The Application of Deep Learning in Airport Visibility Forecast

Lei Zhu, Guodong Zhu, Lei Han, Nan WangAtmospheric 2017 · DOI

Abstract This paper uses Urumqi International Airport’s hourly observation from 2007 to 2016 and builds regression prediction model for airport visibility with deep learning method. From the results we can see: the absolute error of hourly visibility is 706 m. When the visibility ≤ 1000 m, the absolute error is 325 m, and this method can predict visibility’s trend. So we can use this method to provide the airport visibility’s objective forecast guidance products for aviation meteorological services in the future. In this paper, the Urumqi area is as the research object, to explore the depth of learning in the field of weather forecasting applications, providing a new visibility return forecast for weather forecast personnel so as to improve the visibility of the level of visibility to ensure the safe and stable operation of the airport.

Introduction

Low visibility caused by fog, haze, and smoke is a primary disruptor in civil aviation, often resulting in significant delays, cancellations, and safety hazards. Urumqi International Airport, a vital hub for the Xinjiang region and Central Asia, is particularly susceptible to these conditions, averaging 60 low-visibility days annually, primarily during the winter months.

Traditional forecasting methods often rely on empirical rules or statistical models that struggle with the non-linear complexities of atmospheric dynamics. While numerical weather prediction (NWP) models are available, their high computational cost and frequent failure to pinpoint fog onset and dissipation limit their utility. This study investigates whether Deep Learning can offer a more accurate, automated "objective forecast guidance" for aviation meteorologists.

Methodology

The researchers implemented a Multilayer Perceptron (MLP) using the Keras framework to perform time-series regression. Two primary feature engineering strategies were compared:

  1. Single-Factor Models: Utilizing only historical dominant visibility (Vis) data from the past $n$ hours.
  2. Multi-Factor Models: Integrating historical visibility with five additional meteorological variables: temperature (T), dew point temperature (TD), relative humidity (RH), wind direction (WD), and wind speed (WS).

Data was normalized to a $[0, 1]$ range to ensure variables with different scales (e.g., visibility in meters vs. wind speed) contributed appropriately to the model. The team evaluated 28 configurations to determine the optimal "look-back" window for predictions.

Dataset and Distribution

The study leveraged a decade of hourly observations from Urumqi International Airport (2007 to October 2016), focusing on the critical October–March winter window. The dataset comprised 43,752 records, split 80/20 for training and testing.

Figure 6: Distribution of visibility levels at Urumqi International Airport, showing a significant frequency of low-visibility events (0-2 km) and clear conditions (>10 km).

As shown in the figure above, the visibility distribution highlights the prevalence of challenging conditions, with a notable peak in the 1-2 km range and a high frequency of clear conditions exceeding 10 km. This distribution underscores the importance of accurate forecasting in the sub-2km range for operational safety.

Key Results and Insights

The analysis revealed several critical findings regarding model performance:

  • Low-Visibility Accuracy: For the most hazardous conditions (visibility ≤ 1000 m), the single-factor model surprisingly outperformed the more complex multi-factor model. In extremely low visibility (< 300 m), the single-factor approach achieved a mean absolute error of just 86 meters.
  • Stability in Clearer Air: While the single-factor model excelled at the low end, the multi-factor model demonstrated greater stability and accuracy for visibility exceeding 2000 meters, suggesting that broader meteorological trends become more dominant as fog dissipates.
  • Performance Metrics: The overall mean absolute error across all conditions was 706 meters, which dropped to 325 meters for visibility ≤ 1000 meters. The models successfully tracked visibility trends, including rapid drops and improvements.

Conclusion and Future Work

A primary challenge identified was hysteresis, where the model reacts to sudden visibility changes with a slight time lag. Additionally, error rates increased as visibility surpassed 1500 meters due to higher variance in clearer air.

Future research will focus on incorporating more complex atmospheric data, such as high-altitude wind fields and ground pressure fields, to better understand the physical mechanisms of fog dissipation and further reduce prediction lag.

Summary updated

Abstract

This paper uses Urumqi International Airport’s hourly observation from 2007 to 2016 and builds regression prediction model for airport visibility with deep learning method. From the results we can see: the absolute error of hourly visibility is 706 m. When the visibility ≤ 1000 m, the absolute error is 325 m, and this method can predict visibility’s trend. So we can use this method to provide the airport visibility’s objective forecast guidance products for aviation meteorological services in the future. In this paper, the Urumqi area is as the research object, to explore the depth of learning in the field of weather forecasting applications, providing a new visibility return forecast for weather forecast personnel so as to improve the visibility of the level of visibility to ensure the safe and stable operation of the airport.
BibTeX
@inproceedings{zhu2017application,
  author = {Lei Zhu and Guodong Zhu and Lei Han and Nan Wang},
  title = {{The Application of Deep Learning in Airport Visibility Forecast}},
  booktitle = {Atmospheric},
  year = {2017},
  doi = {10.4236/acs.2017.73023},
}
Figure