Prediction of InSAR deformation time-series using improved LSTM deep learning model

Area of Science:

• Geosciences
• Remote Sensing
• Data Science

Background:

• Mining-induced subsidence poses significant risks to mining operations, stakeholders, and the environment.
• Effective management necessitates accurate monitoring and prediction of subsidence.
• Interferometric Synthetic Aperture Radar (InSAR) provides valuable deformation time-series data.

Purpose of the Study:

• To develop and evaluate a modified Long Short-Term Memory (LSTM) model for predicting InSAR deformation time series.
• To assess the model's performance against existing Recurrent Neural Network (RNN) and LSTM models.
• To explore the potential of the modified LSTM for general time-series prediction.

Main Methods:

• A modified LSTM (mLSTM) model was developed for InSAR deformation time-series prediction.
• The model was trained and tested using 26 TSX/TDX datasets from Mine-A in India.
• Performance was evaluated by comparing mLSTM predictions with RNN and LSTM models, using efficiency metrics and Root Mean Square (RMS) error.

Main Results:

• The modified LSTM model achieved the highest prediction efficiency (98.57%) and lowest RMS error (4.22 mm/year).
• mLSTM predictions closely matched observed deformation velocity values compared to RNN and LSTM.
• Five-year predictions indicate overall area stability, with localized deformation near the plant area.

Conclusions:

• The modified LSTM model demonstrates superior performance in predicting InSAR deformation time series for mine subsidence.
• This approach offers a more accurate and reliable method for mine subsidence monitoring and management.
• The mLSTM model's capabilities can be extended to other time-series prediction applications in geosciences.