A mining-area surface-subsidence prediction method based on SBAS-InSAR and STL-XGBoost

Area of Science:

• Geosciences
• Remote Sensing
• Data Science

Background:

• Surface subsidence from mining poses risks to infrastructure and safety.
• Accurate prediction is vital for mitigation but challenging due to complex spatio-temporal patterns.
• Existing models often fail to capture nonlinear dynamics and trends effectively.

Purpose of the Study:

• To develop an advanced surface subsidence prediction model.
• To enhance prediction accuracy by combining seasonal trend decomposition with extreme gradient boosting trees.
• To provide reliable technical support for surface subsidence management in mining areas.

Main Methods:

• Utilized Small Baseline Subset Interferometric Synthetic Aperture Radar (SBAS-InSAR) for subsidence data acquisition (2020-2023).
• Applied Seasonal Trend decomposition (STL) to separate subsidence sequences into trend and non-trend components.
• Developed a hybrid STL-XGBoost model, predicting trend and non-trend terms separately and combining predictions.

Main Results:

• The STL-XGBoost model demonstrated superior performance over single XGBoost.
• Mean Absolute Error (MAE) decreased by 31%, and Root Mean Squared Error (RMSE) decreased by 38%.
• Prediction results showed a strong correlation (>0.9) with the original time series, confirming high accuracy.

Conclusions:

• The proposed STL-XGBoost model effectively captures complex spatio-temporal characteristics of surface subsidence.
• This hybrid approach offers a significant improvement in prediction accuracy and reliability.
• The method provides robust technical support for mitigating risks associated with mining-induced surface subsidence.