[Paper Review] A precise machine learning aided algorithm for land subsidence or upheave prediction from GNSS time series

Mostafa Kiani Shahvandi|arXiv (Cornell University)|Jun 6, 2020

Landslides and related hazards19 references4 citations

TL;DR

This paper presents a machine learning-aided algorithm for predicting land subsidence or uplift using GNSS position time series. It preprocesses data by removing tidal and atmospheric signals and transforms coordinates to a local system, then applies eight machine learning models; Gaussian processes (GP) outperform others, achieving up to 4 mm prediction accuracy with a mean of 2 cm across 14 real-world GNSS time series.

ABSTRACT

This paper is aimed at the problem of predicting the land subsidence or upheave in an area, using GNSS position time series. Since machine learning algorithms have presented themselves as strong prediction tools in different fields of science, we employ them to predict the next values of the GNSS position time series. For this reason, we present an algorithm that takes advantage of the machine learning algorithms for the prediction of positions in a GNSS time series. The proposed algorithm has two steps-preprocessing and prediction. In the preprocessing phase, the periodic tidal and atmospheric signals in the time series are removed and coordinates are transferred to the local coordinate system. In the prediction phase, eight different machine learning algorithms are used, namely, multilayer perceptron, Bayesian neural network, radial basis functions, Gaussian processes, k-nearest neighbor, generalized regression neural network, classification and regression trees, and support vector regression. We show the superiority of the Gaussian processes algorithm, compared to other methods, in 14 different real GNSS time series studies. The proposed algorithm can achieve up to 4 millimeters in accuracy, with the average accuracy as 2 centimeters across all time series.

Motivation & Objective

• To develop a precise prediction method for land subsidence or uplift using GNSS position time series.
• To improve prediction accuracy by removing periodic atmospheric and tidal signals from GNSS data.
• To evaluate and compare the performance of multiple machine learning algorithms in predicting GNSS time series trends.
• To demonstrate that machine learning, particularly Gaussian processes, significantly outperforms traditional statistical methods like Theta in subsidence/uplift prediction.
• To provide a robust, transferable framework for geoscientists and geodetic researchers to forecast land surface deformation.

Proposed method

• Preprocessing involves removing periodic tidal and atmospheric signals using a parametric model based on least squares fitting of sinusoidal components at known frequencies.
• The time series is transformed into a local coordinate system to improve prediction consistency and reduce noise.
• Eight machine learning models are applied: multilayer perceptron, Bayesian neural network, radial basis functions, Gaussian processes, k-nearest neighbor, generalized regression neural network, classification and regression trees, and support vector regression.
• The periodic signal model is defined as $ y_k^p = \sum_{i=1}^m a_i \cos(2\pi f_i t_k) + b_i \sin(2\pi f_i t_k) $, with frequencies from Table 1 (e.g., 0.5, 1, 14, 30, 180.1, 359.5 days).
• Coefficients $ a_i $ and $ b_i $ are estimated via linear least squares using training data, and the fitted signal is subtracted from original data to isolate trend-only components.
• Prediction is performed using supervised learning on the detrended, preprocessed time series, with performance evaluated using MASE, MAE, and RMSE metrics.

Experimental results

Research questions

• RQ1 Can machine learning models improve the accuracy of land subsidence or uplift prediction from GNSS time series compared to traditional statistical methods?
• RQ2 Which machine learning algorithm performs best in predicting GNSS position time series across diverse geophysical and climatic conditions?
• RQ3 How does the removal of periodic atmospheric and tidal signals affect prediction accuracy in GNSS-based subsidence/uplift modeling?
• RQ4 What is the impact of data continuity (continuous vs. discontinuous time series) on prediction performance?
• RQ5 Can a unified preprocessing and prediction pipeline be developed that generalizes across multiple GNSS stations with varying data quality and environmental conditions?

Key findings

• Gaussian processes (GP) achieved the highest prediction accuracy, with a mean RMSE of 0.0221 mm and mean MASE of 0.0229 across 14 GNSS time series.
• The GP model achieved a maximum accuracy of 4 mm in individual time series, significantly outperforming other machine learning models and the traditional Theta method.
• The Theta statistical method showed extremely poor performance, with a mean RMSE of 23.9242 mm and mean MASE of 72.2209, confirming its inadvisability for this task.
• The radial basis function (RBF) network was the least accurate model, with a mean RMSE of 1525.0325 mm and mean MASE of 1848.7074, indicating poor suitability for GNSS time series prediction.
• Prediction accuracy was significantly higher for continuous time series, with millimeter-level accuracy possible if data continuity is maintained, whereas discontinuous data reduced accuracy to the centimeter level.
• The proposed algorithm demonstrates that machine learning, especially GP, is a highly effective tool for forecasting land subsidence or uplift from GNSS data, offering a robust alternative to conventional statistical approaches.