Multi-Source Machine Learning of Tailings Dam Deformation Dynamics: Sentinel-1 MT-DInSAR, Meteorological Forcing, and LightGBM Short-Horizon Forecasts *V. Sanchez1, N. Okada1, Y. Ohtomo1, R. Quispe2, C. Orozco2, S. Gallegos2, S. Morales2, Y. Kawamura1 1 Division of Sustainable Resources Engineering, Graduate School of Engineering, Hokkaido University, Nishi-8, Kita-13, Sapporo, 060-8628, Japan ² Institutional affiliation withheld due to confidentiality obligations under a non-disclosure agreement (*Presenting author: villarrealyadiravanessa.sanchez.f9@elms.hokudai.ac.jp) ABSTRACT Tailings storage facilities (TSFs) are high-consequence earth structures that require continuous deformation surveillance to support risk management during staged construction and operation. Satellite Interferometric Synthetic Aperture Radar (InSAR) provides spatially dense deformation measurements; however, translating these observations into actionable early warning requires short-horizon forecasting. This study develops and evaluates a data-driven framework to predict near-term vertical deformation of a TSF dam by integrating multi-temporal dual-orbit InSAR with meteorological forcing. 85 points were analyzed from a Sentinel-1 MT-DInSAR record spanning January 2020 to July 2025. Ascending and descending line-of-sight time series were decomposed into vertical displacement, dU. Points were grouped into four zones (Crest, Toe, West, and East) to reflect dam geometry and spatially varying deformation regimes. Zone-specific LightGBM models were trained to forecast the one-step-ahead vertical deformation increment, at the next satellite acquisition (12-day horizon) using lagged deformation-memory features, short-term kinematic indicators, and rainfall and temperature aggregates. MT-DInSAR mapping indicates diffuse, non-localised subsidence during 2020-2022, followed by clearer net subsidence by 2023. Settlement intensifies during Construction Stage 4, with the strongest and most heterogeneous deformation concentrated in the West sector and along the Crest. Meteorological forcing showed only a modest and sector-dependent effect, with the clearest improvement in the Toe sector. MAE was adopted as the main optimisation criterion because the target variable was the one-step increment expressed in millimetres, making forecast errors directly interpretable and less sensitive than RMSE to isolated anomalous acquisitions. Best test performance ranged from MAE 2.363.37 mm and RMSE 3.46-5.35 mm across zones, with weaker skill in sectors showing greater heterogeneity and construction-driven non-stationarity. Independent comparison against eight total-station control points showed four points within typical InSAR validation ranges (3-10 mm/year), supporting the suitability of MT-DInSAR inputs for localised forecasting. Overall, the results indicate that zone-specific gradient boosting can provide useful near-term deformation forecasts, while highlighting the need to incorporate operational variables of the TSF, such as: deposition rates, stage elevations, water management, construction activities, changes in drainage, since these factors are usually primary controllers of deformation behavior during construction
RkJQdWJsaXNoZXIy MTM0Mzk2