Dual-Geometry Sentinel-1 SBAS-InSAR for Landslide-Related Deformation Mapping and InSAR-Informed Susceptibility Assessment in Tizi Ouzou, Northern Algeria

This study presents a regional-scale, process-oriented assessment of active landslide-related deformation in TiziOuzou, northernAlgeria, over approximately 2429 km2, using dual-geometry Sentinel-1 SBAS-InSAR time-series analysis. Ascending and descending Sentinel-1 datasets acquired between January 2023 and December 2025 were processed to derive line-of-sight (LOS) deformation products and dual-geometry deformation indicators. A multi-criteria hotspot detection framework was developed by combining absolute LOS velocity, cumulative displacement, temporal coherence, C-index-based geometric visibility, spatial velocity gradients, and approximate two-dimensional deformation amplitude. The final hotspot layer identified 260 deformation polygons, grouped into 190 unique hotspot identifiers, covering approximately 47.9 km2. Of these 190 hotspots, 64 met the size and score criteria for temporal analysis. These hotspots were interpreted as InSAR-derived active deformation candidates rather than confirmed landslide boundaries. Their relevance was assessed through benchmarking against velocity-threshold, Sobel-gradient, and Getis–Ord Gi* strategies, comparison with a compiled bibliographic landslide inventory, temporal rainfall–deformation screening, targeted field verification, and first-order infrastructure exposure analysis. At HS_119, a sustained ascending LOS displacement trend contrasted with predominantly oscillatory descending behaviour. Descending absolute displacement rates showed similar correlations with 60-day antecedent rainfall at the hotspot and a nearby terrain-matched comparison area (r = 0.240 and 0.241, respectively). Targeted field verification at the Boubroun hotspot in the Azazga area provided qualitative support for the InSAR-based interpretation by documenting landslide-related deformation indicators within the inspected sector. To quantitatively integrate environmental conditioning factors, a Random Forest model was trained using hotspot-derived positive samples and pseudo-stable background samples. The RF, including distance to mapped faults, achieved an ROC-AUC of 0.951 and an average precision of 0.899; test recall was 0.606 at a threshold of 0.5 and 0.886 at the training-selected threshold of approximately 0.219. The RF score was combined with normalized InSAR evidence as a prioritization overlay, without a demonstrated improvement in historical-inventory consistency. The proposedworkflowprovides a transferable framework for regional landslide activity screening and susceptibility-oriented prioritization in data-scarce mountainous environments.