Abstract
Rapid developments in satellite remote-sensing technology have enabled the collection of geospatial data on a global scale, hence increasing the need for covariance functions that can capture spatial dependence on spherical domains. We propose a general method of constructing nonstationary, locally anisotropic covariance functions on the sphere based on covariance functions in $\mathbb{R}^3$. We also provide theorems that specify the conditions under which the resulting correlation function is isotropic or axially symmetric. For large datasets on the sphere commonly seen in modern applications, the Vecchia approximation is used to achieve higher scalability on statistical inference. The importance of flexible covariance structures is demonstrated numerically using simulated data and a precipitation dataset.
Introduction
As satellite remote sensing provides increasingly dense global data, researchers face a "spherical" challenge. Traditional geostatistics often approximates small regions as flat planes ($\mathbb{R}^2$), but this fails at a global scale. The core difficulty lies in defining covariance functions—the mathematical tools that describe how much two points are related based on the distance between them—that are valid on a sphere.
Many standard functions used on flat ground become mathematically "invalid" (not positive-definite) when applied to the curvy "great-arc" distance. Furthermore, real-world phenomena like precipitation or atmospheric pressure are rarely stationary (uniform across the globe) or isotropic (the same in every direction). They often exhibit local "anisotropy"—for instance, weather patterns might stretch more along the equator than toward the poles.
Methodology: Constructing Local Anisotropy
The authors propose a general method to build nonstationary, locally anisotropic covariance functions on the unit 2-sphere ($\mathbb{S}^2$).
Chordal Distance and Euclidean Embedding
Instead of the curvy great-arc distance, the framework utilizes "chordal distance"—the straight line that "pierces through" the sphere. This allows the adaptation of valid covariance constructions from 3D Euclidean space ($\mathbb{R}^3$) to the sphere's surface. For points close together, chordal and great-arc distances are nearly identical, allowing the model to gain access to a wider library of mathematically valid functions without sacrificing physical realism.
Tangent Plane Parameterization
To capture directionality, the model defines an anisotropy matrix $\Sigma(s)$ at every location. This matrix represents scaling and rotation on the tangent plane (a flat sheet touching the sphere at that point). This mimics how a local observer on the ground would perceive spatial correlation (e.g., "the rain pattern is stretched 2x more East-West than North-South").
The anisotropy is controlled by three spatially varying parameters: two for scaling ($\gamma_1, \gamma_2$) and one for rotation ($\kappa$). By allowing these parameters to vary across the sphere, the model can represent complex, nonstationary fields.
Model Flexibility and Symmetry
The framework is highly flexible, encompassing several restricted forms as special cases. The authors provide mathematical proofs for the conditions under which the model simplifies into more restricted forms:
- Isotropy: Spatial correlation is independent of direction ($\gamma_1 = \gamma_2$).
- Axial Symmetry: Correlation depends only on latitude, a common assumption in climate modeling.
- General Nonstationarity: Parameters vary freely across the sphere, capturing complex local variations.
Scalability and Numerical Results
Modeling Gaussian Processes (GPs) usually requires computational effort that grows cubically with the number of data points ($O(N^3)$). To handle large global datasets, the authors use the Vecchia approximation, which breaks the problem down into smaller, conditional chunks, reducing the cost to linear ($O(N)$).
Key Findings
- Simulation Performance: In tests where the "true" underlying process was nonstationary, the proposed model significantly outperformed simpler isotropic models across all metrics, including Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE).
- Precipitation Modeling: When applied to CESM precipitation anomalies (13,824 locations), the axially symmetric and general nonstationary models performed much better than the isotropic model. Interestingly, the axially symmetric model was often "good enough," but the general nonstationary model provided the best scores for joint predictive distributions.
- Flexibility is "Cheap": A key finding was that using a more complex, nonstationary model on simple (isotropic) data doesn't significantly hurt prediction accuracy. However, using a simple model on complex, anisotropic data leads to major errors, suggesting that "over-parameterizing" is a safe bet for unknown global data.
Conclusion
This framework provides a robust and scalable tool for global geospatial analysis. While the model is flexible, future work could explore allowing the "smoothness" of the field to vary spatially alongside the anisotropy, or extending the local anisotropy to include the time dimension for dynamic climate science applications.
Summary updated