Comparison of Incremental Machine Learning Methods and Graph Neural Networks for the Field-Calibration of Monitoring Low-Cost Sensor Networks

Air pollution is a major global issue, and although Official Ground-Based Monitoring Stations (OGBMS) provide accurate data, their high-cost associated with purchase, operation, and maintenance limits widespread deployment. Low-Cost Sensor Networks (LCSNs) offer broader spatial coverage but typically suffer from lower accuracy, thus making the improvement of their precision essential for effective air quality monitoring. Since air quality data are time series, they are affected by concept drift due to factors like seasonality, sensor aging, and changing meteorological conditions. Incremental learning offers a more suitable solution for real-time field-calibration than traditional batch machine learning methods. To address this, the study introduces an advanced GNN-Transformer-based framework that operates incrementally, exploiting relationships between sensors to enable accurate, supervised, and continuously adaptive field-calibration. The proposed model is validated using PM10 data collected from 32 LCSN devices and 3 OGBMS in Thessaloniki, Greece, and is compared against other incremental simpler ML models using both forward and spatial (leave-one-station-out) cross-validation. Results show that the GNN-Transformer consistently outperforms alternatives in terms of accuracy and robustness, even when the target station is excluded from training, satisfying the Relative Expanded Uncertainty (REU) criterion and demonstrating superior generalizability for spatial field-calibration tasks.