Abstract: In order to enhance the operational forecasting capabilities of machine learning models for low visibility weather, and alleviate the difficulty of ignoring a few class samples in imbalance data, this paper developed a visibility forecasting model based on the LightGBM algorithm and under-sampling technology. ECMWF model products and the observations from Wuhan during 2020 to 2025 were used to evaluate the model. The forecasting performance of visibility under different positive-to-negative sample ratios was assessed using Probability Of Detection(POD),False Alarm Rate(FAR) and Equitable Threat Score(ETS). The main results are as follow: (1) As the ratio of positive to negative samples increases, POD and FAR increases rapidly. An appropriate ratio can enhance the forecasting ability of the model. At a ratio of 1:7, the ETS is 3.48 times higher than the original ratio;(2) T2M-Td2M、RH700、T2M-TMP000、Wind925_direction、RH850 and T2M are the six most important characteristic quantities for predicting the occurrence of low-visibility weather at Wuhan. They together accounted for 81.8% of the cumulative importance; (3)T2M-Td2M、RH700 and T2M-TMP000 are the three key factors that influence the model’s predictions; (4) The model not only reveals the key factors affecting visibility prediction, but also provides valuable technical support for operational visibility forecasting.