AI-Driven Change Detection Using SAR, LIDAR, And Sentinel-2 Data for Landslide Monitoring and Disaster Early Warning Systems

Abstract

This study addressed the problem that many AI-driven landslide early-warning pipelines generate technically strong change maps but lack auditable, decision-aligned evidence products that stakeholders can trust and act on consistently under uncertainty. The purpose was to evaluate, in a quantitative cross-sectional, case-based design, whether multi-sensor AI change evidence (SAR, LiDAR, Sentinel-2) and socio-technical decision constructs jointly explain perceived Early Warning Effectiveness (EWE). Data were collected from a professional sample of N = 162 screened-valid respondents drawn from operational roles (disaster management 31.5%, GIS/remote sensing 27.8%, engineering 22.2%, planning 18.5%; mean experience 8.6 years) evaluating the case outputs and decision readiness constructs. Key variables included Perceived AI Performance (PAI), Interpretability (INT), Trust in Alerts (TRU), Decision Confidence (DCF), sensor contribution perceptions (PSC_SAR, PSC_LiDAR, PSC_S2), and the fused Multi-Sensor Change Evidence Score (MSCES) as predictors, with EWE as the main dependent variable. The analysis plan applied descriptive statistics, reliability testing (Cronbach’s α), Pearson correlations, and hierarchical regressions comparing single-sensor baselines to fusion and a full decision-aligned model, plus threshold sensitivity and agreement/conflict diagnostics. Construct means were high (PAI M = 4.21, INT 4.02, TRU 4.08, DCF 4.16, EWE 4.12) with strong reliabilities (α = 0.83–0.90). Correlations supported the hypothesized pathway (PAI–EWE r = 0.68, TRU–EWE 0.62, INT–EWE 0.59, all p < .001). Regression results showed LiDAR-only (R² = 0.31) exceeded SAR-only (0.24) and Sentinel-2-only (0.18), while MSCES fusion improved explanatory power (R² = 0.46, β = 0.68, p < .001). In the full model, explained variance increased to R² = 0.62 (Adj. 0.61), with significant unique effects for PAI (β = 0.34), TRU (β = 0.29), MSCES (β = 0.27) and INT (β = 0.18). Operationally, the case pipeline produced 128 mapped change objects, with 46 high-priority zones; agreement classes showed higher confidence when evidence converged (3-sensor agreement 38.4%, confidence M = 4.34 vs single-sensor-only 19.9%, M = 3.61). Trigger reliability further demonstrated robustness: triggered zones dropped from 46 (MSCES 0.60) to 24 (0.80) while a stable core of 24 zones persisted (TSI 0.52 at 0.60, 1.00 at 0.80). These findings imply that deployable landslide early warning should prioritize multi-sensor fusion plus interpretable, graded outputs and threshold-stability reporting to reduce alarm fatigue, improve accountability, and strengthen calibrated trust in alerts across agencies.