Project Details
Projekt Print View

FOrecasting radiation foG by combining station and satellite data using Machine Learning (FOG-ML)

Subject Area Physical Geography
Term from 2019 to 2024
Project identifier Deutsche Forschungsgemeinschaft (DFG) - Project number 424021288
 
Final Report Year 2024

Final Report Abstract

The FOG-ML project aimed to improve the prediction of radiative fog formation and dissipation by combining machine learning (ML) techniques with various data sources, including ground station measurements and satellite data. The overall goal was to develop a powerful and robust fog prediction system that exploits the strengths of both types of data. In the first phase, the main challenges in training and validating ML models for fog prediction, in particular with the eXtreme Gradient Boosting (XGBoost) algorithm, were identified. It was found that problems such as temporal order and autocorrelation in the data lead to an overestimation of model performance. The introduction of baseline models and a more application-oriented evaluation method, the Expanding Window Approach (EWA), provided a more accurate assessment of the model's operational capabilities. In the next phase, the incorporation of pre-processed temporal data and smoothing techniques on the performance of the XGBoost model was investigated. This data preprocessing resulted in a significant increase in the accuracy of the model, with the F1 score increasing from 0.75 to 0.88. This demonstrates the importance of data pre-processing in improving short to medium term fog forecasts. The final phase evaluated the combination of ground-based station data with Meteosat Second Generation (MSG) satellite data for fog forecasting. The investigations showed that models using MSG data in combination with visibility measurements provide comparable forecast performance to models using only station data. The combined data set shows a balanced model performance, especially in the prediction of fog formation and resolution transitions. The use of MSG data thus represents a cost-effective alternative for large-scale fog forecasts. The project has successfully demonstrated the potential of ML in fog forecasting, in particular the combination of station and satellite data to improve the accuracy and efficiency of fog forecasting. The use of satellite data, which provides wide coverage, is a cost-effective alternative to relying on station data alone. The results lay the foundation for a comprehensive, nationwide fog forecasting system in Germany. Future work will focus on extending this approach to multiple stations and further refining the temporal dynamics of the model for longer term forecasts.

Publications

 
 

Additional Information

Textvergrößerung und Kontrastanpassung