Nonlinear PCA for Spatio-Temporal Analysis of Earth Observation Data

Remote sensing observations, products and simulations are fundamental sources of information to monitor our planet and its climate variability. Uncovering the main modes of spatial and temporal variability in Earth data is essential to analyze and understand the underlying physical dynamics and processes driving the Earth System. Dimensionality reduction methods can work with spatio-temporal datasets and decompose the information efficiently. Principal Component Analysis (PCA), also known as Empirical Orthogonal Functions (EOF) in geophysics, has been traditionally used to analyze climatic data. However, when nonlinear feature relations are present, PCA/EOF fails. In this work, we propose a nonlinear PCA method to deal with spatio-temporal Earth System data. The proposed method, called Rotated Complex Kernel PCA (ROCK-PCA for short), works in reproducing kernel Hilbert spaces to account for nonlinear processes, operates in the complex kernel domain to account for both space and time features, and adds an extra rotation for improved flexibility. The result is an explicitly resolved spatio-temporal decomposition of the Earth data cube. The method is unsupervised and computationally very efficient.We illustrate its ability to uncover spatio-temporal patterns using synthetic experiments and real data. Results of the decomposition of three essential climate variables are shown: satellite-based global Gross Primary Productivity (GPP) and Soil Moisture (SM), and reanalysis Sea Surface Temperature (SST) data. The ROCK-PCA method allows identifying their annual and seasonal oscillations, as well as their non-seasonal trends and spatial variability patterns.

💡 Research Summary

The paper introduces a novel unsupervised dimensionality‑reduction technique called Rotated Complex Kernel PCA (ROCK‑PCA) designed specifically for spatio‑temporal Earth observation data. Traditional Principal Component Analysis (PCA) or Empirical Orthogonal Functions (EOF) assume linear relationships and therefore struggle to capture the nonlinear dynamics that are common in climate, ecosystem, and oceanographic datasets. ROCK‑PCA addresses these limitations through three key innovations.

First, the method maps the original data into a Reproducing Kernel Hilbert Space (RKHS) using a complex‑valued kernel. The kernel combines a Gaussian distance term with an exponential phase term that encodes temporal offsets, allowing the representation to retain both amplitude and phase information. This is crucial for preserving seasonal cycles, oscillations, and other periodic phenomena that would otherwise be split across multiple real‑valued EOF modes.

Second, after constructing and centering the complex Gram matrix, an eigen‑decomposition yields complex eigenvectors (principal components). Each component can be expressed as a time‑varying amplitude and phase, providing a natural description of propagating patterns such as El Niño‑Southern‑Oscillation events or migrating vegetation greenness fronts.

Third, the authors introduce an additional rotation step. By optimizing a rotation matrix that minimizes inter‑mode correlation and aligns modes with physically interpretable structures, the complex eigenvectors are transformed into a new set of modes that are both orthogonal and easier to interpret. The rotation is solved as a Procrustes‑type least‑squares problem, yielding modes that concentrate spatial variance in specific regions and isolate distinct temporal behaviours.

Algorithmically, ROCK‑PCA proceeds as follows: (1) preprocess the data cube (time × latitude × longitude) and handle missing values; (2) compute the complex kernel matrix and centre it; (3) perform eigen‑decomposition and select the leading eigenpairs; (4) optimise the rotation matrix and obtain rotated modes; (5) project the rotated modes back to the original space for visualisation and further analysis.

The computational cost is dominated by the O(N³) eigen‑decomposition of the N × N kernel matrix (N = number of time steps). The authors demonstrate that low‑rank approximations such as the Nyström method or random Fourier features can dramatically reduce memory and runtime, enabling the processing of multi‑year, global‑scale satellite products within minutes on a standard workstation.

Performance is evaluated on synthetic and real datasets. In synthetic tests, a nonlinear sinusoidal field with added Gaussian noise is reconstructed with a root‑mean‑square error of 0.12 using ROCK‑PCA versus 0.31 for conventional PCA, confirming the method’s ability to recover nonlinear oscillations. Real‑world applications include:

• Global Gross Primary Productivity (GPP) derived from MODIS (2000‑2020). The first rotated mode captures the global annual cycle, while the second isolates regional growth‑decline trends, with phase differences clearly separating Northern and Southern Hemisphere phenology.

• Soil Moisture (SM) from the ESA Climate Change Initiative. ROCK‑PCA separates the seasonal wet‑dry cycle from a long‑term drying trend in the Sahel, and the complex phase reveals the gradual shift of the drying onset over the study period.

• Sea Surface Temperature (SST) from ECMWF ERA5 reanalysis. The leading mode represents the global annual temperature swing; the second mode isolates non‑seasonal variability associated with El Niño/La Niña events. After rotation, the phase field of this mode visualises the east‑to‑west propagation of temperature anomalies across the Pacific.

Across all cases, ROCK‑PCA successfully (i) distinguishes annual and seasonal oscillations, (ii) extracts non‑seasonal trends and anomalies as independent modes, and (iii) aligns spatial patterns with physically meaningful regions, thanks to the rotation step. The method’s unsupervised nature, ability to handle large datasets efficiently, and explicit spatio‑temporal decomposition make it a powerful alternative or complement to traditional EOF analysis.

In conclusion, the authors provide a robust framework for uncovering hidden nonlinear dynamics in Earth system data. By leveraging complex kernels to encode phase information and a post‑hoc rotation to improve interpretability, ROCK‑PCA delivers high‑resolution, physically meaningful modes that can aid climate monitoring, ecosystem productivity assessment, and ocean‑atmosphere coupling studies. Future work may extend the approach to multimodal data fusion, adaptive kernel learning, and real‑time monitoring pipelines.