Hyperlocal Air Pollution Mapping: A Scalable Transfer Learning LUR Approach for Mobile Monitoring

Many epidemiological studies have traditionally leveraged European maps derived from fixed-site measurements to investigate health effects, primarily emphasizing inter-city variations. Recently, mobile monitoring has been demonstrated to refine the spatial resolution focusing on intra-city variations. Nevertheless, efficiently scaling mobile monitoring campaigns to cover a large spatial area remains challenging.

Tackling this challenge, we explored the transferability of mobile measurements across three European cities. We propose to adapt the traditional land use regression (LUR) models with unsupervised transfer learning algorithms. These models, named CORrelation ALignment (Coral) and its adapted form, inverse distance-weighted Coral (IDW_Coral), aim to estimate air pollution levels in Amsterdam. They rely solely on data from mobile monitoring campaigns in Copenhagen and Rotterdam, bypassing the need for local Amsterdam data itself. The first 30 collection days of mobile campaigns in Copenhagen and Rotterdam were used as the source data (training inputs). By harmonizing the feature space, Coral is designed to minimize the domain difference between the source and target areas. IDW_Coral integrates single Coral models following general geographic principles. Their performance was validated against external routine measurements and compared with a reference LUR model (AMS_SLR), fitted by sequentially increasing amounts of mobile measurements collected in Amsterdam for nitrogen dioxide (NO₂). The proposed models were also compared with our previously published mixed-effect models using all Amsterdam mobile measurements for NO₂ and Ultra Fine Particles (UFP).

For nitrogen dioxide (NO₂), IDW_Coral achieved a balanced performance with an R² of 0.35.  This accounts for 67% of the accuracy of a locally fitted Amsterdam model (AMS_SLR, R² = 0.52), developed using comprehensive mobile monitoring over 160 days in Amsterdam. The difference in absolute errors between the two models was marginal (0.75 for MAE and 0.66 µg/m³ for RMSE). The R² of IDW_Coral matches that of AMS_SLR based on 25 days of data collection, implying that a minimum of 25 days is required to gather city-specific insights through mobile monitoring. If this condition isn't met, IDW_Coral presents a more cost-effective alternative. IDW_Coral correlated strongly (Spearman correlation of 0.72 for NO₂ and 0.76 for UFP) with mixed-effect models fitted with all Amsterdam mobile measurements.

Leveraging Tobler's first law of Geography, our IDW_Coral method proficiently delineates hyperlocal air pollution in areas not directly observed. Further improvements in accuracy and applicability can be achieved by expanding mobile-monitored areas. Requiring no direct measurements in the target area, IDW_Coral has the potential for application across Europe, promising substantial savings in collection efforts.