management (Sánchez et al., 2026). However, observation alone is insufficient for proactive risk management; forecasting is needed to anticipate near-future deformation under changing operational and environmental conditions. TSF deformation reflects interacting internal processes and external forcing. Rainfall is often a key driver because could raise pore pressures and the phreatic level, reduce effective stress, and accelerate settlement or trigger instability (Zhu et al., 2022). Temperature can modulate longer-term behaviour through seasonal drying–wetting cycles and thermo-mechanical effects. Traditional approaches commonly use regression models linking displacement to cumulative rainfall, water-level indicators, or time. Multiple linear regression is widely applied due to simplicity and interpretability (Liu et al., 2023), but it can struggle with nonlinear, time-dependent behaviour. ARIMA models can reproduce short-term trends under near-linear and stationary conditions, yet performance often degrades for longer horizons in nonstationary systems with multiple interacting drivers, as is typical for tailings dams (Hill et al., 2021). Machine Learning (ML) offers an alternative by learning nonlinear relationships directly from data (LeCun et al., 2015). Long Short-Term Memory (LSTM) models, have shown strong performance in dam and landslide forecasting by capturing temporal dependencies (Nava et al., 2023; Yang et al., 2020; Zhu et al., 2022), but they typically require large, consistent datasets for robust training, which are not always available in TSF monitoring (Bansal et al., 2022). Ensemble tree methods, particularly gradient boosting, are therefore attractive for moderate sample sizes and heterogeneous predictors. LightGBM is a computationally efficient implementation that has achieved strong deformation forecasting performance using rainfall histories, hydrological indicators, and displacement “memory” features (Chen et al., 2025). In this study, we present a LightGBM framework for short-term TSF vertical deformation forecasting using Sentinel-1 MT-DInSAR, rainfall, and temperature. Zone-specific models predict the next deformation increment at the native 12-day acquisition interval, preserving operational relevance. Benchmarking against baseline approaches shows that gradient boosting can provide useful near-term forecasts to complement TSF monitoring and support risk-informed decisionmaking. 2. MATERIALS AND METHODS 2.1 Study area and data overview The study was conducted at an operational TSF in South America, located in a humid, mild, high-rainfall climate. The facility is impounded by a downstream embankment with a maximum height of 40 m and a crest length of 545 m. Further site identifiers and operational details are not reported due to confidentiality constraints under a non-disclosure agreement. 2.1.1 Rainfall and temperature Meteorological forcing was derived from a meteorological station, located approximately 1 km from the TSF, and aggregated to the 12-day Sentinel-1 acquisition intervals used in the deformation analysis. Table 1 shows that over February 2020 to July 2025, 12-day accumulated rainfall averaged 104.73 mm (IQR: 70.91-130.53 mm), with a maximum of 267.13 mm. Mean air temperature averaged 18.27 °C while mean maximum and minimum temperatures were 23.28 °C
RkJQdWJsaXNoZXIy MTM0Mzk2