Abstract: Objective　Visibility is an important indicator reflecting atmospheric transparency, and its variation is closely related to daily human activities. Low-visibility weather phenomena, as critical factors affecting road traffic and aviation safety, have received increasing attention in recent years. Consequently, accurately predicting future visibility has become an important research topic and direction. The existing studies on visibility prediction mainly focus on traditional statistical prediction methods and machine learning-based approaches. Although machine learning methods generally outperform conventional statistical techniques in terms of predictive accuracy, they often rely on a large number of meteorological variables and lack systematic comparisons across multiple models. Moreover, visibility prediction typically requires long-term on-site observational data, and obtaining these data is associated with high costs and practical difficulties, which further limits the applicability of machine learning-based visibility prediction methods in real-world scenarios. Therefore, it is essential to identify the key meteorological parameters that have the greatest influence on visibility while maintaining predictive accuracy. By reducing the data requirements of machine learning models, the associated observation costs can be effectively reduced, thereby enhancing the practicality and scalability of visibility prediction systems. In this study, we propose an integrated visibility prediction framework based on optimized meteorological parameter selection. Methods　The proposed integrated visibility prediction scheme based on meteorological parameter optimization utilizes the hourly averaged visibility observations and corresponding meteorological data obtained from a meteorological monitoring experiment conducted in Qingdao from August 2019 to August 2020. Based on a systematic analysis of the influence of meteorological factors on the seasonal variation of visibility, this study focuses on exploring the optimized meteorological parameter configurations for visibility prediction. By comprehensively evaluating the performance of five commonly used machine learning methods under different training parameter schemes, an integrated visibility prediction framework based on optimized meteorological parameters is established. The five machine learning methods include three ensemble learning models, namely Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Random Forest (RF), as well as two conventional machine learning models, namely Support Vector Machine (SVM) and Multiple Linear Regression (MLR). With respect to training parameter selection, four different visibility prediction training schemes are designed based on meteorological parameters to accommodate two distinct application scenarios. During the model development process, all machine learning models are optimized for the final key hyperparameters using the GridSearch approach. Results and Discussion　To more accurately investigate the influence of meteorological parameters on visibility under different seasonal conditions, model training and validation are conducted according to different schemes, followed by visibility prediction and performance evaluation. The visibility prediction results obtained from different models under different parameter schemes are evaluated using multi-season averaged Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and correlation coefficient (CC). The experimental results indicate that: (1) XGBoost and LightGBM outperform the other models and are more suitable for visibility prediction; and (2) employing the LightGBM model with the parameter scheme excluding wind speed in spring and autumn, as well as the XGBoost model with the parameter scheme excluding atmospheric pressure in summer and winter, can further improve visibility prediction performance. Using the proposed meteorological parameter-optimized visibility prediction scheme, the correlation coefficient of the predicted results can be improved to the range of 0.68–0.76. Conclusions　This study proposes an integrated visibility prediction framework based on meteorological parameter optimization. Using five conventional meteorological parameters closely related to visibility, multiple training parameter combination schemes are designed. All five machine learning models are subjected to unified data preprocessing and are trained, predicted, and evaluated on a seasonal basis. The proposed framework demonstrates that, when meteorological parameters are used as input features, selecting different parameter combinations for different seasons can effectively achieve meteorological parameter optimization and improve the accuracy of visibility prediction. The results of this study not only provide practical guidance for the application of machine learning methods in operational visibility forecasting, but also contribute to a deeper understanding of the meteorological factors influencing visibility variability.

Key words: visibility prediction, extreme gradient boosting, light gradient boosting machine, machine learning, ensemble learning