K-Fold Cross-Validation: An Effective Hyperparameter Tuning Technique in Machine Learning on GNSS Time Series for Movement Forecast

In deformation analysis, irregularly spaced data, extreme values, and anomalies in time series can lead to misleading simulations for forecast models, such as overfitting and underfitting. Therefore, K-fold cross-validation is one of the hyperparameter tuning techniques used in machine learning (ML) to deal with these problems. In this study, we use data from 22 permanent GNSS stations to predict the motion trajectory of the Earth’s crust. Lag functions and sampling techniques are applied to generate 924-time series samples. Time series standardization techniques are also performed to improve the quality of data. To test the efficiency of the K-fold cross-validation method, we investigate 26 mathematical models based on six ML algorithms. The optimal K values are selected through trial methods. Root mean squared error (RMSE) of validation and test is the basis for determining the overfitting and underfitting models. The investigations show that the optimal intervals of K-fold range from five to ten folds for the GNSS time series with many anomalies, jumps, and significant variations, from three to ten for stable time series. The sensitivity of cross-validation is more effective on the time series of the Up component than those of the North and East components. In addition, cross-validation can also detect effectively overfitting and underfitting for forecast models in motion of permanent GNSS stations.