GeoFusion-3D: Multi-Scale Geomorphic Feature Fusion for Landslide Scar Detection Using UAV-Mounted LiDAR.

Area of Science:

• Geosciences
• Remote Sensing
• Geomorphology

Background:

• Traditional landslide detection methods often require supervised learning or Digital Elevation Model (DEM) data, limiting their applicability in varied terrains.
• These methods can struggle with rapid deployment and generalization due to data requirements and terrain heterogeneity.

Purpose of the Study:

• To develop a zero-shot, fully unsupervised framework for landslide detection using raw UAV-mounted LiDAR data.
• To eliminate the need for labeled datasets, pre-event baselines, or rasterized terrain representations.

Main Methods:

• A multi-scale approach combining point-level (PCA-based residual depth, local concavity) and cluster-level (geomorphometric descriptors via 3D clustering) instability indicators.
• Utilizing raw 3D point cloud data to identify geometric inconsistencies indicative of landslides.
• Adaptive feature fusion to integrate diverse geomorphometric features for robust detection.

Main Results:

• The framework generates a probabilistic instability field for spatially coherent landslide scar delineation, including rupture and displaced material zones.
• Demonstrated improved detection of subtle terrain disturbances compared to DEM-based and supervised learning pipelines.
• Showcased robustness to noise and terrain variability across diverse geomorphic settings.

Conclusions:

• Unsupervised, geometry-driven inference on raw 3D LiDAR data offers a practical and scalable alternative for near real-time landslide detection.
• The method provides valuable priors for post-event susceptibility analysis without temporal data.
• This approach enhances the capability of Unmanned Aerial Vehicle (UAV)-based systems for rapid landslide monitoring.