Abstract:Based on rotated empirical orthogonal function analysis of the daily observational data of visibility from 2008 to 2018, the objective division of visibility in different seasons is obtained. Taking the global numerical prediction model of ECMWF from 2017 to 2019 as the prediction factor, the visibility prediction model for different regions and seasons is built and the regional model is applied to the station for prediction. Then the ECMWF model forecast data in 2020 are used as an independent sample, and the seasonal forecast of visibility in China is carried out. The results show that using the comprehensive algorithm of multiple linear regression, regression estimate of event possibility and discriminant analysis, the visibility forecast of model output statistics based on regional model output statistics is much better in different seasons and different forecast projections than the model direct output (DMO). The underestimation of DMO is corrected, and the improvement of winter forecast score is the most obvious. The model shows high prediction skills in the prediction of low visibility below 1 km, especially at 05:00 BT. Factor analysis shows the high-frequency factors affecting visibility mainly include temperature, pressure, humidity and wind that are closely related to boundary layer conditions, as well as surface thermal conditions, precipitation related factors and stability. The high-frequency factors selected for visibility prediction of different orders in different seasons are different. Spring is sensitive to temperature-related factors, the factors related to precipitation are selected more frequently in summer， and the unstable factors in autumn and winter are more important.