Airport visibility forecasting method based on deep learning using U-net model

Precise visibility estimation is crucial for various meteorological applications. This study develops a deep learning-based visibility forecasting model utilizing the U-net architecture, applied to airport video data. The model was trained on 6 hours of video footage, achieving a final loss value of approximately 0.025 and mean relative errors of 2.27 m, 4.53 m, and 3.19 m for the Training Set, Test Set 1, and Test Set 2, respectively. By incorporating batch normalization and modified activation functions, the model effectively reduces prediction errors and outperforms traditional statistical models. The results indicate that the U-net model provides a more precise estimation of visibility, with potential applications in enhancing airport visibility forecasting accuracy and ensuring public safety.