Estimation, modeling, and prediction of land subsidence using Sentinel-1 time series in Tehran-Shahriar plain: A machine learning-based investigation

Abstract

Land subsidence (LS) is a major geologic hazard threatening most areas in Iran due to excessive groundwater extractions. The main objective of this study was to estimate LS applying PS-InSAR technique and then model and predict it through machine learning methods. Sentinel-1 time series images were used to retrieve LS from 2014 to 2019. Random forest regression (RFR) and artificial neural networks (ANN) were implemented to model LS of 2014–2017 using some conditioning factors including groundwater level decline, aquifer media, precipitation rate, slope, landuse type, depth to water table, distance from exploiting wells, distance from rivers and distance from faults as independent variables. The best calibrated method was used to predict LS over 2018–2019. In addition, all k-combinations (k=1: n, n is the number of factors) of conditioning factors were evaluated to identify the best combination resulting in the best output. The procedure was examined in the Southwest of Tehran and Shahriar plain, Iran. The maximum cumulative displacement values were 561.78 and 638.5 mm for Tehran and Shahriar, respectively. The accuracy assessment showed the RMSE value of 9.97 mm for LS retrieval. The experimental results of LS modeling indicated that RFR outperformed ANN, representing more robust and reliable results with the RMSE value of 1.35 mm for the both study areas. The investigation of the possible factor combinations showed that using the 4-factor combination consisting of precipitation, ground water table change and depth to water table along with distance from faults for S1 and the thickness of fine-grained soil layers (clay and silt) for S2 have presented promising results in modeling and predicting LS. Using all the conditioning factors, LS was predicted with RMSE value of 2.53 mm and 2.09 mm for S1 and S2, respectively. Using the 4-factor combination, however, the prediction RMSE increased by 0.13 mm and 0.04 mm for S1 and S2, respectively. This 5-percent difference can be ignored against decreasing the total cost of applying all factors. This study demonstrates the usability of processing Sentinel-1 data for LS retrieval and applying RFR for LS prediction.