Unsupervised Landslide Detection Using Multitemporal Sentinel-2 Imagery

Landslides remain a major natural hazard with persistent gaps in global and regional inventories, largely due to the high cost and effort of field-based documentation. To address this, we investigate automated, minimal-input approaches for historical landslide detection using Sentinel-2 NDVI time series (2018–2024) across Bavaria, Germany. The study introduces the Independent Baseline Method (IBM), a novel unsupervised framework leveraging external, landslide-free reference data to mitigate baseline contamination, and compares it with two adapted techniques—Statistical Window Analysis (SWA) and Seasonal-Trend Decomposition (STL).

Evaluation across 15 documented landslide events shows that IBM delivers the most balanced and robust performance. While SWA yielded higher sensitivity, it also generated extensive false positives, whereas STL showed limited detection capacity due to baseline distortion. Detection success was positively correlated with landslide size, confirming the scalability of the approach for medium to large events. A systematic analysis also identified errors in Sentinel-2’s Scene Classification Layer as a dominant source of false detections, primarily linked to atmospheric misclassification.

Despite such constraints, IBM successfully identified previously undocumented landslide occurrences, subsequently confirmed through inventory updates. These results demonstrate that NDVI-based, low-complexity frameworks can meaningfully enhance the completeness of landslide records. The proposed approach, relying solely on open-access EO data and minimal reference information, establishes a scalable, transferable, and cost-efficient foundation for regional landslide monitoring. It also illustrates how strategic use of external baselines can substantially improve unsupervised change detection performance, paving the way for operational applications in risk assessment and environmental management.