Decomposing Spatiotemporal Fields into Data Driven Modes

39 concepts32 papers1 proposal

Earth system sciences—encompassing meteorology, oceanography, and ecology—face a common data challenge: tracking continuous physical properties across both space and time. Whether monitoring sea surface temperatures, atmospheric pressure, or ionospheric electron density, researchers are inundated with high-dimensional datasets. Analyzing these systems grid-point by grid-point is not only computationally exhausting but also obscures the large-scale, coherent mechanisms—such as monsoons, ocean currents, and climate teleconnections—that drive the system.

To extract meaningful signals from this spatiotemporal chaos, researchers rely on Empirical Orthogonal Function (EOF) analysis (closely related to Principal Component Analysis). EOF analysis decomposes a complex, continuous spatiotemporal field into a compact set of static spatial patterns and their corresponding temporal evolutions.

The Intuition: Finding the System's Heartbeat

Imagine a grid of thousands of satellite observation points tracking a variable over several decades. If every grid point fluctuated independently, understanding the system would be impossible. However, physical processes ensure that nearby regions are highly correlated.

EOF analysis exploits these correlations by finding a new coordinate system for the data. Instead of representing the state of the system as a collection of individual grid points, it represents the data as a linear combination of orthogonal spatial maps (the EOFs) weighted by time-varying coefficients (the Principal Components, or PCs).

$$\text{Data}(x, t) = \sum_{i} \text{PC}_i(t) \times \text{EOF}_i(x)$$

Mathematically, this is achieved by computing the eigenvectors of the spatial covariance matrix. The first eigenvector (EOF 1) points in the direction of the maximum possible variance in the dataset. The second eigenvector (EOF 2) captures the next highest variance, subject to being orthogonal to the first.

This data-driven approach is incredibly powerful for dimensionality reduction. For example, when modeling regional ionospheric behavior, Changhao Yu et al. demonstrated that the very first EOF component alone could explain over 90% of the total variation in Total Electron Content (TEC). By focusing on just the leading modes, researchers can filter out localized noise and isolate the dominant physical drivers.

Uncovering Hidden Physical Mechanisms

Because EOFs naturally capture the configurations of maximum variance, the leading modes often correspond directly to real-world physical phenomena. This allows scientists to separate overlapping signals that occur in the same geographic space.

Isolating Intrinsic vs. External Variability

In fluid dynamics, it is often difficult to tell whether a regional change is forced by global climate shifts or generated internally by the fluid's own non-linear behavior. By applying EOF analysis to subsurface ocean data, Terry O'Kane et al. isolated a 20-year "internal heartbeat" of variability in the Southern Ocean that exists independently of atmospheric changes. Similarly, David E. Gwyther et al. used EOFs to show that up to 44% of the interannual variability in Totten Glacier melting is driven by intrinsic ocean processes rather than external climate forcing.

Mapping Teleconnections

Climate phenomena in one part of the world frequently trigger cascading weather effects thousands of miles away. EOF analysis reduces high-dimensional spatial data into simple temporal indices (the PCs), making it possible to track these long-distance relationships.

  • Phong V. V. Le et al. utilized EOF analysis to identify leading modes of sea surface temperature variability, revealing how an intensifying El Niño-Southern Oscillation (ENSO) alters global precipitation predictability.
  • Lejiang Yu et al. used the technique to show how rapid warming in the North Pacific triggers atmospheric "wavetrains" that paradoxically decelerated summer sea ice loss in specific Arctic sectors.
  • Hedi Ma et al. isolated the spatial patterns through which the 11-year solar cycle influences the South Asian jet stream during late winter.

The method is equally effective for tracking ecological and environmental shifts over time. Shilong Piao et al. extracted the dominant spatiotemporal modes of the terrestrial carbon cycle to understand how CO2 fertilization balances land-use emissions. In marine environments, Dingfeng Yu et al. used EOF decomposition to map water transparency variations across the Bohai and Yellow Seas, successfully correlating the dominant modes with sediment dynamics, while Xiangbai Wu et al. applied it to isolate accelerated early spring warming patterns driven by monsoon changes.

Advanced Variants: Rotation, Extension, and Coupling

While standard EOF analysis is a foundational tool, the strict mathematical constraints of orthogonality and global variance maximization can sometimes mix distinct physical processes into a single mode, or create mathematical artifacts. To solve these issues, researchers have developed specialized variations of the technique.

Varimax Rotation for Regional Clarity

Standard EOFs optimize for global variance, which can cause a single mode to span an entire continent, blurring localized behaviors. Varimax rotation relaxes the global constraint to maximize the sparsity of the spatial weights (loadings). This forces the coordinates to align with distinct, localized regional structures.

Ze‐Xin Fan et al. successfully combined EOF analysis with varimax rotation to untangle complex regional climate dynamics on the Yunnan Plateau, separating widespread warming trends from localized cooling in dry valleys. In data-driven causal modeling, Jakob Runge et al. used Varimax-rotated PCA to simplify dense grid-based observations into discrete, interpretable regional nodes, which were then used to map cause-and-effect pipelines across the global climate system.

Extended EOFs for Temporal Evolution

A standard EOF provides a static spatial snapshot; the pattern itself does not move, only its amplitude changes via the PC time series. To capture features that propagate geographically over time, Extended EOF (EEOF) analysis incorporates time-lagged versions of the spatial field into the decomposition.

This allows a single mode to represent a moving sequence. The authors of a study on East Asian winters, Shui Yu et al., leveraged EEOF analysis to capture the "phase-shift mode" of the East Asian trough, mapping exactly how its geographic structure evolves step-by-step from December to February.

Singular Value Decomposition for Coupled Fields

Often, the goal is not just to analyze one variable, but to see how two entirely different spatiotemporal fields interact—such as ocean temperatures and land-based drought. Instead of decomposing a single covariance matrix, Singular Value Decomposition (SVD) can be applied to the cross-covariance matrix between two distinct datasets.

This isolates coupled spatiotemporal patterns shared between the two systems. Keyan Fang et al. used this approach to analyze 700 years of tree-ring data alongside sea surface temperatures, successfully isolating the shared modes of variability that link the Pacific Decadal Oscillation to historical droughts across Central Asia.

Data-Driven vs. Analytical Decompositions

What makes EOF analysis uniquely powerful across these diverse applications is that its basis functions are entirely data-driven. Rather than forcing the data onto predefined mathematical shapes—like Fourier transforms for time or spherical harmonics for space—EOFs naturally adapt to the geometry of the specific physical system being studied.

While analytical expansions remain valuable for highly structured, single-point sensor problems—such as the 4D atmospheric water vapor model built by Jean-Pierre Barriot et al. using trigonometric functions and spherical harmonics—EOF analysis remains the gold standard when dealing with complex, irregular, and interconnected global systems. By letting the data define its own coordinate system, spatiotemporal decomposition allows researchers to look past the noise of thousands of individual grid points and tune into the fundamental rhythms of the Earth system.

Key Foundational Work

Spatial and temporal temperature trends on the Yunnan Plateau (Southwest China) during 1961–2004 (2010) — contributed 2 concept elements to this direction.
Spatiotemporal patterns of terrestrial carbon cycle during the 20th century (2009) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_empirical_orthogonal_function_analysis_900446ef,
  title = {empirical orthogonal function analysis},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/empirical-orthogonal-function-analysis/rdr_900446efe7644924b40b33c0cefb7db9},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_empirical_orthogonal_function_analysis_900446ef_v1,
  title = {empirical orthogonal function analysis},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/empirical-orthogonal-function-analysis/rdr_900446efe7644924b40b33c0cefb7db9/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake