Surface PM_2.5 data have gained widespread application in health assessments and related fields, while the inherent uncertainties in PM_2.5 data persist due to the lack of ground-truth data across the space. This study provides a novel testbed, enabling comprehensive evaluation across the entire spatial domain. The optimized deep-learning model with spatiotemporal features successfully retrieved surface PM_2.5 concentrations in China (2013–2021), with reduced biases induced by sample imbalance.