Integration of geospatial foundation models in unsupervised change detection workflows for landslide identification
This study investigates the integration of Geospatial Foundation Models (GFMs) into an unsupervised change detection workflow for landslide identification. As climate change increases landslide frequency, rapid and automated detection systems are essential for a timely response. However, the high co...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Taylor & Francis Group
2025-08-01
|
| Series: | International Journal of Digital Earth |
| Subjects: | |
| Online Access: | https://www.tandfonline.com/doi/10.1080/17538947.2025.2547292 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | This study investigates the integration of Geospatial Foundation Models (GFMs) into an unsupervised change detection workflow for landslide identification. As climate change increases landslide frequency, rapid and automated detection systems are essential for a timely response. However, the high cost of annotating satellite datasets has driven interest toward unsupervised methods that could operate effectively with limited data. This work addresses two key gaps: (1) the absence of a dedicated dataset for unsupervised landslide change detection, and (2) limited investigation of GFMs as feature extractors in unsupervised frameworks. To address these, we introduce the Global Landslide Dataset for Change Detection (GLaD4CD), comprising 174 Sentinel-2 bi-temporal image pairs of global landslide events, and propose LandslideMetric-CD, an unsupervised model based on Metric-CD by Bandara and Patel [2023. “Deep Metric Learning for Unsupervised Remote Sensing Change Detection.” arXiv:2303.09536 [cs]], adapted to incorporate the SSL4EO DINO GFM. While domain-guided approaches like band-specific thresholding achieve higher F1 scores (48.41% for Band 04), LandslideMetric-CD (F1 = 31.68%) outperforms fully automatic differential thresholding using the full spectral range (F1 = 19.33%) and RGB-based deep learning methods (F1 = 19.66% for Change Detection based on image Reconstruction Loss (CDRL) [Noh et al. 2024. “Unsupervised Change Detection Based on Image Reconstruction Loss.” Remote Sensing Letters 15 (9): 919–929]). These findings underscore the importance of spectral band selection and demonstrate the potential of GFMs for automated, expert-independent landslide detection. |
|---|---|
| ISSN: | 1753-8947 1753-8955 |