Scrutinization of land subsidence rate using a supportive predictive model: Incorporating radar interferometry and...

Area of Science:

• Geosciences and Remote Sensing
• Environmental Science
• Water Resource Management

Background:

• Land subsidence poses a significant challenge for land and water resource management.
• Radar datasets offer a cost-effective method for monitoring subsidence.
• Identifying key drivers is crucial for accurate subsidence modeling.

Purpose of the Study:

• To develop an integrated radar-and-ensemble-based method for simulating land subsidence rates.
• To identify the primary drivers influencing land subsidence using machine learning.
• To compare the performance of individual and ensemble machine learning models for subsidence prediction.

Main Methods:

• Utilized 52 pairs of radar images from 2014-2019 to detect subsidence.
• Employed the simulated annealing (SA) algorithm to identify key subsidence drivers.
• Compared three individual machine learning models (SVM, GP, BART) and three ensemble approaches for subsidence modeling.

Main Results:

• Subsidence rates varied from 0 to 59 cm between 2014 and 2019.
• Radar-derived subsidence showed a strong correlation with geodynamic station data (R² = 0.99).
• Key drivers identified include precipitation, elevation, fine-grained materials, groundwater withdrawal, distance to roads, groundwater decline, and aquifer thickness.
• Ensemble models outperformed individual models, with the nonlinear ensemble (BART combination) showing the best performance (RMSE = 0.061, R² = 0.83).

Conclusions:

• The nonlinear ensemble approach effectively simulates land subsidence rates.
• Fine-grained materials in aquifer systems significantly influence subsidence response to groundwater changes.
• The interaction of groundwater decline, low recharge, and fine-grained materials intensifies land subsidence, posing a considerable risk.