Mapping Earthquake-Triggered Landslides Through Multimodal Sentinel-1 and Sentinel-2 Earth Observation Data

Earth Observation (EO) data play a crucial role in the assessment and management of natural hazards, particularly in post-disaster contexts where rapid, reliable, and spatially consistent information is required to support emergency response and disaster risk reduction strategies. Landslides triggered by major earthquakes represent a typical cascading hazard, often affecting mountainous regions crossed by key infrastructure and occurring under adverse observational conditions such as cloud cover, strong illumination variability, and complex terrain geometry, which limit the effectiveness of conventional optical-based mapping approaches [1].
In this study, we demonstrate the benefit of combining multimodal EO data and vision foundation models for rapid mapping of earthquake-triggered landslides. We exploit the complementary information provided by Sentinel-2 optical imagery and Sentinel-1 Synthetic Aperture Radar (SAR) data within a prompt-free adaptation of the Segment Anything Model (SAM) [2]. The proposed MultiModal SAM (MM-SAM) framework integrates early fusion of optical and SAR observations with a lightweight domain adaptation strategy, enabling the transfer of SAM’s general visual representations to the geohazard mapping domain while keeping most of the pre-trained parameters frozen. This design allows fully automatic, pixel level landslide segmentation with limited labelled data, addressing key limitations of conventional Deep Learning approaches in operational post-disaster scenarios [3].
The approach is evaluated on the 2021 Haiti earthquake case study, that was the focus of a dedicated activation in CEOS Recovery Observatory Demonstrator project [4]. The analysis is conducted using a publicly available multimodal Sentinel-1 and Sentinel-2 dataset specifically developed for earthquake-triggered landslide detection [5]. Result show that the integration significantly enhances mapping robustness under challenging conditions such as cloud cover, complex topography, and heterogeneous surface characteristics. The MM-SAM framework produces accurate and spatially consistent delineation of landslide-affected areas and demonstrates stable performance across independent training, validation, and test subsets.

Overall, this work highlights the added value of multimodal EO data and foundation model-based approaches for scalable and rapid hazard mapping. The proposed MM-SAM framework represents a step toward transferable and operational tools for post-disaster landslide assessment, with potential applications in emergency response, disaster risk reduction strategies, and future multi-hazard monitoring systems.

References

[1] Meng, Shaoqiang, et al. TLSTMF-YOLO: Transfer learning and feature fusion network for earthquake-induced landslide detection in remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 2025.

[2] A. Kirillov, E. Mintun, N. Ravi, H. Mao, et al., “Segment Anything,” arXiv:2304.02643, 2023.

[3] Yu, Junchuan, et al. Landslidenet: Adaptive Vision Foundation Model for Landslide Detection. In: IGARSS 2024-2024 IEEE International Geoscience and Remote Sensing Symposium. IEEE, 2024. p. 7282-7285.

[4] Tapete, D., et al., “SAR-based scientific products in support to recovery from hurricanes and earthquakes: lessons learnt in Haiti from the CEOS Recovery Observatory pilot to the demonstrator”, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-5803, https://doi.org/10.5194/egusphere-egu22-5803, 2022

[5] Bralet Antoine, et al. "Multi-modal Remote Sensing Dataset for Landslide Change Detection in Haiti", IEEE Dataport, July 14, 2024, doi:10.21227/4heb-7h07