Abstract
From a sequence of tracer satellite images, several methods (e.g. optical flow) exist to successfully estimate the main advecting current. Yet, this estimate is limited in resolution. To go beyond, we propose a new parametric estimation method to estimate second-order statistics of the residual small-scale velocity. We first express stochastic transport in a discrete setting to apply standard MLE techniques. Then we propose an efficient method to solve the MLE optimization problem through a fast log-likelihood gradient evaluation algorithm.
1. Introduction
Oceanographic monitoring through satellite imagery provides critical data on tracers such as Sea Surface Temperature (SST) and Sea Surface Salinity (SSS). While standard techniques like optical flow are proficient at capturing large-scale advective currents, they are inherently limited by spatial and temporal resolution. Consequently, subgrid-scale turbulence—the small eddies and fluctuations occurring between pixels or snapshots—remains largely unobserved. Understanding these small-scale motions is vital for accurate modeling of heat and nutrient transport. This paper introduces a rigorous mathematical framework to estimate the statistics of these hidden subgrid flows directly from tracer image sequences.
2. Mathematical Framework
The proposed method bridges fluid dynamics and stochastic statistics through several key components:
2.1 Stochastic Transport Modeling
The framework assumes tracer dynamics are governed by a combination of a smooth, large-scale velocity field and a "white noise" small-scale velocity component. This approach aligns with the Location Uncertainty (LU) and Stochastic Advection by Lie Transport (SALT) frameworks, which are increasingly utilized in stochastic climate modeling.
2.2 Variance Estimation via Quadratic Co-variation
To determine the amplitude of subgrid turbulence, the method employs quadratic co-variation from stochastic calculus. By analyzing high-frequency increments in tracer images, the variance tensor ($a_0$) can be estimated without requiring a complete model of the underlying flow.
2.3 Parametric Covariance Models
The small-scale flow is modeled as a mixture of solenoidal (rotational) and potential (divergent) components. The spatial correlation is represented using Matérn covariances, which offer the flexibility and physical consistency required to model the self-similar nature of turbulence.
3. Maximum Likelihood Estimation (MLE)
Given the non-Gaussian and multiplicative nature of tracer dynamics, standard likelihood functions are inapplicable.
3.1 Girsanov Theorem
The author utilizes the Girsanov theorem to derive a valid log-likelihood function for the stochastic system. This allows for the formal application of MLE techniques to estimate the spectral parameters of the subgrid flow.
3.2 FFT-based Optimization
To ensure computational feasibility for large-scale geophysical datasets, the estimation process is conducted in Fourier space. By leveraging the Fast Fourier Transform (FFT) and the block-diagonal structure of the resulting matrices, the algorithm achieves a complexity of $O(NM \log^2 M)$, where $M$ is the number of pixels. This enables fast gradient evaluation and efficient optimization.
4. Key Insights and Results
A significant finding of this work is the "Rank 1" trap: assuming a purely solenoidal subgrid flow leads to a singular spectrum matrix, causing the MLE to fail. Stability in the statistical estimation requires the inclusion of a divergent component.
Furthermore, the strategy of estimating total variance via quadratic variation prior to performing MLE for spectral parameters ($\beta$ and $\alpha$) mitigates theoretical estimation biases. The resulting algorithm provides a closed-form expression for the log-likelihood gradient, facilitating practical application in calibrating SALT and LU models from observational data.
5. Conclusion and Future Directions
This paper presents a robust method for quantifying subgrid-scale ocean turbulence from satellite imagery. While the current framework assumes clean data and periodic boundary conditions, future work will address measurement noise (aliasing, sensor errors) and adapt the algorithm for complex real-world coastlines. Additionally, extending the method to treat tracers as differential forms could enable the direct processing of surface current observations from high-frequency radars.
Summary updated