Comparing Spatial Interpolation Methods for PM2.5 as Inputs to Urban Decision-Making in Greater Boston

Exposure to ambient P M2.5 is estimated to be associated with ˜3 million deaths globally in 2017. Reducing this number requires targeted health-protective interventions, especially in urban areas with higher pollution burden, however obtaining high-resolution P M2.5 data in urban environments is challenging due to sparse sensor and regulatory monitor distribution as well as the complex spatial heterogeneity of urban air quality. In this study, we evaluate various spatial interpolation methods to estimate PM2.5 concentrations in Brookline, a municipality in Greater Boston, ex-plicitly examining the trade-offs between sensor network size and interpolation performance. Random Forest achieves RMSE of 0.68μg/m3 (MAPE of 7.5%), significantly outperforming other methods. The RMSE for all methods decreased by less than 0.02 μg/m3 when 15 (40%) fewer sensors were used to train the models. These findings highlight the potential of data-driven spatial interpolation techniques in mitigating tradeoffs between cost and sensor network comprehensiveness in complex urban environments.