LIA-adaptive Goldstein phase filtering for improved InSAR displacement estimation in landslide-prone terrain

Abstract Reliable InSAR-based monitoring of ground deformation is often hindered by atmospheric disturbances and phase noise, particularly in regions with rugged topography and high vegetation cover. However, limited work has explored integrated phase filtering and atmospheric correction strategies in complex, landslide-prone terrains like Mizoram, India. This study proposes an enhanced methodology combining a Modified Goldstein phase filter and Atmospheric Phase Screen (APS) removal to improve Line of Sight (LOS) displacement estimation. Mizoram, characterized by frequent landslides and challenging climatic conditions, provides an ideal testbed to evaluate these improvements. Twenty Sentinel-1 SAR images (10 ascending, 10 descending) were processed using SRTM DEM-based co-registration, with corrections addressing height distortions and DEM voids. This study modifies the standard Goldstein filter that incorporates Local Incidence Angle (LIA) information to enhance phase unwrapping stability, combined with Atmospheric Phase Screen (APS) removal applied before phase filtering. Furthermore, an ensemble learning framework using a Random Forest regressor was employed to model and refine time-series displacement trends. The proposed method constrained phase values within a range of -3 to 3 radians, improved displacement map interpretability, and achieved a LOS displacement of±80 mm/month. Validation against ground-collected levelling survey data showed a strong correlation of 76% (n=20 levelling points, p < 0.01) with the modified Goldstein filter applied LOS displacement values, as against 46% (n=20 levelling points, p < 0.05) for the normal Goldstein phase filter applied LOS values. Additionally, 5-fold cross-validation of the Random Forest ensemble model yielded training and testing R² values of 0.892 and 0.856, respectively, confirming the reliability of the displacement estimates.