Landslide Detection with Sentinel-2 Multi-Temporal Composites, Persistence, and U-Net

Extreme precipitation events are increasing, requiring more efficient methods of managing hydrogeological hazards including landslides and mass-transportation by floods. We develop and assess a semi-automatic workflow for post-event rapid landslide detection from Sentinel-2 imagery with a U-Net segmentation model using NDVI and a persistence metric for identifying areas with sustained vegetation loss. The pipeline, including data acquisition, preprocessing, training, and predicting, was implemented in Python (Google Earth Engine) for transparency and reproducibility. Training and validation used well-verified inventories from two major events: Jølster (2019) and Hans (2023). Pixel-based performance assessment gave precision/recall = 0.53/0.53 for Jølster and 0.35/0.44 for Hans (F1 ≈ 0.53 and 0.39, respectively). Object-based evaluation showed 63/120 landslides in Jølster and 16/60 in Hans achieved ≥20% areal overlap with reference polygons. Qualitative diagnosis revealed systematic errors from residual clouds and misclassification of mass-transporting floods, as well as missed detections where the NDVI change is weak including for smaller landslides that are not clear in 10 m resolution images. Despite these limitations, full-tile inference can be completed within three hours, indicating clear potential for improving efficiency compared to manual mapping. We outline avenues to improve robustness and generalization, including improved cloud handling, integrating SAR and higher-resolution optical data, adding context layers, expanding and standardizing training labels. By establishing a workflow that can be continually improved with additional training data, or adapted for use in other regions, the presented workflow constitutes an important foundation toward improving the efficiency of landslide mapping.