Page 199 - 《高原气象》2026年第2期

2 期 殷齐娥等：基于机器学习的机场低能见度短临预报研究 499

Nowcasting of Airport Low Visibility Based on Machine Learning

YIN Qi'e ， NI Changjian ， XIAO An 2， 3
1， 3
1
（1. School of Atmospheric Sciences， Chengdu University of Information Technology， Chengdu 610225， Sichuan， China；
2. Jiangxi Provincial Key Laboratory of Climate Change Risk and Meteorological Disaster Prevention，
Nanchang 330096， Jiangxi， China；
3. Jiangxi Meteorological Observatory， Nanchang 330096， Jiangxi， China）

Abstract： In order to reduce the rate of flight diversion and return caused by low visibility， this study has estab‐
lished a short-term and nowcasting model of low visibility using ground observation data and the upper-air and
surface forecast data of the ECMWF at Jingdezhen Airport based on machine learning algorithms. Comparing the
evaluation indicators， the results find that the XGBoost and LightGBM machine learning algorithms outperform
the SVM machine learning algorithm in nowcasting of the airport low visibility. A detailed comparison of the
evaluation metrics was conducted both before and after feature screening in the same machine learning algo‐
rithms. The study highlights that feature screening significantly boosts the effectiveness of both models. Further‐
more， the SHAP （SHapley Additive exPlanations） method elucidates the contribution of each feature to the
LightGBM model's output. The main conclusions are as follows： （1） The machine learning models established
by LightGBM and XGBoost perform well in airport low visibility forecasting， with the AUC reaching up to
0. 98， and the F1_score for the prediction of current low visibility and the low visibility in the next one hour can
reach up to 0. 92.（2）Data cleaning and feature screening is conducive to improving the prediction accuracy of
the XGBoost algorithm for low visibility in the next hour， according to the principle that "feature engineering in
machine learning requires features to be mutually independent". Moreover the LightGBM model with feature
screening has a lower false negative rate than the LightGBM model without feature selection when forecasting
the current and future one-hour low visibility. For the forecast of the current low visibility， the LightGBM_24_0h
model is the best. For the forecast of low visibility in the next one hour， the XGBoost_24_1h model is the best.
And feature selection has a greater improvement on the performance of the XGBoost algorithm.（3） The splitting
times and SHAP values are used respectively to analyze the feature importance of the LightGBM algorithm mod‐
el. It shows that under different feature importance criteria， nine features， namely the measured relative humidi‐
ty， air temperature， wind， sea level pressure at the airport， and the relative humidity at 1000 hPa， vertical veloc‐
ity and divergence at 925 hPa， and divergence at 850 hPa predicted by ECMWF， are more important for the pre‐
diction of low visibility at the airport. And divergence， as an input feature of the machine learning model， can
greatly improve the performance of the machine learning model.（4） When explaining feature importance based
on SHAP values， the cumulative proportion of the top ten feature importance accounts for 80%. This indicates
that in the nowcasting of low visibility at Jingdezhen Airport where fog is the main factor， the LightGBM model
can output prediction results according to key forecast factors，. And when forecasting whether the low visibility
in the next one hour will continue， more attention should be paid to the changes in 850 hPa divergence，
1000 hPa relative humidity， airport sea level pressure and wind direction.
Key words： airport forecast； low visibility； machine learning； LightGBM