Zero-shot detection for InSAR-based land displacement by the deformation-prompt-based SAM method

verfasst von: Yufang He, Bo Chen, Mahdi Motagh, Yuyan Zhu, Songdong Shao, Jiaye Li, Bing Zhang, Hermann Kaufmann
Abstract: Recently, geological disasters such as land subsidence, landslides, mining-related collapse and others have been occurring more and more, posing serious threats to social and economic stability as well as human safety. InSAR (Interferometric Synthetic Aperture Radar) and advanved InSAR time-series technologies can monitor surface displacements at sub-centimeter level of accuracy, which has great and significant potential in detecting these disasters. Although several automated deformation detection methods based on CNN (Convolutional Neural Networks) have been proposed, these methods are not only time-consuming- requiring specific datasets and models- but also exhibit limited generalizability. In this study, we propose a zero-shot deformation-prompt-based SAM (Segment Anything Model) method, which does not require any pre-training operations or establishing special datasets. It initially maps the surface displacement velocity of the study area by the InSAR time-series analysis. Then the deformation-prompt-based SAM method is proposed to directly identify the exact locations and morphologies of surface deformations on the velocity maps. The method primarily extracts the rectangular boxes and points of deformation areas according to the color display characteristics of the InSAR-based land deformation map. Then these boxes and points of deformation zones serve as the prompt to the SAM, which detects deformation zones in a zero-shot manner. To validate its effectiveness, experiments on deformation extraction for various types of geohazards are conducted by utilizing Sentinel-1 and TerraSAR-X images. The results show that 137 and 93 local urban subsidence areas in Shenzhen City are identified based on different SAR images. Most of them can be automatically detected by the proposed method. Additionally, we have identified 18 landslides near Jichang Town, 36 landslides in Heifangtai and 27 major deformation zones in the Huainan mining area of Anhui Province. Statistical analysis reveals that our method can obtain high pixel accuracy in detecting land subsidence, landslides, and mining-related deformation areas with zero-shot based on different SAR images. Thus, the automatic surface deformation detection method holds considerable promise for compiling and regularly updating inventories of surface deformation data related to different types of geohazards. It has the potential to significantly enhance monitoring efficiency and provide robust technical support for disaster prevention and loss reduction.
Organisationseinheit(en): Institut für Photogrammetrie und Geoinformation
Externe Organisation(en): Dongguan University of Technology
Harbin Institute of Technology
Helmholtz-Zentrum Potsdam Deutsches GeoForschungsZentrum (GFZ)
Hong Kong Polytechnic University
Shandong University (SDU)
Typ: Artikel
Journal: International Journal of Applied Earth Observation and Geoinformation
Band: 136
Anzahl der Seiten: 13
ISSN: 1569-8432
Publikationsdatum: 02.2025
Publikationsstatus: Veröffentlicht
Peer-reviewed: Ja
ASJC Scopus Sachgebiete: Globaler Wandel, Erdoberflächenprozesse, Computer in den Geowissenschaften, Management, Monitoring, Politik und Recht
Ziele für nachhaltige Entwicklung: SDG 11 – Nachhaltige Städte und Gemeinschaften
Elektronische Version(en): https://doi.org/10.1016/j.jag.2025.104407 (Zugang: Offen)

BibTeX

@article{15ef7d50e08749f4baa0e3c60219cc2d,
title = "Zero-shot detection for InSAR-based land displacement by the deformation-prompt-based SAM method",
abstract = "Recently, geological disasters such as land subsidence, landslides, mining-related collapse and others have been occurring more and more, posing serious threats to social and economic stability as well as human safety. InSAR (Interferometric Synthetic Aperture Radar) and advanved InSAR time-series technologies can monitor surface displacements at sub-centimeter level of accuracy, which has great and significant potential in detecting these disasters. Although several automated deformation detection methods based on CNN (Convolutional Neural Networks) have been proposed, these methods are not only time-consuming- requiring specific datasets and models- but also exhibit limited generalizability. In this study, we propose a zero-shot deformation-prompt-based SAM (Segment Anything Model) method, which does not require any pre-training operations or establishing special datasets. It initially maps the surface displacement velocity of the study area by the InSAR time-series analysis. Then the deformation-prompt-based SAM method is proposed to directly identify the exact locations and morphologies of surface deformations on the velocity maps. The method primarily extracts the rectangular boxes and points of deformation areas according to the color display characteristics of the InSAR-based land deformation map. Then these boxes and points of deformation zones serve as the prompt to the SAM, which detects deformation zones in a zero-shot manner. To validate its effectiveness, experiments on deformation extraction for various types of geohazards are conducted by utilizing Sentinel-1 and TerraSAR-X images. The results show that 137 and 93 local urban subsidence areas in Shenzhen City are identified based on different SAR images. Most of them can be automatically detected by the proposed method. Additionally, we have identified 18 landslides near Jichang Town, 36 landslides in Heifangtai and 27 major deformation zones in the Huainan mining area of Anhui Province. Statistical analysis reveals that our method can obtain high pixel accuracy in detecting land subsidence, landslides, and mining-related deformation areas with zero-shot based on different SAR images. Thus, the automatic surface deformation detection method holds considerable promise for compiling and regularly updating inventories of surface deformation data related to different types of geohazards. It has the potential to significantly enhance monitoring efficiency and provide robust technical support for disaster prevention and loss reduction.",
keywords = "Deformation prompts, InSAR, Land displacement identification, SAM, Zero-shot",
author = "Yufang He and Bo Chen and Mahdi Motagh and Yuyan Zhu and Songdong Shao and Jiaye Li and Bing Zhang and Hermann Kaufmann",
note = "Publisher Copyright: {\textcopyright} 2025 The Authors",
year = "2025",
month = feb,
doi = "10.1016/j.jag.2025.104407",
language = "English",
volume = "136",
journal = "International Journal of Applied Earth Observation and Geoinformation",
issn = "1569-8432",
publisher = "Elsevier",
}