Learning Spatiotemporal Manifold Representation for Probabilistic Land Deformation Prediction.

Area of Science:

• Geosciences and Remote Sensing
• Geological Hazard Prediction

Background:

• Landslides pose significant risks to property, economy, and human life, necessitating accurate prediction methods.
• Current land deformation prediction models often fail to capture dynamic surface changes and spatial interactions, limiting their generalization.
• Remote sensing technologies, particularly Interferometric Synthetic Aperture Radar (InSAR), offer continuous terrain monitoring capabilities.

Purpose of the Study:

• To introduce DyLand, a dynamic manifold learning framework for modeling terrain surface dynamics.
• To improve the accuracy and temporal characteristic capture in land deformation prediction.
• To address limitations of prior models that rely on static observations or predefined factors.

Main Methods:

• DyLand utilizes a novel normalizing flow-based method to learn spatial connections from InSAR measurements and estimate conditional distributions on a dynamic terrain manifold.
• The framework incorporates surface permutations to capture innate land surface dynamics, enabling tractable likelihood estimations.
• It formulates spatiotemporal land deformation learning as a dynamic system, unifying spatial embedding and surface deformation learning.

Main Results:

• DyLand effectively models the dynamic structures of the terrain surface, capturing subtle spatial and temporal deformation characteristics.
• Experiments on real-world InSAR datasets demonstrate DyLand's superior performance compared to existing benchmark models.
• The framework shows improved generalization and ability to capture temporal deformation patterns in landslide-prone areas.

Conclusions:

• DyLand represents a significant advancement in land deformation prediction by incorporating dynamic manifold learning.
• The proposed method offers a more robust and generalized approach to predicting landslides and other ground movements.
• Future work can further explore the application of DyLand in various geological hazard monitoring and prediction scenarios.