Analysis of Information Leakage in Decomposition-Based GNSS Coordinate Time Series Prediction


Şimşek M., Taşkıran M., Doğan U.

Advances in Space Research, cilt.1, sa.1, ss.1-27, 2026 (Scopus)

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 1 Sayı: 1
  • Basım Tarihi: 2026
  • Doi Numarası: 10.1016/j.asr.2026.08.090
  • Dergi Adı: Advances in Space Research
  • Derginin Tarandığı İndeksler: Scopus
  • Sayfa Sayıları: ss.1-27
  • Yıldız Teknik Üniversitesi Adresli: Evet

Özet

Signal decomposition methods such as Variational Mode Decomposition (VMD), combined with Long Short-Term Memory (LSTM) networks, have been widely adopted for Global Navigation Satellite System (GNSS) coordinate time series prediction, with studies reporting Root Mean Square Error (RMSE) improvements of 30–75% and coefficient of determination (R2) values approaching 0.99. However, these approaches typically apply decomposition to the entire dataset before train/test splitting, allowing future observations to influence training targets and test inputs. This study quantifies information leakage through a benchmark of 7 GNSS stations (P007, P075, P275, P308, P349, FGST, SPMS) in the western United States, 5 decomposition methods (VMD, Empirical Mode Decomposition (EMD), Singular Spectrum Analysis (SSA), Discrete Wavelet Transform (DWT), Haar Maximal Overlap Discrete Wavelet Transform (MODWT)), and 6 LSTM prediction scenarios. Three findings emerge: First, offline VMD+LSTM forecasting RMSE converges to the VMD reconstruction error floor (ratio of means 1.04 across the 3D coordinate components (E, N, U) of 7 GNSS stations), demonstrating that the LSTM reproduces leaked mode values instead of learning temporal patterns. Second, enforcing temporal causality through sliding-window VMD degrades performance below naive persistence (Skill Score, SS = −0.23), while the strictly causal Haar MODWT+LSTM performs identically to the no-decomposition baseline (SS +0.162 versus +0.159). Third, decomposing training and test sets independently does not resolve the problem, as within-segment leakage persists. Synthetic validation with GNSS colored noise parameters confirms these are structural properties of each decomposition method. The reported accuracy gains in decomposition-based GNSS prediction are therefore attributable to information leakage, not improved forecasting skill.