PM2.5 levels in unmonitored regions. A distance-based graph captures spatial dependencies, while transformer-based temporal encoding improves predictive accuracy. GTA-Net was evaluated against LSTM, GRU, CNN+LSTM, and Vanilla Transformer models, and the tenfold cross-validation results showed that it achieved on average an root mean-square error (RMSE) of $0.679~{\pm }~0.526~\mu $ g/m3 and an $R^{2}$ close to 1.000, performing better than all baseline models. A performance comparison based on different PM embedding sizes revealed that 1-h PM embeddings yield the most accurate estimations. This research highlights the potential of virtual sensors and spatiotemporal modeling to enhance air quality monitoring.