A modified Long Short-Term Memory (mLSTM) deep learning model has demonstrated significant improvements in predicting mining-induced subsidence based on satellite-derived interferometric synthetic aperture radar (SAR) deformation time series data. Researchers from the Indian Institute of Technology aimed to address the challenges of accurately predicting land deformation linked to mining activities, which can pose risks to infrastructure and safety.
Mining subsidence is a serious concern for the mining industry, regulatory bodies, and the environment, requiring detailed monitoring and prediction to manage its effects. Traditional approaches for subsidence prediction often involve complex mathematical statistical models and empirical analyses, which can be impacted by the quality of historical data and the extensive input parameters needed. The necessity for accurate predictive models becomes evident, particularly as the damage caused by subsidence can have detrimental economic and environmental consequences.
To tackle these issues, the authors developed the mLSTM model, which incorporates improvements over standard LSTM frameworks to capture the complex patterns associated with land surface deformation. Utilizing interferometric SAR data from 26 TerraSAR-X and TanDEM-X datasets, the study assessed the effectiveness of mLSTM against other models, including basic LSTM and Recurrent Neural Networks (RNN).
The findings from the study revealed impressive results, with the mLSTM achieving an accuracy of 98.57%, compared to 97.54% for the LSTM model and 82.6% for the RNN model. The root mean square errors for the models were noted as 4.22 mm/year for mLSTM, 5.34 mm/year for LSTM, and 6.58 mm/year for RNN, indicating the superior predictive capabilities of the modified model. Notably, the results also indicated the potential for long-term forecasting, with predictions showing maximum deformation of -20.87 mm/year around mining operations.
Researchers found the mLSTM model not only reduced prediction oscillations but also addressed the common challenges faced by earlier models relating to the vanishing gradient problem. This was achieved through adjustments such as the stacking of multiple LSTM layers and the incorporation of hyperbolic tangent functions, enhancing overall stability and accuracy.
Through the next five years of predictions provided by the mLSTM model, the research findings suggest most areas are expected to remain stable, with only localized deformation noted around operational plant areas. This insight could prove integral for mining operations, helping stakeholders mitigate potential risks associated with subsidence.
Overall, the modification and successful application of the LSTM model for predicting subsidence demonstrates significant advancement over traditional techniques. The authors noted, “The modified LSTM model may also be extended for prediction based on time-series data in general,” highlighting its broader applicability beyond just mining-induced subsidence.
The efficient management of mine subsidence through advanced modeling techniques is not only beneficial for the industry but can also protect the environment and support regulatory compliance. This research paves the way for future explorations and adaptations of deep learning models across various domains, emphasizing their potential for effective monitoring and predictive capabilities.