TL;DR
DyLand is a novel framework that treats the Earth's surface as a dynamic, evolving 3D manifold to forecast land deformation with unprecedented accuracy. By combining Continuous Normalizing Flows with Neural Ordinary Differential Equations (ODEs), it captures the complex co-evolution of terrain shape and movement, significantly outperforming traditional static graph-based models in landslide prediction.
Introduction
Landslides are among the most destructive natural hazards, yet forecasting them is notoriously difficult. Modern geodetic techniques like Interferometric Synthetic Aperture Radar (InSAR) provide massive point cloud datasets of surface displacement, but turning this raw data into accurate forecasts faces three major hurdles: dimensionality loss from 2D projections, static assumptions in graph neural networks, and the mathematical complexity of modeling probability distributions on a moving surface.
Methodology: The DyLand Framework
The authors introduce DyLand to bridge these gaps by modeling land deformation as a dynamic system on a non-Euclidean manifold. The framework consists of two primary innovations:
- Dynamic Manifold Learning via Normalizing Flows (NFs): Instead of a static embedding, the model uses NFs to learn a probability density over the manifold. It incorporates surface "perturbations" (the actual deformations) into the coordinate space, allowing the model to estimate conditional distributions while preserving the innate dynamics of the terrain.
- Co-evolution via Neural ODEs: Surface deformation ($S$) and the latent spatio-temporal representation ($W$) are coupled. They model this relationship as a dynamic system solved using Neural Ordinary Differential Equations. This "lifts" the points into a higher-dimensional space where the learning process is smoother, allowing the model to extrapolate future deformations by solving the system forward in time.
Experimental Results
The model was evaluated on two real-world, high-resolution InSAR datasets from Sichuan, China: HZY (6,700 nodes) and PBG (14,500 nodes). These datasets involve continuous monitoring of slopes that are actively deforming.
DyLand consistently outperformed 11 baseline models across all datasets. On the PBG-West dataset, DyLand achieved an RMSE of 0.013 and an $R^2$ of 0.993, indicating a near-perfect fit to complex deformation trends compared to the best baseline (SA-GNN, RMSE 0.021). Ablation studies confirmed that the unified co-evolution approach is superior to using either spatio-temporal representations or historical observations in isolation.
Key Insights
The "Dynamic Adjacency" Advantage
Traditional models struggle because they assume a point's neighbors always exert the same influence. DyLand demonstrates that as land shifts, the "neighborhood" influence must be re-calculated.
Preserving Topology
While popular manifold techniques like UMAP or t-SNE are excellent for clustering, they often destroy the continuous spatial correlations needed for physical forecasting. DyLand’s approach preserves the local manifold structure, which is essential for understanding how movement in one area triggers a response in another.
Limitations and Future Work
Currently, the model focuses on the internal dynamics of the land. However, landslides are often triggered by external factors like heavy rainfall or seismic activity. Future iterations aim to incorporate meteorological data. Additionally, finding a balance between representation dimension and the computational cost of solving Neural ODEs remains a key challenge.
Reference: Zhou, F., Li, R., Gao, Q., Trajcevski, G., Zhang, K., & Zhong, T. (2022). Dynamic Manifold Learning for Land Deformation Forecasting. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI-22).
Summary updated