Activity-aware landslide susceptibility mapping by InSAR-constrained negative sampling and interpretable tabular GeoAI

Reliable negative samples are essential for landslide susceptibility mapping, yet unmapped or slowly deforming slopes may be mislabeled as non-landslides. This study develops an activity-aware framework using SBAS-InSAR deformation to improve negative-label reliability. The framework was evaluated in the Badong-Zigui section of the Three Gorges Reservoir Area using 121 mapped landslides and Sentinel-1A observations from 2020-2024. It integrates SAR-geometry screening, deformation-constrained negative sampling, spatial holdout, SMOTENC-Tomek, TabPFN, repeated sampling, statistical testing, and SHAP interpretation. At the study-area scale, geometry and deformation screening jointly excluded 56.81% of the area after spatial overlap was accounted for. Under the landslide-object-level holdout, S-T-TabPFN (InSAR) achieved a PR-AUC of 58.66% and a ROC-AUC of 94.64%. Compared with S-T-TabPFN, PR-AUC increased by 2.55 percentage points (approximately 4.55% relative improvement), while ROC-AUC increased by 0.22 percentage points (approximately 0.23% relative improvement). Recall, F1-score, and MCC increased by 12.12, 5.93, and 5.00 percentage points, respectively. High-susceptibility zones were concentrated along reservoir-bank slopes, and SHAP identified distance to river, annual average precipitation, lithology, and fault proximity as the leading contributors to model predictions. At Pengjiawuchang, areas of higher predicted susceptibility broadly overlapped with InSAR-derived deformation and field-observed damage. The displacement record also showed temporal variations associated with rainfall and reservoir-water-level changes. These results indicate that deformation-informed negative sampling improves minority-class recognition, consistent with a potential reduction in pseudo-negative contamination, providing a reproducible approach for more reliable and interpretable landslide susceptibility assessment.