Richter, K. (2025): <b>Machine Learning-supported visibility forecasting by combining station, Meteosat and reanalysis data </b> Philipps University of Marburg, <i>master thesis</i>
Resource Description
Title:
Machine Learning-supported visibility forecasting by combining station, Meteosat and reanalysis data
FOR816dw ID:
580
Publication Date:
2025-06-23
License and Usage Rights:
Resource Owner(s):
Individual:
Kai Richter
Contact:
email:
richter5 <at> students.uni-marburg.de
Germany
Abstract:
Accurate forecasts of radiation fog are an objective of significant relevance due to its impact on traffic, aviation, and transportation. This <br/>
study will explore the adaptation and enhancement of a previously developed Machine Learning-based nowcasting framework for <br/>
radiation fog events. The objective is to explore the expansion potential to a spatial scale and model accuracy improvements through <br/>
application at three distinct weather station locations that experience radiation fog. Further, the effectiveness of Numerical Weather <br/>
Prediction (NWP) data as additional predictor variable source on model performance will be assessed. This will be performed through <br/>
integration of datasets from German Weather Service (DWD) stations, Meteosat Second Generation (MSG) channel properties and <br/>
regional reanalysis variables from COSMO NWP model. Distinct model variants based on different dataset combinations (Station, <br/>
MSG+COSMO, Station+MSG+COSMO, Visibility-Only) will be evaluated. Using eXtreme Gradient Boosting (XGBoost) algorithm, <br/>
the framework forecasts absolute visibility with 60-minute lead time. A persistence model serves as benchmark. Performance will be <br/>
assessed using scoring metrics (Accuracy, Correlation, Percentage bias, Mean Absolute Error) across the full visibility range and three <br/>
visibility threshold bounds (2 km, 1.1 km, 0.4 km). Temporal accuracy of fog formation and dissipation will be determined through <br/>
evaluation of fog formation and dissipation time shifts. XGBoost models mostly outperform PM, with tendencies of <br/>
Station+MSG+COSMO variant performing best and MSG+COSMO variant worst. Prediction difficulties arise in the 0.4 km threshold <br/>
segment due to measurement resolution limitations and value imbalance of visibility data. The model variants reliably predict fog event <br/>
transitions, with the majority forecasted with deviations < 30 minutes and only few events overseen. A consistent tendency towards <br/>
delayed prediction is observed. Variability in model performances across station locations suggests that small-scale environmental <br/>
characteristics contribute to different model robustness at distinct sites. The results indicate strong potential for further spatial framework <br/>
extension. COSMO variables partially contribute to improved model performance. The framework marks a solid foundation for future <br/>
exploitation.
Keywords:
| Radiation fog | fog horizontal visibility | Machine learning | Nowcasting | XGBoost |
