Figure 4 – One-step-ahead forecasts of vertical deformation increments (Δ ) from zone-specific LightGBM models versus MT-DInSAR observations with uncertainty envelopes. Figure 4 compares one-step-ahead LightGBM predictions of Δ with MT-DInSAR observations for representative points in each zone, together with the per-epoch observational uncertainty, . Across all four examples, the model reproduces the oscillatory behaviour of the increment series and captures most changes in sign, with point-level correlations ranging from 0.74 to 0.80 and MAE from 2.43 to 3.30 mm. Agreement is strongest for the Toe example (MAE = 2.43 mm; r = 0.78), while the East example shows the largest magnitude errors (MAE = 3.30 mm; r = 0.75). The West point achieves the highest but still misses the amplitude of several excursions, indicating that event timing is reproduced better than magnitude. The larger forecast misfit in early 2025 is plausibly linked to the anomalous 5 January 2025 acquisition, whose network-wide spike was likely non-physical; because the model uses lagged deformation inputs, this artefact likely propagated into the subsequent one-step predictions for several epochs. 4 DISCUSSION Zone-wise forecasting performance reflects both real differences in deformation behaviour across the TSF and spatial variation in MT-DInSAR reliability under active construction (Table 2). The West and Crest sectors showed the highest mean deformation rates and the largest dispersion, indicating stronger heterogeneity and a less stable deformation regime. These conditions are harder
RkJQdWJsaXNoZXIy MTM0Mzk2