.png){kind=icon}
.png){kind=icon}
Return
A hybrid deep learning model for land subsidence prediction based on TS-InSAR: a case study of Donggugang
.png){kind=icon}
OA
AI
Y
W
K
S
W
R
D
DOI：10.1186/s40677-026-00391-7.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Land subsidence is a severe geological hazard in the North China Plain, especially in Donggugang Town where uneven subsidence has caused ground cracks and building damage. Reliable monitoring and prediction are critical for hazard mitigation. This study aims to develop a high-precision and interpretable framework for land subsidence prediction by integrating time-series InSAR, wavelet decomposition, and hybrid deep learning. We obtained 2016–2022 land subsidence time series using SBAS-InSAR and decomposed signals into trend and periodic components with Db4 wavelet. The Adam-optimized LSTM model predicted the trend component with test R2 > 0.9. The CNN-Attention-BiLSTM model reduced periodic component MAE by 25.2–46.2% and RMSE by 27.3–51.8% compared with benchmark models. The final cumulative displacement prediction achieved R2 = 0.960–0.987. Attention weights revealed that shallow groundwater level variation and rainfall dominate periodic deformation in different subsidence zones. The proposed hybrid framework improves prediction accuracy and enhances interpretability. It provides a robust technical approach for land subsidence early warning and supports sustainable groundwater management and geological hazard prevention in regional plain areas.
Keywords:
SBAS-InSAR
LSTM
CNN-attention-BiLSTM
Land subsidence prediction
Wavelet decomposition
AI Summary
Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.
Journal
IF:
4
Papers:
264
Citations:
1.1K
Organization
Cited Papers