Classification of Solid Earth Tide Observation Data Using GBDT-LGBM Hybrid Stacking Model

Solid Earth tide observation sequences reflect crustal deformation, Earth’s internal structure and tidal dynamic evolution, with important theoretical and engineering value for earthquake precursor monitoring. Conventional standalone machine learning models struggle to handle the nonlinear, non-stationary features of solid tide data, showing limited accuracy, weak generalization and unstable error fluctuation. This paper constructs a GBDT-LGBM hybrid stacked ensemble classification model. GBDT and LGBM serve as dual base learners to extract latent features from tidal time series, and an upper meta-learner fuses their outputs to leverage their complementary advantages. The model classifies five types of solid tide records: observation system failures, natural environmental disturbances, human interference, geophysical events and normal tidal fluctuations. It perfectly discriminates human-activity and system-failure samples; natural disturbances reach AUC = 0.997, while geophysical events and normal tides have AUC values of 0.944 and 0.950. A small number of misclassifications are caused by similar waveforms with gentle baseline drift and minor local disturbances. Benchmark comparisons show this ensemble achieves average macro-AUC, accuracy and macro-F1 above 0.97. Box plots from 40 repeated experiments display narrow metric ranges without outliers, demonstrating low overfitting risk. This model effectively captures temporal variations in tidal data and provides reliable support for solid tide data classification and regional crustal dynamic monitoring.