Land subsidence prediction in Zhengzhou's main urban area using the GTWR and LSTM models combined with the Attention Mechanism
Yonghao Yuan,Dujuan Zhang,Jian Cui,Tao Zeng,Gubin Zhang,Wenge Zhou,Jinyang Wang,Feng Chen,Jiahui Guo,Zugang Chen,Hengliang Guo
DOI: https://doi.org/10.1016/j.scitotenv.2023.167482
2024-01-10
Abstract:In recent years, due to urbanization and human activities, groundwater overexploitation has become increasingly severe, resulting in some degrees of land subsidence and, consequently, causing a series of geological disasters and other environmental issues. Therefore, large-scale and high-precision land subsidence prediction is of great importance for the prevention and control of geological disasters. However, the existing prediction models and methods ignore the effects of the spatiotemporal non-stationary relationships between the influencing factors and the accumulated land subsidence, causing the poor accuracy of the predicted land subsidence results. In this context, a Geographically and Temporally Weighted Regression combined with the Long Short-Term Memory (LSTM)-multivariable and Attention Mechanism (AM) (GTWR-LSTMm-AM) was proposed to more accurately predict the deformation of time series land subsidence in this study. The small baseline subset-interferometric synthetic aperture radar (SBAS-InSAR) was used to reveal the temporal deformation information of Zhengzhou's main urban area, then the GTWR model was used to assess the spatiotemporal non-stationarity relationships between the accumulated land subsidence and its influencing factors monthly groundwater stability level, monthly precipitation and Normalized Difference Vegetation Index (NDVI) data, and to determine the corresponding weight matrix. In addition, we introduced an LSTM model with AM to extract key information from the time-series land subsidence data and adjusted the dynamic weights of the three selected influencing factors to predict the land subsidence in Zhengzhou's main urban area. The prediction accuracy R2 of the GTWR-LSTMm-AM model reaches 0.972, which is higher than 0.929 of the LSTMm model. The prediction accuracy RMSE is less than 3 mm and reaches 2.403 mm. In addition, we determined the importance of the impact factor on the subsidence results by randomly interrupting the impact factor time series, disclosuring that the monthly groundwater level contributed the most to the land subsidence in Zhengzhou's main urban area.