and 15.56 °C, respectively. The meteorological series indicates marked rainfall variability through time, with wetter intervals interspersed with comparatively dry periods, whereas temperature varied within a narrower range. Table 1 – Summary of meteorological conditions used in the analysis Variable Mean Median IQR Minimum Maximum Rainfall accumulation (mm) (12-day) 104.73 100.03 70.91130.53 10.24 267.13 Mean air temperature (°C) 18.27 18.25 17.7518.75 16.24 20.47 Maximum air temperature (°C) 23.28 23.17 22.4124.11 20.74 27.68 Minimum air temperature (°C) 15.56 15.66 14.9816.23 12.9 17.48 2.2 MT-DInSAR time-series Deformation monitoring relied on Sentinel-1A C-band SAR Interferometric Wide (IW) swath single look complex (SLC) data, enabling frequent, all-weather observations. A time series from January 2020 to July 2025 were processed with the SBAS approach implemented in the opensource OpenSARLab framework, producing pixel-wise LOS displacement time series and temporal coherence. For the vertical deformation analysis, pixels with temporal coherence ≥ 0.6 were retained to ensure measurement reliability. Vertical deformation (dU) was estimated by decomposing ascending and descending observations using pixel-wise acquisition geometry. Time-series uncertainty was quantified via coherencebased uncertainty propagation, yielding . 2.3 Prediction framework and model structure To forecast vertical deformation, we implemented a data-driven prediction framework based on gradient boosting. LightGBM was selected as the regressor because it is computationally efficient and performs well on structured data, while capturing nonlinear relationships and feature interactions relevant to the vertical deformation increment Δ +Δ . Separate LightGBM models were trained for each zone (crest, toe, west, and east) to reflect distinct deformation regimes across the TSF. This zone-specific strategy avoids averaging behaviour across areas with different kinematics and is consistent with common practice in deformation forecasting, where nearby locations can respond differently to the same forcing. Each model performs one-step-ahead prediction aligned with the next Sentinel-1 acquisition. Formally, the model learns ( )→Δ +Δ , where is the feature vector at time and Δ +Δ is the predicted deformation increment at the subsequent observation. One-step forecasts can be iterated to extend predictions over longer horizons when required. 2.3.1 Feature engineering for model inputs To train the LightGBM models, a comprehensive set of predictor features was constructed, these were derived from the deformation time-series and auxiliary data, see Table 2.
RkJQdWJsaXNoZXIy MTM0Mzk2