Abstract
Spatio-temporal processes, such as atmospheric dynamics or pollutant dispersion, are frequently characterized using Partial Differential Equations (PDEs). Spectral methods, particularly Fourier series expansions, offer a computationally efficient framework for solving these equations. However, the inherent periodicity assumption of Fourier series leads to the Gibbs phenomenon—distorting "ringing" artifacts—when applied to non-periodic real-world data with boundary discontinuities. This paper introduces a "double data flipping" technique that mirrors spatial data to create a periodic domain, effectively suppressing these artifacts while maintaining physical consistency through extended PDE dynamics.
Introduction
Modeling processes that evolve over both space and time is a cornerstone of environmental and physical sciences. While physics-informed models based on PDEs provide a robust theoretical foundation, their practical implementation often relies on truncated Fourier series to manage computational costs. A significant challenge arises when the underlying data is non-periodic, such as a localized rainstorm captured by radar. In these instances, sharp discontinuities at the spatial boundaries trigger the Gibbs phenomenon, manifesting as spurious oscillations or "ghost signals" that degrade prediction accuracy.
The Gibbs Phenomenon in Spatio-Temporal Modeling
The Gibbs phenomenon occurs because a finite sum of continuous sine and cosine waves cannot perfectly reconstruct a signal with a jump discontinuity. In the context of spatio-temporal modeling, this results in artificial ripples near the edges of the observation window. Traditional mitigation strategies, such as windowing (e.g., Hamming windows), often force boundary values to zero, which can introduce significant bias by distorting valid data near the edges.
Methodology: Double Data Flipping and Extended PDE Dynamics
The authors propose a two-step "Double Data Flipping" method to transform a non-periodic $N_1 \times N_2$ image into a periodic $2N_1 \times 2N_2$ domain. By mirroring the image horizontally and then vertically, the resulting "unfolded" image ensures that every edge matches its mirrored counterpart, satisfying the periodicity requirements of Fourier analysis without discarding data.
To ensure this geometric transformation does not violate the underlying physics, the authors extended the dynamics of advection-diffusion PDEs to the flipped domain. They demonstrated that the temporal dynamics of the spectral coefficients for the flipped process can be derived from original physical parameters (velocity and diffusivity) using a specific transformation matrix ($H$). This allows the use of a Kalman Filter in the flipped spectral domain to estimate and predict the process while preserving the physical interpretation of the PDE.
Experimental Evaluation and Results
The framework was evaluated using both simulated advection processes and real-world precipitation data from a tropical storm.
- Visual Accuracy: In radar reflectivity tests, the proposed method successfully eliminated ghost intensities. While standard methods predicted non-existent rainfall in the southern and eastern boundaries due to ringing, the flipping approach correctly identified zero rainfall in those regions.
- Quantitative Performance: In simulations with massive boundary discontinuities, the proposed model (F400) achieved a Mean Absolute Error (MAE) of 1.7–2.0. In contrast, standard non-flipped models exhibited MAEs ranging from 4.0 to over 10.0 near the boundaries. The method also significantly outperformed Hamming window approaches, which suffered from high bias due to zeroing out edge data.
- Spectral Resolution: By quadrupling the domain size, the method inherently provides four times higher spectral resolution, allowing the model to capture finer spatial details.
Conclusion
The double data flipping technique provides a mathematically rigorous and computationally viable solution to the Gibbs phenomenon in physics-informed spatio-temporal models. While the approach increases the computational burden by expanding the state space, it offers superior accuracy and preserves data integrity at the boundaries. Future work may explore extending this framework to non-linear PDEs and three-dimensional spatio-temporal volumes.
Summary updated