Machine learning-based techniques for land subsidence simulation in an urban area

Abstract

Understanding and mitigating land subsidence (LS) is critical for sustainable urban planning and infrastructure management. We introduce a comprehensive analysis of LS forecasting utilizing two advanced machine learning models: the eXtreme Gradient Boosting Regressor (XGBR) and Long Short-Term Memory (LSTM). Our findings highlight groundwater level (GWL) and building concentration (BC) as pivotal factors influencing LS. Through the use of Taylor diagram, we demonstrate a strong correlation between both XGBR and LSTM models and the subsidence data, affirming their predictive accuracy. Notably, we applied delta-rate (Δr) calculus to simulate a scenario with an 80% reduction in GWL and BC impact, revealing a potential substantial decrease in LS by 2040. This projection emphasizes the effectiveness of strategic urban and environmental policy interventions. The model performances, indicated by coefficients of determination R2 (0.90 for XGBR, 0.84 for LSTM), root-mean-squared error RMSE (0.37 for XGBR, 0.50 for LSTM), and mean-absolute-error MAE (0.34 for XGBR, 0.67 for LSTM), confirm their reliability. This research sets a precedent for incorporating dynamic environmental factors and adapting to real-time data in future studies. Our approach facilitates proactive LS management through data-driven strategies, offering valuable insights for policymakers and laying the foundation for sustainable urban development and resource management practices.