Data-Driven Earth Observation: Evaluating Ensemble Machine Learning and Spatial Interpolation for GPS Velocity Estimation in a Novel Geodetic Network

Accurate measurement of Global Positioning System (GPS) horizontal velocity fields is vital for tracking changes in the Earth's crust, assessing seismic risks, and updating dynamic geodetic reference frames. In the past, researchers have used spatial interpolation methods to find velocity in areas with few stations. However, the complex nature of tectonic boundaries and fault movement often limits the accuracy of these models. This study explores the use of machine learning algorithms, including Random Forest (RF), eXtreme Gradient Boosting (XGB), and LightGBM, alongside traditional methods like Radial Basis Function (RBF) and Inverse Distance Weighting (IDW) in a previously untested GPS velocity field. A clean network of 112 GPS stations for the Marmara region of Türkiye was split into a 70% training set and a 30% testing set to evaluate the algorithms' generalization. The predictive performance varies significantly with the directional velocity component. For the East velocity field, gradient boosting methods showed stronger generalization abilities. LightGBM achieved the lowest Root Mean Square Error (RMSE) of 1.173 mm/yr and Mean Absolute Error (MAE) of 0.942. This performance is nearly twice as good as the best traditional method (RBF Linear, RMSE: 2.354 mm/yr, MAE: 1.749 mm/yr). In contrast, the North velocity component is best modeled using the bagging method of the RF algorithm, which yields the lowest RMSE of 0.909 mm/yr and MAE of 0.678 mm/yr, effectively reducing localized noise. The findings suggest that machine learning approaches can significantly lower prediction errors compared to traditional interpolation methods. However, the choice of algorithm must be carefully matched to the specific geodynamic, anisotropic, and kinematic characteristics of the directional vector being studied.