Numerical model-informed testbed for surface PM2.5 concentration over China and its estimates during 2013-2021

Ground-level PM2.5 data, derived from satellites using machine learning, play a crucial role in health and climate assessments. However, persistent uncertainties stem from the lack of spatially comprehensive observations. To mitigate this issue, we introduce a novel testbed utilizing numerical simulations to assess PM2.5 estimation across the entire spatial domain. This testbed mimics the conventional machine-learning paradigm by training models with grid data from ground monitor sites and evaluating predictive accuracy across other locations. As a result, it facilitates a comprehensive evaluation of various machine-learning methodologies in estimating PM2.5 concentrations throughout the spatial domain. Moreover, the testbed offers an efficient strategy to optimize model structure or training samples, thereby enhancing the performance of satellite-retrieval models. This optimization can involve integrating spatiotemporal features, particularly with CNN-based deep-learning approaches such as the ResNet model as presented here. Additionally, we provide and proposed machine learning method as well as the corresponding estimated daily PM2.5 concentrations spanning the past nine years (2013-2021) over China with 27km by 27km spatial resolution.