Intercomparison of statistical and machine-learning methods for sea-level trend estimation at the Hon Dau tide gauge (1960–2024).

Abstract

This paper presents an intercomparison of statistical, machine-learning, and signal decomposition methods for estimating sea-level trends using the long-term tide-gauge record at Hon Dau tide gauge (northern Vietnam) during 1960–2024. Classical statistical approaches (Mann–Kendall, Ordinary Least Squares), machine-learning models (Random Forest, Support Vector Regression, Artificial Neural Network, Long Short-Term Memory), and signal decomposition techniques (Empirical Mode Decomposition, and Complete Ensemble Empirical Mode Decomposition with Adaptive Noise) were employed to assess linear, nonlinear, and multi-scale sea-level variations. The results indicate a stable diurnal tidal regime with pronounced seasonal modulation, with lowest water levels occurring in March (182.1 cm) and highest levels in October (208.9 cm). All methods consistently detect a statistically significant long-term rise in mean sea level of approximately 3.7–3.9 mm.yr⁻¹ (equivalent to about 25–27 cm over six decades). A higher rate of sea-level rise, on the order of 7–9 mm.yr⁻¹, is identified for the more recent period 2005–2024. Machine-learning and decomposition-based approaches provide complementary insights into nonlinear behavior and multi-scale oscillations associated with ENSO and monsoon variability, while classical statistical methods offer transparent baseline estimates of long-term change. Seasonal analyses further reveal stronger and more stable increases during the dry season and weaker, more variable trends during the wet monsoon. Overall, the consistency of trend estimates across multiple methods highlights the robustness of the observed sea-level rise and underscores the value of a multi-method framework for sea-level monitoring and climate adaptation studies in Vietnam.