Deep learning-based early warning of tailings storage facility instability using Sentinel-1 and Radarsat-2 InSAR calibrated with a geomechanical model.

This study advances the science and application of Interferometric Synthetic Aperture Radar (InSAR) for monitoring tailings storage facilities (TSFs) by integrating multi-scale and multi-source satellite data with geomechanical modelling and deep learning to detect ground movement precursors to failure. A novel calibration strategy combining geomechanical finite element (FE) predictions with InSAR measurements is introduced to dissect the origins and expressions of known InSAR limitations related to phase unwrapping. By investigating different InSAR processing algorithms along with data from both commercial (Radarsat-2) and public (Sentinel-1) sensors, the study demonstrates the extent to which these limitations can be overcome. During periods of rapid or non-linear deformation, such as buttress construction over the TSF, Persistent Scatterer InSAR (PS-InSAR) time series show more sensitivity to phase unwrapping errors, making phase ambiguities more readily observable. The extent to which such ambiguities can be identified and calibrated using physics-based finite element model predictions is demonstrated in this study. In contrast, Intermittent Small Baseline Subset (ISBAS) processing reduces the visible expression of phase ambiguities through inherent signal averaging (i.e. multi-looking), while maintaining a higher spatial density of deformation measurements. An uncertainty analysis for both InSAR approaches links accuracy to SAR signal distortions and uncertainties in FE model parameters. Finally, the capabilities of the deep learning early-warning framework capable of ingesting multi-source InSAR datasets, including different sensors, geometries and decomposed motion components are presented. Embedding-based integration of line-of-sight and decomposed Radarsat-2 (RS2) InSAR data reveals that orbit configuration and resolution directly influence predictive performance. Specifically, the horizontal deformation components provide the clearest precursors to failure, outperforming line-of-sight and vertical motion components. The combined use of RS2-InSAR, Sentinel-1 PS-InSAR, Sentinel-1 ISBAS, FE modelling and deep learning offers a promising framework for TSF monitoring. This integrated approach promises to reduce uncertainties inherent in individual methods and to significantly improve the ability to reduce risks of environmental and economic losses.