Abstract
This paper focuses on the analysis of spatially correlated functional data. The between-curve correlation is modeled by correlating functional principal component scores of the functional data. We propose a Spatial Principal Analysis by Conditional Expectation (SPACE) framework to explicitly estimate spatial correlations and reconstruct individual curves. This approach works even when the observed data per curve are sparse. Assuming spatial stationarity, empirical spatial correlations are calculated as the ratio of eigenvalues of the smoothed covariance surface and cross-covariance surface. Then an anisotropy Matérn spatial correlation model is fit to empirical correlations. Finally, principal component scores are estimated to reconstruct the sparsely observed curves. This framework can naturally accommodate arbitrary covariance structures, but there is an enormous reduction in computation if one can assume the separability of temporal and spatial components. We propose hypothesis tests to examine the separability as well as the isotropy effect of spatial correlation. Simulation studies and applications of empirical data show improvements in the curve reconstruction using our framework over the method where curves are assumed to be independent. In addition, we show that the asymptotic properties of estimates in uncorrelated case still hold in our case if 'mild' spatial correlation is assumed.
Introduction
Functional Data Analysis (FDA) treats observations—such as growth curves or temperature readings—as continuous functions. A standard tool in this field is Functional Principal Components Analysis (FPCA), which decomposes these curves into their primary patterns, or scores. However, traditional FPCA methods, such as the widely used Principal Analysis by Conditional Expectation (PACE), typically assume that individual curves are independent. In many real-world scenarios, such as remote sensing or biological imaging, curves at neighboring locations are likely to be spatially correlated. Ignoring this dependency misses an opportunity to "borrow" information from neighbors, which is particularly critical when data for individual curves are sparse or noisy.
The SPACE Framework
To address this, we introduce SPACE (Spatial Principal Analysis by Conditional Expectation). This framework extends PACE by explicitly modeling spatial dependencies between curves. Instead of modeling the correlation of every raw data point, SPACE models the correlation between the functional principal component (fPC) scores. This approach significantly reduces computational complexity while capturing the essential spatial structure.
A key feature of SPACE is the use of an anisotropic Matérn model. While the Matérn family is a flexible standard for spatial relationships, our inclusion of anisotropy allows the model to account for correlations that vary by direction (e.g., North-South vs. East-West). To estimate these relationships, we utilize local linear smoothers to calculate cross-covariance surfaces, even when data are observed at irregular intervals.
Empirical Evidence and Anisotropy
The necessity of modeling directional correlation is evident in empirical data. For instance, in the analysis of the Harvard Forest Enhanced Vegetation Index (EVI) data—a 25x25 pixel grid of remote sensing data—we observed distinct patterns in how correlation decays with distance depending on the orientation.
As shown in the figure above, the correlation for fPC2 scores decays much more slowly in the North-East/South-West (NESW) direction compared to the North-West/South-East (NWSE) direction. An isotropic model would fail to capture this "diagonal" correlation pattern, leading to less accurate curve reconstructions.
Results and Discussion
Our evaluations using both simulated data and the Harvard Forest dataset demonstrate the advantages of the SPACE framework:
- Superior Reconstruction: SPACE consistently outperforms PACE in curve reconstruction, particularly in high-noise environments. In the Harvard Forest application, SPACE improved reconstruction performance in 100% of the test samples.
- Effective Testing: Our proposed hypothesis tests successfully identified the presence of anisotropy in the real-world data, justifying the use of the more complex spatial model.
- Theoretical Guarantees: We proved that the asymptotic properties of FPCA estimates—the mathematical guarantees that estimates improve with more data—remain valid for spatially correlated data, provided the correlation is "mild."
Conclusion
The SPACE framework bridges the gap between functional data analysis and spatial statistics. By recognizing that spatial correlation is an informative feature rather than a nuisance, SPACE provides a robust method for reconstructing sparse functional data. Future work will focus on scaling these methods to massive global datasets and relaxing stationarity assumptions for larger, more diverse geographic regions.
Summary updated