to forecast because short-term changes associated with earthworks, surface modification, drainage adjustments, or load redistribution are not represented explicitly in the current predictors. By contrast, the East showed intermediate rates and lower variability, supporting a more stable mapping between recent deformation history and subsequent increments. The Toe was the most stable zone, although its smaller increments make residual noise proportionally more important. MAE was the most appropriate optimisation objective for this study because the target variable was the one-step deformation increment, expressed in millimetres. In this setting, MAE provides a direct measure of the typical forecast error in the same physical units as the response variable and is less dominated than RMSE by isolated large residuals. This is particularly relevant for increment forecasting, where short-lived anomalies or acquisition artefacts can disproportionately inflate squared-error metrics. KGE was not used for model selection in the final zone-wise comparison because values remained low, with a maximum of only 0.356 in the Crest, indicating limited ability to reproduce the temporal variability of the increment series. The meteorological analysis indicates that weather was not the dominant control on short-term surface deformation during the study period. High-rain intervals, defined as the upper 10% of 12day rainfall totals did not produce a systematic upward shift in deformation-rate distributions. Instead, the main contrast remained spatial, with the West and Crest showing the widest distributions and highest upper tails. This is consistent with the zone-wise model comparison: including meteorological predictors improved performance only modestly and not uniformly. The clearest gain occurred in the Toe, where MAE decreased from 2.54 to 2.36 mm and RMSE from 3.80 to 3.46 mm, whereas performance worsened in the West and changed only marginally in the East and Crest. The main event-scale mismatch occurred after the 5 January 2025 acquisition, when the network mean increment reached 34 mm in more than 90% of points. This behaviour is more consistent with an acquisition-wide InSAR artefact than with a plausible deformation event. Because the model relies on lagged deformation inputs, anomalous observation or gaps likely impact the subsequent one-step forecasts, producing the temporary mismatch seen in early 2025. However, the recovery of forecast performance within the following acquisitions indicates that this was an episodic disturbance rather than a persistent model bias. Overall, the results support the use of MT-DInSAR-based forecasting for localised TSF deformation monitoring. However, three main constraints: missing operational and internal-state predictors, non-stationarity across construction stages, and residual InSAR noise propagated into . Because the dataset and operational knowledge used in this study were provided for academic research purposes, the model outputs should not be interpreted as standalone indicators of operational risk without integration with site-specific monitoring systems and engineering assessment. Future work should focus on incorporating operational covariates and explicitly accounting for construction-stage regime shifts. 5 CONCLUSION MT-DInSAR captured clear stage-dependent deformation patterns and strong spatial contrasts across the TSF, highlighting its value as a complementary monitoring tool to support existing TSF assessment practices, not to replace conventional instrumentation, inspections, or geotechnical evaluation. Across the 2020–2025 record (166 acquisitions), deformation evolved
RkJQdWJsaXNoZXIy MTM0Mzk2