Abstract Land subsidence is a critical geological hazard with global implications. Timely detection and continuous monitoring of subsidence are essential for effective risk mitigation and disaster response. Interferometric Synthetic Aperture Radar (InSAR) techniques estimate large-scale ground deformation using SAR imagery and update displacement maps with each satellite acquisition. However, conventional InSAR approaches are relatively inefficient in accurately identifying subsidence patterns over extensive areas and long temporal spans. This study introduces the Swin U-Net model, a deep learning framework for the automatic detection of land subsidence directly from raw SAR interferograms. To train the network, we constructed a simulated dataset using distorted two-dimensional Gaussian surfaces and wrapped interferograms generated via differential GACOS atmospheric delays. Compared to conventional convolutional neural networks such as U-Net and DeepLabv3+, Swin U-Net demonstrated superior performance, achieving a precision of 94.44%, recall of 94.90%, mean Intersection over Union (MIoU) of 92.56%, mean pixel accuracy (MPA) of 98.17%, and a Dice coefficient of 0.946 – surpassing U-Net by 0.108 in Dice score. The model was applied to detect subsidence across a 43,659.9 km2 region in central and western Shandong, China, between 2020 and 2022, identifying 153 subsidence locations and retrieving their time-series deformation patterns. The maximum and minimum cumulative subsidence values were –460.2 and –9.2 mm, respectively. Further investigation revealed distinct subsidence evolution trends across different land-use types, including mining zones, urban areas, wetlands, and forested regions. Research highlights <UnorderedList Mark="Bullet"> <ItemContent> Proposed Swin U-Net model outperforms traditional CNNs (e.g., 0.108 higher Dice score), enabling efficient land subsidence detection from raw interferograms. </ItemContent> <ItemContent> Validated technical feasibility via SBAS-InSAR and leveling benchmarks, achieving millimeter-scale accuracy in low-subsidence regions. </ItemContent> <ItemContent> Successfully monitored 43,659.9 km2 in Shandong, China (2020–2022), identifying 153 subsidence points and analyzing spatiotemporal evolution across land-use types. </ItemContent> <ItemContent> Hybrid dataset (simulated 2D Gaussian surfaces + real InSAR data) enhances model robustness and addresses real-world data scarcity. </ItemContent> </UnorderedList>

A deep learning-based method for automated and continuous detection of wide-area land subsidence

期刊论文
发表日期：

2026年1月7日

查看全文

Qiuxiang Tao,
Pengcheng Wang,
Tengfei Gao,
Shihao Zhang,
Yaqi Liu,
Ziting Zhao

摘要

Abstract

Land subsidence is a critical geological hazard with global implications. Timely detection and continuous monitoring of subsidence are essential for effective risk mitigation and disaster response. Interferometric Synthetic Aperture Radar (InSAR) techniques estimate large-scale ground deformation using SAR imagery and update displacement maps with each satellite acquisition. However, conventional InSAR approaches are relatively inefficient in accurately identifying subsidence patterns over extensive areas and long temporal spans. This study introduces the Swin U-Net model, a deep learning framework for the automatic detection of land subsidence directly from raw SAR interferograms. To train the network, we constructed a simulated dataset using distorted two-dimensional Gaussian surfaces and wrapped interferograms generated via differential GACOS atmospheric delays. Compared to conventional convolutional neural networks such as U-Net and DeepLabv3+, Swin U-Net demonstrated superior performance, achieving a precision of 94.44%, recall of 94.90%, mean Intersection over Union (MIoU) of 92.56%, mean pixel accuracy (MPA) of 98.17%, and a Dice coefficient of 0.946 – surpassing U-Net by 0.108 in Dice score. The model was applied to detect subsidence across a 43,659.9 km² region in central and western Shandong, China, between 2020 and 2022, identifying 153 subsidence locations and retrieving their time-series deformation patterns. The maximum and minimum cumulative subsidence values were –460.2 and –9.2 mm, respectively. Further investigation revealed distinct subsidence evolution trends across different land-use types, including mining zones, urban areas, wetlands, and forested regions.

Research highlights

Proposed Swin U-Net model outperforms traditional CNNs (e.g., 0.108 higher Dice score), enabling efficient land subsidence detection from raw interferograms.

Validated technical feasibility via SBAS-InSAR and leveling benchmarks, achieving millimeter-scale accuracy in low-subsidence regions.

Successfully monitored 43,659.9 km² in Shandong, China (2020–2022), identifying 153 subsidence points and analyzing spatiotemporal evolution across land-use types.

Hybrid dataset (simulated 2D Gaussian surfaces + real InSAR data) enhances model robustness and addresses real-world data scarcity.