Modelling geometric geoid of Taiwan island integrating GNSS leveling and gravity data with multi-layer perceptron

This study develops multi-layer perceptron (MLP)-based geoid prediction models for Taiwan using 1′×1′ gravity-anomaly grids, EGM2008 to degree and order 2159, and GNSS/levelling control information. A unified MLP model and an empirical zonal MLP model were constructed to account for spatial differences in terrain and gravity-field characteristics between the Central Mountain Range and coastal regions. The zonal scheme uses the median observed geoid height of the control points, 21.78 m, as a threshold, together with a 0.50 m numerical overlap interval and mean fusion of predictions within the overlap zone. Independent hold-out validation and five-fold cross-validation were used to evaluate predictions at benchmarks excluded from training. The validation results provide evidence of the predictive feasibility of the present experiment under the adopted validation design. Common-grid model-difference analysis shows that the MAEs of the zonal model relative to Twgeoid_hybrid, Stokes_model_tw, LSC_model, and XGM2019e_2159 are 0.10, 0.10, 0.04, and 0.13 m, respectively, which are lower than those of the unified model by 0.01, 0.02, 0.03, and 0.02 m. These results indicate that the zonal model has smaller common-grid numerical differences with the selected comparison surfaces.