Deep learning time-series modeling for assessing land subsidence under reduced groundwater use

Area of Science:

• Geosciences
• Environmental Science
• Data Science

Background:

• Taiwan's Choshui Delta faces worsening land subsidence exacerbated by intensive groundwater extraction and drought.
• Effective predictive modeling is crucial for mitigating land subsidence in vulnerable deltaic regions.

Purpose of the Study:

• To develop and apply a machine learning framework for analyzing land subsidence.
• To identify groundwater-level decline as a primary driver of subsidence using electricity consumption data.
• To evaluate the effectiveness of groundwater pumping reduction strategies for subsidence mitigation.

Main Methods:

• Utilized a Long Short-Term Memory (LSTM) neural network to reconstruct subsidence records and forecast trends.
• Employed an Artificial Neural Network (ANN) to correlate well electricity usage with groundwater level fluctuations.
• Simulated two artificial scenarios to assess the impact of reduced groundwater pumping on subsidence rates.

Main Results:

• The LSTM model accurately reproduced historical subsidence data and provided reliable future predictions.
• Groundwater-level decline due to pumping was identified as a key factor contributing to land subsidence.
• Simulated pumping reductions led to groundwater level recovery and a significant decrease in subsidence rates, e.g., from 2.23 cm/year to 1.34 cm/year at a key monitoring well.

Conclusions:

• Integrating groundwater extraction indicators into subsidence models is essential.
• Curtailing groundwater extraction effectively mitigates land subsidence in vulnerable deltaic areas.
• Machine learning provides a powerful tool for understanding and managing land subsidence.