DINOv3-Driven Semantic Segmentation for Landslide Mapping in Mountainous Regions

Area of Science:

• Geosciences
• Remote Sensing
• Computer Vision

Background:

• Landslide hazard assessment requires integrating heterogeneous remote sensing data.
• Automating this process is challenging due to resolution and texture differences between satellite and aerial sensors.

Purpose of the Study:

• To develop a robust segmentation framework for sensor-agnostic landslide mapping.
• To extract representations that are consistent across different remote sensing data sources.

Main Methods:

• Utilized a DINOv3 transformer encoder for feature extraction.
• Exploited multi-layer transformer representations to capture diverse visual information.
• Developed a framework for sensor-robust representation learning.

Main Results:

• Achieved a Dice coefficient of 0.96 and IoU of 0.938 on the Longxi satellite dataset.
• Achieved a Dice coefficient of 0.965 and IoU of 0.941 on the Longxi UAV dataset.
• Demonstrated consistent segmentation performance across datasets with varying spatial resolutions and textures.

Conclusions:

• The proposed framework effectively addresses challenges in fusing heterogeneous remote sensing data for landslide assessment.
• The DINOv3-based approach provides sensor-robust representations, enabling accurate segmentation across different platforms.