Bamahry, Fikri ; Legrand, Juliette ; Bruyninx, Carine ; Pottiaux, Eric

Poster presented at 4th ESA-ECWMF Workshop on Machine Learning for Earth Observation and Prediction, ESA/ESRIN, Frascati, Italy on 2024-05-09

Abstract: A dense network of GNSS (Global Navigation Satellite Systems) stations has the potential to monitor ground deformations accurately. However, positions may exhibit artificial variations or outliers unrelated to the actual ground motion and can degrade their geodynamic interpretation. We know that data quality degradation is one of the key factors influencing the outliers on GNSS position estimates. Currently, the interpretation process is performed manually by looking at data quality indicator plots, which is time-consuming and subjective to the analyst who interprets the GNSS time series. To have a more robust and efficient system and find the complex correlation between degraded data quality indicators and outliers in daily GNSS position time, we implemented a random forest algorithm to develop a model that can identify degraded GNSS quality that might be causing outliers in position time series. We used the daily position and daily data quality indicators time series of GNSS stations belonging to EUREF GNSS Permanent Network (EPN) to assess the most relevant GNSS data quality indicators that can explain the possible outliers in the GNSS position time series. Furthermore, to understand how the machine learning model predicts the possible outliers, we used the SHapley Additive exPlanations (SHAP) algorithm to explain the prediction that we got from our model. In this presentation, we will present the status of our investigations, the challenges we faced, and the preliminary outcome of this work.

Keyword(s): GNSS quality metrics ; GNSS position time series ; Machine learning
Funding: ESFRIFED/ EF/211/SERVE / SERVE