Land subsidence monitoring based on SBAS-InSAR and prediction using ARIMA-BP model in typical oil exploitation areas
Fengyan Wang,Xiang Wu,Kai Zhou,Qingbo Yu,Jianping Chen,Qing Wang
DOI: https://doi.org/10.5194/egusphere-egu24-21787
2024-01-01
Abstract:The Songnen Plain is a major base for commodity grain production and oil and gas exploitation in China, playing a crucial role in ensuring China's grain production, the development of oil and gas resources, and economic construction. However, the long-term and extensive exploitation of resources such as oil, gas, and groundwater has led to large-scale surface deformation in the Songnen Plain. To investigate the impact of oil and gas exploitation on land subsidence in the Songnen Plain, a study on land subsidence and its prediction models was conducted in the oil and gas exploitation areas of the Songnen Plain. This study used the Small Baseline Subset Interferometry Synthetic Aperture Radar (SBAS-InSAR) to obtain the time-series land subsidence data of the oil and gas exploitation areas for a total of 40 periods during June 2015 to September 2020, and found that the study area showed a trend of subsidence during the monitoring period, and the rate of subsidence was -2.4 mm/year. And analyzed the surface deformation of the study area by combining five factors, such as oil exploitation, water-soil structure environment, groundwater, meteorology, and earthquake, and found that oil exploitation, water-soil structure environment, groundwater level, and cumulative precipitation were all significantly negatively correlated with cumulative deformation in the study area, while the correlation between air temperature and earthquakes and cumulative subsidence in the study area was not obvious. Subsequently, leveraging the advantages of the Autoregressive Integrated Moving Average model (ARIMA) and Back Propagation (BP) neural network, an ARIMA-BP model was proposed to prediction overall ground deformation trends in the study area. Comparing the predictions of the ARIMA-BP model with those of single models such as ARIMA, BP neural network, random forest (RF), support vector machine (SVM), and eXtreme Gradient Boosting (XGBOOST), it was found that the proposed model can accurately prediction the overall surface deformation trend in the study area. The model's performance metrics, including Mean Absolute Error 6(MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and R-squared (R2), were superior, with values of 0.489 mm, 0.73 mm, 20%, and 0.964, respectively, outperforming other single models. This study can provide decision support for the analysis of land subsidence in typical oil exploitation areas of Songnen Plain and its disaster prevention and control.