Abstract

Low-cost air quality sensors can expand PM2.5 monitoring networks but require calibration against regulatory-grade monitors to correct systematic bias towards trustworthy observations. Although interest in predictive uncertainty for air quality estimation has increased, the reliability of uncertainty under spatial sparsity and wildfire-induced distribution shift remains poorly understood. With the increasing use of low-cost sensors during wildfire-smoke episodes, uncertainty quantification during calibration is essential to determine when corrected PM2.5 estimates are reliable enough to support near-real-time alert, public-health, and exposure-related decisions. This study presents a spatial uncertainty evaluation framework for trustworthy AI-enabled environmental sensing. The framework is implemented with a transformer-based PM2.5 calibration model and two uncertainty quantification methods, GeoConformal Prediction (GCP) and Monte Carlo Dropout (MCD), using sensor pairs in California and the Northeast United States. The calibration model achieved strong performance at short spatial distances, with test R2 values of 0.89 and 0.91 at the 1 km threshold in California and the Northeast, respectively, with accuracy declining as sensor separation increased. GCP generally produced calibration curves closer to the ideal diagonal, while MCD generated tighter prediction intervals under normal conditions. During wildfire events, the uncertainty performance depended on sensor separation. At short distances, MCD expanded its uncertainty intervals and captured PM2.5 spikes more effectively than GCP (71% vs. 61% coverage). At larger separations, MCD captured only 44% of elevated observations, whereas GCP widened its intervals and achieved 83% coverage. These results demonstrate that uncertainty reliability is strongly influenced by spatial separation and environmental conditions, highlighting the need for uncertainty-aware and spatially adaptive calibration of low-cost PM2.5 sensors for trustworthy actionable air quality observation.

IPC Classification

Keywords