Ground subsidence prediction model in Xinmi City based on TSC—LSTM
Predicting land subsidence is crucial for conducting in-depth analyses of and providing early warnings for urban land subsidence patterns.However,traditional numerical prediction models frequently experience difficulty in accurately capturing the intricate characteristics of land subsidence data,leading to less precise predictions.This study focuses on Xinmi City and endeavors to improve the accuracy of land subsidence prediction by combining time series feature extraction methods with time series prediction techniques.In this study,232 interference images provided by HyP3 were utilized to acquire land subsidence information in Xinmi City from January 2018 to December 2022,employing Small Baseline Subset(SBAS)-Interferometric Synthetic Aperture radar(InSAR)technology.Recognizing the challenge of achieving high accuracy in directly predicting land subsidence data,this study developed a land subsidence prediction model that integrates trend and seasonal characteristics by using a Long Short-term Memory(LSTM)network,namely,the Trend Seasonal Characteristics-LSTM(TSC-LSTM).The TSC-LSTM model capitalizes on the strengths of weighted regression seasonal trend decomposition(STL)in extracting time series features from settlement data and the LSTM model for addressing the vanishing gradient problem in time series prediction.This fusion of techniques allows for a precise analysis of land subsidence data and enables highly accurate predictions.Distinguishing itself from the conventional LSTM model,the TSC-LSTM model refrains from directly inputting ground subsidence data.Instead,it employs STL to meticulously extract trend and seasonal characteristics from land subsidence data.This approach maximizes the utilization of characteristic information inherent in land subsidence data.Subsequently,these features are fed into the LSTM model for prediction.This unique methodology reduces noise interference and significantly enhances the accuracy of model predictions.This research leverages time-series InSAR data for Xinmi City from 2018 to 2022.It employs the TSC-LSTM model,deep learning architectures(recurrent neural network and LSTM),and conventional machine learning algorithms(multilayer perceptron and support vector regression)to forecast the cumulative subsidence data for five subsidence centers by using SBAS-InSAR.This study identifies the two most optimal models and validates their efficacy in single-point prediction scenarios,utilizing domain-specific terminologies.Research findings indicate the following.(1)Between 2018 and 2022,Xinmi City experienced a land subsidence rate that ranged from-60.3 mm to 51.96 mm per annum,resulting in the identification of five distinct land subsidence center areas.Among these,the highest cumulative settlement and uplift reached 304.9 mm and 197.68 mm,respectively.The universality of the TSC-LSTM model across diverse datasets has been corroborated,demonstrating its high precision,exceptional generalization capability,and stable high performance in the prediction of land subsidence,employing specialized terminologies.(2)The TSC-LSTM model exhibited exceptional performance in predicting the five subsidence center areas.The R2 values for the TSC-LSTM model range from 0.9985 to 0.9992,significantly surpassing the second-best model,i.e.,LSTM,which has an R2 range of 0.9662 to 0.9872.Moreover,the root mean square error values for the prediction accuracy of the TSC-LSTM model are less than 2 mm,achieving a range of 1.2426 mm to 1.7403 mm.(3)Single-point prediction results demonstrate the superior ability of the TSC-LSTM model to accurately capture local changes in the cumulative settlement data.The TSC-LSTM model proposed in this study outperforms the traditional LSTM model in terms of prediction accuracy and model stability,providing robust support for in-depth research on urban land subsidence.
ground settlement predictionTSC—LSTMSBAS—InSARcumulative settlement data decompositionXinmi CityLSTM
万方数据
维普
