Reducing Sampling Effort for Mapping Air Pollution: A Two-Stage Framework

Abstract Mobile monitoring is essential for resolving intraurban air pollution gradients, yet its scalability is constrained by the intensive sampling required to produce stable concentration maps. Here, we develop a screen-and-target two-stage framework to help reduce the effort of mapping hyperlocal air pollution. Using large mobile-monitoring data sets from Shenzhen and Macau, we combined spatially constrained segmentation models with land-use machine learning to evaluate how sampling effort affects air pollution mapping stability. Mean NO2 higher-lower (binary) concentration zone classification agreement first exceeded 80% at 3% of randomly sampled observations in Shenzhen and 15% in Macau. Random selection of whole measurement days required more effort for the same agreement level: 27.7% of observations (8 days) in Macau and 22.6% (30 days) in Shenzhen. Land-use machine learning achieved 80% of the total reduction in prediction error with 25–36% road-network coverage across pollutants, but errors remained high in large unsampled areas. Under the same sampling effort, the screen-and-target strategy reduced road-segment NO2 estimation error by 18–32% relative to conventional uniform sampling, whereas its zone maps improved mainly at low effort (≤20% of the campaign effort). These findings can guide resource-efficient sampling strategies for mapping and mitigation of air pollution in cities.