Mapping the seamless hourly surface visibility in China: a real-time retrieval framework using a machine-learning-based stacked ensemble model

Abstract

Surface visibility (SV), a key indicator of atmospheric transparency, is used widely in the fields of environmental monitoring, transportation, and aviation. However, the sparse distribution and limited number of SV monitoring sites make it difficult to fulfill the urgent need for spatiotemporally seamless fine-scale monitoring. Here, we developed the operational real-time SV retrieval (RT-SVR) framework for China that incorporates information from multiple data sources, including Chinese Land Data Assimilation System meteorological data, in situ observations, and other ancillary data. Seamless hourly SV data with 6.25-km spatial resolution are available in real time via the operational RT-SVR model, which was built using a two-layer stacked ensemble approach that combines multiple machine learning algorithms and a deep learning module. Sample-based cross-validation of the RT-SVR model on approximately 41.3 million data pairs revealed strong robustness and high accuracy, with a Pearson correlation coefficient (R) value of 0.95 and a root mean square error (RMSE) of 3.17 km. An additional hindcast-validation experiment, performed with continuous observations obtained over one year (approximately 20.8 million data pairs), demonstrated the powerful generalization capabilities of the RT-SVR model, albeit with slight degradation in performance (R = 0.85, RMSE = 5.28 km). The seamless hourly SV data with real-time update capability enable tracking of the generation, development, and dissipation of various low-SV events (e.g., fog, haze, and dust storms) in China. The developed framework might also prove useful for quantitative retrieval of aerosol-related parameters (e.g., PM2.5, PM10, and aerosol optical depth).